# 赢得AI竞赛:增强美国在计算与创新领域的能力 在计算与创新领域 【译注:原文首行标题末尾重复出现「IN COMPUTING AND INNOVATION」(在计算与创新领域)字样,系 GPO 源文排版异常,此处照实保留。】 [参议院听证会 119-143] [来自美国政府出版局(GPO)] S. Hrg. 119-143 ## 赢得AI竞赛:增强美国在计算与创新领域的能力 听证会 举行于 美国参议院 商务、科学与交通委员会 第一百一十九届国会 第一次会议 __________ 2025年5月8日 __________ 为商务、科学与交通委员会使用而印行 [图形无 TIFF 格式版本] 在线获取:http://www.govinfo.gov ______ 美国政府出版局 61-426 PDF 华盛顿:2025年 ## 参议院商务、科学与交通委员会 第一百一十九届国会 第一次会议 特德·克鲁兹(Ted Cruz),得克萨斯州,主席 | 共和党 | 民主党 | |---|---| | 约翰·图恩(John Thune),南达科他州 | 玛丽亚·坎特韦尔(Maria Cantwell),华盛顿州,资深成员 | | 罗杰·威克(Roger Wicker),密西西比州 | 埃米·克洛布彻(Amy Klobuchar),明尼苏达州 | | 德布·费舍尔(Deb Fischer),内布拉斯加州 | 布赖恩·沙茨(Brian Schatz),夏威夷州 | | 杰里·莫兰(Jerry Moran),堪萨斯州 | 爱德华·马基(Edward Markey),马萨诸塞州 | | 丹·沙利文(Dan Sullivan),阿拉斯加州 | 加里·彼得斯(Gary Peters),密歇根州 | | 玛莎·布莱克本(Marsha Blackburn),田纳西州 | 塔米·鲍德温(Tammy Baldwin),威斯康星州 | | 托德·扬(Todd Young),印第安纳州 | 塔米·达克沃思(Tammy Duckworth),伊利诺伊州 | | 特德·巴德(Ted Budd),北卡罗来纳州 | 杰基·罗森(Jacky Rosen),内华达州 | | 埃里克·施密特(Eric Schmitt),密苏里州 | 本·雷·卢汉(Ben Ray Lujan),新墨西哥州 | | 约翰·柯蒂斯(John Curtis),犹他州 | 约翰·希肯卢珀(John Hickenlooper),科罗拉多州 | | 伯尼·莫雷诺(Bernie Moreno),俄亥俄州 | 约翰·费特曼(John Fetterman),宾夕法尼亚州 | | 蒂姆·希伊(Tim Sheehy),蒙大拿州 | 安迪·金(Andy Kim),新泽西州 | | 谢莉·摩尔·卡皮托(Shelley Moore Capito),西弗吉尼亚州 | 莉萨·布伦特·罗切斯特(Lisa Blunt Rochester),特拉华州 | | 辛西娅·卢米斯(Cynthia Lummis),怀俄明州 | | 【译注:原文名册为左右两栏(左栏共和党、右栏民主党)逐行交错排印,其中「Ranking」(资深成员)一词因换行被排在威克一行末尾,实属坎特韦尔;此处按两栏原意还原为表格。】 布拉德·格兰茨(Brad Grantz),共和党幕僚长 妮科尔·克里斯图斯(Nicole Christus),共和党副幕僚长 利亚姆·麦肯纳(Liam McKenna),总法律顾问 莱拉·哈珀·赫尔姆斯(Lila Harper Helms),幕僚长 梅利莎·波特(Melissa Porter),副幕僚长 乔纳森·黑尔(Jonathan Hale),总法律顾问 ## 目录 ---------- 页码 2025年5月8日举行的听证会……1 克鲁兹参议员的陈述……1 坎特韦尔参议员的陈述……3 希伊参议员的陈述……29 莫雷诺参议员的陈述……33 克洛布彻参议员的陈述……35 沙茨参议员的陈述……40 巴德参议员的陈述……42 金参议员的陈述……44 施密特参议员的陈述……46 希肯卢珀参议员的陈述……48 柯蒂斯参议员的陈述……50 达克沃思参议员的陈述……53 扬参议员的陈述……55 布伦特·罗切斯特参议员的陈述……56 莫兰参议员的陈述……58 卢汉参议员的陈述……60 卢米斯参议员的陈述……63 罗森参议员的陈述……64 沙利文参议员的陈述……66 马基参议员的陈述……68 彼得斯参议员的陈述……71 费特曼参议员的陈述……73 ### 证人 萨姆·奥特曼(Sam Altman),OpenAI 联合创始人兼首席执行官……5 书面证词……7 苏博士(Dr. Lisa Su),超威半导体公司(Advanced Micro Devices,AMD)首席执行官兼董事长……9 书面证词……10 迈克尔·因特拉托(Michael Intrator),CoreWeave 联合创始人兼首席执行官……12 书面证词……14 布拉德·史密斯(Brad Smith),微软公司(Microsoft Corporation)副董事长兼总裁……21 书面证词……22 ### 附录 萨姆·奥特曼对以下人士所提书面问题的答复: 罗杰·威克阁下……83 玛莎·布莱克本阁下……83 玛丽亚·坎特韦尔阁下……84 埃米·克洛布彻阁下……86 布赖恩·沙茨阁下……86 爱德华·马基阁下……89 塔米·鲍德温阁下……91 杰基·罗森阁下……91 莉萨·布伦特·罗切斯特阁下……92 苏博士对以下人士所提书面问题的答复: 托德·扬阁下……92 玛丽亚·坎特韦尔阁下……93 埃米·克洛布彻阁下……94 布赖恩·沙茨阁下……95 杰基·罗森阁下……96 约翰·费特曼阁下……97 迈克尔·因特拉托对以下人士所提书面问题的答复: 玛丽亚·坎特韦尔阁下……99 埃米·克洛布彻阁下……100 布赖恩·沙茨阁下……101 爱德华·马基阁下……102 杰基·罗森阁下……103 布拉德·史密斯对以下人士所提书面问题的答复: 罗杰·威克阁下……104 托德·扬阁下……104 玛丽亚·坎特韦尔阁下……106 布赖恩·沙茨阁下……109 爱德华·马基阁下……111 塔米·鲍德温阁下……114 杰基·罗森阁下……115 约翰·费特曼阁下……117 莉萨·布伦特·罗切斯特阁下……118 【译注:以上目录页码均照原印刷本保留,未作改动。】 ## 赢得AI竞赛:增强美国在计算与创新领域的能力 在计算与创新领域 【译注:原文此标题块第三行再次重复「IN COMPUTING AND INNOVATION」字样,系源文排版异常,照实保留。】 ---------- 2025年5月8日,星期四 美国参议院, 商务、科学与交通委员会, 华盛顿哥伦比亚特区。 委员会依通知于上午10时02分在拉塞尔参议院办公大楼 SR-253 室召开会议,由委员会主席特德·克鲁兹阁下主持。 出席:克鲁兹参议员[主持]、莫兰参议员、沙利文参议员、扬参议员、巴德参议员、施密特参议员、柯蒂斯参议员、莫雷诺参议员、希伊参议员、卢米斯参议员、坎特韦尔参议员、克洛布彻参议员、沙茨参议员、马基参议员、彼得斯参议员、达克沃思参议员、罗森参议员、卢汉参议员、希肯卢珀参议员、费特曼参议员、金参议员和布伦特·罗切斯特参议员。 ## 得克萨斯州联邦参议员特德·克鲁兹阁下的开场陈述 **克鲁兹主席(Ted Cruz,得克萨斯州,委员会主席):** 早上好。参议院商务、科学与交通委员会现在开会。欢迎各位证人。感谢你们今天上午出席。 过去两年,AI 把美国和世界带到了一个关键的转折点。AI 可能是一项与互联网同样具有变革性、甚至更具变革性的技术。它掀起了一场新的全球工业革命,有望释放改善我们生活质量、创造就业、刺激经济增长的机遇。 在 AI 领域领先的国家将塑造21世纪的全球秩序。无论从经济安全还是国家安全的角度,美国都必须在 AI 竞赛中击败中国。中国已把 AI 置于其国家战略的核心,目标是到2030年在 AI 领域领先世界,并在制造业、国防等行业的 AI 采用上投入巨资。 在这场竞赛中,美国正面临一个岔路口。我们是走拥抱我国创业自由与技术创新传统的道路,还是采纳欧洲那种命令与控制式的政策? 我建议国会借鉴我们可以从互联网诞生之初学到的经验教训。20世纪90年代初,华盛顿拥抱了互联网,并明确采取了一种有意且果断的轻触式监管风格。 国会依据《1996年电信法》(Telecommunications Act of 1996)选择放松管制,而克林顿总统则推动达成保护美国知识产权和技术出口的关税协定与条约。 此外,1998年国会颁布了为期10年的互联网税收暂缓令,以免各州法律把市场割裂成碎片、扼杀电子商务的前景。 这些决定的结果非同凡响。到2000年,美国已连续五年录得生产率增长与投资增长的历史新高。 数十万个新工作岗位被创造出来,美国成为头号技术出口国,巨额私人投资涌入美国数字经济。 相比之下,欧盟国家推行了一系列重手监管,事实证明代价极其高昂。1993年,美国与欧洲的经济体量几乎完全相同。今天,美国经济比欧洲大50%以上。其驱动因素是科技和页岩革命。这两者几乎构成了美国相对欧洲那一巨大增长的全部。 根据欧盟委员会的一份报告,全球 AI 初创企业融资中只有6%流向欧盟企业——6%。这只有流向美国公司金额的十分之一。 该报告把这一巨大鸿沟直接归咎于欧盟恶劣的监管方式。然而,拜登政府却出于令人费解的原因,试图让 AI 政策与欧盟看齐,采纳他们那些失败的政策。 拜登总统那份包罗万象的 AI 行政命令——美国历史上最长的行政命令——把 AI 描绘成危险而不透明的东西,为审计、风险评估和监管认证铺平了道路。 拜登的做法激发了全国各州议会的类似举动,可能让初创企业、开发者和 AI 用户背上沉重的合规成本。 我的一些同事主张,拜登路线的一个更友好的版本是合理的。他们想要一套测试制度来防范 AI 的所谓「歧视」,并让政府人员提供所谓「指导文件」——仿佛出自奥威尔(Orwell)笔下的东西——来引入他们所说的最佳实践,好像 AI 工程师没有官僚就缺乏负责任地构建 AI 的智力似的。 业内许多人愚蠢地支持了这种家长式作风。有害的监管有多种形式。拜登在卸任前夕仓促出台的、针对芯片和模型权重的误导性 AI 扩散规则,本会削弱美国科技公司向全世界销售 AI 的能力。 拜登的计划本会把关键市场拱手让给中国。我们应当希望外国、尤其是我们的盟友购买美国产品。 数月来我一直高声反对这项规则,事实上,资深成员与我曾共同敦促拜登政府不要采纳它;我非常高兴特朗普总统现已确认他计划废除它。 所有这些多管闲事的官僚做派——无论是拜登在芯片出口上的产业政策,还是业界与监管者共同批准的指导文件——都是披着羊皮的狼。 要在 AI 领域领先,美国不能允许监管——即便是所谓无害的那种——扼杀创新或采用。 美国在 AI 领域的主导地位取决于两个因素:创新与采用。创新推动全球竞争力的突破。采用确保这些工具赋能美国工人和企业,使美国成为全球领先的 AI 采用国和出口国。 值得庆幸的是,特朗普总统已在很大程度上扭转了拜登误导性的 AI 议程。事实上,我认为 AI 是刚过去的那次选举中一个未被充分重视的潜在议题。 美国人希望看到特朗普总统、共和党人、乃至所有参议员都力挺聚焦创新与采用的 AI 政策。 对比令人震惊。今年已有超过1万亿美元($1 trillion)的新 AI 项目,包括在得克萨斯州的重大投资,如普莱诺(Plano)的 CoreWeave 数据中心,以及 OpenAI、甲骨文(Oracle)等公司在阿比林(Abilene)投资5000亿美元($500 billion)的「星际之门」计划(Project Stargate)。 对 AI 采取轻触式监管风格,需要国会与总统并肩合作,正如国会当年与克林顿总统合作那样。我们需要推进促进 AI 长期增长与创新的立法。 正因如此,我将很快公布一项新法案,为 AI 设立监管沙盒,其模式借鉴国会与克林顿总统在互联网诞生之初采取的做法,以消除 AI 采用的障碍、防止各州不必要的过度监管,并让 AI 供应链在美国本土迅速成长。 这就是我们加速经济增长、确保美国在 AI 领域主导地位、击败中国的方式。 说到这里,我把发言权交给资深成员坎特韦尔。 - WINNING THE AI RACE: STRENGTHENING U.S. CAPABILITIES IN COMPUTING AND INNOVATION IN COMPUTING AND INNOVATION [Senate Hearing 119-143] [From the U.S. Government Publishing Office] S. Hrg. 119-143 WINNING THE AI RACE: STRENGTHENING U.S. CAPABILITIES IN COMPUTING AND INNOVATION ======================================================================= HEARING before the COMMITTEE ON COMMERCE, SCIENCE, AND TRANSPORTATION UNITED STATES SENATE ONE HUNDRED NINETEENTH CONGRESS FIRST SESSION __________ MAY 8, 2025 __________ Printed for the use of the Committee on Commerce, Science, and Transportation [GRAPHIC NOT AVAILABLE IN TIFF FORMAT] Available online: http://www.govinfo.gov ______ U.S. GOVERNMENT PUBLISHING OFFICE 61-426 PDF WASHINGTON : 2025 SENATE COMMITTEE ON COMMERCE, SCIENCE, AND TRANSPORTATION ONE HUNDRED NINETEENTH CONGRESS FIRST SESSION TED CRUZ, Texas, Chairman JOHN THUNE, South Dakota MARIA CANTWELL, Washington, ROGER WICKER, Mississippi Ranking DEB FISCHER, Nebraska AMY KLOBUCHAR, Minnesota JERRY MORAN, Kansas BRIAN SCHATZ, Hawaii DAN SULLIVAN, Alaska EDWARD MARKEY, Massachusetts MARSHA BLACKBURN, Tennessee GARY PETERS, Michigan TODD YOUNG, Indiana TAMMY BALDWIN, Wisconsin TED BUDD, North Carolina TAMMY DUCKWORTH, Illinois ERIC SCHMITT, Missouri JACKY ROSEN, Nevada JOHN CURTIS, Utah BEN RAY LUJAN, New Mexico BERNIE MORENO, Ohio JOHN HICKENLOOPER, Colorado TIM SHEEHY, Montana JOHN FETTERMAN, Pennsylvania SHELLEY MOORE CAPITO, West Virginia ANDY KIM, New Jersey CYNTHIA LUMMIS, Wyoming LISA BLUNT ROCHESTER, Delaware Brad Grantz, Republican Staff Director Nicole Christus, Republican Deputy Staff Director Liam McKenna, General Counsel Lila Harper Helms, Staff Director Melissa Porter, Deputy Staff Director Jonathan Hale, General Counsel C O N T E N T S ---------- Page Hearing held on May 8, 2025...................................... 1 Statement of Senator Cruz........................................ 1 Statement of Senator Cantwell.................................... 3 Statement of Senator Sheehy...................................... 29 Statement of Senator Moreno...................................... 33 Statement of Senator Klobuchar................................... 35 Statement of Senator Schatz...................................... 40 Statement of Senator Budd........................................ 42 Statement of Senator Kim......................................... 44 Statement of Senator Schmitt..................................... 46 Statement of Senator Hickenlooper................................ 48 Statement of Senator Curtis...................................... 50 Statement of Senator Duckworth................................... 53 Statement of Senator Young....................................... 55 Statement of Senator Blunt Rochester............................. 56 Statement of Senator Moran....................................... 58 Statement of Senator Lujan....................................... 60 Statement of Senator Lummis...................................... 63 Statement of Senator Rosen....................................... 64 Statement of Senator Sullivan.................................... 66 Statement of Senator Markey...................................... 68 Statement of Senator Peters...................................... 71 Statement of Senator Fetterman................................... 73 Witnesses Sam Altman, Co-Founder and Chief Executive Officer, OpenAI....... 5 Prepared statement........................................... 7 Dr. Lisa Su, Chief Executive Officer and Chair, Advanced Micro Devices (AMD).................................................. 9 Prepared statement........................................... 10 Michael Intrator, Co-Founder and Chief Executive Officer, CoreWeave...................................................... 12 Prepared statement........................................... 14 Brad Smith, Vice Chair and President, Microsoft Corporation...... 21 Prepared statement........................................... 22 Appendix Response to written questions submitted to Sam Altman by: Hon. Roger Wicker............................................ 83 Hon. Marsha Blackburn........................................ 83 Hon. Maria Cantwell.......................................... 84 Hon. Amy Klobuchar........................................... 86 Hon. Brian Schatz............................................ 86 Hon. Edward Markey........................................... 89 Hon. Tammy Baldwin........................................... 91 Hon. Jacky Rosen............................................. 91 Hon. Lisa Blunt Rochester.................................... 92 Response to written questions submitted to Dr. Lisa Su by: Hon. Todd Young.............................................. 92 Hon. Maria Cantwell.......................................... 93 Hon. Amy Klobuchar........................................... 94 Hon. Brian Schatz............................................ 95 Hon. Jacky Rosen............................................. 96 Hon. John Fetterman.......................................... 97 Response to written questions submitted to Michael Intrator by: Hon. Maria Cantwell.......................................... 99 Hon. Amy Klobuchar........................................... 100 Hon. Brian Schatz............................................ 101 Hon. Edward Markey........................................... 102 Hon. Jacky Rosen............................................. 103 Response to written questions submitted to Brad Smith by: Hon. Roger Wicker............................................ 104 Hon. Todd Young.............................................. 104 Hon. Maria Cantwell.......................................... 106 Hon. Brian Schatz............................................ 109 Hon. Edward Markey........................................... 111 Hon. Tammy Baldwin........................................... 114 Hon. Jacky Rosen............................................. 115 Hon. John Fetterman.......................................... 117 Hon. Lisa Blunt Rochester.................................... 118 WINNING THE AI RACE: STRENGTHENING U.S. CAPABILITIES IN COMPUTING AND INNOVATION IN COMPUTING AND INNOVATION ---------- THURSDAY, MAY 8, 2025 U.S. Senate, Committee on Commerce, Science, and Transportation, Washington, DC. The Committee met, pursuant to notice, at 10:02 a.m., in room SR-253, Russell Senate Office Building, Hon. Ted Cruz, Chairman of the Committee, presiding. Present: Senators Cruz [presiding], Moran, Sullivan, Young, Budd, Schmitt, Curtis, Moreno, Sheehy, Lummis, Cantwell, Klobuchar, Schatz, Markey, Peters, Duckworth, Rosen, Lujan, Hickenlooper, Fetterman, Kim, and Blunt Rochester. OPENING STATEMENT OF HON. TED CRUZ, U.S. SENATOR FROM TEXAS The Chairman. Good morning. The Senate Committee on Commerce, Science, and Transportation is called to order. Welcome to our witnesses. Thank you for joining us this morning. In the last two years AI has brought the United States and the world to a critical inflection point. AI may be a technology as transformative as the Internet or even more so. It has unleashed a new global industrial revolution with the potential to unlock opportunities that improve our quality of life, create jobs, and stimulate economic growth. The country that leads in AI will shape the 21st century global order. As a matter of economic security, as a matter of national security, America has to beat China in the AI race. China has made AI central to its national strategy and China aims to lead the world in AI by 2030, investing heavily in AI adoption across industries like manufacturing and defense. In this race the United States is facing a fork in the road. Do we go down the path that embraces our history of entrepreneurial freedom and technological innovation or do we adopt the command and control policies of Europe? I would suggest that Congress draw on the lessons we can learn from the dawn of the internet. In the early 1990s Washington embraced the Internet and explicitly adopted a style of regulation that was intentionally and decisively a light touch. Congress chose to deregulate under the Telecommunications Act of 1996 while President Clinton pursued tariff agreements and treaties that protected America's intellectual property and technological exports. Further, in 1998 Congress enacted a 10-year Internet tax moratorium so that state laws would not balkanize and stymie the promise of e-commerce. The results of these decisions were extraordinary. By 2000, the United States has recorded five straight years of historic highs in productivity gains and investment growth. Hundreds of thousands of new jobs were created and the United States became a top tech exporter with massive sums of private investment pouring into the U.S. digital economy. By contrast, EU countries pursued a series of heavy-handed regulations that proved enormously costly. In 1993 the United States and Europe had economies virtually identical in size. Today, the American economy is more than 50 percent larger than Europe's. The drivers of that are tech and the shale revolution. Those two comprise virtually the entirety of that massive growth over Europe. According to one EU Commission report only 6 percent of global AI startup funding flows to EU firms--6 percent. That is one-tenth of the amount that is going to American companies. The report directly blames this yawning chasm on the EU's nasty regulatory approach. And, yet, the Biden administration for inexplicable reasons tried to align AI policy with the EU to adopt their failed policies. President Biden's sweeping AI executive order, the longest executive order in American history, cast AI as dangerous and opaque, laying the groundwork for audits, for risk assessments, and regulatory certifications. Biden's approach inspired similar efforts in state legislatures across the country, threatening to burden startups, developers, and AI users with heavy compliance costs. Some of my colleagues suggest that a friendlier version of the Biden approach makes sense. They want a testing regime to guard against AI, quote, ``discrimination'' and have government agents provide, quote, ``guidance documents,'' seemingly something out of Orwell, that will usher in what they call best practices, as if AI engineers lack the intelligence to responsibly build AI without the bureaucrats. Many in the industry foolishly have supported such paternalism. Harmful regulations take many forms. Biden's misguided midnight AI diffusion rule on chips and model weights would have crippled American tech companies' ability to sell AI to the world. The Biden plan would have handed over key markets to China. We should want foreign countries, particularly our allies, to buy American. I vocally opposed this rule for months and, indeed, the Ranking Member and I together urged the Biden administration not to adopt it and I am very pleased that President Trump has now confirmed he plans to rescind it. All of this busybody bureaucracy, whether Biden's industrial policy on chip exports or industry and regulator- approved guidance documents, is a wolf in sheep's clothing. To lead in AI the United States cannot allow regulation, even the supposedly benign kind, to choke innovation or adoption. American dominance in AI depends on two factors: innovation and adoption. Innovation drives breakthroughs in global competitiveness. Adoption ensures that these tools empower American workers and businesses, enabling the United States to become the world's leading adopter and exporter of AI. Thankfully, President Trump has, largely, reversed Biden's misguided AI agenda. In fact, I think AI was a sleeper issue in this last election. Americans wanted to see President Trump and Republicans and, indeed, all senators champion AI policies focused on innovation and adoption. The contrast has been astounding. This year, there have been over $1 trillion of new AI projects including major investments in Texas like the CoreWeave data center in Plano and the $500 billion project Stargate in Abilene by OpenAI and Oracle and others. Adopting a light touch regulatory style for AI will require Congress to work alongside the President just as Congress did with President Clinton. We need to advance legislation that promotes long-term AI growth and innovation. That is why I will soon release a new bill that creates a regulatory sandbox for AI modeled on the approach taken by Congress and President Clinton at the dawn of the Internet that will remove barriers to AI adoption, prevent needless state overregulation, and allow the AI supply chain to rapidly grow here in the United States. That is how we will accelerate economic growth, secure U.S. dominance in AI, and beat China. And with that, I turn to Ranking Member Cantwell. STATEMENT OF HON. MARIA CANTWELL, U.S. SENATOR FROM WASHINGTON ¶
Senator Cantwell ## 华盛顿州联邦参议员玛丽亚·坎特韦尔阁下的陈述 **坎特韦尔参议员(Maria Cantwell,华盛顿州,资深成员):** 谢谢您,主席先生。 感谢您举行这次听证会,也欢迎在座的各位证人——奥特曼先生、苏博士、因特拉托先生和史密斯先生。 非常高兴各位都能到场,而对太平洋西北地区来说,史密斯先生和奥特曼先生在此尤其是一个值得自豪的时刻,两位都代表着一种 OpenAI 式的路径。 我的意思是这样一种路径:我们希望战胜中国和一个封闭体系,为此要确保在美国本土和全球各地开发出来的,是一种让美国获胜且开放的架构。 要做到这一点,我们需要专注于在算力、算法和稳健的数据来源上取胜。所有这些都将是关键。 就我个人而言,我相信对国家科学基金会(NSF)的持续投入有助于 【译注:原文此句在批次边界处截断(「…a continued investment in NSF helps」),为残句,未完部分见下一批次。】 **坎特韦尔参议员(Maria Cantwell,华盛顿州,资深成员):**……在所有这些领域,与今天在此所代表的产业界形成良好的公私合作伙伴关系。【译注:本批次为连续切片,开头承接上一批次坎特韦尔参议员的开场陈述,原文句子未完,按残句保留,不补写】 我非常自豪我们通过了《芯片与科学法案》(CHIPS and Science Act),因为《芯片与科学法案》也为投资美利坚合众国、把更多供应链带回美利坚合众国奠定了基础,从而在我们——我相信——已经拥有的算力未来领导地位之上继续建设。 但我们也必须明白,我们必须推进《芯片法案》(CHIPS Act),比如华盛顿大学(University of Washington)获得的 1000 万美元($10 million)拨款,用于芯片的多设计集(multi-design sets),也就是超大规模集成电路设计——我相信苏博士今天会向我们介绍这些。 但美国必须继续在未来设计及其实施方面保持领先,这一事实也要求我们在数据中心、电力来源,以及如何构建可能在不久的将来占到电力需求 12% 的供应方面,做得非常明智。 那么我们该怎么做?我注意到各位的证词中都解释了这一点,但我也非常自豪,微软已经与一家公司——位于华盛顿州埃弗里特(Everett)的聚变能源公司(Fusion Energy)——签署了一项电力供应协议——也许奥特曼先生会在他的证词中谈到这一点——他们希望从中在不久的将来获得能源供应。 因此,显然,美国要在电力和发展方面领先。所以,史密斯先生,我非常感谢您在证词中强调这一点:美利坚合众国需要数十万名新电工,这是我们所有人都应该支持的事情。 在美国本土及其他地方拥有电力、电工和数据中心,这一事实将是关键。 虽然我希望看到我们向前推进,正如主席所说——我们签署了一封信,表示我们需要对出口管制给予更广泛的支持——但我想把话说清楚。 出口管制不是一种贸易策略。它们不是美国总统在贸易谈判中随手掏出来的后备议题。 我们将快速行动,因为我们要制定标准。我相信这些标准应当鼓励美国制造和生产的 AI 芯片与技术得到非常广泛的分布,并且我们要求海外伙伴遵守我们确立的规则,比如确保供应不被规避、不会以某种方式落入中国手中,确保我们拥有准入,确保我们能够对此进行核查,同时也确保美国的数据公司和基于云的公司被允许进入那个市场。 我们不应该走到海外市场,结果却被告知,提供来自美国云服务的组织不被允许进入。我相信,这将是一项强有力的倡议,让美国 AI 芯片和美国 AI 开放系统在全球占据主导地位。 我们为什么需要快速行动?我们需要快速行动,因为如果不这样做,我们就会面临另一个华为(Huawei),另一个美国落后的例子,而且还要说我们应该拆除这套如今我们因诸多原因和后门政策而不喜欢的系统。 所以我完全支持取胜。这就是我们通过《芯片与科学法案》的原因。我完全支持取胜,这就是为什么去年我们让这个委员会通过了七项法案,而这些法案可以说在「跛脚鸭」会期(lame duck)卡住了。 我认为委员会主席当时还没有准备好就这七项法案与众议院和参议院的谈判向前推进。但那些法案——比如我与扬参议员(Senator Young)之间关于标准研究所(即 NIST 标准)的法案——我认为我们仍然需要去做。 还有我的同事和我——莫兰(Moran)——关于教育和奖学金、小企业的法案,以及在座我的同事克洛布彻参议员(Amy Klobuchar,明尼苏达州)和图恩(Thune)提出的法案,那也与 NIST 标准有关。 所以一年前我们本有机会快速行动。我们没有做到。那么现在就让我们来做吧。让我们聚在一起把这件事想清楚。美国现在行动得越快越好——我很喜欢保罗·罗默(Paul Romer)这句名言,讲的是协作是创新的下一个阶段。 如果我们在这里不协作,如果我们陷入政治争斗而不是把政策做对,我们就无法快速行动。让我们允许这些人去做他们最擅长的事,让我们确保美国有正确的政策到位,从而使我们的 OpenAI 标准最终胜出。 谢谢您,主席先生。 **克鲁兹主席(Ted Cruz,得克萨斯州,委员会主席):**谢谢。 现在我想介绍今天的各位证人。我们的每位证人及其公司都代表着 AI 基础设施、硬件和软件供应链的关键部分。 我们的第一位证人是萨姆·奥特曼(Sam Altman),OpenAI 的联合创始人兼首席执行官。OpenAI 是世界上最先进的 AI 公司之一,以其 ChatGPT 产品最为人所知。 我们的第二位证人是苏姿丰(Lisa Su),超威半导体(Advanced Micro Devices,AMD)的董事长兼首席执行官。AMD 开发为人工智能提供动力的高性能处理器、图形芯片和 AI 加速器,而且苏博士也是得克萨斯州人。 今天的第三位证人是迈克尔·因特拉托(Michael Intrator),CoreWeave 的首席执行官兼联合创始人,这是一家 AI 超大规模云服务商(hyperscaler)。CoreWeave 是世界上最大的专用 AI 云平台。 我们的最后一位证人是布拉德·史密斯(Brad Smith),微软(Microsoft)的副董事长兼总裁。我相信大家都熟悉他的公司。 奥特曼先生,请您作开场陈述。请把音量打开。 Thank you, Mr. Chairman. Thank you for this hearing, and welcome to the witnesses before us--Mr. Altman, Dr. Su, Mr. Intrator, and Mr. Smith. It is a great pleasure to have all of you here, but it is an especially prideful moment for the Pacific Northwest to have Mr. Smith and Mr. Altman here, both representing an OpenAI approach. By that I mean an approach where we want to win against China and a closed system by making sure that what is developed here in the United States and around the globe is an architecture where the United States wins and is open. To do that we need to focus in winning on computing power, on algorithms, and on robust data sources. All of that will be key. Personally, I believe a continued investment in NSF helps in all of those areas as a good public-private partnership with the industry that is represented here today. I am so proud that we passed the CHIPS and Science Act, because the CHIPS and Science Act also set a foundation for investing in the United States of America and bringing more of the supply chain back to the United States of America to build on a future leadership that we already have, I believe, in the computing power. But we also need to understand that we have to move forward on the CHIPS Act like the University of Washington $10 million grant on multi-design sets for chips, the very large-scale integrated designs I am sure that Dr. Su will tell us about today. But the fact that the United States has to continue to lead on the future designs and the implementation of that also requires us to be very smart about data centers, about sources of electricity, and how we are going to build that supply that could be up to 12 percent of electricity demand in the very near future. So how do we do that? I have noticed in each of your testimonies you all explain this, but I am also very proud that Microsoft has already signed an agreement with one company, Fusion Energy in Everett, Washington, for a power source supply--maybe Mr. Altman in his testimony will talk about this--but that they hope to get very near future energy source from that. So, clearly, the United States leading on electricity and development so, Mr. Smith, very much appreciate in your testimony the accentuation on the fact that the United States of America needs hundreds of thousands of new electricians, something we should all want to get behind. The fact that having electricity and the electricians and the data source centers here in the United States and in other places will be key. While I want to see us move forward, as the Chairman said-- we signed a letter saying we needed a broader support for export controls--I want to be clear. Export controls are not a trade strategy. They are not a back pocket issue that the President of the United States whips out in trade negotiations. We are going to move fast because we are going to set standards. I believe those standards should be encouraging very broad distribution of U.S.-manufactured and -made AI chips and technology, and that we are asking our partners overseas to comply with the rules that we establish, things like making sure that there is no circumvention of the supply that somehow gets into China's hands, making sure that we have access, and making sure that we can verify on that, and also making sure that U.S. data companies and cloud-based companies are allowed to be in that market. We should not be going to markets overseas only to have them tell us that organizations with cloud services from the U.S. would not be allowed. This, I believe, would be a robust initiative on getting U.S. AI chips and U.S. AI open systems dominated around the globe. Why do we need to move fast? We need to move fast because if we do not we are looking at another Huawei, another instance where the United States is behind and also saying we should tear out this system that now we do not like for lots of reasons and back door policies. So I am all for winning. That is why we passed the CHIPS and Science Act. I am all for winning and that is why we have passed seven bills out of this committee last year that, kind of, got stuck in the lame duck. I think the Chairman of the Committee was not ready to move forward in negotiations with the House and Senate on those seven bills. But those bills, a bill between myself and Senator Young on the Institute for Standards--NIST standards--which I think we still need to do. My colleague and I--Moran--on education and scholarships, small business, and the bill by my colleague here, Senators Klobuchar and Thune, which was also related to the NIST standards. So we had an opportunity a year ago to move fast. We did not do it. So let us do this now. Let us get together and figure this out. The faster the United States moves now, I like this great Paul Romer quote, which was about collaboration is the next phase of innovation. If we do not collaborate here, if we throw down on politics instead of getting the policy right, we will not move fast. Let us allow these people to do what they do best and let us make sure the United States has the right policies in place so that our OpenAI standard wins the day. Thank you, Mr. Chairman. The Chairman. Thank you. I would now like to introduce our witnesses for today. Each of our witnesses and their companies represent critical parts of the AI infrastructure, hardware, and software supply chain. Our first witness is Sam Altman, the Co-Founder and CEO of OpenAI. OpenAI is one of the world's most advanced AI companies, known best for its ChatGPT product. Our second witness is Lisa Su, the Chair and CEO of Advanced Micro Devices--AMD. AMD develops high-performance processors, graphic chips, and AI accelerators that power artificial intelligence, and Dr. Su is also a Texan. Our third witness today is Michael Intrator, the CEO and Co-Founder of CoreWeave, an AI hyperscaler. CoreWeave is the world's largest purpose built AI cloud platform. And our final witness is Brad Smith, the Vice Chair and President of Microsoft. I believe everyone is familiar with his company. Mr. Altman, you are recognized for your opening statement. If you could turn on the volume. ¶
altman **奥特曼先生:**抱歉。 **克鲁兹主席:**而且我确实很喜欢教科技人士怎么操作科技设备。 Sorry about that. The Chairman. And I do enjoy telling techies how to operate the tech. ¶
altman 我居然没弄明白,这挺尴尬的。 It is pretty embarrassing that I could not figure that out. ¶
altman STATEMENT OF SAM ALTMAN, CO-FOUNDER AND CHIEF EXECUTIVE OFFICER, OPENAI(萨姆·奥特曼,OpenAI 联合创始人兼首席执行官) 总之,谢谢主席。谢谢坎特韦尔资深成员(Ranking Member Cantwell)。谢谢各位参议员和同席的各位嘉宾。能来到这里真是荣幸。我大约两年前来过这里,那时 ChatGPT 刚刚发布。它是世界上的一个新奇事物。人们不确定它意味着什么、会被用来做什么。今天,我们取得了重大进展。ChatGPT 每周有超过 5 亿人使用。我昨天刚看到,据 Similarweb 统计,它现在已是全球第五大网站,而且增长非常快。但最重要的是,它正被以真正重要的方式使用。它在显著提升生产力。我们听到科学家说,他们的生产力是以前的两三倍。我们听到人们以过去无法做到的方式获取医疗建议或学习。这真的——它不再是未来才会到来的东西,它现在就在这里,人们真的在使用它。我们非常自豪能成为其中的引领者之一。我们非常自豪美国在 AI 上如此显著地领先,我认为这至关重要。克鲁兹参议员(Senator Cruz)谈到美国创新的重要性,谈到我们让互联网时代的故事重演。我相信这将至少与互联网一样重大,也许更大。这需要发生。要做到这一点,基础设施投资至关重要。我相信下一个十年将关乎充裕的智能与充裕的能源,确保美国在这两者上都领先,确保我们能够迎来这两场革命——我认为它们将以极其积极的方式改变我们生活的世界。这至关重要。我昨天去了得克萨斯州阿比林(Abilene),我们正在那里建设将成为全球最大 AI 训练设施的项目。进展非常漂亮。看到它令人超级兴奋。我们需要更多这样的设施。有一整套「AI 工厂」式的供应链:能源、芯片、数据中心拔地而起、机架搭建等等。我们必须在美国把这件事做得非常好,才能持续创新、持续领先、持续塑造这场革命。说到这里,克鲁兹主席的话让我深受启发,所以我想在这里偏离讲稿,讲一个故事。我的书面证词里已经涵盖了基本情况。所以如果可以的话,我想给各位讲个故事。我在圣路易斯(St. Louis)长大,是个电脑迷,那时正值互联网繁荣期,我觉得那是有史以来最酷的事情。我们住在圣路易斯郊区一栋漂亮的老砖房里,我住在阁楼,有一台电脑,我会整夜不睡学习编程,我得以使用互联网,那真是一个创新井喷的疯狂时代。各种各样的事情都在发生。太神奇了,而且这一切都发生在这里。所有互联网公司都在美国。我用的是这里制造的 Mac。我用的芯片,其公司就起步于我现在住处附近。我了解了计算机。我觉得那是最酷的事情,而我可以从那段经历画一条直线,直抵创办 OpenAI、得以参与 Helion 这样的公司。美国创新的精神与对创业的支持。我不认为互联网能在别的任何地方发生;而如果互联网没有发生,我不认为 AI 革命会在这里发生。我是互联网革命的孩子。我有莫大的荣幸,成为 AI 革命众多父母中的一员。我认为这绝非偶然——这样的事一次又一次地发生在美国。但我们需要确保,我们构建系统、制定政策的方式能让这一切持续发生。我认为这是魔法。我也不想住在欧洲。我认为美国就是一个不可思议而特别的存在,它不仅将是 AI 革命发生的地方,也将是之后所有革命发生的地方。我最近回圣路易斯老家,开车经过我们的老房子,那是夜里,我抬头看,顶层那扇窗亮着灯。我想,但愿里面有个孩子正熬夜玩 ChatGPT,琢磨着他或她将如何创办下一家公司——而 AI 之后的下一件大事,也会发生在这里。对我来说,这就是这个国家的魔法。这对我个人极其重要,我希望它一直延续。非常感谢各位邀请我。 Anyway, thank you, Chairman. Thank you, Ranking Member Cantwell. Thank you, all senators and fellow panelists. It is a real honor to be here. I was here about two years ago and at that time ChatGPT had recently launched. It was a curiosity in the world. People were not sure what it was going to mean, what it was going to be used for. Today, we have made significant progress. ChatGPT is used by more than 500 million people a week. I just saw yesterday that according to Similarweb it is now the fifth biggest website on the Internet globally and growing very quickly, but most of all, it is being used in really important ways. It is significantly increasing productivity. We hear scientists say they are two or three times more productive than they could be before. We hear people that are getting medical advice or learning in ways they could not before and it is really--it is no longer this thing that was going to come in the future but it is here now and people are really using it. We are very proud to be one of the leaders of this. We are very proud that America is leading in AI so significantly and I think that is critical. What Senator Cruz said about the importance of innovation in America and that we have the--what happened with the Internet we have happen again. I believe this will be at least as big as the internet, maybe bigger. That needs to happen. For that to happen investment in infrastructure is critical. I believe the next decade will be about abundant intelligence and abundant energy, making sure that those--that America leads in both of those, that we are able to usher in these dual revolutions that will change the world we live in I think in incredibly positive ways. It is critical. I got to go to Abilene, Texas yesterday where we are building out what will be the largest AI training facility in the world. It is coming along beautifully. Super exciting to see. We need a lot more of that. There is a whole sort of AI factory, like, a supply chain of energy chips, standing up data centers, building the racks and more. We have got to do that really well in the U.S. so that we can continue to innovate, continue to lead, and continue to, sort of, shape this revolution. Speaking of that, I was very inspired by what Chairman Cruz said so I would like to deviate from script here and tell a story. In my prepared written testimony I covered the basics. So if it is OK I would love to tell you a story. I grew up in St. Louis and I was a computer nerd, and it was the time of the Internet boom and I thought it was the coolest thing ever. We kind of lived in this beautiful, old brick house in this suburb of St. Louis and I lived in the attic, and I had this computer and I would stay up all night and I would learn to program, and I got to kind of use the Internet and it was, like, a crazy time of tons of innovation. All sorts of stuff was happening. It was amazing and it was all happening here. All the Internet companies were in the U.S. I used a Mac that was built here. I used chips that were started, you know, near where I now live. And I learned about computers. I thought it was the coolest thing ever, and I can draw a straight line from that experience to founding OpenAI and getting to work on companies like Helion. The spirit of American innovation and support of entrepreneurship. I do not think the Internet could have happened anywhere else and if that did not happen I do not think the AI revolution would have happened here. I am a child of the Internet revolution. I have the great honor to be one of the parents of the many parents of the AI revolution and I think it is no accident that that is happening in America again and again and again. But we need to make sure that we build our systems and that we set our policy in a way where that continues to happen. I think this is magic. I do not want to live in Europe either. I think America is just an incredible and special thing, and it will not only be the place where the AI revolution happens but all the revolutions after. I was home visiting St. Louis recently. Drove by our old house and I kind of, like--it was at night and I looked up and in that, like, top floor window the light was on and I thought, you know, hopefully there is some kid in there staying up late at night playing with ChatGPT, figuring out how he or she is going to start whatever company comes next and whatever the next thing is after AI will happen here, too. That is, to me, the magic of this country. It is incredibly personally important and I hope it keeps going. Thank you very much for having me. ¶
Ms. Su ## 苏姿丰博士(Advanced Micro Devices(AMD)首席执行官兼董事长)的陈述 **苏博士(Dr. Lisa Su,AMD 董事长兼首席执行官):**克鲁兹主席、坎特韦尔资深成员、委员会各位成员:能就如此重要的议题来到这里,我深感荣幸。 我是 AMD 的董事长兼首席执行官。我们是一家总部设在美国的半导体公司,创立于56年前,为现代经济打造高性能计算芯片。 每天都有数十亿人依赖由我们的技术驱动的产品和服务,而我们的芯片对于支撑关键使命也极为重要,包括为国防系统和安全通信提供动力,以及促成突破性的科学研究。 不过我得说,我们最自豪的时刻,是看到出色的公私合作之时,而我们在超级计算领域的工作就是一个例子。 通过与能源部十多年的合作,AMD 如今为世界上最快的两台超级计算机提供动力:一台部署在橡树岭国家实验室(Oak Ridge National Labs),于2021年投入使用;另一台在劳伦斯利弗莫尔国家实验室(Lawrence Livermore National Labs),于去年年底刚刚投入运行。 从国家基础设施的角度看,这些系统确实至关重要,它们解决许许多多重大研究课题,也关乎国家安全和科学领先地位。 那么,说到 AI,你知道,关于 AI 已经有太多的论述。我真心感谢克鲁兹主席和坎特韦尔资深成员召开这次听证会。我认为这是一个极好的机会,来谈谈我们如何取胜。 AI 确实是我们这个时代最具变革性的技术。美国今天处于领先,但我想说的是,这是一场比赛。领先地位绝不是理所当然的。这是一场全球竞赛,它将塑造未来数十年国家安全和经济繁荣的结局。 而要保持我们的领先,实际上需要在技术栈的每一层都做到卓越,所以我也很荣幸能与同席的各位证人一起在这里。 我们与 Microsoft(微软)和 OpenAI 有着深度合作,这些合作表明:要取得成功,你需要芯片,需要软件,需要系统,以及——确实如此——应用层。 那么,至于该做什么,我思考了今天最重要的话是什么,并把它们归为五类。 我认为第一条、可能也是最重要的一条,是我们必须继续跑得更快。这是一场比赛,而比赛不会停滞。世界上没有人会原地不动。 我们今天领先,是因为我们做出了大胆的决策,是因为我们拥有的创新经济。但我们需要继续跑得更快,这意味着确保我们有可用的算力。 我认为萨姆关于阿比林(Abilene)的故事就是一个极好的例子:当你允许算力基础设施以私营部门希望的速度和节奏扩张时,你实际上会取得巨大的进展。 我还想谈谈开放生态系统的重要性。我认为开放生态系统确实是美国领先地位的基石,它让——你知道,坦率地说——想法能够从任何地方、从创新周期的每个环节涌现,降低进入门槛,增强安全性,并且坦率地说,为各种想法创造一个竞争性的市场。 第三,我们很高兴看到对强健的国内供应链的重视。对我们半导体行业来说,我们过去并没有得到这么多关注。 如今,由于芯片的重要性,我们得到了大量关注,而事实是我们需要在美国有更多的制造。迄今的努力取得了良好进展,但还有更多可以做的、也应当通过公私合作来做的事情。 第四,我们必须投资于人才。坦率地说,美国应当是学习 AI、从事 AI 工作、真正推进我们所需的一切创新的最佳地方,而我再次认为,这也可以通过大规模的公私合作来实现。 然后第五,当然,在出口管制领域,我们必须——我们作为一个行业完全理解国家安全的重要性,而这是,你知道,不言——作为一家美国公司,这是不言而喻的。 但我们也想确保——正如克鲁兹主席和坎特韦尔资深成员所说,让美国技术得到广泛采用非常重要。我们今天领先,是因为我们拥有最好的技术。 然而,如果我们不能让我们的技术在世界各地得到充分采用,就会有其他技术登场。它们今天也许不如我们的好,但坦率地说,使用确实会激发创新,而这无疑是我们需要通过公私合作来共同推进的事情。 最后,坦率地说,我想说,你知道,和萨姆一样,我成长过程中也有一台电脑。我在纽约长大。我比萨姆年长一点,所以我的第一台电脑是 Commodore 64,后来升级到了 Apple II。但事实是,这里是世界上做计算创新的最佳地方。我们希望它保持这样,拥有一个真正丰富而广泛的生态系统。 再次感谢今天给我来到这里的机会。 [苏博士的书面证词如下:] ### 苏姿丰博士(Advanced Micro Devices Inc.(AMD)董事长兼首席执行官)的书面证词 克鲁兹主席、坎特韦尔资深成员、委员会各位成员:感谢各位让我在如此关键的时刻有机会向各位发言。 我是 AMD 的董事长兼首席执行官。AMD 是一家总部设在美国的半导体公司,创立于56年前。我们打造为现代经济提供动力的高性能计算芯片。 每天,数十亿人依赖由 AMD 技术驱动的产品和服务。AMD 芯片还在推进我国许多最关键的使命中发挥着至关重要的作用,从为国防系统和安全通信提供动力,到促成突破性的科学研究、医疗创新和量子计算。 我们在超级计算领域的工作展示了 AMD 创新与公共影响力的全部实力。通过与能源部十多年的合作,AMD 如今为世界上最快的两台超级计算机提供动力:2021年在橡树岭国家实验室投入运行的「前沿」(Frontier),以及去年年底在劳伦斯利弗莫尔国家实验室投入运行的「埃尔卡皮坦」(El Capitan)。这些系统是美国国家安全和科学领先地位的关键基础设施,支撑着药物发现、医学研究、气候研究、高超音速飞行乃至训练未来几代能力更强的 AI 模型等方面的最新进展。 今天我们在这里讨论 AI。当今没有任何技术比 AI 更能展示高性能计算的力量。 AI 是我们这个时代最具变革性的技术。美国今天处于领先,但领先地位并非理所当然。这是一场全球竞赛,其结果将塑造未来数十年的经济增长、国家安全和技术影响力。 保持我们的领先需要在 AI 技术栈的每一层都做到卓越。AMD 与 Microsoft 和 OpenAI 的合作表明,行业领导者如何能够跨硬件、软件和系统携手合作,推进最先进的 AI。 每一个模型、每一项突破、每一个应用的背后,都是海量的算力。如果我们想在 AI 上领先,就必须在为其提供动力的基础设施上领先。这要求我们在五项国家优先事项上都拿出紧迫感。 第一,我们必须继续跑得更快。美国在行动迅速、敢于构想宏大时就能领先。从半导体到互联网,速度把大胆的美国创意变成了全球性产业。在 AI 领域,速度要求加快芯片和系统创新,以更高的效率交付更强的性能。这也意味着让 AI 算力基础设施在全行业随处可得。这将需要快速大规模建设数据中心,并以可靠、可负担且清洁的能源为其供电。跑得更快还意味着让 AI 走出云端。为确保每个美国人都能受益,AI 必须被内置到我们每天使用的设备中,变得像电力一样易于获取、可靠可用。从车辆和传感器到个人电脑和医疗工具,把 AI 的力量带给每一家企业、每一个美国人,将在最要紧的地方实现更快的决策、更智能的系统和更好的服务。我们今天拥有做到这一点的技术、知识产权和人才,但这是一场全球竞赛,我们必须不断加快步伐。 第二,我们必须倡导开放生态系统。开放标准长期以来一直是美国领先地位的基石。同样的方针也必须指引我们的 AI 之路。开放生态系统让来自不同供应商的硬件、软件和模型能够协同工作。这加速创新,降低进入门槛,通过透明度增强安全性,并创造更健康、更具竞争性的市场。 第三,我们必须为先进半导体制造和封装建立强健的国内供应链。AI 领先地位取决于构建完整、集成系统的能力。这意味着确保我们在最先进节点的晶圆制造和下一代封装技术两方面都拥有国内能力,以及把这一切整合起来所需的先进系统能力。AMD 很自豪能成为首批在 TSMC(台积电)亚利桑那州新晶圆厂生产尖端芯片的合作伙伴之一。迄今宣布的国内半导体制造项目代表了实质性的进展,但我们还能做更多。这是一个强有力的公私合作至关重要的领域。整个半导体行业一致认为,需要携手合作并与政府结成伙伴,大幅扩大美国本土的芯片生产和先进封装能力。 第四,我们必须投资于人才,并确保我们在 STEM(科学、技术、工程与数学)教育、劳动力培训和移民方面的国家战略支撑持续的 AI 领先地位。私营部门当然可以做得更多,包括扩大与大学的合作、投资于技能再培训项目,以及培养取得成功所需的跨学科人才。我们应当激励企业把最关键的 AI 研发工作更多地放在国内,并确保我们的移民政策吸引并留住世界上最优秀的 AI 人才。我们应当让美国成为全世界 AI 人才绝对最佳的去处。 第五,我们必须在国家安全需求与促进美国技术广泛采用的必要性之间取得平衡。当政府考虑「AI 扩散」(AI diffusion)之类的政策时,重要的是要记住:美国今天在 AI 上领先,我们希望世界其他地方在我们的平台上构建。如果我们的国际伙伴无法获取美国平台,他们就会采用替代方案——这些方案今天也许不那么先进,但会随时间成熟。要拿捏好这个分寸,需要政府与产业界更紧密的协作,确保规则清晰、一致,并同时兼顾竞争力与安全。 这是一个关键时刻,是确保美国在 AI 和先进计算领域领先地位的一代人仅有一次的机遇。这不仅仅关乎开发一项变革性技术,更关乎塑造我们经济的未来、捍卫我们的国家安全、增强我们的全球竞争力。 现在正是确保美国不只是跟上步伐,而是采取巩固我们领先地位所需的决定性步骤的时候。 再次感谢。我期待各位的提问。 **克鲁兹主席:**我也有一台 Apple II,还有一鞋盒的软盘,结果不知怎么的我拐错了弯,最后进了政界。因特拉托先生。 Chairman Cruz, Ranking Member Cantwell, members of the Committee, it is a real honor to be here on such an important topic. I am Chair and CEO of AMD. We are a U.S. headquartered semiconductor company founded 56 years ago and we build high- performance computing chips for the modern economy. Every day billions of people rely on our products and services powered by our technologies but our chips are also extremely important to support the critical missions including powering defense systems and secure communications as well as enabling breakthrough scientific research. I have to say our proudest moments, though, are when we see amazing public-private partnerships and our work in supercomputing is an example of that. Through more than a decade of partnership with the Department of Energy, AMD now powers the two fastest supercomputers in the world, one that is housed at Oak Ridge National Labs that was put into place in 2021 and the other at Lawrence Livermore National Labs that was just recently put into commission late last year. These systems are really critical from a national infrastructure standpoint and solve many, many large research issues as well as national security and scientific leadership. Now, in terms of AI, you know, there is so much that has been stated about AI. I really want to thank Chairman Cruz and Ranking Member Cantwell for having this hearing. I think it is a wonderful opportunity to talk about how we win. AI is truly the most transformative technology of our time. The United States leads today, but what I would like to say is it is a race. Leadership is absolutely not guaranteed. It is a global race that will shape the outcome of national security and economic prosperity for many decades to come. Now, maintaining our lead actually requires excellence at every layer of the stack so I am really honored to be here with my panelists as well. We have deep partnerships with Microsoft and OpenAI that demonstrate how you need silicon, you need software, you need systems and, really, the application layer to be successful. Now, in terms of what to do, I thought about what would be the most important things to say today and I put them in five categories. I think the first and probably the foremost is we must continue to run faster. This is a race and the race does not stand still. Nobody in the world stands still. We lead today because of the bold decisions that we have made and because of the innovation economy that we have. But we need to continue to run faster and that means ensuring that we have computing available. I think Sam's story about Abilene is an excellent example of how when you allow the computing infrastructure to expand at the rate and pace that the private sector wants, you actually make tremendous progress. I would also like to mention the importance of open ecosystems. I think open ecosystems are really a cornerstone of U.S. leadership and that allows, you know, frankly, ideas to come from everywhere and every part of the innovation cycle, reducing barriers to entry and strengthening security as well as creating, frankly, a competitive marketplace for ideas. Third, we are very happy to see the focus on a robust domestic supply chain. For us in the semiconductor world we used to not get so much attention. Now we get a lot of attention thanks to the importance of chips, and the fact is we need more manufacturing in the U.S. The efforts so far have made good progress but there is a lot more that can be done and that should be done in public-private partnership. Fourth, we must invest in talent. Frankly, the United States should be the best place to study AI, to work in AI, to really move forward all of the innovations that we need and I think, again, this can also be done in significant public- private partnership. And then fifth, of course, in the area of export controls we must--we totally understand as an industry the importance of national security and that is, you know, without--that goes without saying as a U.S. company. But we also want to ensure--as Chairman Cruz and Ranking Member Cantwell stated, it is important to have widespread adoption of U.S. technologies. We lead today because we have the best technology. However, if we are not able to fully have our technology adopted in the rest of the world there will be other technologies that will come to play. They may not be as good as we are today but, frankly, usage really spurs innovation and this is something that we certainly need to work with in public-private partnership. And I would, frankly, end by saying, you know, like Sam I had a computer when I was growing up. I grew up in New York. I am a little older than Sam so my first computer was a Commodore 64 and then I graduated to the Apple II. But the fact is this is the best place to do computing innovation in the world. We want it to stay that way with, really, a very rich and broad ecosystem. So thank you again for the opportunity to be here today. [The prepared statement of Ms. Su follows:] Prepared Statement of Dr. Lisa Su, Chair and CEO, Advanced Micro Devices Inc. (AMD) Chairman Cruz, Ranking Member Cantwell, Members of the Committee, thank you for the opportunity to speak with you at such a consequential moment. I am Chair and CEO of AMD, a U.S.-headquartered semiconductor company founded 56 years ago. We build high-performance computing chips that power the modern economy. Every day, billions of people rely on products and services powered by AMD technologies. AMD chips also play a vital role advancing many of our Nation's most critical missions, from powering defense systems and secure communications to enabling breakthrough scientific research, medical innovations, and quantum computing. Our work in supercomputing showcases the full strength of AMD's innovation and public impact. Through more than a decade of partnership with the Department of Energy, AMD now powers the world's two fastest supercomputers: Frontier, which went into operation at Oak Ridge National Labs in 2021, and El Capitan, which went into operation at Lawrence Livermore National Labs late last year. These systems are critical infrastructure for U.S. national security and scientific leadership, including the latest advances in drug discovery, medical research, climate research, hypersonic flight, and even training future generations of more capable AI models. Today we are here to talk about AI. No technology today better demonstrates the power of high-performance computing than AI. AI is the most transformative technology of our time. The United States leads today, but leadership is not guaranteed. This is a global race, and the outcome will shape economic growth, national security, and technological influence for decades to come. Maintaining our lead requires excellence at every layer of the AI stack. AMD's collaborations with Microsoft and OpenAI demonstrate how industry leaders can work together across hardware, software, and systems to advance state-of-the-art AI. Underneath every model, every breakthrough, and every application is massive amounts of computing power. If we want to lead in AI, we must lead in the infrastructure that powers it. That requires urgency across five national priorities. First, we must keep running faster. America leads when it moves fast and thinks big. From semiconductors to the internet, speed has turned bold American ideas into global industries. In AI, speed requires accelerating chip and system innovations that deliver more performance with greater efficiency. It also means making AI compute infrastructure readily available across the industry. This will require rapidly building data centers at scale and powering them with reliable, affordable, and clean energy sources. Moving faster also means moving AI beyond the cloud. To ensure every American benefits, AI must be built into the devices we use every day and made as accessible and dependable as electricity. From vehicles and sensors to PCs and medical tools, bringing the power of AI to every enterprise and every American will enable faster decisions, smarter systems, and better services where they matter most. We have the technology, intellectual property, and talent to do that today, but it is a global race and we must keep accelerating our pace. Second, we must champion open ecosystems. Open standards have long been a cornerstone of U.S. leadership. The same approach must guide our path with AI as well. Open ecosystems allow hardware, software, and models from different vendors to work together. This accelerates innovation, reduces barriers to entry, strengthens security through transparency, and creates healthier, more competitive markets. Third, we must build a robust domestic supply chain for advanced semiconductor manufacturing and packaging. AI leadership depends on the ability to build complete, integrated systems. That means ensuring we have domestic capabilities in both wafer manufacturing at the most advanced nodes and next-generation packaging technologies as well as the advanced system capabilities needed to bring it all together. AMD is proud to be one of the first partners producing leading-edge chips at TSMC's new fab in Arizona. The domestic semiconductor manufacturing projects announced to date represent meaningful progress, but there is much more that we can do. This is an area where strong public-private partnerships are critical. The entire semiconductor industry is aligned on the need to work together and partner with the government to significantly scale U.S. chip production and advanced packaging capabilities here at home. Fourth, we must invest in talent and ensure our national strategy for STEM education, workforce training, and immigration supports sustained AI leadership. The private sector can certainly do more, including expanding university partnerships, investing in reskilling programs, and developing the cross-disciplinary talent required for success. We should incentivize companies to increase their most critical AI R&D efforts here at home and ensure our immigration policies attract and retain the world's best AI talent. We should make America the absolute best place for AI talent in the world. Fifth, we must balance the need for national security with the imperative to enable the widespread adoption of U.S. technologies. As the government considers policies like AI diffusion, it is important to remember that the U.S. leads in AI today and we want the rest of the world building on our platforms. If our international partners cannot access U.S. platforms, they will adopt alternatives that may be less advanced today but will mature over time. Threading this needle requires closer collaboration between government and industry to ensure rules are clear, consistent, and aligned with both competitiveness and security. This is a pivotal moment. A once-in-a-generation opportunity to secure U.S. leadership in AI and advanced computing. This is not just about developing a transformative technology. It's about shaping the future of our economy, safeguarding our national security, and enhancing our global competitiveness. Now is the time to ensure the United States doesn't just keep up, but takes the decisive steps needed to cement our leadership. Thank you again. I look forward to your questions. The Chairman. And I had an Apple II as well with a shoe box of floppy disks and somehow I ended up taking a wrong turn and ending up in politics instead. Mr. Intrator. STATEMENT OF MICHAEL INTRATOR, CO-FOUNDER AND CHIEF EXECUTIVE OFFICER, COREWEAVE ¶
Mr. Intrator ## 迈克尔·因特拉托(CoreWeave 联合创始人兼首席执行官)的陈述 **因特拉托先生(Michael Intrator,CoreWeave 联合创始人兼首席执行官):**我是从一台 VIC-20 起步的。 克鲁兹主席、坎特韦尔资深成员、委员会各位尊敬的成员:感谢今天给我作证的机会。我很荣幸能与我的行业同仁和合作伙伴一同出席。 我叫迈克尔·因特拉托。我是 CoreWeave 的联合创始人兼首席执行官,公司创立于7年前。CoreWeave 和许多创新企业一样,起步时很简陋,在一间车库里做实验,最初是用图形处理器(GPU)来挖加密货币。 认识到其中的变革潜力后,我们转型去支持强大的 AI 应用,大幅扩展了愿景和运营规模。 今天,CoreWeave 站在美国 AI 基础设施革命的最前沿,在15个州运营着30多个数据中心。我们管理着超过25万(250,000)块 GPU,目前用电360兆瓦。 在短短两年里,我们的营收激增了12000%,2024年达到19亿美元($1.9 billion)。得益于这一进展,CoreWeave 于2025年3月28日成为一家上市公司。 CoreWeave 的快速增长不仅证明了这项技术,也证明了全球对先进 AI 基础设施激增的需求。我们的基础设施让美国企业能够迅速把 AI 愿景转化为有影响力的经济现实。 通过赋能企业加速创新,我们在增强美国的竞争优势,同时提升生产力和繁荣。 现代 AI 需要专门的基础设施,需要专门打造、在规模和性能上超越传统云计算的计算能力,而今天的通用云是为支持并扩展 AI 工作负载的复杂性而构建的。 我们不能用20世纪的基础设施来运行21世纪的经济。AI 工作负载涉及数万亿次同时进行的计算,需要前所未有的算力、先进的冷却系统、尖端的芯片技术、超高速网络和加速存储。 自2018年以来,先进 AI 模型所需的算力增长了大约10万倍。在 CoreWeave,【译注:跨批次残句,原文以「At CoreWeave」结束、无标点,句意未完,承接下一批次;按源件话轮边界保留,不补写。】 **因特拉托先生(Michael Intrator,CoreWeave 联合创始人兼首席执行官):** 【译注:本批次开头为上一批次末句的续接,上批原文止于「At CoreWeave」,本批自该句中段继续。】 我们的设施象征着美国伟大的创新传统。 我们的数据中心由技艺娴熟的美国工人建造、维护并配备运营人员,体现了现代技术如何不仅刺激经济增长、增强国家安全,也改善人类的生活。 我们正处于全球 AI 竞争的关键关头。在基础设施上领先的国家,将设定全球经济议程,并在未来数十年塑造人类的命运。 我们最大的竞争对手中国认识到其中的利害关系,正在投入大量资源以巩固其地位。我想重点谈四项政策要素,它们将有助于决定美国能否在 AI 竞赛中确保其领导地位。 第一,战略投资稳定性。AI 基础设施是高度资本密集型的,需要产业各利益相关方之间进行相当程度的协调。 稳定、可预期的政策框架,安全的供应链,以及鼓励创新的监管环境至关重要。政策制定者必须在所有司法管辖区提供清晰、一致的政策与法规,以支持长期投资和 AI 技术的快速规模化。 第二,能源基础设施建设。为支持 AI 基础设施的快速部署,美国必须确保充足且可负担的能源供应。有必要对许可审批和监管流程进行审慎改革,以加快基础设施项目,促进更快速的建设、并网互联以及数据中心的能源供应。 第三是全球市场准入。维持美国的领导地位,也意味着确保我们的技术能够公平地进入全球市场。 出口管制和贸易协定可以加以校准,既应对国家安全风险,又支持美国 AI 技术在全球的扩散。 最后,公私合作伙伴关系与劳动力发展。我们强大的公私合作传统,进一步增强了美国在 AI 竞赛中的独特优势。 CoreWeave 很自豪能与 Microsoft(微软)、普林斯顿大学以及新泽西州经济发展倡议共同创立新泽西 AI 枢纽(New Jersey AI hub)。 这样的倡议培养关键的劳动力技能、促进创新,并确保各经济体和社区为 AI 驱动的未来做好准备。 美国已准备好引领这场将带来巨大利益的 AI 革命。这是一个我们必须把握的难得历史时刻。 只要政府、产业界和所有相关方携手合作,美国就能赢得这场竞赛,抓住摆在我们面前的巨大机遇。 再次感谢给我作证的机会。我期待回答各位的问题。 [因特拉托先生的书面证词如下:] ## CoreWeave 联合创始人兼首席执行官迈克尔·因特拉托的书面证词 ### 引言 克鲁兹主席、坎特韦尔资深成员以及委员会各位尊敬的成员,我想感谢各位今天给我机会提供证词,谈谈我们如何确保美国在人工智能(AI)创新领域继续保持全球领先地位。 今天,我们正处于 AI 下一场伟大革命的边缘:这项技术正通过前所未有的复杂系统,以巨大的规模深刻重塑产业、创新与生产力。数百万小时的训练、数十亿次推理查询、数万亿的模型参数,以及持续的动态扩展,都在驱动一种近乎指数级、永不满足的算力与能源渴求。 在 CoreWeave,我们不仅在部署承载当今一些最庞大、最强大 AI 模型的基础设施,而且是在大规模地这样做。 我很感激有机会介绍 CoreWeave 的历程,以及我们在这场历史性竞赛中的角色。 ### CoreWeave 的角色 CoreWeave 通过弥合 AI 雄心与执行之间的鸿沟,为 AI 创新提供动力。CoreWeave 提供专为 AI 打造的云平台,交付释放 AI 全部潜力所需的速度、性能与专业能力。 为了训练下一代 AI 模型,AI 实验室需要算力资源。而为了保持在 AI 创新前沿的领先地位,研究人员要求最新、性能最强的计算基础设施。他们利用这些算力资源运行数万亿次运算,以执行算法、处理数据,训练下一代模型。 「缩放定律」[1]已经证明,增加投入模型的算力会转化为更好的性能。这种关系是指数级的。需要数量级增长的算力,才能解锁模型性能的增量提升。例如,2018年,最流行、最先进的生成式 AI 模型之一需要一定量的算力来训练。仅仅七年之后,到2025年,最新前沿模型所需的算力增长了约10万倍——一个惊人的增幅。[2]性能、实用性和现实世界的采用都大幅提升。这表明,AI 发展的限制边界在于高性能算力——以规模化方式交付、并以峰值效率运行。 [1] 缩放定律描述的是:随着训练数据规模、模型参数或计算资源的增加,AI 系统的性能如何提升。 [2] Epoch AI,《前沿 AI 模型的训练算力每年增长4–5倍》,2024年5月28日。 为了满足不断增长的算力需求、避免落后,企业和实验室持续要求以更大规模部署最新一代芯片。就在英伟达(Nvidia)最近两代芯片之间的短短两年里,训练性能从 Hopper 到 Grace Blackwell 提升了4倍。[3] [3] 英伟达,《NVIDIA GB200 NVL72:驱动计算的新时代》。 CoreWeave 使 AI 实验室、平台和企业能够不断向前推进算力的边界。我们的端到端云平台专为驱动 AI 创新、满足加速计算需求所需的规模、性能与专业能力而打造。我们建设数据中心并为其供电,配置高性能计算基础设施,使企业和实验室能够获取这些关键资源。我们的专长在于管理一个横跨供应链、能源、融资与技术合作伙伴关系的复杂而脆弱的生态系统,从而大规模地构建、优化和部署我们的平台。 截至2024年12月,我们在美国和欧洲的32个数据中心运营,部署了超过25万块GPU,使用360兆瓦(MW)的有功功率。我们现已签约获得1.3吉瓦(GW)的电力。充足、可靠的电力供应,对于推动这场革命、对于美国赢得这场竞赛至关重要。 今年,CoreWeave 是唯一获评铂金级(Platinum)、并经 SemiAnalysis 的铂金级 ClusterMAX™ 评级认定为 AI 云性能第一领导者的云服务商。[4]而且我们已经建立了这样的记录:率先将最新一代硬件推向市场,例如英伟达最新的 GB200 NVL72 芯片——IBM、Mistral AI 和 Cohere 等领先实验室已经在用它来改进和加速其训练任务。[5]我们也支持 OpenAI——我们近120亿美元($12B)的战略协议为其提供算力容量,用于训练其最新模型,并大规模交付给其全球数亿用户。[6] [4] CoreWeave,《CoreWeave 获评第一 AI 云,获 SemiAnalysis 铂金级 ClusterMAX™ 评级背书》,2025年4月10日。 [5] 英伟达,《数千块 NVIDIA Grace Blackwell GPU 现已在 CoreWeave 上线,推动 AI 先驱的开发》,2025年4月15日。 [6] CoreWeave,《CoreWeave 宣布与 OpenAI 达成协议,交付 AI 基础设施》,2025年3月10日。 CoreWeave 专为加速计算的需求而打造。我们以尖端的性能和规模交付这一基础设施,并提供 AI 今天和未来所需的基础设施及与之配套的专业能力。因此,我们的客户能够更早开始训练、更快构建、更快进入市场,这对美国保持 AI 领先至关重要。CoreWeave 是将推动美国在 AI 竞赛中前进的引擎。 ### CoreWeave 的历史 CoreWeave 的起点与许多初创公司和伟大的创业企业一样——在一间车库里,带着一个想法:尝试利用 GPU 进行加密货币挖矿。CoreWeave 的创始人们在2016年购买了第一块GPU,随后变成数百块、数千块,如今已是数十万块。在随后的几年里,CoreWeave 开始寻找机会,将其 GPU 集群用于加密货币挖矿之外的其他高性能用例,例如视觉特效(VFX),然后是 AI。 2020年,CoreWeave 以「全球首家专业化云」的定位正式发布。 2021年,CoreWeave 运营着北美最大的英伟达 A40 集群。 到2022年,世界开始意识到,扩展 AI 模型训练需要更多算力。CoreWeave 开始以更快的速度扩张。 到2023年,公司已有三个数据中心,运行超过1.7万块GPU,有功功率约10兆瓦(MW)。 到2024年,CoreWeave 已有十个数据中心,运行超过5.3万块GPU,有功功率超过70兆瓦。 而一年之后,到2024年底,我们已有32个数据中心,运行超过25万块GPU,有功功率约360兆瓦。 2025年3月28日,CoreWeave 成为一家上市公司。 这些进展发生在短短五年之内。而这正是推动这场技术革命所需的速度。最近,CoreWeave 完成了对领先 AI 开发者平台 Weights & Biases 的收购。我们的愿景是,CoreWeave + Weights & Biases 将交付领先的 AI 云平台——专为更快地开发、部署和迭代 AI 而打造。我们将共同实现更快、更高效的 AI 开发,赋能 AI 开发者快速创新,为 AI 开发提供无缝集成,并支持世界上最先进的 AI 创新者释放 AI 的全部潜力。 ### AI 基础设施与平台重构的需求 AI 需要的计算基础设施与现有基础设施有着根本不同。训练最先进的模型并大规模运行推理,需要跨数十亿参数进行数万亿次并行计算。为满足这一要求,高性能算力基础设施需要更集中的电力占用、更高的冷却需求、最新的芯片、高吞吐网络、加速存储等等。 相比之下,当今服务全球的通用云并不是为满足 AI 的特定需求而构建的。这些云平台是为日常网站托管、数据库管理和运行 SaaS 应用而构建的——这些工作负载依赖简单的固定逻辑计算和轻量级处理。 以 AI 的强度大规模运营算力是高度复杂的。运营 AI 工作负载存在显著的低效问题,从硬件故障到最优使用调度,不一而足。单个3.2万块GPU的集群,可能需要部署约600英里光纤线缆和约8万个光纤连接,还需要高度专业化的热管理能力来支持高功率密度。每一个变量都增加了潜在故障点的数量和复杂性。当集群发生组件故障(GPU、网络、内存、线缆、冷却等)时,可能通过降低训练性能对整个集群造成不利影响,甚至导致整个项目失败。 管理大型集群的困难,导致了我们所称的「AI 效率鸿沟」,我们基于模型浮点运算利用率(Model FLOPs Utilization,MFU)来评估它。它衡量的是观测到的吞吐量与系统在峰值容量下运行时的系统最大值之比。通常,管理 AI 基础设施的复杂性意味着 GPU 中的大部分算力容量可能因系统低效而损失,实证证据表明观测到的性能水平在35%至45%之间。 因此,世界正在经历一场从传统通用云计算基础设施向 AI 云计算基础设施的「平台重构」。为实现这一点,云平台正在被根本性地重新构想,技术栈的每一层都专门针对 AI 工作负载进行优化。这正是支持 AI 工作负载的规模与复杂性所需的专用计算基础设施。 ### CoreWeave 的云平台 我们针对 AI 云计算基础设施的新要求构建了我们的平台。 CoreWeave 的云平台是一个集成解决方案,专为以卓越的性能和效率运行模型训练与推理等 AI 工作负载而打造。它包括基础设施服务、托管软件服务和应用软件服务,所有这些都由我们专有的 Mission Control 和可观测性软件加以增强。这套专有软件支持基础设施的配置供应、工作负载的编排,以及对我们客户的训练与推理 【译注:原文在本批次边界处截断,末句未完(英文原文止于「…the monitoring of our customers' training and inference」),后续内容接下一批次。】 ……环境,以确保高可用性并最大限度减少停机时间。【译注:本段为承接上一批次(batch-004)的残句,原句完整为「This proprietary software enables the provisioning of infrastructure, the orchestration of workloads, and the monitoring of our customers' training and inference environments to ensure high availability and minimize downtime.」(这套专有软件支持基础设施的配置、工作负载的编排,以及对我们客户训练与推理环境的监控,以确保高可用性并最大限度减少停机时间。)】 为了充分释放 AI 基础设施的潜力,CoreWeave 帮助弥合 MFU(模型浮点运算利用率,Model FLOPs Utilization)的「效率差距」——即实测的 35%–45% 与理论上的 100% 之间的差距——实现比公开基准测试高出多达 20% 的性能。[7] 为此,性能优化被内置于平台的每一层,以提升分布式训练的吞吐量。而我们弥合这一差距的能力,显著增强了性能、改善了模型质量、加快了开发进度,并降低了 AI 模型的总体成本。 [7] CoreWeave:《CoreWeave 引领 AI 基础设施效率潮流,GPU 集群性能较替代方案最高提升 20%》("CoreWeave leads the Charge in AI Infrastructure Efficiency, with up to 20 percent Higher GPU Cluster Performance than Alternative Solutions"),2025年3月19日。 这对美国意味着什么?随着我们提升效率、弥合这一差距,美国将在全球 AI 竞赛中保持优势,刺激经济活动、增强国家安全,同时改善面向所有人的基本服务供给。这正是这场竞赛的全部意义所在。 #### 美国在 AI 领域的全球领导地位 我们正处于全球 AI 竞赛的一个关键拐点,这是一个将影响经济繁荣、国家安全、技术标准以及我们如何向全体美国人提供基本服务的关键时刻。AI 代表着技术的下一次重大演进,具有变革社会的潜力。这是美国的 AI 时刻,是美国不容错失的战略机遇。 经济繁荣:据预测,到 2030 年,AI 将产生累计 20 万亿美元($20 trillion)的全球经济影响,相当于全球 GDP 的 3.5%。[8] 引领这场变革的国家将在这条新的经济前沿中获取不成比例的巨大份额。如果美国保持其在 AI 领域的全球领导地位,AI 所创造的生产率提升、新产品、高价值就业岗位,以及从医疗保健到制造业等各行各业的突破,将有助于推动整个美国经济的繁荣,惠及所有人。 [8] IDC:《人工智能对经济与就业的全球影响》("The Global Impact of Artificial Intelligence on the Economy and Jobs"),2024年9月。 国家安全:随着先进 AI 能力成为现代国防不可或缺的一部分——包括武器系统与战场能力的改进、情报以及网络安全系统——保持美国的技术优势已与我们的国家安全密不可分。当其他国家正快速推进自身的 AI 能力、明确以挑战美国的全球经济与军事领导地位为目标时,落后绝不是一个选项。 塑造 AI 的未来:引领 AI 发展的国家将塑造这项技术在全球的演进方式。未来治理 AI 的标准、协议与伦理框架,将反映赢得这场竞赛的那个国家的价值观。 今天正在建设的 AI 基础设施底座,将有助于决定 AI 开发在何处发生。在全球 AI 竞赛中取胜,将越来越依赖于专门构建的 AI 计算基础设施,而不仅仅是大规模部署的通用系统。成功实现「重新平台化」(re-platform)的国家,将在模型能力与开发速度上获得复利式的优势。 CoreWeave 正处于开发这类专用基础设施的前沿,为美国的 AI 能力提供动力。IBM、Mistral、Cohere 等领先企业与 AI 实验室都依赖 CoreWeave 的基础设施。我们的成功通过确保美国保有驱动 AI 所需的全球最先进计算基础设施,支撑了更广泛的国家目标。 #### 美国持续保持 AI 基础设施领导地位的关键因素 美国在 AI 领域的领导地位,部分取决于能否保持其在驱动 AI 的底层基础设施上的优势。基于 CoreWeave 建设和运营 AI 计算基础设施的经验,我想重点指出几个将决定我们国家能否保持领导地位的关键领域。其中许多领域聚焦于将影响这一行业如何演进的关键政策要素。 ##### 战略投资的稳定性 由于这些项目的规模与时间跨度,AI 基础设施投资需要多个行业与政府利益相关方之间的高度协调,代表着开发周期与运营周期长达数年的巨额资本投入。CoreWeave 受益于与领先芯片制造商、原始设备制造商(「OEM」)及软件供应商的稳固合作,由它们向我们供应基础设施组件及其他产品。释放 AI 潜力所需的高度专业化基础设施,其建设与运营极具挑战,在大规模场景下尤其如此。这需要:(i) 数以万计的 GPU;(ii) 数千英里的高速网络线缆;(iii) 数十万个互连连接,汇聚成用于训练和运行 AI 模型的「超级集群」(superclusters);以及 (iv) 数百兆瓦(MW)的电力和大量存储。 要维持美国在 AI 领域的领导地位,美国的 AI 云计算公司必须能够持续接入可靠的供应链,以获取开发和运行尖端 AI 基础设施所需的全部组件,这一点至关重要。获取这些为 AI 工作负载提供动力所必需的高性能组件,需要管理复杂的全球供应链并维持稳固的供应链关系。与全球领先供应商及战略伙伴的持续合作,对于确保美国 AI 基础设施的持续运营、扩张与快速部署,以及维护美国的竞争力,都至关重要。可预期的政策对此不可或缺。 大规模的私营部门投资与开发,已帮助美国在 AI 基础设施领域确立了早期且重要的领先优势。美国约占全球数据中心容量市场的 40%,全球前十大市场中有六个位于美国。[9] [9] Cushman & Wakefield Research:《2024 年全球数据中心市场比较》("2024 Global Data Center Market Comparison")。 我们的竞争对手并非没有意识到 AI 及美国领导地位的重要性。全球围绕 AI 基础设施的竞争日趋激烈,要求我们必须审慎而积极地维护这一初始领先优势。欧盟于 4 月启动了其「AI 大陆行动计划」(AI Continent Action Plan),设定了在未来五到七年内将成员国数据中心容量增至三倍的雄心目标。欧盟还宣布向五座「千兆工厂」(gigafactories)——配备约 10 万颗最先进 AI 芯片的大型高性能计算设施——投资 200 亿欧元,并推进与审批许可、能源问题及用水相关的改革。[10]【译注:原文金额作「=20 billion」,其中「=」为欧元符号(€)在 GPO 排印中的变体,即 200 亿欧元(€20 billion)。】中国则通过协调一致的国家战略和精简的部署时间表——能把数年压缩为数月——明确展现了其在 AI 领域的雄心,力图缩小美国当前的 AI 领先优势。 [10] 欧盟委员会(European Commission):《AI 大陆行动计划》("The AI Continent Action Plan"),2025年4月9日,https://digital-strategy.ec.europa.eu/en/library/ai-continent-action-plan。 世界各国正在积极推进协调一致的 AI 战略,实施补贴基础设施、加快部署时间表的政策。在这一利害攸关的环境中,美国企业能否迅速且有把握地建设 AI 基础设施,将成为 AI 竞赛中的决定性因素,并决定美国能否保住其领导地位。 美国在 AI 基础设施领域的持续领导地位,面临来自多重不确定性的潜在逆风。其中包括关键组件(如先进半导体与网络设备)全球供应链的波动,这可能打乱部署时间表。这些波动可能导致延期或意料之外的成本超支,对美国企业快速扩展 AI 能力的能力造成不利影响。 联邦与州两级监管的变化,给作出投资决策的领先企业带来了巨大的不确定性。出口管制与能源政策的变化、为满足增长的需求而可能需要新增吉瓦(GW)级电力的潜在需求,以及各级政府正在形成的 AI 专项监管格局,也都影响着基础设施部署的速度与规模。监管缺乏清晰度可能抑制投资、拖慢创新周期。此外,美国企业还将受到世界各地正在形成的、将治理 AI 使用的规则与机构的影响——既包括各个国家内部,也包括主要竞争对手参与的重要论坛。作为全球占主导地位的经济体与技术领导者,美国政府必须主导那些塑造美国企业未来竞争环境的规则,这一点至关重要。 最后,监管框架碎片化的可能性——各州要求不一,联邦层面也可能存在差异——构成了一项独特挑战。例如,各司法辖区对界定各类 AI 活动的关键术语定义不一致,可能迫使企业为基本相同的活动去应对一张复杂的合规制度网络。这类政策将要求 AI 基础设施的参与者考虑设计替代性的产品与策略,以便在不同司法辖区开展业务。这种监管拼凑将导致成本上升、运营低效,并最终使美国企业在全球 AI 竞赛中处于竞争劣势。这些不确定性对 CoreWeave 这样的新进入者及其他专业化供应商的影响尤为严重,可能扼杀的恰恰是驱动美国在这一关键行业保持领导地位的创新本身。 要确保美国始终处于 AI 前沿,美国企业必须引领 AI 基础设施的发展。这需要一项协调一致的政策战略,以缓解关键的不确定性(本节已论及其中许多)、维持适当的监督,并营造一个稳定、可预期的政策环境,促进投资、持续增长与创新。 ##### 能源与基础设施发展:为 AI 领导地位提供动力 建设 AI 基础设施的竞赛,从根本上取决于国家能否持续开发驱动创新的新一代数据中心、使其上线运行,并确保有足够的电力为其供电。这将受到以下流程与审批许可制度的影响:用于开发数据中心的流程与制度,用于开发充足、可靠的新增发电供应的流程与制度,以及能够按所需速度和规模输送电力的并网与输电系统。 新增电力供应与精简基础设施发展流程这双重挑战,不仅仅是后勤层面的障碍,更是决定美国能否保持其全球 AI 领导地位的关键因素。若不能有效应对这些挑战,就有可能把阵地拱手让给国际竞争对手——尤其是正积极推进自身 AI 雄心的中国。 能源考量对 AI 基础设施的开发与运营至关重要。据分析,在经历了电力消费长期相对持平的时期之后,美国目前正经历电力需求显著且不断加速的增长。这一激增由若干同时发生的趋势驱动,包括新制造设施的回流本土(onshoring)、交通与供暖的广泛电气化,以及数据中心的增长。 AI 计算是能源密集型的。训练大语言模型、运行复杂模拟、部署 AI 应用,都需要大量电力。即使企业持续创新、提升效率,AI 的广泛采用仍将进一步推高这一需求。根据美国能源部发布的一份报告,2023 年数据中心消耗了全美总电力的约 4.4%。预计这一数字在未来几年还将上升,到 2028 年可能消耗全美总电力的 6.7% 至 12%。[11] 这一预期增长凸显了各级政府政策制定者的紧迫任务:出台政策,使新增电力供应及输送电力的基础设施得以发展。鉴于这些项目可能耗资数亿美元、耗时数年才能建成,启动工作刻不容缓。 [11] 《2024 年美国数据中心能耗报告》("2024 United States Data Center Energy Usage Report"),劳伦斯伯克利国家实验室(Lawrence Berkeley National Laboratory),2024年12月。 这对全球 AI 领导地位的影响是明确而重大的。能够提供充足、可靠且具成本效益的能源的地区,将吸引数十亿美元的 AI 基础设施投资。反之,能源约束——无论是供应有限、输送不可靠、政策不确定,还是成本高得令人却步——都能够而且将会把开发及随之而来的投资推向……【译注:原文至此为残句,句子延续至下一批次(batch-006)开头的「elsewhere」,即「把开发及相关投资推向别处」。】 ……推向别处。【译注:本批次开头为跨批次残句,承接上一批次末尾「…can and will push development and the associated investment」(……可以并且必将把开发活动及相关投资推向……),此处为该句的结尾部分,按源件边界以残句保留,不补写。】 CoreWeave 在数据中心选址方面的考量正体现了这些优先事项: - 充足、可靠的电力供应,并在条件具备时采用无排放能源 - 有竞争力的电价 - 多样化的能源来源 - 容量扩张的路径 - 能够提供时间确定性的高效许可审批流程 然而,获得建设能源基础设施和数据中心所需的各项审批,往往是一个关键瓶颈。在一个创新速度以周而非年计的领域,每延误一个月都意味着丧失阵地。 尤其是能源项目的开发充满挑战。取得必要的批准和许可可能耗费大量时间。许可审批流程中的低效会严重影响能源供应,也影响有吸引力的数据中心开发选址能否推进。各辖区许可审批时限的差异,以及可能出现的多重、依次进行的审查程序和诉讼,会拉长项目开发所需的时间,导致延误,甚至可能使项目停摆。 要精简开发能源和数据中心基础设施所需的许可审批与监管流程——这些基础设施是美国保持其 AI 领先地位所必需的——将会面临挑战。精简这些流程的目标包括: - 维持并扩大由多样化能源驱动的均衡发电组合,以满足不断增长的需求,确保以合理成本实现电力可用性与可靠性 - 扩建并现代化全国输电系统 - 为数据中心容量及相关基础设施的开发商提供可预期的时间表,缩短可行性研究、并网接入和建设的等待时间 - 在保持适当监督的同时精简许可审批流程 CoreWeave 理解,为实现这些重要目标而设立的流程,需要考虑将作出这些投资的关键参与方以及这些设施所在社区的意见。 我们希望,为促成新增发电来源及输送电力的输电基础设施而精简许可审批流程的努力,能够得到本届国会的关注。[12] ---------- [12] CoreWeave 是数据中心联盟(Data Center Coalition,DCC)的成员。DCC 就《AI 行动计划》提交的意见材料中包含关于许可审批和能源基础设施的更多讨论,可在此处获取。[https://static1.squarespace.com/static/63a4849eab1c756a1d3e97b1/t/67d84a70db36cf08e2a329cb/1742228107114/DCC+Comments+-+RFI+AI+Action+Plan.pdf] ---------- 我们认识到,这个问题不可能仅在联邦层面得到解决。各级政府都应在应对新增必要基础设施以满足能源需求的挑战中发挥作用。需要联邦、州和地方政府、产业界及其他受影响各方协同努力,应对这些相互关联的挑战,包括建立高效、透明的流程,使基础设施能够以技术进步所要求的速度和规模建设,从而让美国保持其在 AI 领域的主导地位。 #### 美国 AI 技术栈的全球扩散 全球市场准入是决定哪个国家将在 AI 领域领先的关键因素。出口管制和贸易协定可以被设计用来实现多重目标:既能为美国企业促成正当的市场准入,又能缓解潜在的国家安全风险。然而,未经校准的管制可能适得其反,通过激励在美国境外进行 AI 开发和部署、让竞争对手填补美国企业留下的空白,反而助长外国竞争者。这可能导致技术专长和经济效益的流失。 为巩固美国的 AI 领导地位,出口管制和国际协定应考虑: - **精准定位国家安全风险:**管制应聚焦于对国家安全构成真实且可证实威胁的技术、实体和国家,并设定清晰、具体的参数。 - **支持美国技术领导地位:**对美国技术及其出口目的地施加的限制,应考虑其对美国企业在全球市场竞争能力的负面影响。这包括考虑其他国家采取报复措施的可能性,以及对投资和创新造成「寒蝉效应」的风险。 - **与盟友的战略协同:**与志同道合的国际伙伴密切协调,对于确保出口管制的有效性、防止全球 AI 市场碎片化至关重要。与盟友保持一致还能促进一个协作、安全、可信的 AI 生态系统的扩展。 这些考量对于塑造 AI 创新的未来格局至关重要。校准得当的方针将确保下一代 AI 发展扎根于美国,依托美国的基础设施和专长。 反之,管制也可能无意中产生限制美国技术和专长出口机会的效果,带来不利影响。一项仔细区分不同市场、针对缓解特定风险量身定制出口限制、并促进国际合作的战略,能够在有效保护国家安全的同时,增强美国引领和塑造全球 AI 扩散竞赛的能力。 #### 加速美国创新的公私伙伴关系 美国在 AI 竞赛中的一项独特优势,是我们能够在政府、产业界、学术界和公民社会各方之间建立有效的伙伴关系。这些合作将私营部门的敏捷、创造力、专长和资源,与公共部门的长远愿景以及学术和研究机构的基础与应用研究能力结合起来。这种方式有助于培育一个竞争对手难以复制的创新生态系统。 CoreWeave 很荣幸与 Microsoft(微软)、普林斯顿大学(Princeton University)和新泽西州经济发展局(New Jersey Economic Development Authority)一道,成为新泽西 AI 枢纽(New Jersey AI Hub)的创始合作伙伴。该 AI 枢纽将聚焦研发工作、AI 在若干行业部门的应用,以及 AI 劳动力培养与教育。[13] ---------- [13] 菲尔·墨菲州长(Governor Phil Murphy),《墨菲州长、普林斯顿大学、Microsoft 与 CoreWeave 为重大人工智能枢纽剪彩》(Governor Murphy, Princeton University, Microsoft & CoreWeave Cut Ribbon on Major Artificial Intelligence Hub),2025年3月27日,https://www.nj.gov/governor/news/news/562025/approved/20250327a.shtml。 ---------- CoreWeave 深切致力于支持 AI 教育与研究,以及公私协作伙伴关系。全国各地类似的伙伴关系可以进一步加速美国的 AI 能力,我们鼓励政策制定者探索这种协作模式。这类伙伴关系还能加速美国 AI 和数据中心劳动力的培养。随着技能缺口扩大,数据中心行业的用工短缺正变得普遍。数据中心雇主一直难以找到足够多的受过培训的工人。2020 年接受调查的数据中心管理者中,有一半即 50% 报告难以找到熟练工人填补职位;到 2023 年,仍有 71% 继续报告对找到合格员工感到担忧。[14] 一支熟练、训练有素的劳动力队伍,对于 AI 数据中心的稳定与扩张至关重要——这些数据中心依赖专业的数据中心技术员、网络和电气工程师、网络安全专业人员以及项目经理。CoreWeave 支持培养国内劳动力的各项努力,造就满足不断增长的 AI 需求所需的熟练工人,在加速 AI 创新的同时,创造未来高薪的技能型岗位。 ---------- [14] 白宫,《AI 人才报告》(AI Talent Report),2025年1月14日。 ---------- #### 展望未来:确保美国在 AI 领域的持续领导地位 美国凭借创新型和创业型私营企业、世界一流研究机构、才华横溢的劳动力队伍以及促进蓬勃增长的政策环境这一独特组合,在人工智能领域建立了显著领先优势。这一优势在 AI 基础设施领域尤为突出,CoreWeave 这样的公司已在 AI 技术栈的这一关键层确立了全球技术领导地位。 然而,造就这一地位的各项条件必须得到积极维护,同时保持灵活,以适应新的技术发展和政治考量。世界各国都正确地认识到 AI 的战略重要性,正在协同努力建设 AI 基础设施。我们现在作出的决定,将有助于决定美国能否在未来岁月保持 AI 领导地位。 当前时刻需要一种深思熟虑、透明且可预期的方针,在保持我们竞争优势的同时,抓住未来机遇并回应正当关切。在考虑应对这一充满活力的行业的政策选项时,我们应当关注不同方针对整个 AI 生态系统的影响——从成熟参与者到新进入者,从模型开发商到 CoreWeave 这样的基础设施提供商。 为进一步扩大美国的 AI 领先优势,我们鼓励联邦政府考虑以下政策: - 通过对增强竞争力和驱动创新最为关键的领域实施全国一致的监管框架,营造可预期的投资环境。 - 通过增强新增发电和输电能力、为下一代 AI 基础设施供电的政策,确保以尽可能低的成本获得充足、可靠的电力供应。为此,需要对现有许可审批和监管流程进行审慎改革,使受影响各方能够参与流程。 - 通过经过战略校准的出口政策,在保护国家安全的同时支持美国 AI 技术栈的扩散,从而保持全球竞争力并壮大美国产业。 - 加强像新泽西 AI 枢纽这样的公私伙伴关系,加速跨越研究、产业和政府的创新。 CoreWeave 感谢委员会在这些关键议题上的领导作用,我们期待与委员会合作,协助其制定使美国能够在 AI 竞赛中保持领先地位的政策。 **克鲁兹主席(Ted Cruz,得克萨斯州,委员会主席):**谢谢。史密斯先生。 I started out with a VIC-20. Chairman Cruz, Ranking Member Cantwell, and distinguished members of the Committee, thank you for the opportunity to testify today. I am honored to appear alongside my industry colleagues and partners. My name is Michael Intrator. I am the Co-Founder and CEO of CoreWeave founded 7 years ago. CoreWeave started like many innovative ventures, humbly in a garage, experimenting, initially with graphics processing units, or GPUs, for cryptocurrency mining. Recognizing the transformational potential, we pivoted to support powerful AI applications, dramatically scaling the vision and operation. Today, CoreWeave stands at the forefront of America's AI infrastructure revolution, operating more than 30 data centers across 15 states. We manage more than 250,000 GPUs currently using 360 megawatts of power. Over two short years our revenue has surged by 12,000 percent, reaching $1.9 billion in 2024. As a result of this progress, CoreWeave became a publicly traded company on March 28th of 2025. CoreWeave's rapid growth is a testimony not only to the technology but also to the surging global demand for advanced AI infrastructure. Our infrastructure enables American businesses to rapidly translate AI aspirations into impactful economic realities. By empowering companies to accelerate innovation we are fueling America's competitive edge while improving productivity and prosperity. Modern AI requires specialized infrastructure, purpose built computing capabilities that surpass traditional cloud computing in scale and performance, today's general purpose cloud that was built to support and scale the complexity of AI workloads. We cannot run a 21st century economy on the 20th century's infrastructure. AI workloads involve trillions of simultaneous calculations demanding unprecedented computing power, advanced cooling systems, cutting-edge chip technology, ultra high-speed networks and accelerated storage. Since 2018 the computing power necessary for advanced AI models has multiplied approximately 100,000 fold. At CoreWeave our facilities symbolize America's great tradition of innovation. Our data centers built, maintained, and staffed by skilled American workers embody how modern technology not only stimulates economic growth and enhances national security, but also improves humans' lives. We are at a critical juncture in the global AI competition. The nation that leads in infrastructure will set the global economic agenda and shape human outcomes for decades. Our largest competitor, China, recognizes the stakes and is spending significant resources to strengthen their position. I want to focus on four elements of policy that will help determine whether the U.S. secures its leadership role in the AI race. First, strategic investment stability. AI infrastructure is deeply capital intensive and requires a significant level of coordination across industry stakeholders. Stable predictable policy frameworks, secure supply chains, and regulatory environments that foster innovation are crucial. Policy makers must provide clear and consistent policy and regulations across all jurisdictions that enables long-term investment and rapid scaling of AI technology. Second, energy infrastructure development. To support the rapid deployment of AI infrastructure America must ensure abundant and affordable supplies of energy. Careful reforms and permitting and regulatory process are necessary to accelerate infrastructure projects, facilitate more rapid construction, interconnections, and energy for data centers. Third is global market access. Maintaining America's leadership also means ensuring our technology has fair access to global markets. Export controls and trade agreements can be calibrated to both address national security risks and support global diffusion of American AI technology. And, finally, public-private partnerships and workforce development. America's unique advantage in the AI race is enhanced by our powerful tradition of public-private partnership. CoreWeave is proud to co-found the New Jersey AI hub with Microsoft, Princeton University, and the New Jersey economic development initiative. Initiatives like this develop critical workforce skills, foster innovation, and ensure economies and communities are prepared for the AI-driven future. America stands ready to lead the AI revolution, which will bring enormous benefits. It is a rare moment in time that we must meet. If government, industry, and all affected parties work together the United States can win this race and seize the vast opportunity ahead of us. Thank you again for the opportunity to testify. I look forward to answering your questions. [The prepared statement of Mr. Intrator follows:] Prepared Statement of Michael Intrator, Co-Founder and Chief Executive Officer, CoreWeave Introduction Chairman Cruz, Ranking Member Cantwell, and distinguished Members of the Committee, I want to thank you for the opportunity to provide testimony today on how we can ensure the United States remains the global leader in Artificial Intelligence (AI) innovation. Today we're on the verge of the next great revolution of AI: a technology dramatically reshaping industries, innovation, and productivity at a massive scale through systems of unprecedented complexity. Millions of hours of training, billions of inference queries, trillions of model parameters, and continuous dynamic scaling are all driving an insatiable hunger for compute and energy that borders on exponential. At CoreWeave, we are not only deploying the infrastructure that hosts some of the most massive and powerful AI models in existence, but also doing it at scale. I appreciate the opportunity to describe CoreWeave's journey and our role in this historic race. CoreWeave's Role CoreWeave powers AI innovations by bridging the gap between AI ambition and execution. CoreWeave provides the cloud platform, purpose- built for AI, that delivers the speed, performance, and expertise needed to unleash AI's full potential. To train their next generation of AI models, AI labs require compute resources. And to maintain their leadership position at the forefront of AI innovation, researchers demand the latest and most- performant computing infrastructure. They leverage these compute resources to run the trillions of operations that operate algorithms and process data to train the next generation of models. The ``Scaling Law'' \1\ has demonstrated that increasing the compute deployed against models translates to better performance. The relationship is exponential. Orders of --------------------------------------------------------------------------- \1\ Scaling laws describe how the performance of AI systems improve as the size of the training data, model parameters, or computational resources increases. --------------------------------------------------------------------------- magnitude of increased compute are required to unlock incremental gains in model performance. For example, in 2018, one of the most popular and advanced generative AI models required a certain amount of compute to train. Just seven years later, in 2025, the amount of compute needed for the latest frontier models grew by about 100,000 times--an extraordinary increase.\2\ Performance, usefulness, and real- world adoption has increased dramatically. This demonstrates that the limiting boundary of AI development is high-performance compute, delivered at scale and operated at peak efficiency. --------------------------------------------------------------------------- \2\ Epoch AI, ``Training Compute of Frontier AI Models Grows by 4- 5x per Year,'' May 28, 2024. --------------------------------------------------------------------------- To access the increased need for compute and to avoid being left behind, enterprises and labs are continuously demanding deployments of the latest generation chips at larger scales. In just the two short years between the most recent generations of Nvidia chips, training performance increased by 4x from Hopper to Grace Blackwell.\3\ --------------------------------------------------------------------------- \3\ Nvidia, ``NVIDIA GB200 NVL72: Powering the new era of computing.'' --------------------------------------------------------------------------- CoreWeave enables AI labs, platforms, and enterprises to move the boundaries of compute forward. Our end-to-end cloud platform is purpose-built for the scale, performance, and expertise needed to power AI innovation and meet the demands of accelerated computing. We construct and power data centers, and provision high-performance computing infrastructure which enables enterprise and labs to access these essential resources. Our expertise is in managing a complex and fragile ecosystem across supply chain, energy, financing, and technology partnerships to build, optimize, and deploy our platform at scale. As of December 2024, we operated in 32 data centers in the U.S. and Europe, deploying more than 250,000 GPUs utilizing 360MW of active power. We have now contracted for 1.3GW of power. Adequate, reliable supplies of power are essential to drive this revolution and for the U.S. to win this race. This year, CoreWeave was the sole cloud to be ranked Platinum and as the #1 leader in AI cloud performance, as attributed by SemiAnalysis's Platinum ClusterMAX TM Rating.\4\ And we have established a track record as among the first to market with the latest generations of hardware, such as Nvidia's most recent GB200 NVL72 chip, which leading labs, such as IBM, Mistral AI, and Cohere, are already using to improve and accelerate their training jobs.\5\ We also support OpenAI--our strategic deal of nearly $12B provides compute capacity for training and delivering its latest models at scale to its hundreds of millions of users around the world.\6\ --------------------------------------------------------------------------- \4\ CoreWeave, ``CoreWeave Ranks as #1 AI Cloud, Backed by SemiAnalysis' Platinum ClusterMAX TM Rating,'' April 10, 2025. \5\ Nvidia, ``Thousands of NVIDIA Grace Blackwell GPUs Now Live at CoreWeave, Propelling Development for AI Pioneers,'' April 15, 2025. \6\ CoreWeave, ``CoreWeave Announces Agreement with OpenAI to Deliver AI Infrastructure,'' March 10, 2025. --------------------------------------------------------------------------- CoreWeave is purpose-built for the demands of accelerated computing. We deliver this infrastructure with cutting-edge performance and scale, and provide the expertise with the infrastructure that AI needs today and in the future. And as a result, our customers are able to train earlier, build quicker, and get to market faster, which is critical for the U.S. to maintain its lead in AI. CoreWeave is the engine that will propel the U.S. forward in the AI race. CoreWeave's History CoreWeave began as many start-ups and great entrepreneurial companies do--in a garage with an idea, which was to try leveraging GPUs for crypto mining. CoreWeave's founders purchased their first GPU in 2016, which turned into hundreds, then thousands, and now hundreds of thousands. Over the course of the next few years, CoreWeave began looking for opportunities to use its fleet of GPUs for other high- performing use cases beyond crypto mining, such as visual effects (VFX), and then AI. In 2020, CoreWeave launched as the ``world's first specialized cloud.'' In 2021, CoreWeave operated the largest Nvidia A40 fleet in North America. By 2022, the world began to realize that more compute was required to scale AI model training. CoreWeave began to scale even more rapidly. By 2023, the company had three data centers running more than 17,000 GPUs with approximately 10 megawatts (MWs) of active power. By 2024, CoreWeave had ten data centers running more than 53,000 GPUs with more than 70MW of active power. And, one year later, at the end of 2024, we had 32 data centers running more than 250,000 GPUs with approximately 360MW of active power. CoreWeave became a publicly traded company on March 28, 2025. This progress occurred in five short years. And that is the speed which is required to drive this technological revolution. Most recently, CoreWeave completed its acquisition of Weights & Biases, a leading AI developer platform. Our vision is that CoreWeave + Weights & Biases will deliver the leading AI Cloud Platform--purpose-built to develop, deploy, and iterate AI faster. Together, we will enable faster, more efficient AI development, empower AI developers to innovate quickly, provide seamless integrations for AI development, and support the world's most advanced AI innovators to unleash AI's full potential. The Need for AI Infrastructure and Re-platforming AI requires a fundamentally different computing infrastructure from the existing one. Training state-of-the-art models and running inference at scale requires trillions of simultaneous calculations across billions of parameters. To fulfill this requirement, high- performing compute infrastructure necessitates a more concentrated power footprint, increased cooling needs, the latest chips, high- throughput networking, accelerated storage, and more. In contrast, the generalized clouds that serve the world today were not built to serve the specific requirements of AI. These cloud platforms were built for day-to-day web hosting, database management, and running SaaS applications--workloads that rely on simple, fixed- logic calculations and lightweight processing. Operating compute at scale and at the intensity of AI is highly complex. There are significant inefficiencies associated with operating AI workloads ranging from hardware failures to scheduling optimal usage. A single 32,000 GPU cluster may require the deployment of approximately 600 miles of fiber cables and approximately 80,000 fiber connections, along with highly-specialized heat management capabilities to support high-power density. Each of these variables increases the number and complexity of possible failure points. When a cluster suffers a component failure (GPU, network, memory, cable, cooling, etc), it can adversely impact the entire cluster by reducing training performance, or even causing the entire project to fail. The difficulty in managing large clusters leads to what we call the ``AI Efficiency gap,'' which we evaluate based on Model FLOPs Utilization (MFU). This is a measure of the observed throughput compared to system maximum if the system were operating at peak capacity. Typically, the complexity of managing AI infrastructure means that a majority of the compute capacity in GPUs can be lost to system inefficiencies, with empirical evidence suggesting observed levels of performance in the 35 percent to 45 percent range. As a result, the world is undergoing a ``re-platforming'' from traditional generalized cloud computing infrastructure to AI cloud computing infrastructure. And to achieve this, cloud platforms are being fundamentally reimagined, with every layer of the technology stack being specifically optimized for AI workloads. This is the purpose-built computing infrastructure needed to support the scale and complexity of AI workloads. CoreWeave's Cloud Platform We have built our platform for the new requirements of AI cloud computing infrastructure. CoreWeave's cloud platform is an integrated solution that is purpose-built for running AI workloads such as model training and inference at superior performance and efficiency. It includes infrastructure services, managed software services, and application software services, all of which are augmented by our proprietary Mission Control and observability software. This proprietary software enables the provisioning of infrastructure, the orchestration of workloads, and the monitoring of our customers' training and inference environments to ensure high availability and minimize downtime. To unlock the full potential of AI infrastructure, CoreWeave helps to bridge the MFU ``efficiency gap'' between the observed 35-45 percent and the theoretical 100 percent, driving as much as 20 percent higher performance than public benchmarks.\7\ To achieve this, performance optimizations are built into every layer of the platform to enhance distributed training throughput. And our ability to close this gap significantly enhances performance, improves model quality, accelerates development timelines, and reduces overall AI model costs. --------------------------------------------------------------------------- \7\ CoreWeave, ``CoreWeave leads the Charge in AI Infrastructure Efficiency, with up to 20 percent Higher GPU Cluster Performance than Alternative Solutions,'' March 19, 2025. --------------------------------------------------------------------------- What does this mean for the U.S.? As we improve our efficiency, and close this gap, the United States will maintain its edge in the global AI race, stimulating economic activity, and enhancing national security while improving the provision of essential services for all. This is what the race is all about. U.S. Global Leadership in AI We stand at a critical inflection point in the global AI race, representing a pivotal moment that will influence economic prosperity, national security, technological standards and how we provide essential services to all Americans. AI represents the next major evolution of technology with the potential to transform society. This is America's AI moment, and a strategic opportunity America cannot afford to miss. Economic Prosperity: AI is projected to generate a cumulative global economic impact of $20 trillion, representing 3.5 percent of global GDP, by 2030.\8\ The country that leads this transformation will capture a disproportionate share of this new economic frontier. If America maintains global leadership in AI, the productivity gains, new products, high-value jobs, and breakthroughs across industries from healthcare to manufacturing created by AI will help drive prosperity across the American economy benefitting all people. --------------------------------------------------------------------------- \8\ IDC, ``The Global Impact of Artificial Intelligence on the Economy and Jobs,'' September 2024. --------------------------------------------------------------------------- National Security: As advanced AI capabilities become essential to modern defense including improvements to weapons systems and battlefield capabilities, intelligence, and cybersecurity systems, maintaining America's technological edge becomes inseparable from our national security. Falling behind is not an option when other countries are rapidly advancing their own AI capabilities with explicit aims to challenge American global economic and military leadership. Shaping the Future of AI: The country that leads AI development will shape how this technology evolves globally. The standards, protocols, and ethical frameworks that will govern AI will reflect the values of whichever country wins this race. The foundational AI infrastructure being built today will help determine where AI development occurs. Success in the global AI race will increasingly depend on purpose-built AI computing infrastructure, not just general-purpose systems deployed at scale. Nations that successfully ``re-platform'' gain compounding advantages in model capabilities and development speed. CoreWeave is at the forefront of developing the purpose-built infrastructure that powers America's AI capabilities. Leading companies and AI labs such as IBM, Mistral, and Cohere rely on CoreWeave's infrastructure. Our success supports broader national objectives by ensuring the U.S. maintains the world's most advanced computing infrastructure which is required to drive AI. Factors Critical to Continued U.S. Leadership in AI Infrastructure America's leadership position in AI depends in part on maintaining its edge in the underlying infrastructure that drives it. Based on CoreWeave's experience building and operating AI computing infrastructure, I would like to highlight several critical areas that will determine if our Nation maintains its leadership position. Many of these areas focus on the critical elements of policy which will impact how this sector evolves. Strategic Investment Stability AI infrastructure investment requires a significant level of coordination across multiple industry and government stakeholders due to the scale and timeline of these projects, representing substantial capital commitments with years-long development and operational horizons. CoreWeave benefits from robust collaborations with leading chipmakers, original equipment manufacturers (``OEMs''), and software providers to supply us with infrastructure components and other products. The highly specialized infrastructure that is required to unlock the potential of AI is immensely challenging to build and operate, especially at scale. This requires: (i) tens of thousands of GPUs, (ii) thousands of miles of high-speed networking cables, (iii) hundreds of thousands of interconnects coming together to create ``superclusters'' for training and serving AI models, and (iv) hundreds of MWs of power and substantial amounts of storage. To sustain U.S. leadership in AI, it is important for U.S. AI cloud computing companies to maintain access to a reliable supply chain necessary to access all of the components necessary to develop and run the cutting-edge AI infrastructure. Acquiring these necessary high- performance components to power AI workloads requires managing a complex global supply chain and maintaining robust supply chain relationships. Continued engagement with leading global suppliers and strategic partners is vital to ensuring the continued operation, expansion, and rapid deployment of U.S. AI infrastructure and to uphold U.S. competitiveness. Predictable policy is essential for this. Significant private sector investment and development has helped the United States establish an early and important lead in AI infrastructure. The U.S. accounts for roughly 40 percent of the global market for data center capacity, with six of the top ten markets.\9\ --------------------------------------------------------------------------- \9\ Cushman & Wakefield Research, ``2024 Global Data Center Market Comparison.'' --------------------------------------------------------------------------- The importance of AI and U.S. leadership is not lost on our competitors. Intensifying global competition for AI infrastructure demands that this initial lead must be carefully and actively maintained. The European Union launched its AI Continent Action Plan in April setting out ambitious goals to triple data center capacity across member states in the next five to seven years. The EU also announced a =20 billion investment into five gigafactories-massive high performance computing facilities equipped with approximately 100,000 state-of-the- art AI chips--and reforms related to permitting, energy issues and water usage.\10\ China has made its ambitions regarding AI clear through coordinated national strategies and streamlined deployment timelines that can compress years into months in their effort to shrink America's current AI lead. --------------------------------------------------------------------------- \10\ European Commission, ``The AI Continent Action Plan,'' April 9, 2025, https://digital-strategy.ec.europa.eu/en/library/ai-continent- action-plan. --------------------------------------------------------------------------- Countries around the world are aggressively pursuing coordinated AI strategies, implementing policies which subsidize infrastructure, and accelerate deployment timelines. In this high-stakes environment, the capacity of American companies to build AI infrastructure swiftly and with assurance will be a decisive factor in the AI race and determine whether the United States retains its leadership position. Sustained American leadership in AI infrastructure faces potential headwinds from multiple sources of uncertainty. These include volatility in the global supply chain for critical components, such as advanced semiconductors and networking equipment which can disrupt deployment timelines. These fluctuations can lead to delays or unanticipated cost overruns, adversely affecting American companies' ability to rapidly scale AI capabilities. Changes in regulation at both the Federal and state levels introduce substantial uncertainty for leading businesses making investment decisions. Changes in export controls, energy policies, the potential need to add gigawatts of power to meet increased demand, and the emerging landscape of AI-specific regulations at different levels of government also affect the pace and scope of infrastructure deployment. The lack of regulatory clarity can deter investment and slow down the innovation cycle. Additionally, American companies will be affected by the rules and institutions being developed around the world both in individual nations and in important forums in which key competitors participate that will govern the use of AI. As the world's dominant economic player and technological leader, it is important for the U.S. government to drive the rules which shape the future playing field for American companies. Finally, the potential for a fragmented regulatory framework, with differing requirements from state to state and potentially at the Federal level, poses a unique challenge. For instance, inconsistent definitions of key terms which define various AI activities across jurisdictions could force companies to navigate a complex web of compliance regimes for fundamentally similar activities. These types of policies would require participants in the AI infrastructure to consider designing alternative products and strategies to do business in different jurisdictions. This regulatory patchwork would lead to increased costs, operational inefficiencies, and ultimately, a competitive disadvantage for American companies in the global AI race. These uncertainties disproportionately affect newer entrants like CoreWeave and other specialized providers, potentially stifling the very innovation that drives American leadership in this critical sector. To ensure the United States remains at the forefront of AI, American companies must lead AI infrastructure development. This requires a coordinated policy strategy to mitigate key uncertainties, many of which this section touches upon, maintain appropriate oversight, and create a stable, predictable policy environment that fosters investment, continued growth and innovation. Energy and Infrastructure Development: Powering AI Leadership The race to build AI infrastructure is fundamentally tied to the Nation's ability to continue to develop a new generation of data centers that drive innovation, to bring them online and ensure there is sufficient electricity to power them. This will be affected by the processes and permitting systems that are used to develop data centers and to develop adequate, reliable supplies of new power generation and the interconnection and transmission systems capable of delivering it at pace and scale. The dual challenges of adding new power supplies and streamlining infrastructure development, are not merely logistical hurdles, but critical factors that will determine whether America can maintain its global AI leadership. Failure to address these challenges effectively risks ceding ground to international competitors, particularly China, who are aggressively pursuing their own AI ambitions. Energy considerations are critical to the development and operation of AI infrastructure. After a prolonged period of relatively flat electricity consumption, according to analysis, the U.S. is now experiencing a significant and accelerating increase in power demand. This surge is driven by several concurrent trends, including the onshoring of new manufacturing facilities, the widespread electrification of transportation and heating, and the growth in data centers. AI computation is energy-intensive. Training large language models, running complex simulations, and deploying AI applications all require significant amounts of power. Widespread AI adoption will further increase this demand, even as companies continue to innovate and improve efficiency. According to a report released by the U.S. Department of Energy, data centers consumed approximately 4.4 percent of total U.S. electricity in 2023. This figure is projected to rise in the coming years, potentially consuming between 6.7 percent and 12 percent of total U.S. electricity by 2028.\11\ This projected increase underscores the urgent need for policymakers at all levels of government to put policies in place that will enable the development of new power supplies and the infrastructure to deliver it. Given that these projects can cost in the hundreds of millions of dollars and years to implement, there is no time to lose in getting started. --------------------------------------------------------------------------- \11\ ``2024 United States Data Center Energy Usage Report,'' Lawrence Berkeley National Laboratory, December 2024. --------------------------------------------------------------------------- The implications for global AI leadership are clear and consequential. Regions that can provide abundant, reliable, and cost- effective energy will attract billions of dollars of AI infrastructure investment. Conversely, energy constraints, whether in the form of limited supply, unreliable delivery, policy uncertainty, or prohibitive costs, can and will push development and the associated investment elsewhere. CoreWeave's site selection consideration for data centers illustrates these priorities: Availability of abundant, reliable power, and where available, non-emitting sources Competitive rates Diverse energy sources Pathways for capacity expansion Efficient permitting processes that provide timeline certainty However, obtaining the necessary approvals to build both energy infrastructure and data centers is often a critical bottleneck. Every month of delay represents lost ground in a field where the pace of innovation is measured in weeks, not years. In particular, it is challenging to develop energy projects. Securing the necessary approvals and permits can take a significant amount of time. Inefficiencies in the permitting process can significantly impact both energy availability and whether attractive sites for data center development can move forward. The variability in permitting timelines across jurisdictions and the potential for multiple, sequential review processes and litigation can increase the time required to develop a project leading to delay or potentially stopping projects. There will be challenges in streamlining the permitting and regulatory processes required to develop the energy and data center infrastructure necessary for the U.S. to maintain its leadership in AI. Goals in streamlining these processes include: Maintaining and growing a balanced portfolio of generation powered by diverse energy sources that can meet increasing demand to ensure availability and reliability at reasonable costs Expanding and modernizing the Nation's transmission systems Providing developers of data center capacity and associated infrastructure with predictable timelines and reduced wait times for feasibility studies, interconnections, and builds Streamlining permitting processes while maintaining appropriate oversight CoreWeave understands that the processes put in place to achieve these important objectives need to consider the views of key players that will make these investments and the communities in which these facilities are located. We hope efforts to streamline the permitting process to enable the addition of new sources of generation and the transmission infrastructure to deliver it will receive attention by this Congress.\12\ --------------------------------------------------------------------------- \12\ CoreWeave is a member of the Data Center Coalition. DCC's AI Action Plan submission includes additional discussion related to permitting and energy infrastructure, available here. [https:// static1.squarespace.com/static/63a4849eab1c756a1d3e97b1/t/ 67d84a70db36cf08e2a3 29cb/1742228107114/DCC+Comments+-+RFI+AI+Action+Plan.pdf] --------------------------------------------------------------------------- We recognize that this issue will not be resolved solely at the Federal level. All levels of government have a role to play in addressing the challenges in adding the necessary infrastructure to meet energy requirements. A coordinated effort amongst federal, state, and local government, industry, and other affected parties is required to address these interrelated challenges which include creating efficient, transparent processes that allow infrastructure to be built at the pace and scale that technological advancement requires and for the U.S. to maintain its dominant position in AI. Global Diffusion of the American AI Stack Global market access is a pivotal factor in determining which nation will lead in the AI domain. Export controls and trade agreements can be designed to achieve multiple objectives: they can facilitate legitimate market access for American businesses while also mitigating potential national security risks. However, controls that are not calibrated can inadvertently bolster foreign competitors by incentivizing AI development and deployment outside of the U.S., and competitors filling the void left by U.S. firms. This could result in the loss of technological expertise and economic benefits. To bolster American AI leadership, export controls and international agreements should consider: Precision Targeting of National Security Risks: Controls should be focused on technologies, entities and nations that pose genuine and demonstrable threats to national security, with clear and specific parameters. Supporting American Technological Leadership: Restrictions imposed on U.S. technologies and where they can be exported should consider negative impacts on the ability of U.S. companies to compete in global markets. This includes considering the potential for retaliatory measures from other nations and the risk of creating a 'chilling effect' on investment and innovation. Strategic Alignment with Allies: Close coordination with like-minded international partners is essential to ensure the effectiveness of export controls and prevent the fragmentation of the global AI market. Aligning with allies can also foster the expansion of a collaborative, secure, and trustworthy AI ecosystem. These considerations are crucial in shaping the future landscape of AI innovation. A well-calibrated approach will ensure that the next generation of AI development is anchored in the United States, leveraging American infrastructure and expertise. Conversely, controls could inadvertently have the effect of limiting opportunities for the export of U.S. technology and expertise, with adverse impacts. A strategy that carefully differentiates between markets, tailors export restrictions to mitigating specific risks, and fosters international cooperation, can effectively protect national security while simultaneously enhancing America's ability to lead and shape the global AI diffusion race. Public-Private Partnerships Accelerating American Innovation A unique American advantage in the AI race is our ability to forge effective partnerships between government, industry, academia and elements of civil society. These collaborations combine the agility, ingenuity, expertise and resources of the private sector with the long- term vision of the public sector and the basic and applied research capabilities of academic and research institutions. This approach helps foster an innovation ecosystem that is difficult for competitors to replicate. CoreWeave is proud to be a founding partner of the New Jersey AI Hub, along with Microsoft, Princeton University, and the New Jersey Economic Development Authority. This AI Hub will focus on research and development efforts, applications of AI in several industry sectors, and AI workforce development and education.\13\ --------------------------------------------------------------------------- \13\ Governor Phil Murphy, ``Governor Murphy, Princeton University, Microsoft & CoreWeave Cut Ribbon on Major Artificial Intelligence Hub,'' March 27, 2025, https://www.nj.gov/governor/news/news/562025/ approved/20250327a.shtml. --------------------------------------------------------------------------- CoreWeave is deeply committed to supporting AI education and research, and public-private collaborative partnerships. Similar partnerships across the country could further accelerate America's AI capabilities, and we encourage policymakers to explore this model of collaboration. These types of partnerships also accelerate AI and data center workforce development in the US. Worker shortages in the data center industry are becoming commonplace as the skills gap widens. Data center employers have struggled to find enough trained workers. Half or 50 percent of surveyed data center managers in 2020 reported having difficulty finding skilled workers to fill positions, and 71 percent continued to report being concerned about finding qualified staff in 2023.\14\ A skilled and trained workforce is vital for the stability and expansion of AI data centers--which rely on specialized data center technicians, network and electrical engineers, cybersecurity professionals, and project managers. CoreWeave supports efforts to develop a domestic workforce comprised of the skilled workers required to meet the growing AI demand and to accelerate AI innovation while creating the skilled good-paying jobs of the future. --------------------------------------------------------------------------- \14\ The White House, ``AI Talent Report,'' January 14, 2025 --------------------------------------------------------------------------- Looking Forward: Ensuring Continued American Leadership in AI The United States has built a remarkable lead in artificial intelligence through our unique combination of innovative and entrepreneurial private companies, world-class research institutions, a talented workforce, and a policy environment that fosters dynamic growth. This advantage is especially pronounced in AI infrastructure, where companies like CoreWeave have established global technological leadership in this critical layer of the AI stack. However, the conditions that have enabled this position must be actively maintained while being flexible in order to adjust to new technological developments and political considerations. Countries around the world rightfully recognize the strategic importance of AI and are making coordinated efforts to build AI infrastructure. The decisions we make now will help determine whether America can maintain AI leadership in the years to come. The current moment demands a thoughtful, transparent, and predictable approach that maintains our competitive edge, and seizes future opportunities while addressing legitimate concerns. As we consider policy options to address this dynamic sector, we should be attentive to how different approaches affect the entire AI ecosystem, from established players to new entrants, from model developers to infrastructure providers like CoreWeave. In order to further America's lead in AI, we encourage the Federal Government to consider policies that: Foster a predictable investment environment through the implementation of nationally consistent regulatory frameworks for areas most critical to strengthen competitiveness and drive innovation. Ensure that there are adequate, reliable supplies of power at the lowest possible cost through policies which enhance the ability to add generation and transmission to power next- generation AI infrastructure. Careful reforms of existing permitting and regulatory processes that enable affected parties to participate in the process are needed for this to occur. Maintain global competitiveness and strengthen U.S. industry through strategically calibrated export policies that protect national security while supporting the diffusion of the American AI stack. Strengthen public-private partnerships, like the New Jersey AI Hub, that accelerate innovation across research, industry, and government. CoreWeave appreciates the Committee's leadership on these critical issues and we look forward to working with the Committee as it develops policies enabling the U.S. to maintain its leadership in the AI race. The Chairman. Thank you. Mr. Smith. STATEMENT OF BRAD SMITH, VICE CHAIR AND PRESIDENT, MICROSOFT CORPORATION ¶
Mr. Smith ## 布拉德·史密斯(Microsoft 副董事长兼总裁)的陈述 **史密斯先生(Brad Smith,微软副董事长兼总裁):**克鲁兹主席、坎特韦尔资深成员、委员会各位成员,感谢你们给我今天来到这里的机会。我就接着我的三位同事所说的,谈几点想法。 第一点,我想请大家看这个画架上的图表,它展示了 AI 技术栈,我的书面证词里也有这张图。它说明了一个简单但我认为重要的道理——我们大家都同舟共济。 如果美国要成功引领世界 AI,就需要基础设施,需要在平台层面取得成功,需要创造应用程序的人。 有意思的是,我们 Microsoft 得以与这三位领导者及其公司共事。我们的成功——我们每一家的成功——都依赖于彼此的成功。适用于我们四家的道理,放眼全国和全世界也同样成立:看看那些开源开发者、那些建设发电厂的人、那些每天上班的电工和管道工。 那么,为了成功,我们需要国会和国家提供什么?我认为是三件事。我在书面证词中描述过。 首先,正如克鲁兹主席所说,我们需要创新。正如萨姆所说,有了更多基础设施、更快的许可审批、更多电工,创新就会更快。 正如坎特韦尔资深成员所说,我们需要更多创新,其动力来自我们的大学以及支持全国基础研究的联邦机构所给予的支持——这是这个国家的皇冠明珠之一。 正如克鲁兹主席所说,我们还需要更快的采用,也就是人们所说的 AI 扩散——让 AI 在美国经济的每一个部分发挥作用,以提高生产率、促进经济增长、让人们在工作中创新。历史一再表明,实现这一点的首要要素是技能培训,是投资教育。 最后,我们需要出口。如果美国要引领世界,我们就需要与世界连接。我们需要记住——我认为要永远记住——全世界只有百分之四点五的人口生活在美国。 我们的全球领导地位,依赖于我们以正确的出口管制方针服务世界的能力,也依赖于我们始终——尤其在技术领域——维持世界其他地区信任的能力。 归根结底,我想,那些花时间——如果他们花时间观看或阅读这场听证会——的人可能会问:这到底是为了什么?我们这桌人到底想做什么?AI 这两个字母对他们究竟意味着什么? 我们这些在这个行业工作的人,是在试图制造比人更强的机器,还是在试图制造【译注:本批次到此截断,原文句子未完(…or are we trying to build),接续见下一批次,按残句保留,不补写。】 **史密斯先生:** ……帮助人们变得更好的机器?【译注:本批次承接上一批次史密斯先生口头陈述的末句,原文自此残句开始。】 断然地说,答案是后者,也必须是后者。我们是在试图建造在人们今天所做的所有工作中都胜过人类的机器,还是在试图建造帮助人们在未来追求更好的工作、乃至更有趣的职业生涯的机器? 毋庸置疑,必须是第二种,而不是第一种,而且我相信这正是我们正在共同做、也能够共同做到的事情。 作为一个如今已在这个行业度过近32年的人,有两件事总是让我感触良多。第一件大概不会让你们惊讶。 永远不要低估技术能做什么、它能以多快的速度发展、它能成就什么。但第二件,我认为讨论得太少了,尽管它每天都摆在我们眼前。 永远不要估量人们能做什么。永远不要低估人类的抱负。永远不要低估一个人如果获得更好的技术工具、并有机会学习如何使用它,所能做到的事情。【译注:「永远不要估量」原文为「Never estimate」,疑为「永远不要低估」之笔误,此处按原文直译。】 这就是这个行业的故事。这也是这个国家的故事。正如你们所听到的,这也是萨姆·奥特曼(Sam Altman)的故事。不是每个人都能成为萨姆·奥特曼、萨提亚·纳德拉(Satya Nadella)或比尔·盖茨(Bill Gates),但每个人都值得拥有尝试的机会。 今晚,在全美各地,无论是在房屋的阁楼、地下室,还是一间普通的卧室里,都有孩子拥有电脑、拥有手机、能够接入互联网,并且现在能够让 AI 为他们所用。 让我们投资于他们的教育。让我们投资于美国公众所需要的技能。然后,让我们投资于创造美国公众应得的未来。 谢谢。 [史密斯先生的书面证词如下:] ### 微软公司副董事长兼总裁布拉德·史密斯(Brad Smith)的书面证词 克鲁兹主席、坎特韦尔资深成员及委员会各位成员: 感谢有机会就人工智能这一关键议题作证。我是布拉德·史密斯,微软公司(Microsoft Corporation)副董事长兼总裁。 AI 有潜力成为人类有史以来发明的最有用的工具。与此前的通用目的技术——如电力、机床和数字计算——一样,AI 将影响我们经济的每一个部分。从现在到本世纪中叶,它不仅将塑造我们工作和生活的方式,还将塑造我们作为一个国家如何竞争、繁荣并保持安全。 本次听证会的通知恰如其分地提到了「AI 竞赛」。今天我想谈谈赢得这场竞赛需要什么。 AI 竞赛既涉及技术,也涉及经济。它既需要创新,也需要扩散。它既是短跑,也是马拉松。一个国家可能赢下一圈,却输掉整场竞赛——如果它未能把成功所需的全部要素汇聚起来的话。 这是一场没有任何公司或国家能够独自获胜的竞赛。 要赢得 AI 竞赛,美国需要在 AI 技术栈的每一层支持私营部门。国家需要与世界各地的美国盟友和朋友合作。 在今天的证词中,我将聚焦于三个战略优先事项,本届国会和联邦政府将在这些方面发挥重要作用。 第一,国家必须赢得 AI 创新竞赛。这需要大规模的数据中心和 AI 基础设施,而它们需要联邦支持来扩建和现代化其所依赖的电网。国家必须招募和培训电工、管道工等供不应求的技术工人。我们大家都必须召集国家实验室和大学中最优秀的研究人员,并依托已成为世界羡慕对象的联邦基础研究计划和合作伙伴关系给予支持。我们需要继续擅长把创新想法从学术实验室带入公司和新产品。我们还需要支持 AI 开发者,让他们能够开放、广泛地获取公共数据。 第二,国家必须赢得 AI 扩散竞赛。这要求我们推动 AI 的广泛采用,从而带动经济各部门的生产率增长。最重要的是,这需要新的举措来推动美国劳动力的 AI 技能培训。这将包括在我们学校中的 AI 基础素养教育,以及我们社区学院中新的 AI 培训项目。还将包括代表下一代计算机科学学位的高等 AI 教育、将在全国商学院中研习掌握的组织技能,以及全国法学院的新课程。这些结合起来,将使企业、非营利组织和政府机构都能有效使用 AI。联邦、州和地方各级政府然后可以通过采用 AI 服务来提高其向公众提供的服务的有效性和效率,从而帮助加速这一扩散。 第三,美国必须向美国的盟友和朋友出口 AI。没有任何公司或国家强大到可以没有朋友就掌控 AI 的未来。美国和中国不仅在竞争创新,还在竞争把各自的技术传播到其他国家。竞赛的这一部分很可能由最快的先行者赢得。美国需要一个明智的出口管制战略,既保护我们的国家安全,又向其他国家保证它们能够可靠、持续地获得关键的美国 AI 组件和服务。也许不亚于任何其他因素的是,这要求我们共同维持国际社会对我们的产品、我们的公司以及这个国家本身的信任。 #### 作为通用目的技术的人工智能 经济学家有时把技术分为两类:通用目的技术和单一用途工具。世界上大多数东西都是单一用途工具,比如烟雾探测器或割草机。它们把一件事做得非常好。但在历史进程中,某些所谓的通用目的技术会影响、有时甚至重新定义经济的几乎每一个部门。电力是最典型的例子,因为只要想一想就会发现,电力改变了每个经济部门的运作方式。 掌控 AI 未来的关键,部分始于理解技术在过去所扮演的角色。过去三个世纪给世界带来了三次工业革命,每一次都由这些通用目的技术驱动。第一次是英国的铁加工,始于18世纪。然后是19世纪的电力和机床,当时美国通过比任何其他国家都更广泛地使用这些技术而超越了英国。再就是过去50年间由计算机芯片和软件驱动的第三次工业革命。 毫无疑问,在推进通用目的技术方面成为全球领导者,会给一个国家带来重大优势。但历史的一个教训是,受益最多、前进最快的国家,未必是技术被发明出来的国家,而是技术被扩散——或者说被采用——得最快、最广泛的国家。这是有充分理由的。如果一项技术能提高生产率并改变经济的每个部分,那么最广泛、最迅速地使用它的国家将受益最多。 这既框定也定义了美国面临的 AI 机遇与挑战。作为一个国家,我们需要同时专注于推进创新和推动扩散——既在国内,也作为一项领先的美国出口。 #### AI 技术栈 同时推动创新与扩散的关键,在于认识到 AI 与所有通用目的技术一样,建立在我们业内所称的「技术栈」之上——即一组配合使用的技术。每一项伟大的通用目的技术都是如此。例如,如果我们回到过去,想想电力,就能看到这一点。托马斯·爱迪生(Thomas Edison)于1878年首次成功地用电点亮了灯泡。但要让灯光照亮整座城市,很快就需要建设发电厂、提供运转所需的燃料、创建电网、统一电路标准,以及灯泡本身之外的各种电器。简言之,一个电力的技术栈。 人工智能同样建立在一个 AI 技术栈之上。从根本上说,它分为三层:基础设施层、平台层和应用层。下图展示了这一点。 [图表无法以 TIFF 格式提供] 基础设施层规模庞大。微软本财年在这一层所需的资本投资上支出超过800亿美元($80 billion),其中一半以上花在美国。这些资金用于购置土地、投资电力和宽带连接、采购 GPU 等芯片,以及安装液冷系统。这些投入带来数据中心的建设——往往是由多栋建筑组成、可能容纳数十万台计算机的数据中心园区。这些基础设施既支持新 AI 模型的训练,也支持其部署,使它们能够用于世界各地基于 AI 的服务。 在基础设施之上是平台层。这一层的核心是 AI 基础模型,包括 OpenAI 等公司创建的前沿模型,以及来自众多其他公司的开源模型和其他模型——包括 Anthropic、Google、Mistral、DeepSeek 和微软自身。平台层依靠数据来训练模型并为其提供事实依据。它还包括新一代基于软件的 AI 平台服务,用于帮助构建 AI 应用。 最终,基础设施层和平台层共同支撑应用层。应用层是利用 AI 为人们提供更好服务的设备和软件应用。ChatGPT 和微软的 Copilot 都是 AI 应用的例子。应用层令人惊叹的一点在于,创造 AI 应用的不只是公司——无论大小、无论老牌还是初创——而是每一个人。是研究人员使用融入 AI 的新应用来改变药物发现。是非营利组织改变其提供服务的方式。是教师把 AI 当作工具,改进他们准备课堂材料的方式。是政府让从申报纳税到驾照续期的一切事务变得更简单、更高效。 要建设新的 AI 经济,关键是让这三层都运转起来,并让整个生态系统的飞轮转动起来。必须建设基础设施层,人们才能在平台层开发和部署模型。必须使用 AI 模型,人们才会在其之上构建应用。必须让客户采用这些应用,市场才能增长,并带动更多投资以进一步扩建基础设施。这个过程不断重复。一个新经济就是这样诞生的。 #### 成功需要整个生态系统 飞轮效应清楚地表明,成功不仅需要国家在技术栈的某一层取得进展,而是需要在每一层都取得进展。这正是私营部门目前在美国所追求的,而且做得比任何其他国家都好。这也是本届国会和行政部门可以通过一项贯穿整个技术栈上下、同时促进 AI 创新与扩散的战略来支持的事情。 国家层面的 AI 领导地位不仅需要少数公司的成功,而是需要众多公司的成功。今天的证人小组包括 OpenAI、AMD、CoreWeave 和微软等领先企业,反映了新 AI 经济的几个重要侧面。AI 经济需要一个多层面、一体化的生态系统,包括「大型科技公司」(Big Tech)和「小型科技公司」(Little Tech)、初创企业和较成熟的企业、开源和专有开发者、供应商和客户、产生数据的公司和使用数据的公司,所有各方协同合作。政府作为监管者和领先的 AI 采用者,都有关键角色要发挥。 评论者有时聚焦于这个技术生态系统中不同参与者之间的紧张关系。这些值得关注。但常被忽视的是,不同参与者也相互依赖。这意味着 AI 生态系统的各个贡献者都需要保持健康。 像微软这样的大型科技公司拥有独特的机会——也是责任——与技术栈每一层的参与者合作并给予支持。我们努力推进的不仅是创新,还有有助于 AI 市场长期增长的经济架构、商业模式和负责任的实践。不仅为了美国,也为了这个国家的朋友和盟友。 #### 赢得创新竞赛 尽管 AI 经济主要由私营部门建设,政府政策和举措需要发挥关键作用。这始于为创新提供动力所需的工作。有几个领域值得本次听证会特别关注。 ##### 为数据中心的增长提供电力 正如没有发电厂家里就不可能有可靠的电力一样,没有数据中心和 AI 基础设施就不可能有 AI。而这些数据中心的建设需要庞大的供应链,运转需要大量电力。 美国先进的经济依赖已有50年历史的基础设施,它无法满足 AI、制造业回流和电气化程度加深所驱动的日益增长的电力需求。美国需要投资更多输电设施和能源资源,把供应链迁回国内,并现代化电网,以支撑预测中的电力负荷增长。微软自身正在这些领域投资。 我们敦促联邦政府精简联邦许可审批流程,以加速所有这些领域的增长。当前的联邦许可审批流程往往涉及多个机构以及复杂的、 【译注:原文此处为残句,以逗号结束,句子延续至下一批次。】 【译注:本批次为史密斯先生(Brad Smith,微软副董事长兼总裁)陈述的续段,承接 batch-007;开头为上一批次末尾句子的后半部分,按残句译出。】 ……不可预测、长达数年的审查。这阻碍了进展。联邦政府应立即采取措施,为许可审批决定确立可靠、合理且透明的时间表。这也可以通过标准化联邦许可审批流程、并指定一个牵头机构来引导许可走完整个流程来实现。此外,许可审批机构应利用AI和数字工具,为申请人改善审批时限与透明度,并确保审批机构能够快速获取信息,以协助其审查和决策过程。 我们很高兴看到特朗普总统近期签署的行政命令《更新21世纪的许可审批技术》(Updating Permitting Technology for the 21st Century),该命令指示各机构在环境审查和许可审批过程中最大限度地利用技术。国会还应关注「联邦—州现代电网部署倡议」(Federal-State Modern Grid Deployment Initiative),这是一个已被证明行之有效、可以借力以取得成果的项目。 这只是实现美国能源电网现代化与扩容所需工作的开端。我们需要认识到,今天对电网的新投资,与一个世纪前同样重要——当时美国在私营和公共部门对电力的支持方面领先世界。 #### 壮大AI基础设施劳动力队伍 美国数据中心扩张面临的最大单一挑战,或许是全国性的人手短缺——包括熟练电工和管道工。以电工为例,他们对数据中心建设至关重要,负责安装由配电盘、变压器和备用电源系统组成的复杂系统。我们已在全国各地雇用了数千名电工,包括在亚利桑那州、佐治亚州、弗吉尼亚州、华盛顿州和威斯康星州。但美国的电工数量不足以满足不断增长的需求。我们估计,未来十年,美国将需要招募并培训50万名新电工,才能满足全国日益增长的用电需求。我们需要一项国家战略,确保美国工人抓住这一机遇。 这些是优质岗位,将为全国各地的人们提供出色的长期职业发展。我们建议动用现有的联邦教育与培训资金以及税收激励,来扩大这些机会。具体措施可包括:把现有的联邦学徒制投资定向投向已确定重大AI基础设施项目的地区,并支持现有培训中心迅速增加以电工为重点的注册学徒岗位数量。 我们赞赏特朗普总统近期签署的行政命令《让美国人为未来的高薪技能工种岗位做好准备》(Preparing Americans for High-Paying Skilled Trade Jobs of the Future),它凸显了技能工种在建设AI基础设施中的重要性,并为把握当下这一时机铺平了道路。在联邦机构着手落实该命令的过程中,行业预测机构与工会培训中心通力合作以最大化成效,将至关重要。 归根结底,我们需要各级政府以及全国各社区采取新的举措。例如,作为一个国家,我们需要做更多工作,重振美国高中的工业技艺课和工艺实习课(shop classes)。这应成为地方教育委员会和州政府的优先事项。同样,全国的社区学院也需要付出更多努力,支持一项帮助培养新一代技能劳动力(包括电工和管道工)的国家行动。 #### 投资AI研发 要维护美国作为全球科学领导者的地位,就必须加强联邦对基础科学研究的投资。美国在运用公私合作伙伴关系方面有着载入史册的悠久历史。几十年前作出的以公共资金资助科研基础设施、并向有才华的科学家和企业家提供财政支持的决定,为美国的技术领先地位铺平了道路。通过联邦、州和地方政府的各项行动,资金投向了区域经济和各类项目,押注于美国人民的聪明才智。20世纪那些著名的创新孵化器——例如贝尔实验室(Bell Labs)和联邦国家实验室网络——正是有意联合产业界、政府和学术界以推动科学进步的成果。今天,我们必须对AI和量子技术采取类似的战略。对这些领域的投资,对于推进旨在应对复杂全球挑战的创新技术解决方案的开发至关重要。 要在竞争中胜过中国这样已大幅增加研发(R&D)投资的国家,美国必须加快对未来技术科研的战略性投资。专家预测,未来十年,中国将继续在AI、先进制造、清洁能源、量子计算和半导体等下一代技术上投入大量资源。 自第二次世界大战以来,美国的技术创新一直由研发驱动,而这种研发建立在两项关键要素之上,令世界其他国家既研究又羡慕。第一项是对基础研究的持续支持。虽然少数科技公司在基础研究上投入巨资——正如我们通过微软研究院(Microsoft Research,MSR)所做的那样——但大多数世界领先的基础研究是由美国大学的学者开展的,往往依托国家科学基金会(National Science Foundation)及其他联邦机构的资助。这类研究由好奇心而非利润动机驱动,常常带来意想不到却影响深远的发现,并公开发表。 第二项要素是各种规模的企业对产品开发投资的持续投入。美国比任何其他国家都更娴熟地掌握了把新想法迅速从大学转移到私营部门的过程。这一成功建立在对「研」与「发」两方面的良性投资之上——认识到基础研究通常由公共资金资助、且一般在大学进行,而产品开发则通过企业获得强劲的私人资金支持。正是两者的结合,使美国的研发如此成功。 2019年,特朗普总统批准了一项旨在巩固美国在人工智能领域领先地位的行政命令。它正确地聚焦于联邦对AI研究的投资,以及让联邦数据和计算资源更易获取。六年后,总统和国会应在这些努力的基础上加以扩展,以支持推进美国的AI领导地位。为国家科学基金会的基础研究、并通过我们的大学提供更多资金,是一个很好的起点。 #### 确保公共数据开放且可获取 数据是驱动人工智能的燃料。数据的质量、数量和可获取性,直接决定AI模型的强大程度与精细程度。虽然互联网一直是训练数据的主要来源,但联邦政府仍是最大的尚未开发的高质量、大规模数据来源之一。然而今天,其中许多数据集要么无法获取,要么无法用于AI开发。 通过让政府数据可随时用于AI训练,美国可以显著加快AI能力的进步,推动创新与发现。开放这些数据集的访问,将使得跨广泛数据集分析主题、模式和洞见成为可能,推动美国走到全球AI发展的最前沿。 重要的是,可获取的公共数据能营造公平的竞争环境。它不仅赋能大公司,也让初创企业、学术机构和非营利组织能够训练和改进AI模型。这将培育一个更具竞争性、更具包容性的AI生态系统,让创新由想法和才智驱动——而不仅仅是专有数据。 相比之下,中国和英国等国家已在大力投资其数据资源,认识到国家规模数据管理的经济与战略价值。中国把数据集作为战略资源加以管理的综合体系,以及英国的国家数据图书馆(National Data Library),都凸显了一个日益增长的全球趋势:把数据视为服务于经济竞争力的公共财富。 ### 赢得AI扩散竞赛 历史告诉我们,一项通用技术的真正影响,不仅取决于其领先发明的水准,更取决于这些发明在全社会被采用的速度、广度和有效性。但现实是,技术扩散需要时间、投资、伙伴关系和健全的公共政策。 电力的历史为AI提供了重要启示。1878年托马斯·爱迪生(Thomas Edison)证明电可以点亮灯泡之后,谁还会选择在夜里坐在仅靠蜡烛或煤油灯照明的房间里?然而就在今晚,近150年过去了,地球上仍有超过7亿人家中没有通电。扩散不仅需要有出色的技术,还需要有合理的经济性。 技术扩散的经济性始于技能培训。各国需要投资于使用新技术所需的技能,无论是个人层面还是各类组织层面。人们很容易低估技能培训所发挥的作用,也容易低估公共政策为其提供支持的必要性。但在每一次工业革命中,最善于驾驭当时领先通用技术的国家,正是最迅速、最广泛地对其人口进行技能培训的国家。 #### 培训美国劳动力 在新的AI经济中,各种背景的美国人都会需要关键的AI技能才能参与竞争。要应对技能培训挑战的全部规模,国家必须确立一个新的全国目标:让每一位美国人都能获得并用上AI技能培训。这将需要极其广泛的伙伴关系和新的政策思路,跨越地域、组织、经济和政治的分野。 特朗普总统近期聚焦AI教育与劳动力的行政命令,为建立全国性的AI技能培训战略迈出了关键步伐。《推进美国青少年人工智能教育》(Advancing Artificial Intelligence Education for American Youth)行政命令确立了一项明确政策:通过负责任地把AI融入面向教师和学生的教育,来提升AI素养。通过促成这种早期接触,全国青少年将为AI赋能的工作做好更充分的准备。国会也可以考虑利用现有的联邦对各学区的拨款,鼓励K-12(幼儿园至高中)教育中的AI学习与AI素养培养。 企业和非营利组织也有重要作用可发挥。在微软,我们正努力为应对这一技能培训挑战尽一份力。仅2025年一年,我们就有望为250万美国人提供基础AI技能培训。我们正与美国全国未来农民协会(National Future Farmers of America,FFA)合作,通过我们的「Farm Beats for Students」计划,培训每个州的教育工作者把AI融入农业课堂。我们正与美国教师联合会(American Federation of Teachers,AFT)——美国代表全国教育工作者的最大组织——合作,向1万名AFT会员提供共同开发的培训项目。我们还正与新泽西州、普林斯顿大学(Princeton University)和CoreWeave合作,在新泽西建设一个AI枢纽(AI Hub),其中将包括对当地社区学院AI教育的支持。 说到AI技能培训,我们需要做的最重要的事,是认识到这是一个关键领域,亟待关注、学习、合作与创新。它将对在我们的经济和社会中扩大这项技术的可及性产生巨大影响。生成式AI是一项新兴的年轻技术。我们对于AI技能培训项目和支持的需求全貌的认识,也同样处于起步阶段。这是一项头等优先事项,理应获得与AI技术创新本身同等程度的关注与支持。 #### 鼓励采用AI 联邦政府自身使用AI,也将在AI扩散中发挥关键作用。政府各部门都有机会利用AI来提高面向公民的公共服务的质量与效率。 令人鼓舞的是,管理与预算办公室(OMB)近期发布了聚焦联邦政府使用与采购AI的M备忘录(M-Memos)。两份备忘录都强调:消除创新障碍、最大限度使用本土开发的AI产品、并鼓励联邦政府内部的AI负责人推动负责任的AI采用,这些都很重要。 我们也看到各州层面的行动。我们与得克萨斯州交通部(Texas Department of Transportation)合作,启动了一个为期六周的试点项目,旨在提升生产力并改善各处室的决策。该项目取得了强劲成果:97%的参与者在试点期间使用了AI数字助理,68%的参与者已将其融入日常工作流程,参与者报告称在常规任务上平均每周节省12小时。 ### 出口美国AI 出口我们的AI的能力,对于维持我们的全球竞争力、并确保我们的技术进步不仅惠及本国、也惠及世界各地的盟友与伙伴,至关重要。在近期AI外交努力的基础上,美国在全球技术格局中提供了一个有吸引力且值得信赖的价值主张。 包括微软在内的美国科技公司,正在世界各地对AI基础设施进行前所未有的投资。仅微软一家就在40多个国家建设AI基础设施,其中包括中国重点投资的地区。我们迫切需要一项国家政策,为这些投资在出口管制与贸易支持之间提供恰当的平衡。 虽然美国政府正确地聚焦于通过出口管制保护安全数据中心内的敏感AI组件,但AI竞争中一个更为重要的要素,将涉及一场竞赛…… 【译注:原文在本批次末尾处中断,末句未完(…will involve a race),延续至下一批次(batch-009)。】 **史密斯先生:** ……在美国与中国之间展开,以将各自的技术推广到其他国家。鉴于技术市场的性质及其潜在的网络效应,这场美国与中国之间争夺国际影响力的竞赛,很可能将由最快的先行者获胜。美国需要一项明智的国际战略,以迅速在全球范围内支持美国的人工智能。 【译注:本批开头为残句,承接上一批末尾未完成的句子(原文句子在批次边界处切分)。】 这一根本教训来自过去二十年的电信设备出口。最初,朗讯(Lucent)、阿尔卡特(Alcatel)、爱立信(Ericsson)和诺基亚(Nokia)等美国和欧洲公司打造了定义国际标准的创新产品。但随着华为(Huawei)投入创新,以及中国政府补贴其产品销售——尤其是在整个发展中世界——这些中国产品的采用速度超过了竞争对手,并成为众多国家电信网络的骨干。这为后来成为特朗普政府2020年一个重要议题的问题奠定了技术基础:当时政府正努力应对华为5G产品的存在及其对国家安全与网络安全的影响。 早期迹象表明,中国政府有意复制其成功的电信战略。中国正开始向发展中国家提供补贴,使其能够获得稀缺芯片,并承诺在当地建设人工智能数据中心。中方明智地认识到,如果一个国家以中国的AI平台为标准,它未来很可能会继续依赖该平台。 国际伙伴关系将至关重要。正因如此,Microsoft(微软)已与阿联酋的G42等实体以及贝莱德(Blackrock)和MGX等投资基金建立合作,目标是为人工智能基础设施和供应链筹集至多1000亿美元($100 billion)资金。美国科技公司与私人资本市场正在与中东的关键国家及主权投资者建立更紧密的联系,其力度超过了以往在电信领域对抗中国补贴的努力,体现了我们对创新与合作的承诺。尽管中国政府可能会补贴其技术在发展中地区的采用,但它将难以匹敌美国私营部门投资的规模与影响。 务实的美国出口管制政策至关重要,要在安全保护与快速扩张的能力之间取得平衡。通过防止对手获取先进AI技术来保护国家安全,这一点至关重要。规则应包含针对安全数据中心部署的定性标准,以防止芯片被转运至中国,并确保先进AI服务得到保护。我们支持这类做法。 然而,我们已对1月发布的《人工智能扩散临时最终规则》(interim final AI Diffusion Rule)对GPU出货量施加的数量上限表示关切。这些上限将美国的关键盟友与伙伴置于「第二层级」(Tier Two)类别,对人工智能数据中心的扩张施加限制。这包括瑞士、波兰、希腊、新加坡、印度、印度尼西亚、以色列、阿联酋和沙特阿拉伯等国家。这些国家的客户如今担心获取美国AI技术的渠道受限——这可能使他们转向替代方案,从而让中国的AI产业受益。 特朗普政府有机会修订该规则,取消数量上限,保留定性标准。这种做法能确保美国的盟友与伙伴对获取美国AI产品保持信心。 归根结底,我们需要认识到,世界各国只有在能够信任美国AI的情况下才会使用它。这就赋予美国公司责任,要以负责任的方式开发和部署满足当地需求的人工智能基础设施与产品。这也要求各国对持续、不间断地获取关键AI组件与服务抱有信心。美国长期以来建立了可信赖技术的声誉,这是中国一直无法企及的。但这一声誉,如同所有真正重要的事物一样,需要持续的关注与呵护。 Chairman Cruz, Ranking Member Cantwell, members of the Committee, thank you for the opportunity to be here today. Let me just build on what my three colleagues have said and offer a few thoughts. The first is I just wanted to refer to the chart on this easel that shows the AI tech stack that was also in my written testimony. It makes a simple but, I think, important point--we are all in this together. If the United States is going to succeed in leading the world in AI it requires infrastructure. It requires success at the platform level. It requires people who create applications. Interestingly, we at Microsoft get to work with all three of these leaders and their companies. Our success, each of our success, depends on each other's success, and what is true of the four of us is true when you look across the country and around the world at open source developers, people who are building power plants, electricians and pipe fitters who are going to work every single day. So what do we need from the Congress and the country in order to succeed? I think it is three things. I described them in my written testimony. First, as Chairman Cruz said, we need innovation. Innovation will go faster with, as Sam said, more infrastructure, faster permitting, more electricians. We need more innovation fueled, as Ranking Member Cantwell said, by support from our universities and the Federal agencies that support basic research across the country, one of this country's crown jewels. We also need, as Chairman Cruz said, faster adoption, what people refer to as AI diffusion--the ability to put AI to work across every part of the American economy to boost productivity, to boost economic growth, to enable people to innovate in their work, and the number-one ingredient for that history shows time and time again is skilling, investing in education. And, finally, we need to export. If America is going to lead the world we need to connect with the world. We need to remember--I believe always--that as a country only 4 and a half percent of the world's people live in the United States of America. Our global leadership relies on our ability to serve the world with the right approach to export controls and always, especially in technology, in our ability to sustain the trust of the rest of the world. Ultimately, I think people who take the time--if they take the time to watch or read about this hearing may wonder what is this all about--what are we at this table trying to do? What do these two letters, AI, really mean to them? Are we who are working in this industry trying to build machines that are better than people or are we trying to build machines that will help people become better? Emphatically, it is and needs to be the latter. Are we trying to build machines that will outperform people in all the jobs that they do today or are we trying to build machines that will help people pursue better jobs and even more interesting careers in the future? Indisputably, it needs to be the second, not the first, and I believe that is what we are and can do together. As somebody who has now spent almost 32 years in this industry there are two things that always strike me. The first probably will not surprise you. Never underestimate what technology can do, how quickly it can move, what it can accomplish. But the second is one that I think is too seldom discussed even though every day it stares us in the face. Never estimate what people can do. Never underestimate human ambition. Never underestimate what a person can do if given a better technology tool and the ability to learn how to put it to use. That is the story of this industry. It is the story of the country. It is, as you heard, the story of Sam Altman. Not everybody becomes a Sam Altman or a Satya Nadella or a Bill Gates but everybody deserves the opportunity to try. Tonight across America, whether it is the attic of a house or the basement or just an everyday bedroom there are kids with computers, with phones, with access to the Internet and now the ability to put AI to work. Let us invest in their education. Let us invest in the skills that the American public needs. Let us then invest in creating the future that the American public deserves. Thank you. [The prepared statement of Mr. Smith follows:] Prepared Statement of Brad Smith, Vice Chair and President, Microsoft Corporation Chairman Cruz, Ranking Member Cantwell, and Members of the Committee, Thank you for the opportunity to testify on the critical issue of artificial intelligence. I am Brad Smith, the Vice Chair and President of Microsoft Corporation. AI has the potential to become the most useful tool for people ever invented. Like the general purpose technologies that preceded it, such as electricity, machine tools, and digital computing, AI will impact every part of our economy. It will shape not just how we work and live, but how we compete, prosper, and stay secure as a nation between now and the middle of this century. The notice for this hearing aptly refers to an ``AI race.'' I would like to talk today about what is needed to win this race. The AI race involves both technology and economics. It requires both innovation and diffusion. It is both a sprint and a marathon. The country can win a lap but lose the race if it fails to bring together all the ingredients needed for success. It is a race that no company or country can win by itself. To win the AI race, the United States will need to support the private sector at every layer of the AI tech stack. The nation will need to partner with American allies and friends around the world. In my testimony today, I will focus on three strategic priorities where this Congress and the Federal government will make a difference. First, the country must win the AI innovation race. This will require massive datacenters and AI infrastructure that need Federal support to expand and modernize the electrical grid on which they depend. The country must recruit and train skilled labor like electricians and pipefitters that are in short supply. We all must summon the best of our researchers at national labs and universities, supported by Federal basic research programs and partnerships that have become the envy of the world. We will need to continue to excel in moving innovative ideas from academic labs into companies and new products. And we will need to support AI developers with open and broad access to public data. Second, the Nation must win the AI diffusion race. This will require that we promote broad AI adoption that will enable productivity growth across every sector of the economy. More than anything, this requires new initiatives to promote the AI skilling of the American workforce. This will involve basic AI fluency in our schools and new AI training programs in our community colleges. It will also include advanced AI education that will represent the next generation of computer science degrees, organizational skills that will be mastered in the country's business schools, and new courses in the nation's law schools. When combined, these will enable companies, non-profits, and government agencies alike to put AI to effective use. Governments at the federal, state, and local levels can then help accelerate this diffusion by adopting AI services to improve the effectiveness and efficiency of the services they provide to the public. Third, the United States must export AI to American allies and friends. No company or country is so powerful that it can master the future of AI without friends. The United States and China are competing not only to innovate but to spread their respective technologies to other countries. This part of the race likely will be won by the fastest first mover. The United States needs a smart export control strategy that protects our national security while assuring other countries that they will have reliable and sustained access to critical American AI components and services. Perhaps as much as anything, this requires that we collectively sustain international trust in our products, our companies, and the country itself. AI as a General Purpose Technology Economists sometimes put technologies into two categories, general purpose technologies and single-purpose tools. Most things in the world are single-purpose tools, like a smoke detector or a lawn mower. They do one thing very well. But over the course of history, certain so- called general purpose technologies impact and sometimes even redefine almost every sector of the economy. Electricity is the prototypical example, because when you think about it, electricity changed the way every economic sector works. The key to mastering the future of AI starts in part by understanding the role technology has played in the past. The past three centuries have brought the world three industrial revolutions, each driven by these general purpose technologies. First, it was iron working in the United Kingdom, starting in the 1700s. And then it was electricity and machine tools in the 1800s, when the United States overtook the United Kingdom by putting these technologies to work more broadly than any other country. And then there was the third industrial revolution during the last 50 years, driven by computer chips and software. Without question, being a global leader in advancing a general purpose technology gives a country a major edge. But one lesson of history is that the countries that benefit the most and advance the fastest are not necessarily the countries where the technology is invented. Rather, it's where the technology is diffused--or adopted-- the most quickly and broadly. This is for good reason. If a technology improves productivity and changes every part of an economy, then the country that uses it the most broadly and quickly will benefit the most. This both frames and defines the AI opportunity and challenge for the United States. As a nation, we need to focus both on advancing innovation and driving diffusion, both domestically and as a leading American export. The AI Tech Stack The key to driving both innovation and diffusion is to recognize that AI, like all general purpose technologies, is built on what we in the industry call a tech stack--a stack of technologies that are used together. This is true for every great general purpose technology. You can see this, for example, if we go back in time and think about electricity. Thomas Edison first succeeded in 1878 in using electricity to light a lightbulb. But the illumination of lights across a city quickly required the construction of power plants, the fuel to run them, the creation of an electrical grid, the standardization of circuits, and a wide range of electrical appliances beyond the lightbulb itself. In short, a tech stack for electricity. Artificial intelligence similarly is built on an AI tech stack. Fundamentally, it is divided into three layers, infrastructure, the platform layer, and applications. You can see this illustrated below. [GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT] The infrastructure layer is massive. Microsoft is spending more than $80 billion this Fiscal Year on the capital investment needed for this layer, with more than half this amount being spent in the United States. This goes to buying land, investing in electricity and broadband connectivity, procuring chips like GPUs, and installing liquid cooling. These lead to the construction of datacenters--or often datacenter campuses with many buildings with potentially hundreds of thousands of computers. This infrastructure supports both the training of new AI models and their deployment, so they can be used for AI-based services around the world. On top of this infrastructure, there is the platform layer. The heart of this layer consists of AI foundation models, including frontier models created by companies like OpenAI, as well as open source and other models from a wide variety of other firms--including Anthropic, Google, Mistral, DeepSeek, and Microsoft itself. The platform layer relies on data to train and ground models. And it includes a new generation of software-based AI platform services that are used to help build AI applications. Ultimately, both the infrastructure and platform layers support the applications layer. These are devices and software applications that use AI to deliver better services to people. ChatGPT and Microsoft's Copilot are both examples of AI applications. One of the amazing things about the applications layer is it's not just companies--large or small or established or startup--that are creating AI applications. It's everybody. It's researchers using new AI-infused applications to change drug discovery. It's non-profits changing the way they deliver services. It's teachers using AI as a tool to improve the way they prepare material for a classroom. It's governments making everything from the filing of a tax return to the renewal of a driver's license easier and more efficient. To build a new AI economy, it's critical to get all three of these layers working and to get a flywheel turning across the ecosystem. It's essential to build the infrastructure layer so people can develop and deploy the models at the platform layer. It's essential to use the AI models so that people will build the applications on top of them. And it's essential for customers to adopt the applications, so the market can grow, and drive increased investment to expand the infrastructure further. The process repeats itself. This is how a new economy is born. Success Requires an Entire Ecosystem The flywheel effect makes clear that success requires not only national progress at one layer of the tech stack, but at every layer. That is what the private sector currently is pursuing in the United States better than in any other country. And it's what this Congress and the Executive Branch can help support with a strategy that promotes both AI innovation and diffusion up and down this stack. National AI leadership requires not only success by a few companies, but by many. Today's panel, involving leading firms such as OpenAI, AMD, CoreWeave, and Microsoft, reflects important slices of the new AI economy. The AI economy requires a multifaceted and integrated ecosystem that includes ``Big Tech'' and ``Little Tech,'' startups and more established firms, open source and proprietary developers, suppliers and customers, firms that create data and firms that consume it, all working together. Governments as both regulators and leading AI adopters have critical roles to play. Commentators sometimes focus on the tensions between different participants in this tech ecosystem. These deserve attention. What's often overlooked is that the different participants also depend on each other. And this means that the different contributors to the AI ecosystem all need to be healthy. A large technology company like Microsoft has a unique opportunity--and responsibility--to partner with and support the participants at every level of the tech stack. We strive to advance not just innovation but an economic architecture, business models, and responsible practices that will help grow the AI market on a long-term basis. Not just for the United States, but the country's friends and allies. Winning the Innovation Race Although the AI economy is being built mostly by the private sector, government policies and initiatives need to play a critical role. This starts with work needed to help fuel innovation. A few areas deserve particular attention in this hearing. Power the growth of datacenters Just as you can't have reliable electricity in your home without a powerplant, you can't have AI without datacenters and AI infrastructure. And these datacenters require a vast supply chain to construct and large amounts of electricity to operate. America's advanced economy relies on 50-year-old infrastructure that cannot meet the increasing electricity demands driven by AI, reshoring of manufacturing, and increased electrification. The United States will need to invest in more transmission and energy resources, onshore our supply chains, and modernize our electric grid to support forecasted increases in electrical loads. Microsoft is investing in these areas itself. We urge the Federal government to streamline the Federal permitting process to accelerate growth in all these areas. The current Federal permitting processes often involve multiple agencies and complex, unpredictable, multi-year reviews. This hinders progress. The Federal government should take immediate steps to establish reliable, reasonable, and transparent timelines for permitting decisions. This can also be done by standardizing Federal permitting processes and designating a lead agency to shepherd the permits through the process. Further, the permitting agencies should utilize AI and digital tools to improve timelines and transparency for applicants and ensure the permitting agencies have quick access to information to assist them in their review and decision-making process. We were pleased to see President Trump's recent Executive Order, ``Updating Permitting Technology for the 21st Century,'' directing agencies to make maximum use of technology in the environmental review and permitting process. The Congress should also look to the Federal- State Modern Grid Deployment Initiative as a proven program that can be leveraged to deliver results. This is just the start of what is needed to modernize and expand America's energy grid. We need to recognize that new investments in the grid are just as important today as they were a century ago, when the United States led the world in private and public sector support for electricity. Grow the AI Infrastructure workforce Perhaps the single biggest challenge for data center expansion in the United States is a national shortage of people--including skilled electricians and pipefitters. Electricians, for example, are essential to datacenter construction, installing a complex system of electrical panels, transformers and backup power systems. We have hired thousands of electricians across the country, including in Arizona, Georgia, Virginia, Washington, and Wisconsin. But the United States doesn't have enough electricians to fill the growing demand. We estimate that over the next decade, the United States will need to recruit and train half a million new electricians to meet the country's growing electricity needs. We need a national strategy to ensure we meet this opportunity for American workers. These are good jobs that will provide great long-term careers for people across the country. We recommend making existing Federal education and training funds, as well as tax incentives, available to scale up these opportunities. These could include targeting current Federal apprenticeship investments in regions that have identified major AI infrastructure initiatives and supporting existing training centers to quickly increase the number of registered apprenticeships focused on electricians. We commend President Trump's recent Executive Order, ``Preparing Americans for High-Paying Skilled Trade Jobs of the Future,'' for highlighting the importance of skilled trades in the building of AI infrastructure and for paving the way to meet this moment. As Federal agencies work to implement the order, it will be critical that industry forecasters and union training centers work together to maximize impact. Ultimately, we need new steps at every level of government and in communities across the country. For example, we need to do more as a nation to revitalize the industrial arts and shop classes in American high schools. This should be a priority for local school boards and state governments. Similarly, the nation's community colleges will need to do more to support a national initiative to help train a new generation of skilled labor, including electricians and pipefitters. Invest in AI research and development To uphold America's position as a global scientific leader, it is imperative to enhance Federal investment in fundamental scientific research. The United States boasts a storied history of employing public-private partnerships. The decisions made decades ago to publicly fund research infrastructure and provide financial support to talented scientists and entrepreneurs paved a pathway to American technological leadership. Through federal, state and local government initiatives, investments were made in regional economies and programs, betting on the ingenuity of the American people. Notable incubators of the 20th century--such as Bell Labs and the network of Federal national laboratories--were the result of deliberate efforts to unite industry, government, and academia to propel scientific advancement. We must deploy a similar strategy today for AI and quantum technologies. Investments in these areas are critical to advancing the development of innovative technological solutions that address complex global challenges. To outcompete nations like China, which have significantly boosted their research and development (R&D) investments, the United States must accelerate strategic investments in scientific research for future technologies. Experts predict China will continue to invest substantial resources in next-generation technologies such as AI, advanced manufacturing, clean energy, quantum computing, and semiconductors over the next decade. Since the Second World War, America's technological innovation has been driven by R&D based on two critical ingredients that the rest of the world has both studied and envied. The first is sustained support for basic research. While a few tech companies invest substantial sums in basic research, as we do through Microsoft Research (MSR), most world-leading basic research is pursued by academics at American universities, often based on funding from the National Science Foundation and other Federal agencies. Driven by curiosity rather than a profit motive, this research often leads to unexpected but profound discoveries that are published publicly. The second ingredient is a sustained commitment to investments in product development by companies of all sizes. The United States, more than any other country, has mastered the process of moving new ideas quickly from universities to the private sector. This success rests on healthy investments in both R and D, recognizing that basic research is often publicly funded and typically in universities, while product development is robustly and privately funded through companies. It's the combination of the two that makes American R&D so successful. In 2019, President Trump approved an executive order designed to strengthen America's lead in artificial intelligence. It rightly focused on Federal investments in AI research and making Federal data and computing resources more accessible. Six years later, the President and Congress should expand on these efforts to support advancing America's AI leadership. More funding for basic research at the National Science Foundation and through our universities is one good place to start. Ensure public data is open and accessible Data is the fuel that powers artificial intelligence. The quality, quantity, and accessibility of data directly determines the strength and sophistication of AI models. While the Internet has been a major source of training data, the Federal government remains one of the largest untapped sources of high-quality and high-volume data. Yet today, many of these datasets are either inaccessible or not usable for AI development. By making government data readily available for AI training, the United States can significantly accelerate the advancement of AI capabilities, driving innovation and discovery. Opening access to these datasets would allow for the analysis of themes, patterns, and insights across broad datasets, propelling the country to the forefront of global AI development. Importantly, accessible public data levels the playing field. It empowers not only large companies but startups, academic institutions, and nonprofits to train and refine AI models. This fosters a more competitive and inclusive AI ecosystem, where innovation is driven by ideas and ingenuity--not just proprietary data. In comparison, countries like China and the United Kingdom are already investing heavily in their data resources, recognizing the economic and strategic value of national-scale data management. China's comprehensive system to manage datasets as a strategic resource and the UK's National Data Library underscore a growing global trend of treating data as a common good for economic competitiveness. Winning the AI Diffusion Race History teaches us that the true impact of a general purpose technology is not measured solely by the caliber of its leading inventions, but by how quickly, widely, and effectively these are adopted across society. But the reality is that technology diffusion takes time, investment, partnerships, and sound public policy. The history of electricity offers an important insight for AI. Once Thomas Edison proved in 1878 that electricity could power a lightbulb, why would anyone choose to sit at night in a room illuminated by a candle or kerosene? Yet tonight, almost 150 years later, more than 700 million people on the planet still live without electricity in their homes. Diffusion requires not only great technology, but sound economics. The economics of tech diffusion start with skilling. Countries need to invest in the skills needed to use new technology, both as individuals and across organizations. It is easy to underestimate both the role that skilling plays and the need for public policy to support it. But in each industrial revolution, the country that best harnessed the leading general purpose technology of its time was the Nation that skilled its population the most quickly and broadly. Skill the American workforce In the new AI economy, Americans of all backgrounds will need critical AI skills to compete. To meet the totality of the skilling challenge, the country must pursue a new national goal to make AI skilling accessible and useful for every American. This will require a very broad range of partnerships and new policy ideas, spanning across geographic, organizational, economic, and political divides. President Trump's recent executive orders focused on AI education and the workforce provide critical steps towards a national skilling strategy for AI. The ``Advancing Artificial Intelligence Education for American Youth'' EO establishes a clear policy to promote AI literacy by responsibly integrating AI into education for teachers and students. By fostering this early exposure, the Nation's youth will be better positioned for AI-enabled work. Congress can also consider leveraging existing Federal funding to the Nation's school districts to encourage AI learning and literacy in K-12 education. Businesses and non-profits have important roles to play. At Microsoft, we are seeking to do our part to meet this skilling challenge. In 2025 alone, we are on a path to train 2.5 million Americans in basic AI skills. We're partnering with the National Future Farmers of America (FFA) to train educators in every state to integrate AI into the agricultural classroom through our Farm Beats for Students program. We are partnering with the American Federation of Teachers (AFT), the largest organization representing the Nation's educators in America, to deliver a co-developed training program to 10,000 AFT members. And we're partnering with the State of New Jersey, Princeton University, and CoreWeave on an AI Hub in New Jersey that will include support for AI education in local community colleges. When it comes to AI skilling, the most important thing we need to do is recognize that this is a critical field that is ripe for attention, learning, partnership, and innovation. It will have a huge impact on broadening access to this technology across our economy and society. Generative AI is a new and young technology. So is our knowledge of the full extent of need in terms of AI skilling programs and support. This is a first-class priority that deserves as much attention and support as innovation in AI technology itself. Encourage AI adoption The Federal Government also will play a critical role in AI diffusion by using AI itself. There are opportunities across the government to use AI to improve the quality and efficiency of public services for citizens. It's encouraging to see the recent OMB publication of M-Memos focused on Federal government use and procurement of AI. Both memos emphasized the importance of removing barriers to innovation, maximizing the use of domestically developed AI products, and encouraging AI leaders within the Federal government to facilitate responsible AI adoption. We're seeing activity in the states as well. We partnered with the Texas Department of Transportation to launch a six-week pilot program aimed at boosting productivity and improving decision-making across various departments. The program saw strong results with 97 percent of participants using the AI digital assistant during the pilot, 68 percent have integrated it into their daily workflow, and participants reporting saving an average of 12 hours a week on routine tasks. Exporting American AI The ability to export our AI is essential to sustaining our global competitiveness and ensuring that our technological progress benefits not only our nation, but also our allies and partners around the world. Building on recent AI diplomacy efforts, the United States offers a compelling and trusted value proposition in the global technology landscape. American tech companies, including Microsoft, are making unprecedented investments in AI infrastructure around the world. Microsoft alone is building AI infrastructure in more than forty countries, including regions where China has focused its investments. We urgently need a national policy that provides the right balance of export controls and trade support for these investments. While the U.S. government rightly has focused on protecting sensitive AI components in secure datacenters through export controls, an even more important element of AI competition will involve a race between the United States and China to spread their respective technologies to other countries. Given the nature of technology markets and their potential network effects, this race between the United States and China for international influence likely will be won by the fastest first mover. The United States needs a smart international strategy to rapidly support American AI around the world. This fundamental lesson emerges from the past twenty years of telecommunications equipment exports. Initially, American and European companies such as Lucent, Alcatel, Ericsson, and Nokia built innovative products that defined international standards. But as Huawei invested in innovation and China's government subsidized sales of its products, especially across the developing world, adoption of these Chinese products outpaced the competition and became the backbone of numerous countries' telecommunications networks. This created the technology foundation for what later became an important issue for the Trump Administration in 2020, as it grappled with the presence of Huawei's 5G products and their implications for national and cybersecurity. Early signs suggest the Government of China is interested in replicating its successful telecommunications strategy. China is starting to offer developing countries subsidized access to scarce chips, and it's promising to build local AI datacenters. The Chinese wisely recognize that if a country standardizes on China's AI platform, it likely will continue to rely on that platform in the future. International partnerships will be critical. This is why Microsoft has partnered with entities like the UAE's G42 and investment funds like Blackrock and MGX, aiming to raise up to $100 billion for AI infrastructure and supply chains. American tech companies and private capital markets are forging stronger ties with key nations and sovereign investors in the Middle East, surpassing previous efforts to counter Chinese subsidies in telecommunications and reflecting our commitment to innovation and cooperation. While China's government may subsidize its technology adoption in developing regions, it will struggle to match the scale and impact of America's private sector investments. Pragmatic American export control policies are essential, balancing security protections with the ability to expand rapidly. Protecting national security by preventing adversaries from acquiring advanced AI technology is crucial. Rules should include qualitative standards for secure datacenter deployments to prevent chip diversion to China and ensure advanced AI services are safeguarded. We support this type of approach. However, we have expressed our concerns about the quantitative caps imposed on GPU shipments by the interim final AI Diffusion Rule issued in January. These place key American allies and partners in a Tier Two category, imposing limits on AI datacenter expansion. This includes countries like Switzerland, Poland, Greece, Singapore, India, Indonesia, Israel, the UAE, and Saudi Arabia. Customers in these countries now fear restricted access to American AI technology-- potentially benefitting China's AI sector by turning to alternatives. The Trump administration has an opportunity to revise the rule, eliminating quantitative caps and retaining qualitative standards. This approach ensures American allies and partners remain confident in accessing American AI products. Ultimately, we need to recognize that countries around the world will use American AI only if they can trust it. This creates responsibilities for American companies to develop and deploy AI infrastructure and products in a responsible manner that meets local needs. And it requires that countries have confidence in sustained and uninterrupted access to critical AI components and services. The United States has long built a reputation for trustworthy technology that China has been unable to match. But this reputation, like everything that truly matters, requires constant attention and care. STATEMENT OF HON. TIM SHEEHY, U.S. SENATOR FROM MONTANA ¶
Senator Sheehy ### 尊敬的蒂姆·希伊(Tim Sheehy)陈述 ### 蒙大拿州联邦参议员 感谢各位证人的证词。我将开始第一轮提问,然后沿席位依次请到我们的少数党首席成员。 谢谢你们的证词。知道是你们几位坐在方向盘后面,我晚上睡觉确实踏实了些——虽然还是会担心「终结者」和「天网」来找我们。 但请用五个词以内回答,从史密斯(Smith)先生开始:你们最需要从我国政府那里看到的第一——最需要看到的五个词是什么,才能确保我们赢得这场人工智能竞赛? Thank you, witnesses, for your testimony and I will start off with the first round of questions and move down dais to our Ranking Member here. Thank you for your testimony. It certainly makes me sleep better at night worried about Terminator and Skynet coming after us, knowing that you guys are behind the wheel. But in five words or less, starting with you, Mr. Smith, what are the first--what are the five words you need to see from our government to make sure we win this AI race? ¶
Mr. Smith 多培养电工。这才两个词。 〔笑声。〕 More electricians. That is two words. [Laughter.] ¶
Mr. Smith 更广泛的 AI 教育。 Broader AI education. ¶
Senator Sheehy 而且不许拿 ChatGPT 当「电话求助」用。 And no using ChatGPT as a phone a friend. ¶
Mr. Intrator 谢谢。我要说的是,我们需要着力简化「建造大型设施」的能力。 Thank you. I would say that we need to focus on streamlining the ability to build large things. ¶
Ms. Su 能让我们在创新竞赛中跑得更快的政策。 Policies to help us run faster in the innovation race. ¶
altman 允许供应链。明智的政策。【译注:碎片话轮,原文仅两短句(Allow supply chain. Sensible policy.),上下文未完整保留;按原文直译,不补写】 Allow supply chain. Sensible policy. ¶
Senator Sheehy 说得不错。所以我听到的,和我们过去赢过的那些竞赛很像——比如核能。 您知道,德国人和奥地利人确实在那项创新的早期领先,但我们赢了那场竞赛,因为我们投入了大规模的政府力量,与大学和其他方面协作,从而赢了下来。 太空——您知道,苏联人先送上了第一颗卫星、先把人送上了太空,但我们赢了太空竞赛,因为我们确立了一套确保我们能赢的框架。航空、汽车等等,都是如此。 所以我从你们那里听到的是:你们确实需要我国政府的支持,但你们也需要政府别挡着路,好让你们创新、赢下这场竞赛。 我们该如何激励企业在美国做生意,以确保我们在这场竞赛中于美国取胜、让美国不仅领先中国、也领先其他非国家行为体? 我的意思是,我认为从能力角度看,人工智能最可怕的地方在于:赢下这场竞赛并不一定非得是国家行为体。它不像核能,不像太空技术。一个非国家行为体同样可以轻松赢下这场竞赛,并掌握比任何人都大的权力。 那么,我们该如何鼓励创新者的投资落在美国,以确保我们赢下这场竞赛? 奥特曼先生,您要先说吗? That was good. So what I hear there is something pretty similar to the races we have won before-- nuclear energy, for example. You know, the Germans and Austrians really led the innovation around that but we won the race because we put a massive government effort, collaborating with our universities and others to win that race. Space--you know, the Soviets put the first satellite up, put the first man in space, but we won the space race because we adopted a framework to ensure that we won. Aviation, automobiles, et cetera. So what I hear from you is you do need support from our government but you also need the government to stay out of your way so you can innovate and win this race. How do we incentivize companies to do business here in America to make sure we win this race in America and America leads not just China but other nonstate actors, too? I mean, I think that the scariest thing about AI from a capability standpoint is it does not have to be a state actor to win this race. It is not like nuclear energy. It is not like space technology. A nonstate actor could just as easily win this race and wield more power than anyone else. So how do we encourage innovators' investment to happen here in America to ensure we win this race? Mr. Altman, do you want to start? ¶
altman 我们很荣幸在今年 1 月宣布了「星际之门」计划(Project Stargate)——一笔 5000 亿美元($500 billion)的对美国基础设施投资。正如我提到的,它现在正顺利推进,我昨天在阿比林亲眼见到了。第一个站点令人难以置信。我们需要更多这样的项目。我们需要在建设整条供应链、建设数据中心、电力审批方面的确定性。我们很乐意把芯片生产、网络设备生产、服务器机架生产都带到这里。我认为世界确实想来投资。大量全球投资正流入美国来做这件事。我们也想确保其他国家能够用我们的技术构建、使用我们的模型,某种程度上处于我们的轨道之内,你知道,受益于美国这里的技术扩散。所以这真的很重要。我们需要确保想来美国公司工作的最高技能研究人员能够来这里工作。我们需要确保 OpenAI 这样的公司对我们如何运营有法律上的明确性。当然,会有规则。当然,需要一些护栏。这是一项影响巨大的技术。但我们需要能够在全球竞争。我们需要能够训练。我们需要能够理解我们将如何提供服务、道路规则大致是什么,所以那里需要明确性。我认为像互联网那样的路径——它曾以非常重大的方式促成这个国家的繁荣——我们需要再来一次。 We were honored to announce back in January, Project Stargate, a $500 billion investment in United States infrastructure. That is now well underway, as I mentioned, getting to see it yesterday in Abilene. The first site was incredible. We need a lot more of that. We need certainty on the ability to build out this entire supply chain, build the data centers, permit the electricity. We would love to bring chip production here, network production here, server rack production here. And I think the world does want to invest. We have a lot of global investment flowing into the U.S. to do this. We also want to make sure that other countries are able to build with our technology, use our models, and sort of, like, be in our orbit and, you know, use U.S. diffusion of technology here. So that is really important. We need to make sure that the highest skilled researchers that want to come work at U.S. companies can come here and do that. We need to make sure that companies like OpenAI and others have legal clarity on how we are going to operate. Of course, there will be rules. Of course, there need to be some guardrails. This is a very impactful technology. But we need to be able to be competitive globally. We need to be able to train. We need to be able to understand how we are going to offer services and sort of where the rules of the road are going to be, so clarity there. And I think an approach like the internet, which did lead to flourishing of this country in a very big way, we need that again. ¶
Senator Sheehy 苏(Su)博士。 Dr. Su. ¶
Ms. Su 我要补充一点:我认为算力是这一切的基础。我们希望有更多算力由美国公司在美国境内建成,并确保我们为此提供良好的环境。我们也希望我们在世界各地的技术能被广泛使用、并以正确的方式使用。 所以我认为,关于出口管制和规则的讨论应当简单、易懂、易执行,并让美国的人工智能平台成为基础。 当然,关于把制造业带回国内、确保我们拥有合适人才基础的这些意见,都是其中极为重要的要素。 I would add I think computing is a foundation to all of this. We want to have more compute built in the U.S. by U.S. companies and ensure that we have a great environment for that. We want to ensure that our technology around the world is also used broadly and in the right ways. So I think the conversation about export controls and rules should just be simple, easy to follow, easy to enforce, and enable U.S. AI platforms to be the foundation. And then, certainly, the comments around bringing manufacturing back home and ensuring that we have the right talent base are all extremely important elements of that. ¶
Senator Sheehy 企业是否在权衡「在美国做 AI」还是「在中国做 AI」?它们是否在做这种并排比较? Are companies weighing doing business in AI in America versus China? Are the companies making that side by side comparison? ¶
Ms. Su 我认为,如果你看遍全世界,会有一些国家和公司问这些问题。您知道,如果获取美国技术很困难——尽管美国技术是最好的,但如果难以获取,而人们对人工智能又有渴求,他们就会选择能拿到的东西;如果中国的东西能拿到,那当然会出现我们不愿看到的结果。 I think if you look across the world there are countries and companies that will ask those questions. You know, if it is hard to obtain U.S. technology--although U.S. technology is the best, if it is hard to obtain then there is a hunger for AI and they will choose what is available and if China is available that will certainly be a outcome that we would not like to see. ¶
Senator Sheehy 嗯,我听到的词是基础设施、电工、大学、监管框架,我认为这些是我们能帮上忙的。 我听到的词是创新、人才,然后我说——我听到苏博士说「跑得更快」。这些不是——政府没法制造人才。我们没法让你们跑得更快。 但我们可以给你们做到这一点的工具;我认为现在是时候建立一套框架,让你们拥有赢下这场竞赛所需的工具,因为最终赢下它的是你们,而不是我们。 谢谢你们的证词。少数党首席成员坎特韦尔(Cantwell)。 Well, I think I hear the words infrastructure, electricians, universities, regulatory framework, and I think those are things we can help with. I hear words like innovation and talent and I say--I hear Dr. Su, run faster. Those are not things--the government cannot manufacture talent. We cannot make you run faster. But we can give you the tools to do that and I think it is time that we create a framework so that you have the tools you need to win this race because you are going to be the ones that win it, not us. Thank you for your testimony. Ranking Member Cantwell. ¶
Senator Cantwell 谢谢您,主席先生。我想大体上延续同一个主题,谈竞争力。我们需要国家标准与技术研究院(NIST)来制定标准吗? 请只回答是或否,然后依次往下。 Thank you, Mr. Chairman. I would like to continue that same theme generally about competitiveness. Do we need NIST to set standards? If you could just yes or no and just go down the line. ¶
altman 我不认为我们需要它。它可能会有帮助。 I do not think we need it. It can be helpful. ¶
Ms. Su 是。 Yes. ¶
Mr. Intrator 是。 Yes. ¶
Mr. Smith 是。 Yes. ¶
Senator Cantwell 好。那么在我们这里讨论的语境中,我们其实只是在说——我不确定,史密斯先生、因特拉托(Intrator)先生或苏博士,哪位都行。 问题在于,如果我们想快速推进,我们就要像电力那样,建立让我们得以快速推进的标准。 在这里,我把它叫做「代码对代码」,这正是我们想要的,对吧?我们希望 NIST 在标准制定上做点什么,让我们能推进得快得多。是这样吗? 史密斯先生或因特拉托先生,哪位都行。 OK. So in the context of what we are talking about here, we are really just talking--I do not know, Mr. Smith or Mr. Intrator or Dr. Su, any. The issue here is if we want to move fast we want to create just like with electricity the standards by which we want to move fast. Here, I would just call it code for code is what we want, right? We want NIST to do something in the standard setting that will allow us to move much faster. Is that right? Either Mr. Smith or Mr. Intrator. ¶
Mr. Smith 我要说的是这样。 首先,NIST 是标准「被采用」的地方,但它未必是标准「最初被创造」的地方。所以我们已经有—— What I would say is this. First of all, NIST is where standards go to be adopted but it is not necessarily where they first go to be created. So we have got what---- ¶
Senator Cantwell 对。谢谢您把这一点讲清楚。我们说的是产业界的 IEEE,您知道,很多不同的组织,产业界参与,然后标准被采用。 所以是的,我们把这一点说清楚。我们把这一点说清楚。 Right. Thank you for that clarity. We are talking about a industry IEEE, you know, lots of different organizations, industry input, and then they are adopted. So yes, let us clarify that. Let us clarify that. ¶
Mr. Smith 是的。我认为运作方式就是这样。 Yes. I think that is the way it works. ¶
Senator Cantwell 是的。但您认为我们需要做这件事,特别是如果美国想领先的话? Yes. But you think we need to do that, particularly if the United States wants to lead? ¶
Mr. Smith 我们将需要产业标准。我们将需要美国采用这些标准。而且您说得对,我们也将需要美国付出努力,真正确保全世界接受这些标准。 We will need industry standards. We will need American adoption of standards. And you are right, we will need U.S. efforts to really ensure that the world buys into these standards. ¶
Senator Cantwell 好。因特拉托先生。 OK. Mr. Intrator. ¶
Mr. Intrator 我认为重要的是:当你与标准打交道时,标准所带来的是一个共同语汇,而共同语汇带来加速。 所以,只要我们能进入那个角色、建立起各方都能就具体事项达成一致的接触点,就会在国内和国外都带来加速。 I think it is important that when you are working with standards what that allows for is a common vocabulary which allows for acceleration. And so to the extent that we can step into that role and establish touch points where everyone can agree on specific things that will lead to an acceleration both domestically and abroad. ¶
Senator Cantwell 我不知道是否——您知道,再深入谈您认为那些具体是什么;但总的来说,您知道,当我想到互联网和 HTTP、HTML,或者任何那些——TCP/IP——我们谈的是让我们得以更快前进的东西,而这些标准的建立帮助我们做到了这一点。 关于出口问题,因特拉托先生,云服务来源的问题不应被遗漏。如果我们说「那就去马来西亚吧」,马来西亚会告诉我们,他们能证明这些——您知道,这些算力不会被转运到——您知道,到中国去;而我们基本上有一套办法来确保这一点被理解、被监督,那我们同时也想要准入,对吧?我们想要美国公司的准入。 And I do not know if, you know, drilling down more on what you think those are but in general, you know, when I think about the Internet and HTTP or HTML or any of the--TCPIP--we are talking about things that allowed us to move faster and getting those standards established helped us do that. On the export issue, Mr. Intrator, the issue of cloud sources should not be left out. If we say let us go with Malaysia, Malaysia is going to tell us that they can certify that there is no, you know, diversion of these ships to--you know, to China and we basically have a way that we can make sure that this is understood and monitored then we also want access, right? We want access by U.S. companies. ¶
Mr. Intrator 是的。我认为莉萨(Lisa)那点说得非常好,对吧?归根结底,世界希望能够以非常广泛的方式构建和部署人工智能,而如果我们——您知道,自然厌恶真空。 如果我们不进入那个角色,其他技术就会进入那个角色。如果它是次优的,那就次优吧。有次优的东西也好过什么都没有,所以这就是—— Yes. I think Lisa's point was excellent, right? At the end of the day, the world wants to be able to build and deploy artificial intelligence in a very broad way and if we--you know, nature abhors a vacuum. If we do not step into that role other technology will step in that role. If it is suboptimal so be it. It is better to have something that is suboptimal than have nothing and so that is what---- ¶
Senator Cantwell 嗯,我们不希望再出现一个「华为」:发展得更快,然后带着政府的后门,最后我们都得起来反对它。我支持一个「技术北约」。 我支持由五个最成熟的民主国家与技术国家来制定道路规则,并且说:你们就该从这些国家买。别从任何有政府后门的人那里买。那不是好主意。 这就是我们获得杠杆的方式。您知道,正因如此,我对总统的关税议程并不太热心,因为我们没有在建立联盟。 我们在制造敌人。而我想做的是把供应链放到这里,把半导体的流动放到这里,降低成本,尽可能快地推进。 Well, we do not want a reoccurrence of a Huawei that develops faster and then has a government back door and then we all have to raise opposition. I am for a tech NATO. I am for the five most sophisticated democracies and tech nations setting the rules of the road and saying, this is who you should buy from. Do not buy from anybody else who has a government back door. Not a good idea. So that is how we get leverage. You know, I am not so hot on the President's tariff agenda for this very reason because we are not building the alliances. We are creating the enemies, and what I want to do is get the supply chain here, get the semiconductor flow here, lower the cost, and go as fast as we can. ¶
Mr. Intrator 是的,我同意这一点。我不认为那是——我不认为会有人不同意这一点,对吧?我认为那是一个极好的目标。 我只是认为,在那五个北约国家公司之外,会发生的情况是:对人工智能会有需求,而他们会用他们能用的东西继续推进。 Yes, I agree with that. I do not think that that is--I do not think anybody is not going to agree with that, right? I think that is an excellent objective. I just think that what will happen beyond the five NATO companies is that there will be a demand for artificial intelligence and they will proceed with what they can proceed with. ¶
Senator Cantwell 苏博士,您怎么看:我们如何取胜、如何保护我们的目标,同时在出口战略上更为进取? Dr. Su, what is your view of this about how we win, how we protect our objectives, but we are more aggressive on the export strategy? ¶
Ms. Su 嗯,我认为大家清楚地认识到我们需要一套出口战略,所以进行——您知道,进行这场对话非常重要;从我们的角度看,思路是确保我们的盟友——说实话,我用「盟友」这个词是在最宽泛的意义上——能够获得我们所拥有的卓越美国技术,同时配套适当的管制,我认为这两件事可以同时做到。 回应您先前的话,坎特韦尔首席成员,关于需要让美国技术进入那些国家:我认为那些国家其实非常有兴趣这么做,因为我们今天确实拥有最好的技术;利用这一点真正构建起这个广泛的 AI 生态,正是我们的机会。 Well, I think there is a clear recognition that we need an export strategy and so having--you know, having this conversation is very important, and from our perspective the idea is to ensure that our allies--and, frankly, I use allies in the very broadest sense--get access to the great American technology that we have with the appropriate controls in place, and I think you can do both. To your earlier comment, Ranking Member Cantwell, about the need to have U.S. technologies in those countries I think those countries are actually very interested in doing that because we do have the best technology today, and using that to really build this broad AI ecosystem is really our opportunity. ¶
Senator Cantwell 我同意。非常感谢。 I agree. Thank you so much. ¶
Senator Sheehy 请俄亥俄州资深参议员发言。 The senior senator from Ohio. STATEMENT OF HON. BERNIE MORENO, U.S. SENATOR FROM OHIO ¶
Senator Moreno ### 尊敬的伯尼·莫雷诺(Bernie Moreno)陈述 ### 俄亥俄州联邦参议员 谢谢您,希伊主席。请务必让克鲁兹(Cruz)参议员也听到这一句。 那么,首先,感谢各位到场、抽出时间。我想快速确认一下,我听到各位相当一致的说法是:我们需要在这个国家大幅增加发电能力。对吗? 好。那么,苏博士,您最近刚与台积电(TSMC)合作,在美国本土制造你们的芯片。谢谢您。我认为这有点姗姗来迟。我希望您——我们希望我们早些时候能多做些这类事情。 那些半导体晶圆厂是高能耗用户吗? Thank you, Chairman Sheehy. Make sure Senator Cruz heard that one. So, first of all, thank you for being here and taking the time. If I could just real quickly just confirm that I have heard what you said pretty unanimously, which is we need dramatically more power generation in this country. Is that correct? All right. So, Dr. Su, you just recently did a partnership with TSMC to manufacture your chips here in America. Thank you. I think it is a little bit long overdue. I wish you had--we had done more of that earlier. Are those semiconductor fabs high energy users? ¶
Ms. Su 谢谢您,参议员。 我们非常高兴能与政府一道努力,把更多制造环节带回美国。 回答您的问题:当然,半导体制造厂是高能耗用户,正如您提到的,我们既需要更多电力用于制造,也需要更多电力用于数据中心。 Thank you, Senator. We are very pleased with our efforts together with the government on bringing more manufacturing back to the United States. To your question, certainly, semiconductor manufacturing plants are high energy users and we do need more power for both manufacturing as well as for data centers, as you mentioned. ¶
Senator Moreno 而没有芯片,这一切就无从谈起。比如,如果我们不能在美国本土制造出性能最高的芯片,这一切就不会在这里发生,对吗? And without chips this just does not work. Like, if we do not have the highest performance chips made here in the United States this is not going to happen here, correct? ¶
Ms. Su 我们绝对需要性能最高的芯片,我们还需要芯片制造所需的整个生态。晶圆是其中一环,但还有很多其他环节。 We absolutely need the highest performing chips and we also need the entire ecosystem for chip manufacturing. So wafers are one piece but there are many other pieces as well. ¶
Senator Moreno 那些芯片是靠太阳能和风力发电驱动的吗? And are those chips powered by solar power and windmills? ¶
Ms. Su 今天还不是,但我认为显然有机会这样做。 Today they are not but I think there are opportunities to certainly do that. ¶
Senator Moreno 那么,您是否认为这很离谱:去年由于拜登(Biden)政府的政策,我国新增发电量的 90% 是风机和太阳能板,我们实实在在地把美国的能源能力打残了? 我们在宾夕法尼亚、俄亥俄和西弗吉尼亚脚下有可用一千年的天然气,然而去年我国发电量的 90% 却是太阳能板和风机。这让我们这个国家更有竞争力,还是更没有竞争力? 这个问题谁想答都可以接。 So do you think it is outrageous that last year because of the policies of the Biden administration that 90 percent of new power generation in this country was windmills and solar panels and we absolutely kneecapped American energy? We have a thousand years of natural gas sitting in Pennsylvania, Ohio, and West Virginia, and yet 90 percent of power generation in this country last year was solar panels and windmills. Does that make this country more competitive or less competitive? Anybody can jump into that one that wants to answer that. ¶
Mr. Smith 让我说两点。第一,您说得对,我们需要更多电力。我认为值得记住的是,我们这个行业只会占我国新增电力总需求的 15%。 我们将需要来自多种来源的电力。今天在美国,我们 56% 的电力来自含碳能源,44% 来自无碳能源,也就是核能、风能或太阳能。我们需要一个广泛基础的做法,需要来源的多样性。 Let me say two things. One, you are right, we need more electricity. I think our industry, it is worth remembering, is only going to account for 15 percent of the total additional electricity the country is going to need. We are going to need electricity from a variety of sources. Today in the United States 56 percent of our electricity comes from carbon. Forty-four percent comes from carbon-free energy, meaning nuclear wind, or solar. We need a broad based approach and we need a diversity of sources. ¶
Senator Moreno 而且,再一次,那 90% 是不便宜、不充裕、也不可靠的能源。 我换个话题。奥特曼先生,首先感谢你以开放的方式创建了你的平台,并同意坚守非营利地位的原则。我认为这非常重要。 你认为互联网时代,从上世纪 90 年代初到 2000 年代,在保护儿童方面做得如何? And, again, 90 percent was energy that is not affordable, it is not abundant, and it is not reliable. Let me just shift gears. Mr. Altman, thank you for, first of all, creating your platform in an open basis and agreeing to stick to the principles of nonprofit status. I think that is very important. Do you think that the Internet age did a good job between the beginning of the 1990s through the 2000s of protecting children? ¶
altman 我想说,倒不是特别需要。 I would say not particularly. ¶
Senator Moreno 是的。而且你刚当上父亲,对吗? Yes. And you are a new father, correct? ¶
altman 是的。 Yes. ¶
Senator Moreno 恭喜。 Congratulations. ¶
altman 非常感谢。 Thank you very much. ¶
Senator Moreno 他一切都好? He is doing well? ¶
altman 他是。这是有史以来最神奇的事情。【译注:原文 He is. 所回应的问句未在提取上下文中完整保留(上下文仅到莫雷诺参议员「你刚当爸爸了,对吧?」,其间可能涉及其他嘉宾话轮);按原文直译,不推测补写】 He is. It is the most amazing thing ever. ¶
Senator Moreno 是的。我不认为你会希望你孩子最好的朋友是一个 AI 机器人。 Yes. I do not think you want your best-- your child's best friend to be an AI bot. ¶
altman 我不希望。 I do not. ¶
Senator Moreno 那我们能做什么?我们如何携手保护儿童? So what can we do? How can we work together to protect children? ¶
altman 我们已经谈了很多我们正在这里做的一些事情。我们努力吸取上一代的教训,你知道,事情大致就是这样:人们犯错,然后下一次做得更好。我们内部常说的一句话是:我们想把成年用户当成年人对待。我们想给他们很大的灵活性。我们想让他们以很大的自由度使用服务;而对儿童,则需要高得多的保护水平,这意味着服务不会做他们可能想要的一些事。现在,我们还处于早期,所以有时人们说,哦,你们的规则太严了——只是我们没法完美地判断这个。但如果我们能划出一条线,如果我们能确切知道用户是儿童还是成人,我们会允许对成年人宽松得多,而对儿童实行更严格的规则。 We have talked a lot about some of the things we are doing here. We are trying to learn the lessons of the previous generation and, you know, that is kind of the way it goes. People make mistakes and you do it better next time. One thing we say a lot internally is we want to treat our adult users like adults. We want to give them a lot of flexibility. We want to let them use the service with a lot of freedom, and for children there needs to be a much higher level of protection, which means the service will not do things that they might want. Now, we are still early so sometimes people say, oh, you are being too strict on the rules and it is just we cannot perfectly, like, tell this. But if we could draw a line and if we knew for sure when a user was a child or an adult we would allow adults to be much more permissive and we would have tighter rules for children. ¶
Senator Moreno 所以我想请求的是:能否让你的团队承诺与我们的团队合作,确保我们尽早搭好正确的框架——我认为这是我们向前推进的最好方式,因为我们不想过度监管,但我们不能重犯互联网和社交媒体时代的错误,那时候孩子们受到了伤害。 So I think what I would ask is if you could have your team commit to having your teams work with our teams to make certain that we put together the right framework early on I think is the best way we can move forward, because we do not want to overregulate but we cannot repeat the mistakes of the Internet and social media era where children got harmed. ¶
altman 我们很乐意与你们所有人合作。我认为这极其重要。 We would be delighted to work with you all. I think it is super important. ¶
Senator Moreno 谢谢。 Thank you. ¶
altman 我可以说一件关于你所说的话的事吗? Can I say one more thing about what you said? ¶
Senator Moreno 当然。 Of course. ¶
altman 这种 AI 与社会关系的想法,我认为这是一件新事物,我们需要高度关注。人们越来越依赖 AI 获取人生建议、某种程度上是情感支持之类的东西。这是近几个月才出现的新事物,但我——我并不认为它全是坏事,但我认为我们必须,比方说,理解它并非常仔细地观察它。 This idea of AI and social relationships I think this is a new thing that we need to pay a lot of attention to. People are relying on AI more and more for life advice, sort of, emotional support, that kind of thing. It is a newer thing in recent months, but I--and I do not think it is all bad, but I think we have to, like, understand it and watch it very carefully. ¶
Senator Moreno 谢谢,也谢谢你的这项承诺,非常感激。我已经和你的团队谈过了。都是很好的人。 Thank you, and thank you for that commitment. It is very appreciated. I have talked to your team already. Good people. ¶
altman 很好。 Great. ¶
Senator Moreno 因特拉托先生,很快问一句:能否谈谈在美国建立一个稳健的稳定币生态的重要性,它与支付的未来有什么关系,以及 AI 将如何参与其中?因为我认为人们还没看清这如何嵌入更大的图景。 Mr. Intrator, real quickly, can you talk about the intersection between the importance of a robust stablecoin ecosystem here in America and how that has a future with payments and how AI will factor into that? Because I do not think people see how this fits into the broader puzzle. ¶
Mr. Intrator 谢谢您的问题。 我们最初其实是一个加密领域的副业公司,后来有点跑偏成了现在这样。 您看,我认为稳定币、加密货币、AI 共享某些共同的基因,那就是它们都试图建起一个新技术会让社会变得更好的未来;而我们有一个巨大的潜力,可以把稳定币、加密货币和 AI 结合起来,取得更好的结果。 So thank you for the question. And we did start out as a crypto-based company hobby that kind of got away from us a little bit. Look, I think that stablecoins, crypto, AI, they share certain DNA in common which is that they are attempts to build into a future where new technology will make things better for society and there is a huge potential for us to use stablecoins, crypto, and AI in a combination for better outcomes. ¶
Senator Moreno 好。谢谢。 这是 1959 年以来最快的一次政变。 〔笑声。〕 **The Chairman〔主持〕.** 克洛布查尔参议员。 All right. Thank you. And that was the quickest coup since 1959. [Laughter.] The Chairman [presiding]. Senator Klobuchar. STATEMENT OF HON. AMY KLOBUCHAR, U.S. SENATOR FROM MINNESOTA ¶
Senator Klobuchar ### 尊敬的艾米·克洛布查尔(Amy Klobuchar)陈述 ### 明尼苏达州联邦参议员 非常感谢,克鲁兹参议员。 AI 有很多令人兴奋的东西,尤其是对像我这样的州——梅奥诊所(Mayo Clinic)就在我们州,它有着释放科学研究潜力的可能:我们已经绘制了人类基因组,还有很多罕见病可以被攻克。 所以有很多积极面,但我们都知道,正如各位所表达的,也有一些挑战需要通过许可改革去应对。我对此深信不疑。还有能源开发。 谢谢您,史密斯先生,提到风能、太阳能以及更多聚变与核能的潜力;但风能和太阳能的价格在过去几年大幅下降,而要达到目标,我们还得做得更好。 我认为戴维·布鲁克斯(David Brooks)说得最好,他说:「我发现写关于 AI 的东西难得要命,因为这项技术究竟在把我们带向天堂还是地狱,实在是不可知的。」 我们希望它把我们带向天堂,而我认为做到这一点的办法,是确保我们有一套道路规则,使它不会因为诈骗、或者因为想害我们的人的使用而受阻或被拖回去。 正如坎特韦尔参议员提到的,我和图恩(Thune)参议员已经联手提出立法,为风险最高的非国防类 AI 应用设立基本防护栏。 奥特曼先生,你是否同意:以风险为基础进行监管,是在不为扼杀创新的前提下为 AI 设置必要防护栏的最佳方式? Thank you very much, Senator Cruz. A lot of exciting things with AI, especially from a state like mine that is home to the Mayo Clinic with the potential to unleash scientific research while we have mapped the human genome and we have rare diseases that can be solved. So there is a lot of positive but we all know, as you have all expressed, there are challenges that we need to get at with permitting reform. I am a big believer in that. energy development. Thank you, Mr. Smith, for mentioning this with wind and solar and the potential for more fusion and nuclear, but wind and solar the price going down dramatically in the last few years and to get there we are going to have to do a lot better. I think David Brooks put it the best when he said, ``I found it incredibly hard to write about AI because it is literally unknowable whether this technology is leading us to heaven or hell''. We want it to lead us to heaven and I think we do that by making sure we have some rules of the road in place so it does not get stymied or set backward because of scams or because of use by people who want to do us harm. As mentioned by Senator Cantwell, Senator Thune and I have teamed up on legislation to set up basic guardrails for the riskiest nondefense applications of AI. Mr. Altman, do you agree that a risk-based approach to regulation is the best way to place necessary guardrails for AI without stifling innovation? ¶
altman 我同意。这对我来说很有道理。 I do. That makes a lot of sense to me. ¶
Senator Klobuchar 好,谢谢。那是在你的阁楼里想出来的吗? OK, thanks. And did you figure that out in your attic? ¶
altman 不,那是一个较近的发现。 No, that was a more recent discovery. ¶
Senator Klobuchar 谢谢。很好。我只是想确认一下。 我们的法案指示商务部,史密斯先生,开发教育消费者的方式,教他们如何安全使用 AI 系统。你是否同意消费者需要更多教育?这是你回答「五个词」时的答案之一,所以我假定你是同意的。 Thank you. Very good. Just want to make sure. Our bill directs, Mr. Smith, the Commerce Department to develop ways of educating consumers on how to safely use AI systems. Do you agree that consumers need to be more educated? This was one of your answers to your five words so I assume you do. ¶
Mr. Smith 是的,而且我认为,作为公司和整个商界,我们也有责任为这种教育作出贡献。 Yes, and I think it is incumbent upon us as companies and across the business community to contribute to that education as well. ¶
Senator Klobuchar 好,很好。 回到奥特曼先生。如我们所知,美国人越来越依赖 AI 来处理一些高影响的问题。要让他们能够信任,我们就必须确保我们可以信任模型的输出。 《纽约时报》本周早些时候报道说,AI 的「幻觉」——这对我来说是个新词——即模型生成不正确或误导性结果的情况,正在变得更糟。这是他们的说法。 OpenAI 用什么标准或指标来评估其训练数据的质量,以及模型输出的正确性? OK, very good. Back to Mr. Altman. Americans rely on AI, as we know, increasingly on some high impact problems. To make them be able to trust that we need to make sure that we can trust the model output. The New York Times recently reported earlier this week that AI hallucinations--a new word to me--where models generate incorrect or misleading results are getting worse. That is their words. What standards or metrics does OpenAI use to evaluate the quality of its training data and model outputs for correctness? ¶
altman 总的来说,AI 的幻觉(hallucinations)正在明显好转。我们还没有完全解决问题,但在过去几年里我们取得了相当显著的进展。当我们最初推出 ChatGPT 时,它会时不时地产生幻觉。这种鲁棒性(robustness)的概念——确信你可以信任这些信息——我们在那里取得了巨大进展。我们引用来源。模型变得更聪明了。很多人一直在用这些系统,我们曾担心如果它不是 100.0% 准确——这对这些系统来说仍然是个挑战——就会引发一连串问题。但用户是聪明的。人们理解,你知道的,这些系统擅长什么,什么时候用它们,什么时候不用,而随着这种鲁棒性的提升——它会持续提升——人们会把它用于越来越多的事情。但我们已经——作为一个行业,在过去几年里在那个方向上取得了相当显著的进展。 On the whole, AI hallucinations are getting much better. We have not solved the problem entirely yet, but we have made pretty remarkable progress over the last few years. When we first launched ChatGPT it would hallucinate things all the time. This idea of robustness, being sure you can trust the information, we have made huge progress there. We cite sources. The models have gotten much smarter. A lot of people use these systems all the time and we were worried that if it was not 100.0 percent accurate, which is still a challenge with these systems, it would cause a bunch of problems. But users are smart. People understand, you know, what these systems are good at, when to use them, when not, and as that robustness increases, which it which it will continue to do, people will use it for more and more things. But we have made--as an industry we have made pretty remarkable progress in that direction over the last couple of years. ¶
Senator Klobuchar 我知道我们会持续关注这一点。另一个挑战是——我们已经看到,克鲁兹参议员和我一起推进一项法案很久了,那就是《下架法案》(Take It Down Act);情况是,我们越来越多地看到这样的互联网活动:孩子想找个男朋友或女朋友,也许发出了一张自己的真实照片。 结果照片在他们学校被传播,或者有人试图借此向他们诈骗钱财,或者是 AI——我们越来越多地看到,甚至不是某人自己的照片,而是有人把一张假的身体接上去;我们一年之内就有大约 20 多起年轻人自杀,因为他们觉得自己的生活被毁了,因为这件事——他们会以这种方式被曝光。 所以我们推动这项法案在参众两院通过。第一夫人支持它,它正在送往总统签署。史密斯先生,能否谈谈我们如何构建能更好识别有害深度伪造的模型? I know we will be watching that. Another challenge that has been--we have seen, and Senator Cruz worked and I worked on a bill together for quite a while and that is the Take It Down Act, and that is that we are increasingly seeing Internet activity where kids looking for a boyfriend or a girlfriend, maybe they put out a real picture of themselves. It ends up being distributed at their school or they somehow they--someone tries to scam them for financial gain, or it is AI, as we have increasingly seen where it is not even someone's photos but someone puts a fake body on there and we have had about over 20 suicides in one year of young people because they felt like their life was ruined because this was-- they were going to be exposed in this way. So this bill we passed and through the Senate and the House. The First Lady supported it and it is headed to the President's desk. Could you talk about how we can build models that can better detect harmful deep fakes, Mr. Smith? ¶
Mr. Smith 是的。我是说,我们正在做这件事。OpenAI 在做。我们中不少人都在做;我认为目标是先识别出由 AI 生成的内容,然后往往是识别出哪类内容是有害的,而我认为我们在这两方面的能力上都取得了很大进展。 私营部门有大量工作正在进行,并与全国失踪与受虐儿童中心(NCMEC)等机构合作,以协同识别这类内容,从而将其下架。 自互联网出现以来,我们在某些方面已经做了 25 年,我们还需要做得更多。 Yes. I mean, we are doing that. OpenAI is doing that. A number of us are and I think the goal is to first identify content that is generated by AI and then often it is to identify what kind of content is harmful, and I think we have made a lot of strides in our ability to do both of those things. There is a lot of work that is going on across the private sector and in partnership with groups like NCMEC to then collaboratively identify that kind of content so it can be taken down. We have been doing this in some ways for 25 years since the Internet and we are going to need to do more of it. ¶
Senator Klobuchar 还有这个问题——主席先生,最后一个问题。既然上一个问的是你们的法案,我想这个应该也可以。史密斯先生,各家报纸——你曾在参议院司法委员会就那项法案作证。 肯尼迪参议员和我仍然认为,这里存在一个关于内容费率谈判的问题。我们最近在加拿大和其他地方看到了一些动作。 能否谈谈这些演变中的——AI 开发者之间的动态,以及这里正在发生什么,以确保内容提供者和记者为其作品获得报酬? And on the issue--last question, Mr. Chair. Since the last one was about your bill I figured it is OK. The newspapers, and you testified before the Senate Judiciary Committee, Mr. Smith, about the bill. Senator Kennedy and I still think that there is an issue here about negotiating content rates. We have seen some action recently in Canada and other places. Can you talk about those evolving--the dynamics with AI developers and what is happening here to make sure that content providers and journalists get paid for their work? ¶
Mr. Smith 是的,这是个复杂的话题,但我只说几点。 第一,我认为我们都应当希望报纸以某种形式在全国各地兴旺,包括那些日益变成「新闻荒漠」的乡村县——报纸已经消失了。 第二,这也是我们在司法委员会讨论过的问题:应当给报纸一个机会,让它们联合起来集体谈判。我们支持这一点。这将使它们大体上能做得更好。 第三,每当出现新技术,就会出现新一代的版权辩论。这场辩论现在正在进行。其中一部分大概会由国会决定,一部分由法院决定。 很多也正在通过协作行动来处理;我们应当希望所有这些,我只想说,达成一种平衡。我们希望人们能靠创作内容谋生,我们也希望 AI 能通过获取数据而进步。 Yes, it is a complicated topic but I will just say a couple of things. First, I think we should all want to see newspapers in some form flourish across the country including, say, rural counties that increasingly have become news deserts. Newspapers have disappeared. Second, and it has been the issue that we discussed in the Judiciary Committee, there should be an opportunity for newspapers to get together and negotiate collectively. We have supported that. That will enable them to basically do better. Third, every time there is new technology there is a new generation of a copyright debate. That is taking place now. Some of it will probably be decided by Congress, some by the courts. A lot of it is also being addressed through collaborative action, and we should hope for all of these things to, I will just say, strike a balance. We want people to make a living creating content and we want AI to advance by having access to data. ¶
Senator Klobuchar 好,谢谢。其余问题我会以书面形式提出。谢谢您,主席先生。 **The Chairman.** 谢谢。 您知道,克洛布查尔参议员问 AI 会把我们带向天堂还是地狱。这让我想起耶鲁法学院教授格兰特·吉尔摩(Grant Gilmore)的一句著名评论:在天堂里没有法律,狮子会与羔羊同卧;在地狱里除了法律一无所有,而正当程序被一丝不苟地遵守。 让我问这个问题,问四位证人每一位:在 AI 竞赛中,谁在赢,美国还是中国?如果答案是美国,中国离我们有多近?我们要做什么才能确保答案仍然是「美国会赢」? 奥特曼先生,从你开始。 OK, thanks. I will ask other questions on the record. Thank you, Mr. Chair. The Chairman. Thank you. You know, Senator Klobuchar asked whether AI will lead us to heaven or hell. It reminded me of a famous observation by Yale Law Professor Grant Gilmore that in heaven there is no law and the lion will lie down with the lamb. In Hell there is nothing but law and due process is meticulously observed. Let me ask you this, and this is to each of the four witnesses. In the race for AI who is winning, America or China? If the answer is America how close is China to us, and what do we do to make sure the answer remains America will win? Mr. Altman, we will start with you. ¶
altman 我们相信,美国的模型,包括我们公司 OpenAI 以及 Google 等公司的一些模型,是世界上最好的模型。很难说我们领先多少,但我会说时间上不是很多,而且我认为,为了维持这种领先地位、以及随之而来的影响力、以及世界使用美国技术产品和服务所带来的所有不可思议的益处——我的同事们在这里谈到的那些东西,在基础设施上取胜的必要性、不拖慢我们的明智监管、那种我认为是世界上美国所特有的创新与创业精神——这些都不是什么高深莫测的事(rocket science)。我们只需要继续做那些长期有效的、正确的事情,不要犯愚蠢的错误。 It is our belief that the American models, including some models from our company OpenAI and Google and others are the best models in the world. It is very hard to say how far ahead we are but I would say not a huge amount of time, and I think to continue that leadership position and the influence that comes with that and all of the incredible benefits of the world using American technology products and services, the things that my colleagues have spoken about here, the need to win in infrastructure, sensible regulation that does not slow us down, the sort of spirit of innovation and entrepreneurship that I think is a uniquely American thing in the world, none of this is rocket science. We just need to keep doing the things that have worked for so long and not make a silly mistake. ¶
Ms. Su 我就芯片这个领域来回答。我要说,美国今天在芯片上领先。我们拥有世界上最好的人工智能加速器。 我认为中国,尽管受到限制,考虑到他们使用先进技术的能力——有一点我们所有人都必须记住:做成一件事有很多种办法。您知道,拥有最好的芯片很棒,但即使你没有最好的芯片,你也能做成很多事。 所以我认为,关于「中国落后多少」的这场讨论——他们当然正在追赶,因为做成一件事有很多条路。 相对于我们能做的事,我会继续强调:真正要确保我们的创新精神得以发挥,而这需要有非常支持性的政府政策,需要有非常一致的政策,让我们能做我们最擅长的事,那就是在技术栈的每一层创新。 **The Chairman.** 因特拉托先生。 I will answer in the realm of chips. I would say America is ahead in chips today. We have the best AI accelerators in the world. I think China, although they have restrictions, given their ability to use advanced technologies, the one thing that is very important for us all to remember is there are multiple ways to do things. You know, having the best chips is great, but even if you do not have the best chips you can get a lot done. So I think this conversation about how far behind China is they are certainly catching up because there are many ways to do things. I think relative to what we can do I will continue to say really ensure that our spirit of innovation is allowed to work and that is having very supportive government policies to do that, having very consistent policies and allowing us to do what we do best, which is innovate at every layer of the stack. The Chairman. Mr. Intrator. ¶
Mr. Intrator 那我就从支撑这一切的物理基础设施和软件栈来讲。 美国领先,但从能力的角度看,这正是软肋——正如我开始说的——好一些了吗?抱歉。 所以,要建造非常大规模的解决方案来支撑这里的算力基础设施,这正是我们会吃力的地方:从许可审批、到建造大型项目、到能够提供电力,让那些构建人工智能的人能在这场竞赛中尽可能快地推进。 **The Chairman.** 史密斯先生。 So I will speak to it from the physical infrastructure and software stack to deliver that. America is ahead, but it is the Achilles heel from the perspective of the ability, as I started to--better? Sorry about that. So the ability to build very large solutions to the computing infrastructure component of this is an area that we are going to struggle with from a permitting and building large projects to be able to deliver the power to allow those building artificial intelligence to continue to move as fast as they can in the race that we are in. The Chairman. Mr. Smith. ¶
Mr. Smith 我认为美国今天在一场胶着的竞赛中领先,而这场竞赛很可能仍将胶着。 决定美国还是中国赢下这场竞赛的第一要素,是谁的技术在世界其他地区被最广泛地采用。 这是一个全球市场,而它正如技术市场通常那样,将由网络效应来界定。世界上 18% 的人生活在中国,4% 生活在美国,78% 生活在别处。 华为和 5G 的教训是:谁先到那里,谁就难以被取代。我们需要带着恰当类型的管制去出口。我们需要赢得世界其他地区的信任。 我们需要有金融架构,不仅覆盖工业化国家,也覆盖比如非洲各地的国家——在那里,中国和华为通常做得非常好。 **The Chairman.** 我的几位同事把标准说成是值得追求的东西,而我要说,「标准」常常是「监管」的代号;事实上,欧盟在互联网方面那些令人窒息的标准,正是杀死欧洲科技的原因。 我们现在看到各州议会模仿欧盟,比如加利福尼亚州的 S.B. 1047——所幸它被压倒性地否决了,但它本质上会为 AI 模型登记造出一个「加州车管所」。 如果美国走上欧盟这条路,为 AI 设立一套笨重的事前审批政府监管流程,那对赢下 AI 竞赛会有多大伤害? I think the United States has a lead today in what is a close race and a race that will likely remain close. The number-one factor that will define whether the United States or China wins this race is whose technology is most broadly adopted in the rest of the world. This is a global market and it will be defined as technology markets typically are by network effects. Eighteen percent of the people of the world live in China. Four percent live in the United States. Seventy-eight percent live somewhere else. The lesson from Huawei and 5G is whoever gets there first will be difficult to supplant. We need to export with the right kinds of controls. We need to win the trust of the rest of the world. We need to have the financial architecture that gets not only to the countries that are industrialized but the nations, say, across Africa where typically China and Huawei have done so well. The Chairman. So some of my colleagues have made reference to standards as something that is desirable, and I will say standards is often code word for regulations and, indeed, the EU stifling standards concerning the Internet is what killed tech in Europe. We are seeing now state legislatures mimicking the EU such as California's S.B. 1047 which, thankfully, was overwhelmingly defeated but would have created essentially a California DMV for AI model registration. How harmful would it be to winning the race for AI if America goes down the road of the EU and creates a heavy handed prior approval government regulatory process for AI? ¶
altman 我认为那将是灾难性的。为了更具体地回答你先前的问题——我认为它触及了为什么那会如此糟糕——这些 AI 系统有三个关键输入。一是算力,我们所说的一切基础设施;二是我们都在研究的算法;三是数据。如果你缺少其中任何一项,你就不可能成功造出最好的模型,而且正如 Brad 所说,美国在这里影响世界的方式,就是拥有人们最想使用、最广泛采用的技术。世界使用 iPhone、Google 和微软的产品,这很棒。比如,这就是我们拥有影响力的方式。我们不希望这种局面停止。所以,那些在任何这些领域阻止我们的制度,你知道的,如果我们就可以使用什么数据训练制定出不具全球竞争力的规则,那么事情就可能崩盘。如果我们无法建设基础设施,特别是如果我们无法在这个国家制造芯片,规则就可能崩盘。如果我们无法制造人们想要、并自然会在市场中胜出的产品——我认为人们确实想要使用美国产品。我们能把它们做成最好的。但如果阻止我们这样做,人们就会去使用别人做的更好的产品,而别人没有受到同样方式的——你知道的——阻碍。所以这是——我对过早设定标准感到紧张。我完全接受我的一些同事们所持的立场,即一旦行业弄清楚标准应该是什么,让政府机构采纳并使其更加正式是可以的。但我相信行业正在迅速朝着弄清正确的协议和标准前进,而我们需要创新的空间和快速行动的空间。主席:那么,如果你们每位能简要回答那个问题,因为我的时间到了。所以我想尊重这一点。 I think that would be disastrous. To give a more specific answer to your previous question, which I think touches on why it would be so bad, there are three key inputs to these AI systems. There is compute, all the infrastructure we are talking about, there is algorithms that we all do research on, and there is data. If you do not have any one of those you cannot succeed in making the best models and, as Brad said, the way for America to influence the world here is to have the technology that people most want to use and most adopt. The world uses iPhones and Google and Microsoft products, and that is wonderful. Like, that is how we have our influence. We do not want that to stop happening. So systems that stop us on any of these areas, you know, if we have rules about what data we can train on that are not competitive with the rest of the world then things can fall apart. If we are not able to build the infrastructure and particularly if we are not able to manufacture the chips in this country the rules can fall apart. If we cannot build the products that people want that naturally win in the market--and I think people do want to use American products. We can make them the best. But if we are prevented from doing that people will use a better product made from somebody else that does not have the sort of--you know, that is not stymied in the same way. So it is--I am nervous about standards being set too early. I am totally fine, you know, at the position that some of my colleagues took that standards, once the industry figures out what they should be, it is fine for them to be adopted by a government body and sort of made more official. But I believe the industry is moving quickly toward figuring out the right protocols and standards here and we need the space to innovate and to move quickly. The Chairman. So if each of you could briefly answer that question because my time has expired. So I want to be respectful of that. ¶
Ms. Su 我同意萨姆(Sam)刚才的那些说法。 I agree with the comments that Sam put up. ¶
Mr. Smith 我同意。我只想说,我认为您所指出的是:我们必须非常小心,不要有这些事前审批要求,包括在州一级,因为那会真正拖慢这个国家的创新。 I agree, and I would just say and I think the point you are making is we have to be very careful not to have these preapproval requirements including at state levels because that would really slow innovation in the country. ¶
Mr. Intrator 我认为,监管层层叠加、彼此不统一,会在我们建设和扩展所做之事的能力上造成摩擦。 **The Chairman.** 谢谢。 柯蒂斯(Curtis)参议员?沙茨(Schatz)? 沙茨参议员?抱歉。 I think that a patchwork of regulatory overlays will cause friction in the ability to build and extend what we are doing. The Chairman. Thank you. Senator Curtis? Schatz? Senator Schatz? Apologies. STATEMENT OF HON. BRIAN SCHATZ, U.S. SENATOR FROM HAWAII ¶
Senator Schatz ### 尊敬的布莱恩·沙茨(Brian Schatz)陈述 ### 夏威夷州联邦参议员 没问题,主席先生。 感谢各位到场。我只想接着主席的问题,以及委员会里某种——也许是正在形成的共识——说几句。 好。我不认为即便在席位这一侧,有谁在提议某种欧洲式的事前审批。 我认为有些人希望在监管领域什么都不做,但我认为多数人明白,一些防护栏——奥特曼先生,这是你用的词——规则和防护栏是必要的。 你是在说,在当前时刻自我监管已经足够了吗? No problem, Chairman. Thank you for being here. I just want to follow up on the Chairman's question and a sort of--maybe an emerging consensus on the Committee. OK. I do not think there is anybody even on this side of the dais that is proposing a sort of European style preapproval. I think there are some people who would like to do nothing at all in the regulatory space but I think most people understand that some guardrails--those are the words that you use, Mr. Altman--rules and guardrails are necessary. Are you saying that self-regulation is sufficient at the current moment? ¶
altman 不。我认为某些政策是好的。我认为它很容易走过头,而随着我对世界运作方式了解得更多,我更担心它可能走过头并产生非常糟糕的后果。但人们想要使用总体上是安全的产品。你知道的,当你坐上飞机时,你基本上不会考虑自己做安全测试。你会觉得,这是——嗯,也许现在不是用飞机举例的好时候——但你有点想直接相信你可以登上一架——【译注:碎片话轮,原文以破折号结尾,句意未完;按源件边界保留。原文「The Chairman.」为主席插话,被采集为一句话轮内文本,按源件保留不译。】 No. I think some policy is good. I think it is easy for it to go too far, and as I have learned more about how the world works I am more afraid that it could go too far and have really bad consequences. But people want to use products that are generally safe. You know, when you get on an airplane you kind of do not think about doing the safety testing yourself. You are, like, this is--well, maybe this is a bad time to use the airplane example but you kind of like want to just trust that you can get on an---- ¶
Senator Schatz 现在正是用飞机那个例子的好时机。但我认为你这一点说得完全正确:看,确实有一场竞赛,但我们需要明白我们在为什么而竞赛,对吧? 这也关乎美国价值。这不仅仅是一场商业竞赛,好让我们在公共部门和私营部门都压过我们的同级竞争对手。我们是在为让美国价值在国际上占据主导而赢一场竞赛。 史密斯先生,我想换一个话题。在我看来,在消费者一侧,互联网用户最基本的权利之一,是理解自己在看什么、在听什么,以及那究竟完全由人创作、由人借助 AI 创作,还是由 AI 自动生成。 你是否认为一套标识制度——不是禁止使用 AI,而只是披露,特别是在图像、音乐、创意方面——你认为标识对消费者会有帮助吗? It is an excellent time to use the airplane example. But I think your point is exactly right is that, look, there is a race but we need to understand what we are racing for, right? And it also has to do with American values. It is not just a sort of commercial race so we can edge out our near peer competitor both in the public sector and the private sector. We are trying to win a race so that American values prevail internationally. Mr. Smith, I want to move on to another topic. It seems to me that on the consumer side that one of the most basic rights of a user on the Internet is to understand what they are looking at or listening to and whether or not it was created solely by a person, a person using an AI, or automatically generated using AI. Do you think a labeling regime--not a prohibition on the use of AI but just the disclosure, especially as it relates to images, music, creativity--do you think a label would be helpful for consumers? ¶
Mr. Smith 总的来说,是的;我认为这正是我们业界一直在努力建立的东西。我认为您做出区分是对的,尤其是聚焦于图像、视频、音频文件。 有一个叫 C2PA 的标准,我们和不少公司现在都在推进它。它带有内容凭证,能让人知道某个东西是在哪里创建的、由谁创建的;我认为——您说得对——也能知道它是由人创建、由 AI 创建,还是由人借助 AI 创建的。 Generally, yes, and I think that is what we in the industry have been working to create. I think you are right to make the distinction and focus especially on, say, images, video, audio files. There is a standard called C2PA that we and a number of companies now have been advancing. It has content credentials. It enables people to know where something was created, who created it, and I think--you are right--to know whether it was created by a person, by AI, or a person with the help of, say, AI. ¶
Senator Schatz 我只想用日常语言,而不是你们所有人都在用、或者我们都被训练去用的那套语言。 当你把数据说成让模型运转的三个要素之一时,数据其实就是知识产权。它是人类的创新、人类的创造力;我确实认为我们之间可能存在分歧——我同意克洛布查尔参议员所说的,需要认识到这些模型是在数据上训练的,但我们真正谈的是直到此刻为止的人类成就。 我有一个深切的担忧——看,在能源领域和公共服务领域,当然还有健康创新领域,我其实是个乐观主义者。这里有很多非常令人兴奋的机会。 但我们必须为人们的知识付酬;我担心这些模型在吐出海量看似知识的东西上会如此成功,以至于在末端,我们不会为所有的输入付酬,而这些模型会就此停滞。 你谈到了某种张力,但我想弄清楚这种张力究竟是什么——除了「你希望为这些输入付得尽可能少」之外。 请讲,史密斯先生。 I just want to use sort of common language, not the language that all of you use or that we have all learned to use. When you talk about the data as one of the three elements that makes a model work, data really is intellectual property. It is human innovation, human creativity, and I do think we may have a disagreement--and I agree with Senator Klobuchar about the need to understand that these models have been trained on data but what we are really talking about is human achievement all the way up to now. And I have a deep worry--look, I am actually an optimist in the energy space and the public service space, certainly in health innovation. There are a lot of really exciting opportunities here. But we got to pay people for their knowledge and I am concerned that these models are going to be so successful in spitting out what appears to be knowledge that we are going to, on the back end, not pay people for all of the inputs and we will have a sort of stalling out of these models. And you talked about a tension but I am trying to figure out what the tension really is other than you would like to pay as little as possible for these inputs. Go ahead, Mr. Smith. ¶
Mr. Smith 嗯,您说到最后一句之前我都被您说服了。 Well, you had me until the last sentence. ¶
Senator Schatz 我知道。 I know. ¶
Mr. Smith 看,我们创造知识产权,我们尊重知识产权。所以我们坚决认为,知识产权以及它的创造应当得到回报。 归根结底,知识产权法永远是关于划线的。就是您说的那条线。在版权里,受保护的是「表达」。如果你写了一本书、有人照抄,那你就有权获得报酬。 但还有「思想」。如果有人读了你的书,如果有人记得莎士比亚写过一个两个少年相爱的故事—— Look, we create intellectual property. We respect intellectual property. So we are emphatically of the view that intellectual property and the creation of it should be rewarded. Ultimately, intellectual property laws are always about drawing the line. It is really the line that you refer to. In copyright, there is expression that is protected. If you write a book and somebody copies it then you are entitled to be paid. But there are ideas. If someone reads your book, if someone remembers that Shakespeare wrote a story about two teenagers who fell in love---- ¶
Senator Schatz 当然。那属于合理使用。 Sure. Then that is fair use. ¶
Mr. Smith 那属于合理使用。这正是这个国家和国会创设它的原因。 Then that is fair use. That is why this country and Congress created it. ¶
Senator Schatz 好。张力就在这里。 OK. That is where the tension is. ¶
Mr. Smith 这正是我们需要聚焦的地方。 That is what we need to focus on. ¶
Senator Schatz 主席先生,征得您同意,我想问最后一个问题。 **The Chairman.** 请讲。 With your permission, Chairman, I want to ask one final question. The Chairman. Proceed. ¶
Senator Schatz 谢谢。其实我对这样一个前景相当兴奋:20 年后人们会说,还记得那时候你得打电话排队,才能跟凯撒医疗(Kaiser Permanente)或者退伍军人事务部(VA)说上话吗? 所以我——也许奥特曼先生和史密斯先生,我想请你们——您知道,我有个哥们儿以前常说,画一幅画,把我画进去。 好。就政府实际提供服务这件事,我想请你们描述一下,一个 AI 智能体或 AI 能做什么,来减轻那些我们在与政府打交道时当作生活常态接受的痛点。 在我看来,让我们对政府恼火的很大一部分原因,是明明某处存在数据,却缺乏整理,我们拿不到。 那么请很快地说,每人 15 秒,来点鼓劲的话。 Thank you. I am actually quite excited about the prospect that in 20 years people are going to say, remember when you had to wait on the phone to talk to Kaiser Permanente or the VA? So I just--maybe Mr. Altman and Mr. Smith, I want you to-- you know, a buddy of mine used to say, paint a picture and paint me in it. OK. For the government actually delivering services I want you to describe what an AI agent or AI can do to kind of reduce those pain points that we accept as a fact of life in interacting with the government. It seems to me so much of what makes us irritated with the government is the lack of sorting data that exists somewhere but we cannot get access to it. So just very quickly, you have 15 seconds each for some cheerleading. ¶
altman 我可以设想这样一个未来:美国政府提供一种 AI 驱动的服务,让你能够极其便利地使用所有政府服务,获得优质的医疗、优质的教育。你口袋里有这么一个东西,如果你有任何医疗问题,你就能得到答案;如果你需要,比如说,就你正在经历的某个政府流程提出申诉,或者报税之类的,你只要立刻就做。你口袋里有一个与政府完全整合的 agent(代理),生活就很轻松。 I can imagine a future where the U.S. Government offers a AI-powered service that makes it really easy to use all government services to get great health care, to get great education. You have this thing in your pocket and if you have any medical problem you get an answer if you need to, you know, like, appeal something on some process you are having with the government or file your taxes or whatever. You just do it instantly. You have an agent in your pocket fully integrated with the U.S. Government and life is easy. ¶
Senator Schatz 有什么要补充的吗? Anything to add? ¶
Mr. Smith 还记得你得排队去换驾照的时候吗?还记得你不知道该怎么报告你家街上那个需要修补的坑洞的时候吗? 还记得你车子轻微追尾,得填一大堆表格、跟一大堆人说话才能拿到保险赔付的时候吗? 现在,你可以用一套 AI 系统全部搞定。你可以用手机;顺便说一句,你今天在阿布扎比就能这么干。我们需要把它带到美国来。 Remember when you had to stand in line to renew your driver's license? Remember when you did not know how to report a pothole that needed to be repaired on your street? Remember when you had a fender bender in a car and you had to fill out all these forms and talk to all these people to get insurance coverage? Now you can do it all with one AI system. You can use your phone and, by the way, you can do this today in Abu Dhabi. We need to bring it to America. ¶
Senator Schatz 谢谢。 Thank you. ¶
Senator Young 巴德(Budd)参议员。 Senator Budd. STATEMENT OF HON. TED BUDD, U.S. SENATOR FROM NORTH CAROLINA ¶
Senator Budd ### 尊敬的泰德·巴德(Ted Budd)陈述 ### 北卡罗来纳州联邦参议员 谢谢您,主席先生。 再次感谢各位到场。我和各位都有过各种愉快的交谈。 美国新增发电能力、升级电网的能力,在很多方面是与中国的竞赛的关键。 能源是我们取胜的方式,也是我们落败的方式。这个国家的许可审批耗时太长。中国的指令控制体制意味着,他们部署所需能源、达到开发最先进模型所需的规模,不会失败——而这些模型将带来 AI 的全部好处。 所以我很高兴能与卢米斯(Lummis)参议员共同推进《FREE 法案》,它将建立一套「依规则许可」的结构,让大型项目在前期就满足全面标准,而不是被拖进逐案审批的流程。 我们都想保护环境,我们都想维持美国的经济与技术领导地位。 那么,因特拉托先生,CoreWeave 在电力合同方面的经验如何?你是否担心现行许可制度会让美国更难获得赢下这场 AI 竞赛所需规模的投资? Thank you, Chairman. Again, thank you all for being here. I have enjoyed various conversations with each of you. The ability for the U.S. to deploy new energy generation capacity and upgrade its grid it is in many ways the key to the race against China. Energy is how we can win and it is also how we can lose. Permitting in this country takes too long. China's command and control system means that they will not fail to deploy the energy needed to achieve the scale necessary to develop the most advanced models which will drive all the benefit of AI. So I am glad to be working with Senator Lummis on the FREE Act, which would set up a permit by rule structure which would let large projects meet comprehensive standards at the front end instead of dragged out on a case by case process. We all want to protect the environment and we all want to maintain U.S. economic and technological leadership. So, Mr. Intrator, what has CoreWeave's experience been in contracting power and are you concerned that the current permitting system can make it harder for the U.S. to achieve capital investment in the scale needed to win this AI race? ¶
Mr. Intrator 正如您所说,获得电力、获得规模化的电力,确实是我们赢下这场竞赛能力的关键之一。还有其他关键,但这是我花很多时间思考的一个。 我把这个说法拆成「获得电力」和「获得规模化电力」,因为我确实认为,我们正走向这场竞赛的一个阶段:把人工智能——构建它的实验室、构建它的公司——以必要速度向前推进所需的基础设施,其规模与体量,将成为一个真正需要深思的具体挑战。 我们组织中有很大一部分专注于不仅获得电力,而且获得足以支撑基础设施建设的电力规模——您知道,达到阿比林(Abilene)那样的规模或接近它——以便让这件事向前推进。 这很难,对吧;而且随着时间推移会更难,因为现有基础设施中那些还有余量的部分会被消耗掉;一旦被消耗完,你就得回到第一原理:现在怎么把电接上网?在目前配置的监管环境下,这真的会很有挑战。 So, as you said, access to power, access to scale power is certainly one of the keys to our ability to win this race. There are others but it is one that I spend a lot of time thinking about. I separated the comment into access to power and access to scale power because I do think that we are moving toward a period of this race where the size, the magnitude of the infrastructure that is being required to move our artificial intelligence--the labs that are building it, the companies that are building it--forward at the velocity that is necessary is going to be a specific challenge that really requires a lot of thought. We have a huge part of our organization focused on not just getting access to power but getting access to the size and scale of power that is going to be able to build the infrastructure, you know, at the scale of Abilene or close to it in order to, you know, allow this to move forward. It is tough, right, and it will get harder as we move through time because the existing infrastructure that does have opportunities it has some level of elasticity, is going to be consumed, and once that is consumed you are going to get down to kind of a first principle how do we get power online now, and that is really going to be challenging within the regulatory environment as it currently is configured. ¶
Senator Budd 谢谢。 史密斯先生,一个类似的问题。微软在如何为其数据中心锁定电力?我是说,我们最近在新闻里读到过;但联邦政策需要聚焦什么,才能确保我们不会因为拿不到足够的能源而输掉这场竞赛? Thank you. Mr. Smith, a similar question. How is Microsoft trying to secure power for its data centers? I mean, we read about that in the news recently but what does Federal policy need to focus on to make sure that we do not lose this race because we cannot get enough energy? ¶
Mr. Smith 嗯,我们投资把更多发电能力接入电网,然后把它经由电网输送到我们的数据中心。我们在更多国家拥有的许可申请,很可能比这个星球上任何一家公司都多。 我上次看的时候,是 40 多个国家的 872 份申请。有意思的是,在美国,许可方面头号挑战不是地方的,也不是州的,而是由陆军工程兵团(Army Corps of Engineers)管理的联邦湿地许可。 我们通常能在 6 到 9 个月内拿到地方和州许可。那个全国性的——湿地许可往往要 18 到 24 个月。 即将离任的拜登政府和即将上任的特朗普(Trump)政府都关注过这件事,但如果我们能把这个解决掉,我们就能在这个国家大幅加速。 Well, we invest to bring more electricity generation onto the grid and then to bring it through the grid to our data centers. We probably have more permitting applications in more countries than quite possibly any company on the planet. Last time I looked at it, it was 872 applications in more than 40 countries. The number-one challenge in the United States when it comes to permitting, interestingly enough, is not local. It is not state. It is the Federal wetlands permit that is administered by the Army Corps of Engineers. We can typically get our local and state permits done in about six to 9 months. The national--the wetlands permit is taking often 18 to 24 months. Both the outgoing Biden administration and the incoming Trump administration have focused on this, but if we could just solve that we could accelerate a lot here in this country. ¶
Senator Budd 非常有帮助。谢谢。 奥特曼先生,关于中国像 DeepSeek 这样的开源模型,已经有很多讨论。我们一两个月前谈过这个。 我的一个担忧是:由中国共产党推动的、易于获取的中国模型,对 AI 应用开发者来说可能是一个有吸引力的选择,尤其是在发展中经济体。 那么,美国在开源或闭源 AI 模型上的领导地位有多重要? Very helpful. Thank you. Mr. Altman, much has been made about the Chinese open source models like DeepSeek. We spoke about that a month or two ago. A concern that I have is that accessible Chinese models promoted by the Chinese Communist Party might be an attractive option for AI application developers to build on top of, particularly in developing world economies. So how important is U.S. leadership in either open source or closed AI models? ¶
altman 我认为在两者上都领先相当重要。我们意识到 OpenAI 可以在这方面做更多,所以我们将发布一个开源模型,我们相信它将成为今年夏天领先的模型,因为我们希望人们在美国的技术栈(stack)之上构建,包括闭源模型。世界上很多人使用我们的技术以及我们同道的技术。我们认为我们在那方面处境良好。 I think it is quite important to lead in both. We realize that OpenAI can do more to help here so we are going to release an open source model that we believe will be the leading model this summer because we want people to build on the U.S. stack in terms of closed source models. A lot of the world uses our technology and the technology of our colleagues. We think we are in good shape there. ¶
Senator Budd 那么联邦政策还能怎样进一步帮助鼓励 AI 生态在美国本土发展? So how could Federal policy further help encourage the AI ecosystem to be developed right here in the U.S.? ¶
altman 嗯,你在能源方面触及了一个很好的点。我认为怎么强调能源对这里未来的重要性都不为过。你知道的,最终芯片、网络设备,那将由机器人制造,我们会让它非常高效,让它越来越便宜。但电子就是电子。最终,智能的成本、AI 的成本,将收敛到能源的成本,而它将取决于你能拥有多少能源。它的丰裕程度将受到能源丰裕程度的限制。所以,就美国要做的长期战略投资而言,我想不出有什么比能源更重要的了。你知道的,芯片和所有其他基础设施也是,但我认为这——我认为这一件事最终会归结到能源上。 Well, you touched on a great point with energy. I think it is hard to overstate how important energy is to the future here. You know, eventually chips, network gear, that will be made by robots and we will make that very efficient and we will make that cheaper and cheaper. But an electron is an electron. Eventually, the cost of intelligence, the cost of AI, will converge to the cost of energy and it will be how much you can have. The abundance of it will be limited by the abundance of energy. So in terms of long-term strategic investments for the U.S. to make I cannot think of anything more important than energy. You know, chips and all the other infrastructure also but energy is where this--I think this ends up. ¶
Senator Budd 谢谢。主席先生? Thank you. Chairman? ¶
Senator Young 金(Kim)参议员。 Senator Kim. STATEMENT OF HON. ANDY KIM, U.S. SENATOR FROM NEW JERSEY ¶
Senator Kim ### 尊敬的安迪·金(Andy Kim)陈述 ### 新泽西州联邦参议员 谢谢。 史密斯先生,我想先从你开始,因为我觉得你关于「这场竞赛到底是什么」的那一点——您知道,我们一直在谈这场竞赛,而你用一种特定的方式框定了它,说它关乎在世界其他地区、那 78% 的人当中的采用。 我猜我只是想请你再多展开一些,谈谈在理解我们能在国会、在政府中扮演什么角色方面——去加速并倡导那种国际上的 AI 采用? Thank you. Mr. Smith, I think I would like to start with you because I thought your point about what exactly is the race, right--you know, we keep talking about the race, and you framed it in a particular way saying that it is about adoption in the rest of the world, the 78 percent. I guess I just wanted to ask you to tease that out some more in terms of understanding what role we could play in Congress, in government, in terms of trying to accelerate and champion that AI adoption internationally? ¶
Mr. Smith 我认为有两点。第一,这恰好凸显了把出口管制做对的重要性,也就是眼下正在讨论的「AI 扩散规则」;我认为它表明,作为一个国家,我们想要的是那种确保——比如说——芯片不会被转运到中国、或者不会被中国境内错误的使用者出于错误的原因获取的国家安全管制。 这是商务部里有人起草的东西。与此同时,我认为我们需要取消为所有这些「第二层级」国家设定的数量上限,因为它们所做的是向 120 个国家发出一个信号:你们未必能指望我们提供你们想要、也需要的 AI。 想一想就知道了。如果这是你们国家基础设施的关键部分,而你又不确定供应商能否满足你的需求,你怎么敢押注在他们身上? 所以我认为,你们在国会和参议院可以帮助白宫和商务部把这件事做对。 I think there is two things. The first is it just shines light on the importance of getting it right for export controls, which is the AI diffusion rule that is being discussed right now, and I think what it shows is we want to have, I believe, as a country the kinds of national security controls that ensure that, say, chips do not get diverted to China or get accessed by the wrong users, say, in China for the wrong reasons. And that is something that people have drafted in the Department of Commerce. At the same time, we need, I believe, to say get rid of the quantitative caps that were created for all of these tier two countries because what they did was send a message to 120 nations that they could not necessarily count on us to provide the AI they want and need. And just think about it. I mean, if this is a critical part of your country's infrastructure how can you make a bet on suppliers if you are not confident that they will be able to fulfill your needs? So I think you in Congress and the Senate can help the White House and the Department of Commerce get this right. ¶
Senator Kim 奥特曼先生,我想听听你的看法。这是对这场竞赛的正确框定吗?它关乎在其他国家、在国际上的采用吗?我猜我是在试着把这件事想透。 比如,你刚才在前一个回答里说的部分内容是,我们希望其他国家能够在美国的 AI 技术栈之上构建。这是正确的框架吗?这就是我们在考虑的东西吗? 还是说它更多关乎消费者?更多关乎让世界其他地区、那 78% 的人口采用美国的 AI 应用?还是说这些是相互关联的? Mr. Altman, I wanted your thoughts on this. Is that the right framing of the race? Is it about the adoption internationally in terms of other countries? I guess I am trying to think through it. Like, part of what you just said in your previous response was that we want other nations to be able to build upon the U.S. AI stack. Is that the right framework? Is that what we are thinking about? Or is it more about the consumer? Is it more about getting the rest of the world and the 78 percent of the population to adopt AI applications that are U.S.? Or is it interrelated? ¶
altman 我认为它是高度相互关联的。对我来说,这个技术栈(stack),你知道的,从底部的芯片到顶部的应用,我们希望全世界都使用美国的技术栈。我们希望他们使用美国的芯片。我们希望他们使用像 ChatGPT 这样的服务。 I think it is heavily interrelated. To me, the stack is, you know, from the chips at the bottom to the applications on the top, and we want the whole world on the U.S. stack. We want them to use U.S. chips. We want them to use services like ChatGPT. ¶
Senator Kim 让其他国家在技术栈的基础设施层面之上构建——这是否或多或少保证、或者至少有很大可能,那个国家的消费者随后会使用我们的产品和应用?这就是那个「案件理论」吗? Does having other nations building on the infrastructure component of the stack--does that more or less than guarantee or at least have a high likelihood that then the consumers in that country will be using our products and applications? Is that the sort of theory of the case? ¶
altman 它很可能确实让事情略微更有可能一些。但我也认为——如果有人使用一个我们不信任的技术栈来训练模型,比如,谁知道它会做什么?谁知道可能有什么样的后门(back doors)?谁知道可能有什么样的——你知道的——数据破坏问题?我认为 AI 技术栈将越来越成为一个从芯片一直到终端消费者产品、以及其间许多东西的联合设计系统。我认为把它拆开在实践中效果不会很好,我们也不应该想要那样做。比如,再说一次,我认为这一点——这是一个非常关键的点,美国从让 iPhone 成为人们最想要的移动设备、以及 Google 成为全世界人们最想要的搜索引擎中所获得的杠杆和力量是巨大的。我们或许较少谈论有多少人想要使用这里开发的芯片和其他基础设施,但我认为它同样重要,我们应该力争让整个美国的技术栈被尽可能多的世界所采用。 It probably does make it marginally more likely. But I also think the--if someone is using a stack that we do not trust to train models, like, who knows what it is going to do? Who knows what sort of back doors would be possible? Who knows what sort of, you know, data corruption issues could be possible? I think the AI stack is increasingly going to be a jointly designed system from the chip all the way up to the end consumer product and, you know, lots of stuff in between. I think separating that will not work that well in practice and we should not want to. Like, again, I think this point-this is a very critical point that the leverage and the power the U.S. gets from having iPhones be the mobile device people most want and, you know, Google being the search engine that people most want around the world is huge. We talk maybe less about how much people want to use chips and other infrastructure developed here but I think it is no less important, and we should aim to have the entire U.S. stack be adopted by as much of the world as possible. ¶
Senator Kim 是的。我是说,当我们看——您知道,你谈的是我们对模型的投入和那类构建。 不过,在应用层的开发上——那些试图嵌入人们生活的 AI 工具和应用——我们做得怎么样? 您知道,不一定只是那些统领性的模型;你觉得我们在这类开发上投入了所需的强度吗? Yes. I mean, when we are looking at--you know, you are talking about our investment into models and building of that nature. How are we doing in terms of development of the application--the AI tools and applications, though, that are trying to embed in people's lives? You know, not necessarily just the overarching models but do you feel like we are putting the level of intensity that we need to in terms of that type of development? ¶
altman ChatGPT 是世界上采用最广泛的 AI 服务——不仅在美国,而且在全世界——以相当大的幅度领先。我们非常自豪人们喜欢它,我们需要继续在这方面推进。我认为出于你刚才讨论的所有原因,这很重要。还有许多其他美国公司正在构建令人难以置信的产品和服务,它们也在全球范围内被采用。这是美国最擅长做的事。 ChatGPT is the most adopted AI service in the world--not just in the United States but in the world--by a quite significant margin. We are very proud that people like it and we need to keep pushing on that. I think it is important for all the reasons you just discussed. There are many other U.S. companies building incredible products and services that are also getting globally adopted. This is what the U.S. does best. ¶
Senator Kim 苏博士,我只想最后问你一点。你们每一位都反复谈到人才是这种不可思议的力量,但也可能成为我们的瓶颈。 在这个国家的人才培养方面,我们做得怎么样?如果要你给我们打个分,就我们目前的培养状况,你会给我们打多少分? Dr. Su, I want to just ask one last point to you. Over and over again each of you is talking about talent as this incredible power but also could be a bottleneck to us. How are we doing when it comes to development of talent in this country? If you were to give us a grade what would you grade us at in terms of our development right now? ¶
Ms. Su 谢谢您的问题,参议员。您看,我认为最聪明的工程师在美国。我们有很好的人才基础。 但我要说的是,我们需要更多。我们需要更多硬件开发者、软件开发者、应用开发者。 Thank you, Senator, for the question. Look, I think the smartest engineers are in the United States. We have a great base of talent. But what I will say is we need more. We need more hardware developers, software developers, application developers. ¶
Senator Kim 这个缺口有多大?如果我们像你那样把它说成一场竞赛,您知道,这是一个我们缺口更大的领域,还是一个也在迅速收窄的领域? How wide is that delta? If we are talking about this as a race, as you did, you know, is that a space where we have a larger amount of delta or is that a place where it is closing rapidly, too? ¶
Ms. Su 嗯,我认为我们的整体人才基础确实非常有才华,但我们同时也渴望拥有最顶尖的人,而这不仅包括——您知道——美国公民,也包括拥有最好的国际学生。 Well, I think we do have a very talented overall talent base but we also have the desire to have the best and that includes not only, you know, U.S. nationals but also having the best international students. ¶
Senator Kim 把人才吸引过来—— Drawing the talent from---- ¶
Ms. Su 对。我认为高技能移民正是这样一个领域:我们希望世界上最好的人在美国做他们的工作。 而且,参议员,请允许我就您先前关于循环、关于我们想赢哪场竞赛的那一点补充一句。 您知道,技术是那种可以形成非常强烈的正循环的东西。换句话说,当我们领先、更多人采用时,就意味着有更多开发者让我们的技术变得更好。 这就扩大了我们的领先。所以我们要的,就是让我们的领先随着时间不断扩大。 That is right. I think high-skilled immigration is one of those areas where we want the best people in the world to be doing their work in the United States. And, Senator, if I can just add something to your previous point about the cycle and what race we are trying to win. You know, technology is one of those things where you can have a very vicious positive cycle. So, in other words, when we lead and more people adopt that means more developers that make our technology better. That increases our lead. So that is what we want is to have our leadership just increase over time. ¶
Senator Young 施密特(Schmitt)参议员。 Senator Schmitt. STATEMENT OF HON. ERIC SCHMITT, U.S. SENATOR FROM MISSOURI ¶
Senator Schmitt ### 尊敬的埃里克·施密特(Eric Schmitt)陈述 ### 密苏里州联邦参议员 谢谢您,主席先生。 奥特曼先生,我从你开始。你在圣路易斯长大的家里亮着灯的那个故事,我真的很喜欢,也受到启发;你谈到了创新精神。那就是圣路易斯精神。 作为同为圣路易斯土生土长的人,这是个好听的故事;我们也期待你的公司在圣路易斯有更多投资。那也会很棒。所以我在这里替它说句话。 但我想具体问你:关于美国和欧洲的监管环境之间的比较,人们谈了很多。 具体是什么——我也想开放给其他人——具体是欧洲哪里出了问题,我们能从中得出一些结论? Thank you, Mr. Chairman. Mr. Altman, I will start with you. I really enjoyed and was inspired by your story with the light on in the home you grew up in in St. Louis and you talked about the spirit of innovation. That is the Spirit of St. Louis. As a fellow St. Louis native that is a good story to hear, and we just look forward to more investment in St. Louis from your company. That would be great, too. So I will put a plug in for that. But I do want to ask you specifically, there is a lot made of sort of the comparison between the United States and the regulatory environment and what exists in Europe. What specifically--and I will open this up too--what specifically has gone wrong in Europe that we can draw some conclusions from? ¶
altman 首先,我们很想弄清楚如何在圣路易斯(St. Louis)更多地投资。我很想有个借口更常回家。我想指出一个我认为对用户非常痛苦的例子。当我们推出一项新功能或一个重大的新模型时,我们现在有一种有点像内部笑话的说法,我们说我们有这个很棒的新东西,但不会在欧盟和少数其他国家提供,因为它们在模型推出之前有一个漫长的流程。我相信,将会有非常安全、非常稳健的优秀模型和服务,我们将无法在其他监管制度下提供,而如果你试图在这个新世界里保持竞争力,并且你持续地比其他国家的人能获得的东西落后几个月,这就是一个对用户极其痛苦的例子。 First of all, we would love to figure out how to invest more in St. Louis. I would love an excuse to get to go home more often. I will point out one example that I think is just very painful to users. When we launch a new feature or a major new model we have what is now considered a little bit of like an in joke where we say we have this great new thing not available in the EU and a handful of other countries because they have this long process before a model can go out. And there will be, I believe, great models and services that are quite safe and robust that we will be unable to offer in other regulatory regimes, and if you are trying to be competitive in this new world and if you are consistently some number of months behind what other people in other countries get access to, that is an example that is extremely painful to users. ¶
Senator Schmitt 你提到了你的一个观察:AI 技术栈可能会让它更加垂直整合。那这会怎样运作? 因为现在最好的估计,我猜,是这样——我不知道,中国在大语言模型上大概落后 2 到 6 个月。希望我们在美国看到的一些进展,也许能拉开一定差距。有 DeepSeek,很难确切知道,对吧。 但再往下到芯片,那个优势大概更像几年,差不多是那样。 那么,如果我们正朝那个方向走,在你看来,这会扩大美国的优势,还是说随着我们越来越垂直整合,这反而让中国追得更快? And you mentioned sort of your observation that the AI stack may make it more vertically integrated. So how does that work then? Because right now the best estimates, I suppose, right, is that--I do not know, with China is 2 months to 6 months behind maybe on large language models. Hopefully, some of the advances we are seeing in the U.S. maybe there is a degree of separation. It is hard to know exactly, right, with DeepSeek. But then you get down to the chips and that advantage is more like a couple of years probably, something like that. So if that is where we are headed does that increase the U.S.' advantage, in your view, or does that sort of allow China to catch up quicker as we get more vertically integrated? ¶
altman 我认为有很多事情可以提升美国的领导力。但我们之前谈到了这一点——我认为它如此重要。世界各地会造出优秀的芯片。世界各地会训练出优秀的模型。如果美国公司能在产品上取胜,以及那种来自「一旦真实用户在他们的日常生活中、用他们最困难的任务使用你的产品,你就能改进它」的所有正反馈循环,那是某种特别的东西,不是只靠做好的芯片和好的模型就能轻易追上的。所以确保美国能在这个层面、在产品层面取胜。显然,我,比如,有点是在自卖自夸(talking my book),但我确实相信这一点——这相当重要,而且这是在芯片、算法或基础设施、算法和数据之外额外的一个领域。我认为这是一个美国真正在取胜、并且有非常强的复合效应的新领域。 I think there are a lot of things that can increase U.S. leadership. But we touched on this earlier--I think it is so important. There will be great chips made around the world. There will be great models trained around the world. If the United States companies can win on products and the--sort of all of the positive feedback loops that come from how you can improve this once, you know, real users are using your products in their daily lives for their hardest tasks, that is something special that is not so easy to catch up with just by doing good chips and good models. So making sure that the U.S. can win at the product level here. Obviously, I am, like, talking my book a little bit, but I really do believe it--is quite important, and that is in addition to all of the chips algorithm or the infrastructure algorithms and data. I think this is a new area where the U.S. is really winning and has a very strong compounding effect. ¶
Senator Schmitt 因特拉托先生——顺便问一下,我念对了吗? Mr. Intrator--did I pronounce that correctly, by the way? ¶
Mr. Intrator 是的。 Yes. ¶
Senator Schmitt 好,谢谢。 我想稍微转一下,还是留在这个监管环境的话题上。我认为从欧洲出来的最令人担忧的事情之一,是这种审查体制,它不仅存在于线上,也存在于现实生活中。 但当然,它正在线上发生。我是说,人们因为在网上说的话而被逮捕;而我对 AI 的一个担忧,我猜是:如果我们最终走到一个它不知怎地竟在「警务化」所谓「虚假信息」的地步,您知道,我认为即便在 NIST 最近那份自愿性标准里,要警惕的风险之一也是虚假信息的传播。 所以这个问题的要点是:我们如何确保——我认为欧洲出问题的一部分原因,是它在某种程度上——它在给信息做漏斗过滤;而在我看来,无论我是否同意那个观点,它都应当被摆出来。人们可以自己做决定。对抗你不认同的言论,不是靠审查它,而是靠——更多的言论。 这方面有什么教训可以吸取,以确保这种事不会在这里发生? OK, thank you. I want to turn a little bit, sort of staying on this regulatory environment, one of the things I think that is most concerning that is coming out of Europe is this sort of censorship regime that exists not just online but in real life. But, certainly, it is happening online. I mean, people are being arrested for things that they say online, and one of the concerns I have with AI, I suppose, is that if we end up with a place where it is somehow policing, quote/unquote, ``misinformation'' and, you know, I think even in NIST's most recent voluntary standards one of the risks to be on the lookout for was the spread of misinformation. So the point of the question is how do we make sure that--I think part of what is going wrong in Europe is it is a sort of a--it is funneling information and, in my view, whether I agree with the point of view or not it ought to be out there. People can make their own decisions. You combat speech you do not agree with not by censoring it but by--with more speech. What are some lessons to be learned there and make sure that does not happen here? ¶
Mr. Intrator 所以欧洲正以欧洲的方式推进它的监管体制;从我们这种必须做出巨额资本投资的立场看,欧洲做法中让我们每天都深感担忧的一点,是那种割裂——用这个词吧——他们处理信息流动的方式、监管信息的方式,以及他们联盟中每个组成部分各有自己一套规则的方式;随着时间推移,这在欧洲会极具挑战,因为要做出我们那种量级的投资——而我们正处在—— So Europe is moving forward with its regulatory regime in a European way, and from our seat where we have to make these enormous capital investments one of the things about the approach that Europe is taking that we are deeply concerned about every day is the balkanization to use-- of how they go about allowing information to flow and how they go about regulating it, how they go about with each component of their union having its own set of rules, which will be tremendously challenging in Europe as time goes on because it is really hard to make the magnitude of investments that we-- where we are---- ¶
Senator Schmitt 不过,在那之外,我现在说的是管辖权层面的内容问题。 Beyond that, though, jurisdictionally I am talking about content now. ¶
Mr. Intrator 所以我们不是——我们公司的角色其实在那一层之下,您知道——而且,您知道,萨姆和微软会得到更多针对内容层面的关注,因为他们在技术栈中所扮演的角色。那并不是我们主要聚焦的地方。我们真正聚焦的是投资这一侧。 So we are not--the role of our company is really kind of below that, you know--and, you know, Sam and, you know, Microsoft you are going to get a lot more attention paid to the content level because of the role that they play in the stack. It is not really where we are primarily focused. We are really focused on the investment side of it. ¶
Senator Schmitt 是的。如果各位中有谁愿意——萨姆,如果你——或者,奥特曼先生,如果你愿意回应这一点,我想得到一些答案。 Yes. If any of you would like to--Sam, if you--or, Mr. Altman, if you would like to respond to that I would like to get some answer. ¶
altman 我认为——嗯,首先,我强烈同意,人们因为他们在网上说的话而被——比如——关进监狱是非常——不美国的,也不应该是我们做的事。AI 与社交媒体相当不同,至少在它目前的演进阶段是这样。人们以一种一对一的方式使用这些工具,而不是那种网上的大规模的东西。所以我认为很容易做太多类比,但试图用我们谈论社交媒体的同样方式来谈论 AI 和我们将在这里面对的事情,有点危险。但我们的立场是,我们需要给成年用户很大的自由,让他们以自己想要的方式使用 AI,并信任他们会负责任地使用这个工具。我知道世界其他地方、以及美国一些地方有越来越大的压力要求不要这样做,但我认为这是,比如——这是一个工具,我们需要让它成为一个强大而有能力的工具。我们当然会设置一些护栏和非常宽泛的边界,但我认为我们需要在这里给予很大的自由。 I think--well, first of all, I strongly agree that people getting, you know, like, put in jail for stuff they say online is very--not American and not what we should be doing. AI is quite different than social media, at least in its current evolution. People are using these tools in this sort of one on one way instead of this massive thing online. So I think it is easy to make too many analogies but it is a little bit dangerous to try to talk about AI and the things we are going to face here in the same way that we did for social media. But our stance is that we need to give adult users a lot of freedom to use AI in the way that they want to use it and to trust them to be responsible with the tool. And I know there is increasing pressure in other places around the world and some in the U.S. to not do that but I think this is, like--this is a tool and we need to make it a powerful and capable tool. We will, of course, put some guardrails and very wide bounds but I think we need to give a lot of freedom here. ¶
Senator Schmitt 是的。我时间到了,但那里还有很多问题,我们会跟进。 谢谢您,主席女士。 Yes. I am out of time but there is a lot more questions there that we will follow up with. Thank you, Madam Chair. ¶
Senator Cantwell 谢谢。谢谢。 希肯卢珀(Hickenlooper)参议员。 Thank you. Thank you. Senator Hickenlooper. STATEMENT OF HON. JOHN HICKENLOOPER, U.S. SENATOR FROM COLORADO ¶
Senator Hickenlooper ### 尊敬的约翰·希肯卢珀(John Hickenlooper)陈述 ### 科罗拉多州联邦参议员 我很欣赏那一路的提问。我本来也准备好让你继续说下去。我本可以给你一两分钟。 史密斯先生,微软有着悠久而深厚的传统,通过软件改造世界各地的工作场所,从 Windows 操作系统到 PowerPoint、Excel 这类办公应用,再到如今由 AI 驱动的 Copilot 应用。 在软件开发中,生命周期似乎越来越短,更新越来越频繁。 微软遵循哪些内部流程来评估 Copilot 在发布前的准确性和性能?除了微软自己的产品开发者之外,还涉及哪些独立评审团队?你们请谁来帮忙做这件事? I appreciate that line of questioning. I was ready for you to continue as well. I could have given you a minute or two. Mr. Smith, Microsoft has a long and deep history in transforming workplaces all over the world through software, from Windows operating system to its office applications like PowerPoint, Excel, and now the AI-powered Copilot application. In software development, life cycles seem to be becoming increasingly shorter, updates becoming more frequent. What are the internal processes that Microsoft follows to evaluate Copilot's accuracy and performance before it was released and what kind of independent review teams other than Microsoft's own product developers are involved in that? Who do you bring in to help with that? ¶
Mr. Smith 嗯,首先,既然我们这里谈的 Copilot 大部分都始于在 OpenAI 开发的模型,我要说 OpenAI 有它自己的内部流程。 然后有一个联合的——叫做 DSB,部署安全委员会(Deployment Safety Board),顾名思义,我们在那里共同决定某个东西是否可以安全部署。 随后在应用层,我们有自己内部的部署安全委员会。我们有一套工程工具来评估这些功能。我们测试这些功能。 我们有红队,也就是彼此竞争的团队,常常专门去攻击这些功能;最终,当这些测试完成、结果令人满意时,产品才会发布。 Well, first of all, since most of what we are talking about here when you are talking about our Copilot's start with models that are developed at OpenAI, I would say OpenAI has its internal process. There is then a joint--what is called the DSB, a Deployment Safety Board, where we decide together whether something is safe to deploy, as the name implies. We then at the applications level have our own internal Deployment Safety Board. We have a variety of engineering tools that we use to assess these features. We test these features. We have red teams, meaning sort of competing teams that often go to work to sort of attack the features, and then ultimately the product is released when those tests are completed and the results are satisfactory. ¶
Senator Hickenlooper 我喜欢这个。那让我转向奥特曼先生。显然,你们都有天然的动力去确保产品高质量且安全。 但这个领域竞争如此激烈,而且,您知道,在应用研究和严格测试中,这些持续的改进确实是模型性能的根本步骤。 所以风险评估就是那个关键工具,而我是循证技术标准的坚定信奉者。有人说过我是参议院里唯一真正发表过同行评议论文的科学家。 那么,奥特曼先生,你是否认为,在适当情况下,由合格评估者依据标准开展的、自愿进行的独立评估,能够帮助验证你们在内部以及与同行公司共同进行的测试? I like that. Well, let me go over to Mr. Altman. Obviously, you all have a natural incentive to ensure that the products are high quality and safe. But the field is so competitive and, you know, in applied research and with rigorous testing these constant improvements really are fundamental steps to the performance of a model. So risk assessments are that key tool, and I am a big believer in evidence-based technical standards. I have been accused of being the only real scientist who has published peer-reviewed papers in the Senate. So, Mr. Altman, do you believe that under appropriate circumstances independent evaluations based on standards performed by qualified evaluators and done voluntarily could help validate the testing that you are performing internally and in conjunction with peer companies? ¶
altman 谢谢您,参议员,我认为您——发表过同行评议论文——这太棒了,很乐意看到更多这样的人。 Thank you, Senator, and I think it is awesome that you are--have published peer-reviewed papers and would love to see more of that. ¶
Senator Hickenlooper 我在——在科学的马斯洛三角上,我位置挺低的。我是地质学家。所以那不算高—— I was--on the Maslow's triangle of science I was near the bottom. I was a geologist. So that is not high up in that---- ¶
altman 地质学很棒。是的,我认为你说的非常重要。它是我们今天流程的重要组成部分。外部测试帮助我们发现我们内部可能遗漏的东西,而且——我们总体上对我们的安全记录非常自豪,并不是说我们——你知道的,我们并非一直完美,而且我们还在持续学习新东西,但我认为我们确实有一个流程,引领着走向公众普遍认为使用起来安全、稳健的模型,而且我们已经开发了很多技术来能够持续交付这一点。但外部测试人员和红队测试人员(red teamers)是那个流程的关键部分,我认为他们帮助我们在模型中发现了许多需要改进的东西。 Geology is great. Yes, I think what you say is very important. It is an important part of our process today. External testing helps us find things that we may have missed internally and--we are very proud of our safety record on the whole, not that we--you know, we have not been perfect and we are continuing to learn new things, but I think we do have a process that is leading toward models that the public generally thinks are safe and robust to use, and we have developed a lot of techniques to be able to continue to deliver that. But external testers and red teamers are a critical part of that process and I think they have helped us find many things in the models to improve. ¶
Senator Hickenlooper 史密斯先生,你要补充什么吗? Mr. Smith, would you add anything to that? ¶
Mr. Smith 没有。 No. ¶
Senator Hickenlooper 好,明白了。一个来作证却说没什么可补充的人——这是一个科学反思的时刻。 苏博士,两党共同的《芯片与科学法案》是一项历史性努力,试图维持美国在半导体等新兴技术上的领先地位,也包括其他领域;随着技术军备竞赛在全球持续,你也谈到了这一点,AMD 在交付最先进设计方面扮演关键角色——为我们的电子设备、为让 AI 走向全球的设备提供动力的新芯片,要做到最好。 当科学家们夜以继日地开发新突破、试图提升并改善性能、同时缩短研发周期时,你认为芯片技术在能效方面的下一个前沿是什么?以及——这不仅相对于中国竞争者——我们如何携手改进高性能计算中的直接芯片冷却(direct-to-chip cooling)? OK. Got it. Someone giving testimony who does not have something to add--it is a moment of scientific reflection. Dr. Su, the bipartisan CHIPS and Science Act, an historic effort to try and maintain U.S. leadership in emerging technologies like semiconductors but others as well, as the technology arms race continues globally, and you were talking about this, AMD plays a key role in delivering state-of-the-art designs, the best for the new chips that are going to power our electronics and the devices that are going to allow AI to become global. As scientists work around the clock to develop new breakthroughs and to try to increase and improve performance but at the same time shorten R&D timelines, what do you see as the next frontier of chip technology in terms of energy efficiency and how can--and that is not just based on the Chinese competitors but how can we work together to improve direct to chip cooling for high-performance computing? ¶
Ms. Su 嗯,谢谢您的问题,参议员。 我要说,您看,今天半导体领域有大量的创新正在进行。《芯片与科学法案》确实有助于提升芯片在美国的受关注度。 相对于——您知道——我们为跑得更快、造得更好、更节能的芯片在做什么:坦白说,我们在整个芯片开发周期中广泛使用 AI,它确实让我们能够增强那些通常非常长的周期——很多年,您知道,我们要花好几年开发芯片。 我们可以缩短时间,也能提高效率;在冷却技术方面有许许多多很好的新技术,对我们建造今天早些时候谈到的那种大规模系统极为重要。所以谢谢您的问题。 Well, thank you for the question, Senator. I would say, look, there is a tremendous amount of innovation that is going on in the semiconductor sector today. The CHIPS and Science Act was certainly helpful in raising the profile of chips in the United States. Relative to, you know, what are we doing to go faster and build better and more power efficient chips, frankly, we are using AI extensively through our chip development cycles and it does allow us to augment what are typically very long cycles, many years--you know, several years for us to develop chips. We can shorten the time and also improve the efficiency, and there are lots and lots of great new technologies in terms of cooling technologies that are super important for us to build the large-scale systems that we talked about earlier today. So thank you for the question. ¶
Senator Hickenlooper 好。我时间到了。我把时间交还主席。谢谢各位。 **The Chairman〔主持〕.** 而且,希肯卢珀参议员,我要说,作为一个父母做过石油天然气生意的得克萨斯人,我认为地质学家很了不起。 All right. I am out of time. I will yield back to the Chair. Thank you all. The Chairman [presiding]. And, Senator Hickenlooper, I will say as a Texan whose parents were in the oil and gas business I think geologists are awesome. ¶
Senator Hickenlooper 我们达成共识了。 **The Chairman.** 柯蒂斯参议员。 We have a consensus. The Chairman. Senator Curtis. STATEMENT OF HON. JOHN CURTIS, U.S. SENATOR FROM UTAH ¶
Senator Curtis ### 尊敬的约翰·柯蒂斯(John Curtis)陈述 ### 犹他州联邦参议员 谢谢您,主席先生。很高兴来到这里。 奥特曼先生,你挑起了关于电脑的「比惨」;我告诉你,1985 年你出生的那个月,我正在杨百翰大学(Brigham Young University)上课,带了一台笔记本电脑进去,差点被赶出来。 Thank you, Mr. Chairman. It is a delight to be here. Mr. Altman, you started kind of a one-upmanship on computers and I will just tell you in 1985 the month you were born I was attending a class at Brigham Young University and carried in a laptop and was almost kicked out. ¶
altman 什么笔记本电脑? What laptop? ¶
Senator Curtis 是一台 TRS-80—— It was a TRS 80---- ¶
altman 哦,太棒了。 **Senator Curtis〔继续〕.** ——Radio Shack 造的。我把内存从 40K 升级到 80K。用四节 AA 电池供电,而且—— Oh, awesome. Senator Curtis.--made by Radio Shack. I upgraded the memory from 40K to 80K. Ran on four AA batteries and------ ¶
altman 太不可思议了。 That is incredible. ¶
Senator Curtis 是的。所以我非常羡慕你们这一代。如果可以,我从你开始。 我认为,您知道,犹他州会希望以数据中心和先进技术领跑。能否请你谈谈,对各州、特别是对犹他州来说,是什么让它们对「星际之门」(Stargate)这样的项目具有吸引力? Yes. So I am very envious of your generation. Let me start with you, if I would. I think, you know, Utah would aspire to lead out with data centers and advanced technologies. Could you just address for states and Utah specifically what it is that makes them attractive to projects like Stargate? ¶
altman 是的,我知道我们正在就犹他州(Utah)的一些潜在场址进行富有成效的讨论。电力、冷却、快速的许可流程、能够建设这些东西的劳动力——电工、建筑工人、整个技术栈。一个愿意,比如,与我们合作快速推进的州。得克萨斯州在这方面的表现真的令人难以置信。我认为那会是其他州值得研究的好事情,但我们很高兴能设法想出一些方案。 Yes, and I know that we are having productive discussions about some potential sites in Utah. Power cooling, fast permitting process, labor force that can build these things--the electricians, the construction workers, the entire stack. A state that wants to, like, partner with us to move quickly. Texas really has been unbelievable on this. I think that would be a good thing for other states to study but we would be excited to try to figure something out. ¶
Senator Curtis 谢谢。我想我能代表我们州的领导人说话。我们也会很兴奋。但如你所知,这也带来挑战,其中之一就是能源需求;关于我们如何保护电费缴纳者、如何在他们之间设一道防火墙,你有什么想法? Thank you. I think I could speak for our state leaders. We would be excited as well. But as you know, this also brings challenges and one of those challenges are the demands for energy, and what are your thoughts on how we protect rate payers and kind of put a little bit of a firewall between them? ¶
altman 我是说,我认为最好的办法就是大幅增加供给、增加发电量。你知道的,比如,我认为如果你让以合理利润创造大量额外发电容量变得容易,市场就会那样做。这不仅不会因为 AI 负载而推高电价——希望它会因此压低所有方面的电价。我们已经谈了很多能源对 AI 的重要性。能源对生活品质真的非常重要。在我看来,历史上最一以贯之的事情之一是,每当能源成本下降,生活质量就会上升,所以在短期内为使能源更便宜做很多事——我认为这可能看起来像是更多天然气,尽管在某些应用上我认为太阳能在中期内确实能帮上忙。我希望是先进的核裂变和核聚变。更多能源的重要性远远超出 AI。你知道的,在某种意义上,我们正在进行 AI 和能源的双重革命——拥有新想法的能力,以及把这些想法做出来、在我们都生活的物理世界中让它们成真的能力。比如,这些就像是繁荣的限制性试剂(limiting reagents),让我们拥有更多这样的东西吧。 I mean, I think the best way is just much more supply, more generation. You know, like, I think if you make it easy to reasonably profitably create a lot of additional generative capacity the market will do that. That will not only not drive up rates because of the AI workload--hopefully it will drive it down for everything. And we have talked a lot about the importance of energy to AI. Energy is just really important to quality of life. One of the things that seems to me the most consistent throughout history is every time the cost of energy falls the quality of life goes up, and so doing a lot to make energy cheaper in the short term--I think this probably looks like more natural gas, although there are some applications where I think solar can really help in the medium term. I hope it is advanced nuclear fission and fusion. More energy is important well beyond AI. You know, in some sense we have these dual revolutions going of AI and energy, the ability to have new ideas and the ability to get them done, to make them happen in the physical world where we all live. Like, these are kind of the limiting reagents of prosperity and let us have a lot more. ¶
Senator Curtis 谢谢。 史密斯先生,我们谈过电力对这里取得成功有多重要。你认为微软和其他科技领袖在发展能源、特别是发展正确类型的能源方面,应扮演什么角色? Thank you. Mr. Smith, we have talked about how significant power was-- is to the success here. What role do you think Microsoft and other tech leaders have in developing energy and particularly the right type of energy? ¶
Mr. Smith 我认为我们有一份巨大的责任去为解决方案作贡献;我认为萨姆已经用他那个清单帮了忙。 我要强调两点,我想我会用我们在各地所做的一切来说明,但最近是威斯康星州东南部的一个大型站点。我们基本上是从零起步,变成了该州最大的工业用电户——大约 400 兆瓦。 于是我们与当地电力公司合作。我们投入资金去帮助、并真正使他们得以扩大发电能力。 接下来,那些电需要从他们的电厂经由电网输送到我们的数据中心。我们去了公共事业委员会,提出对我们自己加价,因为我们认为重要的是由我们来为电网的这项改进付费,好让所谓的「邻居们」不必承担。 我认为这真正体现出的,是所需的那种协作伙伴关系:提供资本、进行建设、改善电网,并且我认为,要对整个社区非常敏感。 I think we have a tremendous responsibility to contribute to the solution and I think Sam helped with his list. I would highlight two things, and I just would, I guess, illustrate it with what we do everywhere but most recently with a major site in southeastern Wisconsin. We went from zero, basically, to becoming the largest industrial user of electricity in the state--roughly, 400 megawatts. And so we worked with the local utility. We made the investment to help and really enable them to expand their electricity generation. Now, that electricity then needed to be delivered from their power plant through the grid to our data center. We went to the Public Utilities Commission and we proposed a rate increase on ourselves because we thought it was important that we pay for that improvement to the grid so that the neighbors, so to speak, would not have to. And I think what it really illustrates is the collaborative partnerships that are needed to provide the capital, to do the construction, to improve the grid, and to be, I think, very sensitive to the community as a whole. ¶
Senator Curtis 谢谢。 奥特曼先生,让我回到你这里。我曾是个小企业主。我对小企业主有一份特别的感情。 我们能否谈一下 ChatGPT,以及它可能如何帮助小企业主?让我把画面拉宽一点。我们听到很多关于其他一些工具的讨论——它们或许不那么受青睐,特别是对美国政府部门而言——但那些工具对小企业非常有帮助。 但我不确定小企业是否完全了解你拥有的这个平台,以及他们可以如何用它做营销、做数据研究,用各种方式帮助自己的小生意成功。 Thank you. Mr. Altman, let me come back to you. I was a small business owner. I have a special spot in my heart for small business owners. Can we talk a little bit about ChatGPT and how that might assist small business owners? And let me paint a little broader picture. We have heard a lot about other tools that are, perhaps, out of favor, particularly with the U.S. Government, that are very helpful for small businesses. But I do not know if small businesses are fully understanding the platform that you have and how they might use it for marketing, for data research, and ways to help their small business be successful. ¶
altman 其中——在 ChatGPT 开始腾飞的时候,有所有这些时刻,我们会,比如,哦,我们手上可能有一个大热门。比如,那是——有人把它用于这个、这个和那个,你知道的,陌生人在谈论它。你看到有人在咖啡馆里用它。但对我来说真正突出的一件事是,在 ChatGPT 推出后很快,比如,头六个月左右,我当时在一辆优步(Uber)车里,司机在搭话。他说,你听说过一个叫 ChatGPT 的东西吗?太棒了。我说,是的。比如,你觉得它怎么样?而他基本上在用经营他的整个小生意。他说,我——他经营着一家自助洗衣店,他说,我有所有这些问题,你知道的,找不到好人帮我写广告,没法让法律文件得到审核,没法回复客户支持的邮件。他说,他是一个超级早期采用者,但他是那种用 AI 来,比如,让一家小生意运转起来的人,那是——我们当时经常谈论那个故事,但现在回头再想想它真好。我们现在从很多人那里大规模地听到了这样的事情,但那是早期的那些时刻之一,我们当时想,哦,这也许真能成。 One of--there were all these moments as ChatGPT was beginning to take off where we would be, like, oh, we may have, like, a hit on our hands. There is, like, that is--someone is using it for this and this and that, you know, strangers talking about it. You see someone using it in a coffee shop. But one of the ones that really sticks out for me is pretty quickly after ChatGPT launched, like, in the first six months, say, I was in an Uber and the driver was making conversation. He is, like, have you heard of this thing called ChatGPT? It is amazing. And I was, like, yes. Like, what do you think about it? And he was using it to run basically his entire small business. He was, like, I had--he ran a Laundromat and he was, like, I had all these problems, you know, like, could not find good people to write my ads, could not get, like, legal documents reviewed, could not, like, answer customer support emails. And he was, like, a mega early adopter but he was one of these people that was using AI to, like, make a small business work and that was--we talked about that story a lot at the time but it is nice to reflect on it again now. We have now heard that at scale from a lot of people, but that was one of those moments early on we were, like, oh, this is maybe going to work. ¶
Senator Curtis 所以我时间到了,但只想指出:这不仅仅是帮你校对邮件的东西,对吧?你不需要回应,因为我时间到了,但我想我们都会同意这一点。 So and I am out of time but just to mark this is more than just something that helps proofread e-mails, right? And you do not need to comment because I am out of time, but I think we would all agree with that. ¶
altman 是的。 It is. ¶
Senator Curtis 期待看到这些应用向前推进。 主席先生,我把时间交还。 **The Chairman.** 达克沃斯(Duckworth)参议员。 And look forward to seeing these applications move forward. Mr. Chairman, I yield my time. The Chairman. Senator Duckworth. STATEMENT OF HON. TAMMY DUCKWORTH, U.S. SENATOR FROM ILLINOIS ¶
Senator Duckworth ### 尊敬的塔米·达克沃斯(Tammy Duckworth)陈述 ### 伊利诺伊州联邦参议员 谢谢您,主席先生。感谢您,也感谢各位今天到场。 我想先谈私营部门与我国国家实验室之间的伙伴关系,对维持美国在 AI 领域领导地位的重要性。 伊利诺伊州有幸拥有国家实验室体系中的两颗明珠:费米实验室(Fermilab),美国首屈一指的粒子物理与加速器实验室;以及阿贡国家实验室(Argonne National Laboratory),那里有「极光」(Aurora)超级计算机,将加速 AI、癌症研究和基础物理领域的突破。 没有什么比维持并放大对我国这套了不起的国家实验室网络的投入更重要。 然而,唐纳德·特朗普(Donald Trump)和埃隆·马斯克(Elon Musk),在一些国会共和党人的支持下,正谋划用一把电锯去砍向全国各地正在开展的关键研究计划。 这是一场自我破坏的攻击,就这么简单;如果任其推进,特朗普和马斯克将对我们的创新能力与产能造成持久伤害——那是我们的敌人做梦都想达到的。 真的有人相信,如果「政府效率部」(DOGE)几十年前就存在,他们不会把创造互联网的那个项目当作「浪费公帑的研发」给砍掉吗? 所以我向各位小组成员提出的问题是:能否解释国家实验室体系对维持我们研究优势的重要性,并谈谈你们已经建立或正在推进的任何伙伴关系,特别是那些因国家实验室研究遭大规模削减而受威胁的伙伴关系? Thank you, Mr. Chairman. I thank you, the panel, for all of you being here today. I want to begin by talking about the importance of partnerships between the private sector and our National Laboratories in maintaining United States leadership in AI. Illinois is the proud home of two crown jewels of the National Laboratories, Fermilab, America's premier particle physics and accelerator laboratory, and Argonne National Laboratory, home to the Aurora supercomputer that will accelerate breakthroughs in AI, cancer research, and fundamental physics. There is nothing more important than sustaining and amplifying investments in our Nation's incredible network of National Labs. Yet, Donald Trump and Elon Musk, with the support of some Republicans in Congress, are plotting to take a chainsaw to the vital research initiatives being carried out across our country. This is a self-sabotaging attack, plain and simple, and if allowed to proceed Trump and Musk will inflict lasting harm on our innovative capabilities and capacity that our enemies could only dream of achieving. Does anyone truly have confidence that had DOGE been around decades ago they would not have cut the project that created the Internet as an example of wasteful publicly funded research and development? So my question to any member of the panel is the following. Can you explain the importance of the National Labs system to maintaining our research edge and discuss any partnerships you have established or are currently pursuing, especially those threatened by massive cuts to the National Labs' research? ¶
altman 我们与国家实验室(National Labs)合作,所以也许我可以先尝试回答这个。 We partner with the National Labs so maybe I could take a first cut of this. ¶
Senator Duckworth 请。 Please. ¶
altman 还有,参议员,我很想有一天能去参观费米实验室(Fermilab)。那将是,比如,不真实的(unreal)。 Also, Senator, I would love to get to visit Fermilab someday. That would be, like, unreal. ¶
Senator Duckworth 那正是我的下一个问题。不用谢。 That was my next question. You are welcome. ¶
altman 那将成为人生的一大亮点。那会很酷。AI 将为世界做很多美妙的事情,但我个人最兴奋的是 AI 将对科学发现产生的影响。我相信,新的科学发现是世界变得更好、人们生活质量随时间提高的最重要投入。很难夸大因为科学进步我们身在何处、以及没有它我们会身在何处。所以我们很高兴能与国家实验室在这件事上合作。我认为科学的效率还没有达到它可以达到的程度,而且我们也很高兴听到科学家们说他们现在的效率是以前的数倍,我认为 AI 工具将意味着我们能在某个时候以相同或更低的成本,在一年内取得相当于十年的科学进步。在我看来,这将是 AI 对世界做出的最重要的贡献之一。而这不再是理论上的。比如,国家实验室就是一个很好的例子。这是我们唯一一次把我们的模型权重副本交给另一个机构的合作。这对我们来说是一段非常深厚和重要的伙伴关系,我预计它真的会结出果实。 That would be a real life highlight. That would be very cool. There are many wonderful things that AI is going to do for the world but the one that I am personally most excited about is the impact AI will have on scientific discovery. I believe that new scientific discovery is the most important input to the world getting better and people's quality of life getting better over time. It is hard to overstate where we would be if--where we are because of scientific advancement and where we would be without it. So we are thrilled to get to partner with the National Labs on this. I think science has not been as efficient as it can be, and we are also thrilled to hear from scientists that they are, you know, multiples more effective than they used to be and I think that AI tools will mean we can accomplish at some point a decade worth of scientific progress in a year for the same cost or even less. This will be one of the most important contributions, in my opinion, that AI makes to the world. And it is no longer theoretical. Like, the National Labs are a great example. It is the only partnership where we have given a copy of our model weights to another organization. It is a very deep and important partnership to us and I expect that that will really bear fruit. ¶
Senator Duckworth 谢谢。小组里还有别人吗? Thank you. Anybody else on the panel? ¶
Mr. Smith 是的,我认为你指出了一个非常重要的问题。这个国家有 17 个国家实验室,由能源部管理,还有大约 85 到 90 所研究型大学;它们共同构成了大量科学发现的肌理,自二战时期的曼哈顿计划以来一直如此。 我们科技界、我们微软,与其中大多数、几乎全部都合作;美国掌握了一种特殊的创新循环。你在这些机构里有好奇心驱动的研究,然后进展从这些机构流向初创公司,再流向更大的公司。 而我一直觉得有意思的是,当我会见世界各地官员时,他们研究过这套东西,他们想效仿它;而我一直担心,在美国我们自己有把它视为理所当然的风险。 我们永远不应把它视为理所当然。它是这个国家技术领导地位的根基。 Yes, I think you highlight a very important issue. This country has 17 National Labs administered by the Department of Energy and about 85 to 90 research universities, and together they are the fabric of much of scientific discovery and have been since the Manhattan Project in World War II. We in the tech sector, we at Microsoft, work with most, almost all of them, and there is a particular cycle of innovation that the United States has mastered. You have curiosity-driven research in these institutions and then the advances move out of those institutions into startups and into larger companies. And what I always find interesting, as I meet with officials around the world they have studied this. They seek to emulate it, and I always worry that in the United States we run the risk of taking it for granted. We should never take this for granted. It is the foundation for the country's technological leadership. ¶
Senator Duckworth 确实如此。 苏博士。 Very much so. Dr. Su. ¶
Ms. Su 我只想补充一点。我们也是国家实验室公私伙伴关系的坚定支持者。 我认为国家实验室在某种意义上一直试图走在曲线前面,您知道,那是我们投资的好地方。 我们认为它们是关键的一环。过去十年里,我们也与所有国家实验室建立了合作,您知道,那里仍然是我认为可以有重要公私伙伴关系的地方。 I just wanted to add to that. We are also very large supporters of the public-private partnerships with the National Labs. I think the National Labs have, you know, in a way always tried to look ahead of the curve and, you know, that is a great place for us to invest. We think they are a key piece. We have partnered with all of the National Labs as well, you know, over the last decade and that continues to be a place where I think there can be significant public-private partnership. ¶
Senator Duckworth 谢谢。因特拉托先生。 Thank you. Mr. Intrator. ¶
Mr. Intrator 我只是觉得,如果能来到这些「AI 工厂」,在这些机构里走一走、穿行其间,找出所有那些科学的片段——它们回溯到、并最终由这些机构产出的东西所驱动和奠基——那会非常有意思。实际上,这很惊人。 I just think it would be really interesting to come to these AI factories and to walk or travel through these institutions and identify all the different pieces of the science that leads back and was ultimately driven and founded on something that came out of those institutions. It is amazing, actually. ¶
Senator Duckworth 谢谢。那么你们剩下三位中,有谁愿意来伊利诺伊州的实验室,费米或者阿贡?我亲自带你们参观。 〔笑声。〕 Thank you. And would any of the remaining three of you like to come to a lab in Illinois, either Fermi or Argonne? I will give you personal tours. [Laughter.] ¶
Senator Duckworth 好。你们四位都来。就这么定了。 谢谢您,主席先生。 **The Chairman.** 谢谢。 杨(Young)参议员。 All right. All four of you. It is done. Thank you, Mr. Chairman. The Chairman. Thank you. Senator Young. STATEMENT OF HON. TODD YOUNG, U.S. SENATOR FROM INDIANA ¶
Senator Young ### 尊敬的托德·杨(Todd Young)陈述 ### 印第安纳州联邦参议员 谢谢您,主席先生,召开这场关于赢得 AI 竞赛的重要听证会。很高兴看到各位小组成员在场。 每当谈这个话题,我喜欢强调的一点是:我们讨论的不仅仅是一场创造就业的竞赛,也不仅仅是如何从我们的经济中多榨出一些增长的竞赛——尽管那很重要。 也不仅仅是在设法弄清楚,人类、特别是美国人,如何通过以各种方式把 AI 解决方案应用于日常生活而更加兴旺。 这是一个国家安全与经济安全问题。我希望国内的民众能明白这一点。我知道我们所有小组成员对此都非常熟悉、非常了解。 在与你们以及许多其他人的讨论中,我听到的是:我们需要与志同道合的伙伴和盟友合作才能赢下这场竞赛,而且只能靠集体来完成。 今天我从你们中不少人那里听到,这场竞赛部分关乎取得市场份额、把我们的 AI 模型和解决方案扩散到其他国家。 在我看来,正是通过那种方式——也许可以请你们展开谈谈——我们才能看到自己的价值得到推进。 这些模型想必会被嵌入我们关于隐私、透明度、财产权、言论与宗教自由的价值观,而不是中国共产党在所有这些方面的价值观。 而如果我们能与这些其他国家建立数字贸易规则、数字跨境数字贸易协定,我们就有可能为那些没有嵌入我们标准的模型——比如说,中国共产党认可的那些模型——竖起更高的进入壁垒。 所以这整个对话之上有一层地缘政治与国家安全的考量,这也是为什么我认为主席强调不要过度限制创新或部署非常重要。 但这也是为什么我认为,重要的是我们要思考如何与其他国家在它们的标准制定上合作。 那么,这就是我想开始提问的地方。我从史密斯先生开始。 如果美国不通过某个实体——无论是 NIST、其他联邦实体,还是联邦认可的实体——采用一些标准,那么其他国家难道不会径自觉得有必要在完全不与美国协商的情况下采用它们自己的标准吗? Thank you, Mr. Chairman, for holding this important hearing on winning the AI race. It is good to see our panelists here. One of the things that I like to underscore whenever I talk about this issue is we are not just discussing a race to create jobs, not just discussing a race to figure out how to eke out more growth from our economy, although that is important. Not just trying to identify how humans can flourish more, especially Americans, through application of AI solutions to our daily lives in various ways. But this is an issue of national and economic security. I want folks at home to get that. I know all our panelists are highly conversant and knowledgeable about that. In my discussions with you and many others I have heard we need to work with like-minded partners and allies to win this race, and it is only going to be done collectively. I have heard here today from a number of you that this race is in part about getting market share, diffusion of our AI models and solutions into other countries. It is through that means for me to, perhaps, elaborate on your thoughts that we can see that our own values are advanced. These models presumably they will be embedded with our values related to privacy and transparency and property rights and freedom of speech and religion, not the values of the Chinese Communist Party on each of those various fronts. And then if we can establish digital trade rules, digital cross-border agreements on digital trade with these other countries, we could conceivably erect higher barriers to entry for models that do not come embedded with our standards, models of, say, the Chinese Communist Party has given sanction to. So there is a geopolitical national security overlay to this entire conversation, which is why I think the Chairman's emphasis on not overly constraining innovation or deployment is very important. But it is also why I think it is important that we be thinking about how to work with other countries in their standards development. And so that is that is where I want to begin asking questions. I will start with Mr. Smith. If the United States does not adopt some standards through some entity, whether it is NIST or another Federal entity or federally sanctioned entity, then won't other nations go ahead and feel the need to adopt their standards without any consultation with the United States? ¶
Mr. Smith 我认为你这一点提得非常重要,这也是隐私法演变给出的教训。美国没有通过一部全国性的隐私法。 欧洲通过了两次;而今天任何规模的美国公司,在全美范围内都适用符合欧洲隐私法的工作方式。那样就是更有效率。 所以我认为美国必须在国际上参与进来,才能影响世界其他地区;如果你什么都不做,你就无法参与。你必须做点什么。 所以你可以采纳克鲁兹参议员的想法——一种轻量级的做法—— I think it is a really important point you make and it is the lesson from the evolution of privacy law. The United States did not adopt a national privacy law. Europe did twice, and most American companies of any size today apply across the United States work that complies with European privacy law. It is just more efficient. So I think the United States needs to be in the game internationally to influence the rest of the world, and you cannot be in the game if you do nothing. You must do something. So you take Senator Cruz's idea--a lightweight approach---- ¶
Senator Young 是的。 **Mr. Smith〔继续〕.** ——然后在它周围建立起支持。 Yes. Mr. Smith.--and then you build support around it. ¶
Senator Young 那么,就把这一点展开——我继续问史密斯先生,请其他人原谅,因为我时间有限——如果我们已经通过了一套标准,是不是就更容易去塑造那些与我们共享大部分价值观的其他大型经济体的标准? So just to unpack that--and I will stick with Mr. Smith with apologies to everyone else because my time is limited--would it be easier to shape the standards of other large economy countries that share most of our values if we already have a set of standards adopted? ¶
Mr. Smith 总的来说,是的。我认为我们始终要小心,因为如果你动得太早,就会在标准真正成形之前出手。但你必须有一个可以向世界其他地区展示、并为之赢得支持的样板。 Generally, yes. I think we always have to be careful because if you go too soon you go before the standards have really come together. But you have got to have some kind of model that you can show the rest of the world and win support for. ¶
Senator Young 那么想必标准是可以协调统一的,对吧?它们并不是刻在石板上、一成不变的,可以这么说,对吧? And then presumably standards could be harmonized, right? They are not set in--and chiseled onto a tablet, so to speak, right? ¶
Mr. Smith 那是不可或缺的。我是说,如果我们的技术要走向世界,我们就需要一套法律或法规,实际上为互惠与互操作性创造基础。 That is indispensable. I mean, if our technology is going to go around the world we need a set of laws or regulations that, in effect, create that basis for reciprocity and interoperability. ¶
Senator Young 好。我只剩 25 秒了。对史密斯先生的立场有谁强烈反对吗?因为在我看来,那非常合理。 看起来与「轻触式」的做法一致,但同时也显示出某种紧迫感:美国需要行动。 我在剩下的 10 秒里要说的最后一件事是:我计划今天与几位同事一道提出一项名为《AI 公众认知与教育宣传法案》的立法,目标是采取全政府的方式来促进更高的 AI 素养认知、扩大 STEM 机会,以培养我们下一代的劳动力,并期待把它推进下去。 所以它将可供公众审阅、批评,甚至赞誉;主席先生,我交还时间。 **The Chairman.** 谢谢。 布朗特·罗切斯特(Blunt Rochester)参议员, ### 尊敬的莉萨·布朗特·罗切斯特(Lisa Blunt Rochester)陈述 ### 特拉华州联邦参议员 **Senator Blunt Rochester.** 谢谢您,克鲁兹主席,也非常感谢各位证人。这是一场如此重要的听证会。五分钟对我不够用。我会提交一些书面问题。 我注意到,当「给我画一幅未来的画」这个问题提出来时,奥特曼先生和史密斯先生——你们其实在椅子上坐直了。那是一种兴奋,而我对未来也正是这样。 2017 年我进入众议院时,我发起了一个两党「工作未来」核心小组,因为我担心:第一,有些群体将被落在后面;第二,作为一个国家,我们也可能被落在后面。 我发起——我办过一场活动,让每个人走进房间时用一个词云告诉我,听到「工作的未来」你会想到什么。进门时最大的词是「恐惧」。出门时最大的词是「机会」。 所以对我来说,这场对话如此重要:既要思考机会,也要确保我们留意伦理、留意诈骗、留意技术不要反过来支配人。 所以我非常感谢这场对话;奥特曼先生,我听过大约一年前你接受莱斯特·霍尔特(Lester Holt)的一次采访,你在那次采访中谈到 OpenAI 最初甚至并不是为了做一个产品。 不是为了钱。所以我知道你是在特拉华州注册的,我也了解在先前提出的立法过程中——不是立法,是在转为营利性实体的过程中——你一直在与我们的总检察长合作;而这个周一,OpenAI 决定改为申请成为一家公益公司(PBC),并由这家 PBC 来管理你们的非营利部门。 这个决定是怎么做出的?哪些考虑影响了这次组织变更的时机? OK. I only have 25 seconds left. Are there any violent objections to Mr. Smith's position? Because that seems eminently reasonable to me. Seems consistent with the light touch approach but it also shows a certain sense of urgency that the United States needs to act. The last thing I will say in my remaining 10 seconds is that I am planning on introducing legislation today called the AI Public Awareness and Education Campaign Act with several of my colleagues and our aim is to have a whole of government approach to foster greater awareness of AI literacy and grow STEM opportunities to create the next generation of our workforce, and looking forward to moving that forward. So it will be available for public review, critique, even accolades and, Mr. Chairman, I yield back. The Chairman. Thank you. Senator Blunt Rochester, STATEMENT OF HON. LISA BLUNT ROCHESTER, U.S. SENATOR FROM DELAWARE Senator Blunt Rochester: Thank you, Chairman Cruz, and thank you so much to the witnesses. This is such an important hearing. Five minutes will not suffice for me. I will be submitting some questions for the record. I notice that for Mr. Altman and Mr. Smith when the question of paint me a picture of the future came up there was--you actually leaned up in your chair. There was a level of excitement, and that is how I am about the future. When I came into the House of Representatives in 2017 I started a Future of Work bipartisan caucus because I had a concern that, number one, there were certain groups of people that were going to be left behind but there--also as a country that we could be left behind. And I started--I had an event where we had everyone walk into the room and use a word cloud and tell me what you think of when you hear the future of work. The biggest word coming in the door was fear. The biggest word walking out the door was opportunity. And so, to me, this conversation is so vital to think about the opportunities but also making sure that we are watching out for ethics, watching out for scams, watching out that technology does not take over the human. And so I am just grateful for the conversation and, Mr. Altman, I listened to an interview about--that you gave with Lester Holt maybe a year or so ago and you talked in that interview about how OpenAI--it was not initially even about making a product. It was not about the money. And so I know you are incorporated in Delaware and I understand you have been working with our attorney general during the previously proposed legislation to transition to a for profit--not legislation but to transition to for profit--and this Monday, OpenAI decided to apply to become a public benefit corporation instead and to have the PBC govern your nonprofit arm. What went into this decision and what considerations influenced the timing of the organizational change? ¶
altman 所以我们从未——谢谢您的问题,参议员,也感谢您给我解释这件事的机会。这是一件复杂的事情,我认为它一直被歪曲了。所以这是一个谈论它的绝佳场合。我们从未计划让非营利组织转变成任何别的东西。这个非营利组织本来就会一直是非营利组织,而且我们从一开始就计划成立一家公益公司(PBC)。还有一堆其他的考量,比如是不是由 PBC 的董事会以某种方式控制非营利组织,或者,你知道的,我们的资本结构将如何运作,其中很多在媒体上有大量猜测,而这些猜测大多是不准确的。但我们的计划始终是拥有一个强健的非营利组织。我们希望我们的非营利组织将成为世界上最好的、也许有一天是资源最丰富的非营利组织之一,以及一家拥有相同使命的 PBC,它将使我们能够筹集所需的资本,以人们想要使用的质量水平和可得水平来交付这些工具和服务,同时仍然坚守我们的使命,对于过去近十年我们在这一使命上的进展,我们一直感到自豪。所以我们就如何最好地做到这一点,与很多利益相关方、很多律师和很多监管者进行了很多富有成效的对话。这比我们想象的花了更长时间。你知道的,我本以为我们去年就会在谈论这件事。但现在我们有一个人们似乎相当兴奋的提案,我们正努力推进它。Blunt Rochester 参议员:而且,苏博士,您的公司主要在 AI 技术栈的物理硬件部分运营。我与 Cantwell 和 Blackburn 参议员有一项法案,叫作《促进韧性供应链法案》(Promoting Resilient Supply Chains Act),它授权商务部加强美国在关键行业和新兴技术方面的供应链。苏博士和其他人,半导体和芯片制造对于推进 AI 的发展至关重要,但我们正面临这些全球供应链的制约。我们需要采取哪些具体政策——我知道您也提到了供应链方面的政策——来帮助美国公司克服供应链问题,并在国际上与对手竞争?【译注:本 turn 原文后半段(从「Blunt Rochester 参议员:」起)为议员发问内容,被采集为同一话轮文本;按源件边界 1:1 保留,不拆分、不改写。】 So we never--thank you for the question, Senator, and the chance to explain this. It is a complicated thing that I think has gotten misrepresented. So this is a wonderful forum to talk about it. We never planned to have the nonprofit convert into anything. The nonprofit was always going to be the nonprofit, and we also planned for a PBC from the very beginning. There were a bunch of other considerations about is it the PBC board that would control the nonprofit somehow or, you know, how our capital structure was going to work that there was a lot of speculation on most of it, inaccurate in the press. But our plan has always been to have a robust nonprofit. We hope our nonprofit will be one of the best, maybe someday the best resourced nonprofit in the world, and a PBC with the same mission that would make it possible for us to raise the capital needed to deliver these tools and services at the quality level and availability level that people want to use them at but still stick to our mission, which we have been proud over the last almost decade of our progress toward. So we had a lot of productive conversations with a lot of stakeholders and a lot of lawyers and a lot of regulators about the best way to do this. It took longer than we thought it was going to. You know, I would have guessed that we would have been talking about this last year. But now we have a proposal that people seem pretty excited about and we are trying to now advance. Senator Blunt Rochester. And, Dr. Su, your company primarily operates in the physical hardware portion of the AI stack. I have a bill with Senators Cantwell and Blackburn called the ``Promoting Resilient Supply Chains Act'', which authorizes the Department of Commerce to strengthen American supply chains for critical industries and emerging technologies. Dr. Su and others, semiconductor and chips manufacturing is critical to advancing the advancement of AI but we are facing these global supply chain constraints. What specific policies--and I know you mentioned policies as well for supply chains--would we need to adopt to help American companies overcome the supply chain issues and compete in international with our rivals? ¶
Ms. Su 谢谢您的问题,参议员。 毫无疑问,半导体供应链和整体供应链对我们赢下 AI 竞赛确实至关重要。从半导体的角度看,我认为把制造环节迁回美国所做的努力是积极的。 我认为那只是一个开始。我们还能做很多,而其中最重要的方面之一,是真正端到端地去考虑它。 从半导体供应链的起点到终点有那么多步骤,包括先进晶圆、包括封装、包括后端与系统测试。 所有这些环节都需要在美国有落脚点;同时我们在世界各地有很多盟友,您知道,它们是供应链全球韧性中非常出色的伙伴,我们希望这些伙伴关系继续兴旺。 **Senator Blunt Rochester.** 如果可以,最后一个问题。 史密斯先生,你如何看待 AI 技术栈各部分之间的相互依存,在 AI 供应链中究竟造成脆弱性还是机会? Thank you, Senator, for the question. There is no question the semiconductor supply chain and overall supply chains are really critical for us to win the AI race. I think from a semiconductor standpoint the efforts that have been made to move manufacturing back to the United States have been positive. I think they are a start. There is a lot more that we can do, and one of the most important aspects of it is really to think about it end to end. There are so many steps to go from beginning to end in a semiconductor supply chain including advanced wafers, including packaging, including the back ends and system tests. All of those avenues need to have a footprint in the United States, and then we have many allies around the world which are, you know, very excellent partners as part of the global resiliency in the supply chain and we would like to see those partnerships continue to flourish. Senator Blunt Rochester. Last question, if I can. Mr. Smith, how do you see the interdependence between the AI stack sections creating either vulnerabilities or opportunities in the AI supply chain? ¶
Mr. Smith 我认为它们创造的机会多于脆弱性,因为它让公司能做自己最擅长的事,让我们能协同工作。 今天这个世界,你买的任何东西背后都有一条整合的供应链。我们只是去杂货店时不会去想它。 我认为科技行业的力量之一在于,我们拥有——我称之为「一串珍珠」,在技术栈的每一层都有伟大的公司;而且坦白说,我们将需要更多伟大的公司,特别是在应用层,这就是我们协作的方式。 **Senator Blunt Rochester.** 非常感谢。我时间到了,但我们会以书面形式、也会单独跟进提问。谢谢,我交还时间。 I think they create more opportunities than vulnerabilities because it enables companies to do what they do best and that we can work together. And the world today has an integrated supply chain for anything that you buy. We just do not think about it when we go to the grocery store. I think one of the strengths of the tech sector is that we have--I will call it a string of pearls, great companies in very--in every layer of the stack and we are going to need, frankly, more great companies, especially at the applications layer, and that it is how we work together. Senator Blunt Rochester. Thank you so much. I am out of time but we will be following up with questions for the record as well as individually. Thank you, and I yield back. ¶
Senator Lummis 莫兰(Moran)先生。 Mr. Moran. STATEMENT OF HON. JERRY MORAN, U.S. SENATOR FROM KANSAS ¶
Senator Moran ### 尊敬的杰里·莫兰(Jerry Moran)陈述 ### 堪萨斯州联邦参议员 卢米斯主席,非常感谢。 史密斯先生提到了数据隐私,这是我长期关注的话题,而我们在立法上一直没有成功。但我仍然抱着一个目标:确保消费者能掌控自己的数据。 我本来想问你,奥特曼先生:我们如何能让消费者对 AI 公司如何使用其数据拥有更多控制权,同时保持 AI 系统的实用性?也就是说,如何既得到更多隐私,又得到那些好处? Chairman Lummis, thank you very much. Mr. Smith mentioned data privacy, which has been a topic of mine for a long time, and we have been unsuccessful in legislation being adopted. But I still have the goal of making certain that consumers have control over their own data. And I was going to ask you, Mr. Altman, how can we provide consumers with more control over how their data is used by AI companies while preserving the utility of the AI system? So how do you get more privacy and still get the benefits? ¶
altman 所以,有所有那些像我们这样的公司以及其他公司都会建立、也应该建立的标准隐私控制,但有一个新的领域,我很乐意提请你们考虑,那就是:人们正在向 AI 系统分享的信息,比我认为他们向之前几代技术分享的更多,而这些系统的最大效用,发生在模型能够高度个性化地适配你的时候。所以这是一件美妙的事情,我们应该找到办法实现它,但这些 AI 系统将在你的一生中如此深入地了解你,我认为这为我们在 AI 世界中如何看待隐私提出了一个新的挑战和新的重要性层级——当一个人就生活中发生的任何事情与 AI 系统交谈时,我们将如何考虑保障人们的隐私——当另一个系统连接到另一个系统时,我们如何确保它分享的是恰当的信息、而不分享其他信息,并且用户对此拥有控制权。我相信,随着人们将这项技术更多地融入生活,这将成为未来几年 AI 最重要的问题之一,我认为这是你们所有人值得思考并相当认真对待的一个领域。 So there is all of the standard privacy controls that companies like ours and others build and should, but there is a new area that I would love to flag for your consideration, which is people are sharing more information with AI systems than I think they have with previous generations of technology, and the maximum utility of these systems happens when the model can get very personalized to you. So this is a wonderful thing and we should find a way to enable it, but the fact that these AI systems will get to know you over the course of your life so well I think presents a new challenge and level of importance for how we think about privacy in the world of AI--how we are going to think about guaranteeing people privacy when they talk to an AI system about whatever is happening in their lives--how we make sure that when one system connects to another it shares the appropriate information and does not share other information and that users are in control of that. I believe this will become one of the most important issues with AI in the coming years as people come to integrate this technology more into their lives, and I think it is a great area for you all to think about and take quite seriously. ¶
Senator Moran 我们确实有。我们只是在找到结论上没有取得成功。不过谢谢你的鼓励。 我担任商业、司法、科学拨款小组委员会主席,该小组委员会为司法部提供资金,而司法部在我国网络安全中扮演重要角色。 我刚从一场与联邦调查局局长帕特尔(Patel)博士的预算听证会过来,我们谈到了网络安全威胁。 AI 可以——我认为这是真的——AI 可以被用在网络攻击的两端,可以被用来自动化钓鱼、生成恶意软件。但机器学习也能提升我们检测和响应网络威胁的能力。 国会应当如何考虑联邦网络安全资源的配置?在 AI 方面我们应当考虑什么? We do. We just do not have any success in finding the conclusions. But thank you for the encouragement. I chair a Commerce, Justice, Science Appropriations Subcommittee that funds the Department of Justice and it plays a significant role in cybersecurity of our country. I just came in from a budget hearing with the FBI Director Dr. Patel in which we covered cybersecurity threats. AI can--and I think this is true--AI can be used on both sides of a cybersecurity attack and it can be used to automate phishing, malware creation. But machine learning can also increase our ability to detect and respond to cyber threats. What should Congress think about allocation of Federal resources for cybersecurity and what should we consider when it comes to AI? ¶
Mr. Smith 我要说,正如您所说,在网络安全方面,AI 既是进攻性武器,也是防御性盾牌;而如同许多其他事情一样,过去几年这条战线的前沿在乌克兰,因为俄罗斯拥有如此老练的网络攻击能力。 您知道,我们发现的是——我们这家公司从战争爆发那一刻起就一直参与支持乌克兰——AI 是一个改变游戏规则的东西。 我们拦截针对乌克兰的攻击,比人类能够察觉的速度更快,我们阻止了那些攻击得逞。 所以你要把 AI 部署到——姑且称之为产品本身的前线。我们必须认识到,最终是人在防御,不只是国家,还有公司和政府,是那些首席信息安全官(CISO)。 所以我们做了一款叫「网络安全副驾驶」(cybersecurity copilot)的东西,基本上为那些人把大量占用他们时间的工作流自动化,让他们能更有效、更高效。 谈到联邦拨款,我认为简单说:美国政府必须始终处在拥有自身所需网络安全能力的最前沿,以保卫政府,而且每天都是如此——我是说,就在这场听证会进行时,我们正在政府机构里把中国人清出去之类的事情;从今往后,大概直到永远,每年每一天都会发生这种事。 所以我们必须让美国政府在这个领域资金充足;我认为我们也需要让情报机构、特别是国家安全局(NSA)资金充足,好让他们在这个领域的全球领导地位上保持最前沿。 I would say that AI, as you said, is both an offensive weapon and a defensive shield when it comes to cybersecurity, and as with many other things the front line of this the last few years has been in Ukraine because Russia has such a sophisticated cyberattack capability. And, you know, what we have found is a company that has been involved in supporting Ukraine since literally the moment that war began is that AI is a game changer. We have intercepted attacks against Ukraine faster than a human could detect them and we block those attacks from taking place. So you deploy AI into--call it the front line of the products themselves. We have to recognize that it is ultimately the people who defend not just countries but companies and governments, the chief information security officers, or the CISOs. So we have created what is called a cybersecurity copilot that basically automates for those individuals much of the workflow that takes their time so that they can be more effective and efficient. When it comes to Federal appropriations I think that, to put it simply, the U.S. Government must remain at the forefront of having for itself the cybersecurity capabilities that it needs to defend the government and every day--I mean, we are in government agencies today during this hearing, you know, pushing Chinese out of agencies and the like, and this will happen every day of every year from now to probably eternity. So we must keep the U.S. Government well funded in this space and I think we also need our intelligence agencies and especially the NSA to be well funded so they can remain at the forefront when it comes to global leadership in this field. ¶
Senator Moran 谢谢你的观察和鼓励。 我最后一个问题——乡村地区,那是我来自的地方,常常缺乏高速宽带;而由于许多 AI 工具依赖连接性,我担心全国很多地方、堪萨斯很多地方可能无法获得 AI 将带给企业、学校、医疗等的益处。 联邦政府能做什么来支持「端侧」或低带宽 AI 系统的开发与可获得性——那种不依赖持续连接的 AI 系统? Thank you for your observations and encouragement. My final question--rural areas, a place I come from, often lack high-speed broadband, and since many AI tools rely upon connectivity I am concerned that many parts of the country and many parts of Kansas may not be able to access the benefits that AI will bring to business, schools, health care, et cetera. What can the Federal Government do to be supportive of development and availability of on device or low broadband width AI systems that do not rely on constant connectivity? ¶
altman 我总体上对 AI 在这里能做什么相当兴奋,因为你可以把大量的处理工作转移到云端,然后只传输相对少量的数据。如果你想一想,比如,ChatGPT——文本进来,有一个「大脑」非常努力地思考它,然后一些文本返回,我们就能以同样的服务质量很好地支持低连接性地区的人们。除此以外,我认为让各处都获得良好的连接很重要,但就在 AI 这个具体领域,我认为我们实际上能很好地解决那个差距。 I am generally pretty excited about what AI will do here because you can offload so much of the processing to the cloud and then ship a relatively small amount of data. If you think about, you know, ChatGPT as text comes in there is, like, a brain that thinks about it really hard and some text comes back, we can support people in low connectivity areas quite well with the same quality of service. Separately to that, I think getting great connectivity everywhere is important but in the specific area of AI I think we can actually address that gap quite well. ¶
Senator Moran 知道这一点很好。 非常感谢。 That is good to know. Thank you very much. ¶
Senator Lummis 卢汉(Lujan)先生。 Mr. Lujan. STATEMENT OF HON. BEN RAY LUJAN, U.S. SENATOR FROM NEW MEXICO ¶
Senator Lujan ### 尊敬的本文·雷·卢汉(Ben Ray Lujan)陈述 ### 新墨西哥州联邦参议员 谢谢您,主席女士。 首先,我想表彰并感谢奥特曼先生和史密斯先生,你们的机构持续参与 NIST 美国 AI 安全研究所(USAI Safety Institute);也感谢苏博士和奥特曼先生与我国国家实验室的持续合作。 那么,苏博士和奥特曼先生,能否解释你们与国家实验室的伙伴关系如何支持科学研究?你们在回答达克沃斯参议员的问题时也解释过,但能否简要谈谈? Thank you, Madam Chair. And, first, I want to begin by recognizing and thanking Mr. Altman and Mr. Smith for your organizations' ongoing involvement in the NIST USAI Safety Institute, as well as Dr. Su and Mr. Altman for your ongoing partnerships with our National Laboratories. Now, Dr. Su and Mr. Altman, can you explain how your partnership with the National Labs support scientific research? You explained this to a question that was asked by Senator Duckworth as well but if you could just touch on that quickly. ¶
altman 我们最新的模型,比如,2003,擅长科学推理,因此科学家们能够用它们来帮助他们审阅文献、提出新想法、设计实验、以前几代模型根本做不到的方式分析数据。我们曾让国家实验室和其他科学家花时间使用以前的模型,他们说,哦,这是,你知道的,有点酷。很有意思。但它并没有改变事情。这些新模型是我们第一次从国家实验室和其他地方的科学家那里听到,说这对他们的研究产出是一个真正改变游戏规则的东西(game changer)。【译注:原文「like, 2003」疑为口误或采集误差(可能指某模型代号或年份),按原文照译并标记待核;不自行推测为具体模型。】 Our latest models, like, 2003 are good at scientific reasoning and so scientists are able to use these to help them review literature, come up with new ideas, propose experiments, analyze data in a way that the previous generations of models just could not. We have had the National Labs and other scientists spend time with previous models and they say, oh, this is, you know, kind of cool. It is interesting. It is not transforming things. These new models are the first time we are hearing from scientists at the National Labs and elsewhere that this is a legitimate game changer to their research output. ¶
Senator Lujan 我很感谢。苏博士? I appreciate that. Dr. Su? ¶
Ms. Su 是的,我要补充同样的内容。我认为我们与国家实验室的伙伴关系看到了巨大的机会。我们在各国家实验室拥有大规模算力,而真正开发新应用的能力——利用我们所说的传统高性能计算,加上我们刚才谈到的新 AI 模型能力——我认为这是大幅推进科学发现能力的绝佳机会。 Yes, I would add the same. I think our partnerships with the National Labs have seen just tremendous opportunity. We have large-scale compute across the National Labs and the ability to really develop new applications that take advantage of, let us call it traditional high-performance computing, together with the new AI model capability that we just talked about is, I think, a great opportunity to substantially move forward the ability for scientific discovery. ¶
Senator Lujan 再问你们两位:能否解释为什么联邦对基础研究和标准机构的投资对你们的公司至关重要? To both of you, again, can you explain why Federal investment in foundational research and standards bodies are crucial to your companies? ¶
altman 我认为标准有助于加快创新速度,但重要的是行业首先弄清楚标准应该是什么。我认为一个糟糕的标准真的会让事情倒退,我们在历史上已经看到了很多这样的例子。我确实认为这里有一个新的协议有待发现,其重要性层级与 HTP 相当。这只是一个例子。还有很多其他事情。我相信行业会通过一些反复试错弄清楚这一点,然后我认为正式采纳它可能是有帮助的。【译注:原文「HTP」为缩写,未在源件中展开,按原文保留并标记待核;不自行猜测其全称。】 I think standards can help increase the rate of innovation, but it is important that the industry figure out what they should be first. I think a bad standard can really set things back and we have seen many examples of that in history. I do think there is a new protocol to discover here at the level of importance of HTP. This is just one example. There is many other things, too. I believe the industry will figure that out through some fits and starts and then I think officially adopting that can be helpful. ¶
Senator Lujan 苏博士。 Dr. Su. ¶
Ms. Su 我相信公私伙伴关系确实让我们能够思考——姑且说——走在曲线之前。所以我们在产业界做很多事,而且做得非常非常好。 然而,国家实验室和联邦研究的美妙之处在于,它确实允许——姑且说——更多「蓝天」研究,而我认为那是一份非常积极的补充。 所以我认为关键在于,我们如何确保——您知道——一美元联邦资金能通过叠加其上的私人投资,发挥出大得多得多的效用。 I believe public-private partnerships really enable us to think, let us call it, ahead of the curve. So there are lots of things that we do in industry and we do them very, very well. However, the beauty of the National Labs and Federal research is it does allow let us call it a bit more bluesky research, and I think that is a very, you know, positive add. So I think the key is how we can make sure that, you know, one Federal dollar goes much, much further than that with a private investment on top of that. ¶
Senator Lujan 昨天我重新提出了一项名为《测试 AI 法案》的立法,它有两党支持,目的只是提升联邦政府在这个领域测试与评估的能力。所以非常感谢你们两位的回答。 但我认为,这只是确保美国保持领先所需的众多步骤之一。现在,尽管全国有强烈支持——包括今天在座的业界领袖——特朗普总统却在砍掉基础研究的预算,很多人心中充满疑问。 我要说,这会摧毁我们国家的竞争优势。我只是呼吁所有同事,看看对国家科学基金会、国家卫生研究院和能源部科学办公室的投入。 让我们合作。如果有疑问,让我们找到回应的办法。但让我们确保这些投入产生积极的差别,好让我们有更多成功、更多像今天这样值得庆祝的听证会。 现在,除了与联邦政府的伙伴关系之外,我想更多了解你们在建设中心时如何与当地社区合作。 数据中心会给能源和水资源带来压力。然而,与其他企业不同,它们未必带来很多长期就业和经济收益。 那么,史密斯先生,你们有多少工程师专门从事模型或硬件优化以减少能耗?当你们建一个中心时,有哪些减少用水的举措? Yesterday I reintroduced a piece of legislation called the ``Test AI Act'', which has bipartisan support, which would simply improve the Federal Government's capacity to test and evaluate in this area as well. So very much appreciate both your responses. But this is just one of many steps I would argue that is needed to ensure that the United States stays ahead. Now, despite strong support across the country including from industry leaders here today President Trump is annihilating budgets for basic research, and there are questions abound by so many. I will argue that this will destroy our Nation's competitive advantage. I simply just call on all my colleagues that we look at the investments to the National Science Foundation, National Institutes of Health, and Department of Energy, Office of Science. Let us work together. If there is questions that we have let us find ways to address those. But let us ensure that these investments are making a positive difference so that we have more successes and more hearings celebrating what we are celebrating today. Now, beyond your partnership with the Federal Government I would like to know more about how you partner with local communities when building out centers. Data centers put a strain on energy and water resources. However, unlike other businesses they do not introduce many long-term jobs and economic benefits necessarily. So, Mr. Smith, how many engineers do you have dedicated to model or hardware optimization to reduce energy use, and when you build a Center what initiatives do you have in place to reduce water use? ¶
Mr. Smith 我一时想不起我们有多少工程师在做优化,但我很乐意去查一个答案给你。 用水是一个巨大的优先事项,尤其是在——您知道——数据中心,比如美国西南部和世界上其他水资源紧缺的国家。 如果你看我们今天的数据中心,它们采用液冷,是闭环系统。那种液体坦白说是水与其他化学品的混合物,但基本上一旦开始运行,几乎全部的水都会被回收。所以我们消耗的水量通常远、远小于大多数人的估计。 我们还有水资源回补的承诺。我们的目标是「水正效益」,也就是我们向社区提供的水多于我们消耗的水。 比如,今天在全美我们有 90 多个水资源回补项目,其中包括一个聚焦于你们新墨西哥州圣胡安河(San Juan River)的项目,着眼于该河流的水安全。 所以我认为这是一个很好的例子,说明我们如何在应对一个日益重要的问题上发挥负责任的作用。 I do not know off the top of my head the number of engineers we have working on optimization but I would be happy to track down an answer and get it to you. Water use is a huge priority especially, you know, in data centers, for example, in the southwestern United States and other countries around the world where water is in short supply. If you look at our data centers today they run on liquid cooling. It is a closed loop system. The liquid is a combination of, frankly, water and other chemicals but basically once it starts running almost all of the water is recycled. So the amount of water that we consume is typically far, far smaller than what most people would estimate. We also have a commitment to water replenishment. Our goal is to be water positive, meaning that we are providing more water to the community than we are consuming. So, for example, across the United States today we have more than 90 water replenishment projects including one that focuses on the San Juan River in your state of New Mexico, which focuses on water security for the river. So I think it is a good example of how we can play a responsible role in addressing an issue that is of growing importance. ¶
Senator Lujan 我很感谢。 因特拉托先生,同样的问题。 I appreciate it. Mr. Intrator, same question. ¶
Mr. Intrator 是的,我无法回答我们有多少工程师专注于此;但我要说,从既定的每兆瓦电力中榨出更多算力的能力,对我的公司、对在座所有人都是至关重要的;我们花大量时间整合最前沿的技术,那些技术在算力产出上的效率,比传统技术历来所能做到的,是阶跃式的提升。 您知道,所以转向液冷在效率上是不可思议的改进;归根结底,我们是在一个既定的数据中心、既定的功率包络内面对这个问题:我们能把计算资源向前推进多少?那确实是我们所做工作中很重要的一部分。 Yes, I cannot answer the question of how many engineers we have focused on it but I will say that the ability to extract more computational power out of a given megawatt is of paramount importance to my company, to all of us in this room, and we spend an enormous amount of time integrating the most bleeding-edge technology, which is a step function more efficient in terms of its computational output than the legacy technology has historically done. You know, so moving to liquid cooling has just been an incredible improvement in efficiency and, ultimately, we face this problem from, you know, within a given data center, within a given power envelope. How much can we move the computational resources forward, and that is really an important part of what we do. ¶
Senator Lujan 我很感谢。 主席先生,我还有其他问题会以书面形式提交。 莫兰先生问过一个问题,奥特曼先生回应了。但能否请各位只回答是或否:为了让人工智能达到它的充分显要地位,确保全国各地的人都能接入快速、可负担的互联网,是否重要? 苏博士。 I appreciate it. Mr. Chairman, I have other questions I will submit into the record. Mr. Moran did ask one question. Mr. Altman, you responded to it. But can you all just answer yes or no, is it important to ensure that in order for AI to reach its full prominence that people across the country should be able to connect to fast, affordable internet? Dr. Su. ¶
Ms. Su 是。 Yes. ¶
Mr. Intrator 是。 Yes. ¶
Mr. Smith 是。 Yes. ¶
Senator Lujan 谢谢。很感谢。我交还时间。谢谢。 **The Chairman〔主持〕.** 谢谢。 卢米斯参议员。 Thank you. Appreciate it. I yield back. Thank you. The Chairman [presiding]. Thank you. Senator Lummis. STATEMENT OF HON. CYNTHIA LUMMIS, U.S. SENATOR FROM WYOMING ¶
Senator Lummis ### 尊敬的辛西娅·卢米斯(Cynthia Lummis)陈述 ### 怀俄明州联邦参议员 谢谢您,主席先生;也感谢各位今天前来。 我真的对这个领域持续取得的卓越进展感到惊叹;我已经在怀俄明看到有人用 ChatGPT 或 Claude 改善他们的生意,无论是医疗、采矿、石油天然气、教育,甚至牧场经营。我真的很兴奋于这个机遇给美国带来的一切。 现在,在我看来,世界给了我们两条路。一方面,欧盟选择先监管、后提问。GDPR 已经在限制欧洲获取最强 AI 模型。 另一方面,中国似乎在为 AI 发展开快车道,非常迅速地建起大量能源设施,试图压过美国。 所以我想问几个问题:我们如何确保充分获得这项技术的好处并加速其发展。 第一个问题:过去一年,我们看到包括加利福尼亚和得克萨斯在内的许多州在考虑自己的 AI 框架,每一个本身都相当繁重。与此同时,我们对中国的领先正在缩小到大约只有 6 个月。 那么首先,奥特曼先生,能否请你勾勒一下:如果美国出现一套拼盘式的监管框架,世界会是什么样子?那会如何影响我们的竞争力? Thank you, Mr. Chairman, and thank you all for coming today. I really have been amazed at the outstanding progress that continues to be made in this field and I am already seeing people in Wyoming that are using ChatGPT or Claude to improve their businesses, whether it is health care or mining or oil and gas or education, ranching, even. I am just really excited about what this opportunity brings to America. Now, as I see it, the world has presented us with two paths. On one hand, the EU has chosen to regulate first and ask questions later. The GDPR is already limiting European access to the most capable AI models. On the other hand, China appears to be fast tracking AI development, standing up large amounts of energy very quickly in an attempt to outcompete America. So, I would like to ask a few questions about how we can make sure we get the full benefit of this technology and accelerate its development. So first question, over the past year we have seen many states including California and Texas consider their own AI frameworks, each one significantly burdensome in their own right. At the same time, our lead against China is shrinking to about only 6 months. So, first of all, Mr. Altman, could you please sketch out what the world could look like if the U.S. were to have a patchwork regulatory framework and how that could impact our competitiveness? ¶
altman 我认为那会相当糟糕。我认为很难想象我们能弄清楚如何遵守 50 套不同的监管规定,而在其中许多州,已经有几十项不同的法案被提出,据我理解其中有几项可能会通过。那会在我认为不符合任何人利益的时点拖慢我们。一个统一的、轻量接触(light touch)的、我们能理解的联邦框架,让我们能够,你知道的,以这个时刻所要求的速度行动,这看起来很重要也很好,但那种每个州在这里采取不同做法的情形,我认为会相当繁重,并显著损害我们做我们需要做的事情的能力,也希望那是你们希望我们做的事。 I think it would be quite bad. I think it is very difficult to imagine us figuring out how to comply with 50 different sets of regulation and in many of these states there have been dozens of different bills proposed that I understand several of which could be passed. That will slow us down at the time where I do not think it is in anyone's interest for us to slow down. One Federal framework that is light touch that we can understand and that lets us, you know, move with the speed that this moment calls for seems important and fine, but the sort of every state takes a different approach here I think would be quite burdensome and significantly impair our ability to do what we need to do and, hopefully, you all want us to do, too. ¶
Senator Lummis 有人不同意奥特曼先生对「拼盘」的判断吗? 谢谢。我有一些关于领先并参与 AI 竞争所必需的基础设施的问题,所以我接下来的问题给我们的基础设施提供方,史密斯先生和——是叫因特拉托吗? Does anyone disagree with Mr. Altman's assessment of a patchwork? Thank you. I have some questions about the infrastructure that is going to be necessary to lead and compete in AI so my next questions are for our infrastructure providers, Mr. Smith and Mr.--is it Intrator? ¶
Mr. Intrator 没错。 That is correct. ¶
Senator Lummis 因特拉托。谢谢。能否详细谈谈现行许可流程如何影响了你们快速部署 AI 基础设施的能力?越具体越好。 Intrator. Thank you. Could you elaborate on how current permitting processes have impacted your ability to rapidly deploy AI infrastructure? The more specific you can be the better. ¶
Mr. Intrator 那我先就「拼盘」简单说一句,然后进入正题。我们在基础设施一侧做的投资是巨大的;而这样一种设想——你做了一笔投资,结果它可能被困在一个有某类特定监管的辖区里,让你无法充分利用它——那真的非常、非常次优,也让围绕基础设施的决策变得困难。 就许可而言,每当这个话题出现,围绕许可的讨论都令人痛苦;它之所以令人痛苦,是因为你无法快速建设、无法大规模建设;我认为这只是从数据中心往前算,甚至还没开始谈从数据中心往回、贯穿为这些大规模投资供电所必需的能源基础设施——而正是那个规模让它们对推进人工智能有意义。我很乐意花更多时间深挖细节,但也许直接单独谈更好。 So a quick comment on the patchwork and then I will dive in here. The investment that we are making on the infrastructure side is enormous, and the idea that you can make an investment that could then become trapped in a jurisdiction that has a particular type of regulation that would not allow you to make full use of it is really very, very suboptimal and makes the decisionmaking around infrastructure challenging. As far as the permitting goes, whenever this topic comes up the discussion around permitting is excruciating and it is excruciating from the ability to quickly build and to build large, and I think that is kind of from the data center forward without even beginning the discussion from the data center back through the energy infrastructure that is necessary to be able to power these large investments at the scale that make them of relevance to moving artificial intelligence forward. I am happy to spend more time kind of digging into more details but probably do that directly. ¶
Senator Lummis 好。我期待那次谈话,因为我担心怀俄明州非常清洁的天然气会成为你们行业担忧的东西,因为特朗普总统喜欢天然气,而拜登总统不喜欢。 而如果你建了巨大的数据中心,之后又来了一位反天然气的总统,那在你们决定如何配置资本时就是个顾虑。 史密斯先生,你同意吗? OK. And I will look forward to that conversation because I am worried about Wyoming's very clean natural gas being something your industry is concerned about because President Trump likes natural gas but President Biden did not. And if you build huge data centers and another President comes along who is anti-natural gas that is a concern for you as you are deciding how to deploy capital. Mr. Smith, do you agree? ¶
Mr. Smith 总的来说,我同意。我是说,我要说我们需要这个国家在各届政府之间保持一致性。我们需要找到更多两党共识的机会;我只想指出,在夏延(Cheyenne),我们长期有一个数据中心综合体,您知道,我们确实有烧天然气的备用发电机。所以我们有很多方式把不同的能源供应善加利用。 Generally, I do. I mean, I would say we need consistency across administrations in this country. We need to find more opportunities for bipartisan agreement, and I will just highlight that in Cheyenne where we have long had a data center complex, you know, we do have backup generators that run on natural gas. So there are a variety of ways for us to put different energy supplies to good use. ¶
Senator Lummis 你们在探索小型模块化核反应堆吗? Are you exploring small modular nuclear? ¶
Mr. Smith 是的,包括与怀俄明的人合作。 Yes, including with people in Wyoming. ¶
Senator Lummis 谢谢。 奥特曼先生,我很高兴听到你们要发布一个开——哦,我时间到了。抱歉。时间过得真快。 Thank you. Mr. Altman, I am pleased to hear you are releasing an open--oh, my time is up. Excuse me. It goes so fast. ¶
altman 我很乐意改个时间跟你谈这件事。我们对此也非常兴奋。 I would love to talk to you about it another time. We are very excited about it, too. ¶
Senator Lummis 是的,谢谢。我交还时间。 **The Chairman.** 谢谢。 罗森(Rosen)参议员。 Yes, thank you. I yield back. The Chairman. Thank you. Senator Rosen. STATEMENT OF HON. JACKY ROSEN, U.S. SENATOR FROM NEVADA ¶
Senator Rosen ### 尊敬的杰基·罗森(Jacky Rosen)陈述 ### 内华达州联邦参议员 谢谢您,克鲁兹主席。我准备好按按钮了;总之,时间确实过得很快。谢谢您召开这场听证会。 我真的相信 AI 的前景。太令人兴奋了;我们必须提出正确的问题,一方面促进它的成长、探索并创造这些新的可能与路径,另一方面尽我们当时所能知道的,最好地保护自己免受恶意行为者或不良后果之害。 还有,奥特曼先生,谢谢你昨天花时间与我交谈。我期待在这件事上继续与你合作。 那么,我今天想从 DeepSeek 这个对抗性 AI 开始,因为二月份我与赫斯特德(Husted)参议员提出了两党立法,禁止在政府设备上使用 DeepSeek;本周早些时候,卡西迪(Cassidy)参议员和我提出了一项法案,把这些禁令扩大到联邦承包商。 那么,史密斯先生,对于在中华人民共和国这类对抗性国家开发、或由其开发的 AI 模型,我们应该采取什么方针? 我们是否应当担心我们的对手挪用 AI 来推广某种特定意识形态、收集美国的敏感数据?你们在如何应对这一威胁? Thank you, Chairman Cruz. So I am ready to push the button and, anyway, time does go by very fast. Thank you for having this hearing. I really believe in the promise of AI. So exciting, and we have to ask the right questions in order to promote its growth on one hand, and how can explore and create these new possibilities and pathways and also how do we protect ourselves from bad actors or outcomes as best as we can know at the time. And, Mr. Altman, thank you for spending some time with me yesterday. I look forward to continuing to work with you on this. So I want to start a little bit today at DeepSeek, an adversarial AI, because in February I introduced bipartisan legislation with Senator Husted to prohibit using DeepSeek on government devices, and earlier this week Senator Cassidy and I introduced a bill that would expand those prohibitions to include Federal contractors. So, Mr. Smith, what should our approach be to AI models that are developed in or by adversarial countries like the PRC? Should we be concerned about our adversaries co-opting AI to promote a particular ideology, collect sensitive U.S. data, and how are you combating this threat? ¶
Mr. Smith 嗯,我认为你可以拿 DeepSeek 这个例子来说明,它很好地说明了问题;我认为值得想一想这样一个事实:DeepSeek 产出了两样东西。他们有一个模型,是开源的模型;他们还有一个应用,DeepSeek 应用。 在微软,我们不允许员工使用 DeepSeek 应用。我们没有把 DeepSeek 应用放进我们的应用商店,正是出于你提到的那些担忧,也就是数据回流到中国,以及这个应用生成那种人们会说是与中国宣传相关的内容。 与此同时,因为这个模型本身是开源模型,我们有可能进去分析它、修改模型中的代码——我们和其他人都有这个许可——以移除有害的副作用。 所以我认为我们必须始终考虑这项技术的不同方面。我要说,把安全放在第一位,然后从那里往前走。 Well, I think you can take the DeepSeek example and it illustrates it well, and I think it is just worth thinking about the fact that DeepSeek produced two things. They have a model that is an open source model and they have an application, the DeepSeek app. At Microsoft we do not allow our employees to use the DeepSeek app. We did not put the DeepSeek app in our app store because of the kinds of concerns that you mentioned, namely, data going back to China and the app creating the kinds of content that I think people would say were associated with Chinese propaganda. At the same time, because the model itself is an open source model it was possible for us to go in it, analyze it, and change the code in the model, which we and other people have the permission to do to remove the harmful side effects. And so I think we have to always think about the different aspects of the technology. I will say put security first and then go forward from there. ¶
Senator Rosen 谢谢。我想我们都知道,在我们当前的世界里,数据才是真正的权力。他、她,或者不管是谁拥有数据,就真的能控制我们所做的很多事情。 但我想换个话题,奥特曼先生,稍微谈谈 AI 与反犹主义,因为今年早些时候,反诽谤联盟(ADL)发布了一份报告,显示几个主要的生成式 AI 模型延续了危险的反犹刻板印象,而且令人难过的是,还有阴谋论。 那么,奥特曼先生,业界正在采取什么措施确保 AI 模型不延续反犹主义?你是否会考虑与公民社会合作,建立一种与反犹主义相关的 AI 标准基准,把它作为一种评估形式,然后也许我们也能把这些用于其他形式的仇恨? Thank you. I think we all know that data is the real power in our current world. He or she or whomever owns the data really can control a lot of what we do. But I want to move on and speak with you, Mr. Altman, about AI and anti-Semitism a little bit because earlier this year ADL released a report showing that several major generative AI models have perpetuated dangerous anti-Semitic stereo tropes and, sadly, conspiracy theories. So, Mr. Altman, what steps is industry taking to ensure that AI models do not perpetuate anti-Semitism? Will you consider collaborating with civil society to create kind of a standard benchmark for AI related to anti-Semitism, use it as a form of evaluation, and then maybe we could use those for other forms of hate as well? ¶
altman 当然,我们确实在这个话题上与公民社会合作,我们很高兴继续这样做。我们希望我们的用户有用他们想要的方式使用模型的自由,但我们也不希望他们损害某种社会结构或特定群体。总会有一些争论,而在 AI 语境下的言论自由问题是一个新问题,我认为它与我们以前面对的不同。我们确实把这些视为用户一对一使用的工具,但当然,我们在这里不是为了,你知道的,制造可怕的反犹太(anti-Semitic)产品。 Of course, we do collaborate with civil society on this topic and we are excited to continue to do so. We want our users to have freedom to use models in the way they want, but we also do not want them to be damaging to sort of the fabric of society or particular groups. There will always be some debate and the question of free speech in the context of AI is novel and I think it is different than what we faced before. We really do view these as tools for users one-on-one but, of course, we are not here to, you know, make horrible anti- Semitic products. ¶
Senator Rosen 谢谢。我想转到——卢汉参议员谈过数据中心的能耗、用水,这是我们都很关心的。我想在此基础上再加一点数据中心安全,把它加进这个组合里。 上一届国会,我实际上推动一项法案成了法律,我的两党法案《联邦数据中心增强法案》。它为联邦数据中心建立了网络安全与韧性标准。 那么,问史密斯先生,或者——抱歉,苏博士。谢谢。 苏博士,我想问你一点关于硬件的问题。像 AMD 设计的芯片、我们希望去构想的那些新芯片,是否有办法——我知道我作为软件开发者的职业生涯里,我们只知道东西变得越聪明、越快,而且它们——越凉快,我们就能算得越好。 那么我们如何让芯片更凉?如何让我们的数据中心、我们的算力更安全?我知道互操作性有时是个因素。但你能稍微谈谈这个吗? Thank you. I want to move on to--Senator Lujan talked about data center energy use, water use, something we are all really concerned about. I want to put on top of that a little bit about data center security, add that to the mix. So last Congress I actually got a bill passed into law, my bipartisan ``Federal Data Center Enhancement Act''. It establishes cybersecurity and resiliency standards for Federal data centers. And so to Mr. Smith, or--I am sorry, Dr. Su. Thank you. Dr. Su, I want to ask you a little bit about hardware. Are there ways the hardware like the chips AMD designs, new chips that we are hoping to think about--I know my career as a software developer we just know things have gotten smarter, faster, and they just--the cooler they can be the better we can compute. So how can we make our chips cooler? How can we make our data centers, our computing power, more secure? And I know interoperability is sometimes a factor. But can you talk about this a little bit? ¶
Ms. Su 当然。谢谢您的问题,参议员。 您看,我认为正如您所说,所有这些都是极其重要的。所以在能效这一块——您知道,从芯片角度我们面临的功率约束——您知道,我们的工作是继续让我们的芯片每年都更高效。 您知道,过去几年我们已经看到 30 倍的改进,我们将继续聚焦于那个领域。 然后关于您说的安全、确保我们的芯片是安全的、别人无法以某种方式侵入它们,这些在我们未来几代芯片的整体开发周期中也是非常高的优先事项。 Sure. Thank you for the question, Senator. Look, I think all of those things are extremely important, as you said. So in our part of the energy efficiency, you know, power constraints that we have from a chip standpoint, you know, our job is to continue to make our chips more and more efficient every year. We have seen, you know, 30 times improvement over the last few years and we will continue to focus, you know, in that area. And then to your comments about, you know, security and ensuring that our chips are secure and people are not somehow breaking into them, those are also very high priorities in our overall development cycle for future generation chips as well. ¶
Senator Rosen 谢谢。我期待在这些重要问题上再次与各位合作。 主席先生。 **The Chairman.** 谢谢。 沙利文(Sullivan)参议员。 Thank you. I look forward to working with all of you again on these important issues. Mr. Chairman. The Chairman. Thank you. Senator Sullivan. STATEMENT OF HON. DAN SULLIVAN, U.S. SENATOR FROM ALASKA ¶
Senator Sullivan 谢谢您,主席先生。我要感谢各位证人今天的证词。感谢主席召开这场听证会;我同意克鲁兹参议员的开场陈述:就我们这场竞赛——无论你怎么称呼它——与中国的竞争而言,这是国家经济与国家安全的问题。 我知道这个话题已经被反复强调,但我只想——我想再挖深一点。各位都同意这一点吗? 我就问几个简短的问题:相对于中国,这是一个巨大的国家安全、经济安全问题,而我们美国需要在这方面取胜。非常重要。 大家都在点头。我知道这一点已经被提到过,但证人之间的共识是否是:我们现在领先,但这是一种暂时性的领先? 那会是什么——很快地说,从你开始,奥特曼先生。你对此的评估是什么?我知道你已经谈过,我只是想为后面一些问题设定背景。 Thank you, Mr. Chairman. I want to thank the witnesses for the testimony today. I appreciate the Chairman calling this hearing, and I agree with Senator Cruz's opening statement about this is a matter of national economic and national security in terms of our race, however you want to call it--competition with China. So I know this topic has been pressed but I want to just get--I want to dig down a little bit deeper. Do you agree with that, all of you? I am just going to ask some quick questions, that this is a huge issue of national security, economic security, relative to China, and we as America need to win in that regard. Very important. Everybody nodding their head. And then I know that it had been touched, but is the consensus among the witnesses that we are ahead right now but as a kind of tentative lead? What would be--very quickly we will start with you, Mr. Altman. What is your assessment on that? I know you have already talked about it. I just want to set the context for some of the questions. ¶
altman 是的,我相信我们现在正在引领世界。我相信我们会继续这样做。我们希望在美国制造 AI,我们希望全世界都从中受益。我认为这对美国是最有力的事情。我认为这对全世界所有人来说也是正确的事情。我真的很感谢你们今天在这里——与我们在一起,因为我认为我们需要你们的帮助,而你所说的每一件事,或者几乎每一件事,听起来都很棒。 Yes, I believe we are leading the world right now. I believe we will continue to do so. We want to make AI in the United States and we want the whole world to get the benefit from that. I think that is the strongest thing for the United States. I think it is also the right thing to do for all the people of the world. And I really appreciate you all being with here--with us here today because I think we will need your help, and everything you are saying or almost everything you are saying sounds great. ¶
Senator Sullivan 那么在我问这个问题时,我会问各位是否认为我们领先;但关键的是,当你说「我们需要你们的帮助」时,那会是什么——非常简洁地说,有时候我们这里的人也没那么聪明——你们最需要美国政府做的关键事情是什么,以帮助我们维持这种领先并主导这个领域?我认为这是我们必须做的。 奥特曼先生,还是你,快速说说。 So as I ask this question I will ask if you guys think we are ahead, but then the key things when you say we need your help what would--very succinctly, sometimes we are not so smart up here--what would the key things be that you would need from the U.S. Government to help us maintain that lead and dominate this space, which is what I think we need to do? Mr. Altman, again, to you real quick on that. ¶
altman 我们谈了一点基础设施,但我认为我们怎么强调它的重要性都不为过,以及把整条供应链、或尽可能多的部分放在美国的能力。之前的技术革命也关乎基础设施和供应链,但 AI 在我们需要的资源规模上有所不同。所以我们正在美国做的像「星际之门」这样的项目,比如把芯片制造、当然还有芯片设计带到美国,快速审批电力许可。比如,这些是至关重要的。如果我们这件事没做对,我认为我们做的其他任何事都帮不上忙。在模型创造方面,我们已经谈到了我们在训练能力上需要确定性、并与世界其他地方拥有公平竞争环境,以确保我们能保持竞争力。在合理、公平、轻量接触的监管框架下提供产品的能力,让我们能在市场中取胜,因为产品将是那种反馈循环以及让产品越来越好的关键,以及快速部署它们、在产品和模型、基础设施、数据领域之外在产品层面取胜的能力,真的相当重要。把世界上最有才华的人、最有才华的研究人员带到这里来的能力。我们在美国有大量这样的人。世界上还有更多。我们应该设法把他们全部带到这里,在这里改进模型。我认为这些就是一些具体内容。 We have talked a little about infrastructure but I think we cannot overstate how important that is, and the ability to have that whole supply chain or as much of it as possible in the United States. The previous technological revolutions have also been about infrastructure and the supply chain, but AI is different in terms of the magnitude of resources that we need. So projects like Stargate that we are doing in the U.S., things like bringing chip manufacturing, certainly, chip design to the U.S., permitting power quickly. Like, these are critical. If we do not get this right I do not think anything else we do can help. On the model creation side, we have talked about the need for certainty in our ability to train and to have fair footing with the rest of the world to make sure we can remain competitive. The ability to offer products under a reasonable, fair, light touch regulatory framework where we can go win in the market, because the products will be so key to the sort of feedback loops and making them better and better, and the ability to deploy them quickly and win at the product level in addition to the model and infrastructure and data area is really quite important. The ability to bring the most talented people in the world here, the most talented researchers. We have a ton in the United States. There are more out in the world. We should try to get them all here, improving models here. I think those are some of the specifics. ¶
Senator Sullivan 好。这非常有帮助。 让我问,史密斯先生,另外两个我想谈的点。我完全同意卢米斯参议员。 我相信克鲁兹参议员也有同样的看法。在我看来,我们相对于中国的比较优势之一必须是能源——「全都要」的能源。 希望你们看到了,在阿拉斯加我们有一个非常大规模的液化天然气(LNG)项目,我认为我们就要让它落地了,我们为此努力了很久。 我们将有可供应一百年的天然气。所以我们希望你们各位带着数据中心到阿拉斯加来。我们天气冷。我们有很多很冷的天气。我们有天然气。我们有土地。我们有水。我们什么都有。 Good. That is very helpful. Let me ask, Mr. Smith, two other ones that I want to touch on. I agree fully with Senator Lummis. I am sure Senator Cruz has the same view. One of our comparative advantages over China, in my view, has to be energy--all of the above energy. Hopefully, you have seen in Alaska we have a very large- scale LNG project that I think we are going to get off the ground here we have been working on for a long time. We will have a hundred years supply of natural gas. So we want you guys all to come up to Alaska with your data centers. We have got cold weather. We got a lot of cold weather. We got gas. We got land. We got water. We got it all. ¶
altman 那非常有说服力。 That is very compelling. ¶
Senator Sullivan 所以,是的,上来吧。这个项目完成后,一百年的天然气供应。比得克萨斯稍微冷一点。 那么两个与我们的比较优势有关的问题,史密斯先生,然后谁想加入都可以。 能源——我们认为是吗?我认为是。第二个,我认为在某种程度上是个劣势。这让我很沮丧。也许你们几位没看到这一点。 我们有一些美国金融公司、风险投资机构、银行等等,令人惊讶的是,我们在美国拥有的所有机会,反而在帮助资助中国的一些项目。 我一直是坚决反对那些有机会在其他地方投资的美国人去投资中国 AI、中国量子的,因为我们都知道他们会用这些让自己的军队更致命。我是说,他们就是这么干的。 我最近读到这个 Benchmark Capital。我不认识这些人,但他们显然为中国某家 AI 公司做了一轮 7500 万美元的融资。这是不是另一个问题,史密斯先生? 优势是能源,问题是美国公司在资助我们的竞争对手? So, yes, come on up. When this project is done, 100-year gas supply. A little colder than Texas. So two questions that relate to our comparative advantage, Mr. Smith, and then any others who want to jump in. Energy--do we think that is? I think it is. And then second it is, I think, somewhat of a disadvantage. It frustrates me. Maybe you guys do not see this. We have had American finance companies, venture capital firms, banks, others, that, remarkably, all the opportunities we have in America are helping fund some of these projects in China. I have been a real staunch opponent of Americans who have opportunities to invest in other places investing in Chinese AI, Chinese quantum, because we all know they are going to use that to help make their military more lethal. I mean, that is what they do. I was reading recently about this Benchmark Capital. I do not know these guys but they evidently did a $75 million round for some--an AI company in China. Is that another problem as well, Mr. Smith? Advantage energy problem--American companies financing our competition? ¶
Mr. Smith 我会把三件事连起来:能源、人才和资本的可获得性。美国在能源上有巨大资源,但永远不要低估中国建设大量发电厂的能力——也许比其他任何国家都更多、更快。 所以我们最好带着这样的心态去做:我们必须跟上,不能把任何事视为理所当然。但接着我要说,在定义数字技术的这 50 年里,美国头号比较优势一直是把世界上最好的人带到我们国家,并让他们获得风险资本;我们应当继续打磨这两者。 而且我认为你问「风险资本还流向哪里」是对的。我只想说这一点:如果我们能继续把最好的人带到美国,如果我们能继续在美国培养最好的人,我相信钱就会在这里,让他们成功。 但让我们确保我们继续把世界上最好的人带进来,给他们在美国建立伟大公司的机会。 I would connect three things: energy, people and access to capital. The U.S. has huge resources in energy, but never underestimate the ability of China to build a lot of electrical power plants, maybe more and faster than any other country. So we are better off going into that with the mindset that we have to keep up and not take anything for granted. But then I would say the number-one comparative advantage of the United States throughout the 50 years that have defined digital technology has been bringing the world's best people to our country and giving them access to venture capital, and we should continue to burnish both of those. And I think you are right to ask where else is venture capital going. I will just say this. If we can keep bringing the best people to the United States and if we can keep educating the best people in the United States, I believe the money will be here to enable them to succeed. But let us make sure we are continuing to bring the best people in the world and giving them the opportunity to build great companies here in the United States. ¶
Senator Sullivan 而美国的风险投资基金资助中国 AI,这符合我们的国家利益吗? And American venture capital funds funding Chinese AI, is that in our national interest? ¶
Mr. Smith 我认为「是否符合」这个问题确实很值得问,我也认识到你们各位非常正确地在关注这一点。 我只想不停地说:把人带到这里来。他们就能获得资金,我们就能在竞争中胜过世界。 I think there is a really good question about whether it is and I recognize that you all are quite rightly focused on that. I will just keep saying, bring the people here. They will have access to the money and we will outcompete the world. ¶
Senator Sullivan 很好。谢谢。谢谢您,主席先生。 **The Chairman.** 谢谢。 马基(Markey)参议员。 Great. Thank you. Thank you, Mr. Chairman. The Chairman. Thank you. Senator Markey. STATEMENT OF HON. EDWARD MARKEY, U.S. SENATOR FROM MASSACHUSETTS ¶
Senator Markey ### 尊敬的爱德华·马基(Edward Markey)陈述 ### 马萨诸塞州联邦参议员 非常感谢,主席先生。 我想谈谈人工智能的环境影响。人工智能可以通过改进天气预报、让我们更好地预测电力供需,来帮助我们应对气候变化。但设计、训练和部署 AI 模型,也对我们的环境构成真实风险。 对 AI 开发至关重要的那些巨型数据中心需要大量电力,给电网带来压力,并可能推高消费者的成本。 这些数据中心还会产生大量热量。为它们降温需要巨量水资源,而这些地区往往已经因气候变化面临干旱;一些数据中心还有现场备用柴油发电机,可能引发呼吸和心血管问题,并增加周边社区的癌症风险。 事实是,我们对 AI 的环境成本与环境收益都了解得太少。 史密斯先生,你是否同意:由政府开展一项关于人工智能环境影响的综合研究,会有帮助? Thank you, Mr. Chairman, very much. I would like to talk about the environmental impact of artificial intelligence. Artificial intelligence can help us combat climate change by improving weather forecasts and enabling us to better predict power supply and demand. But designing and training and deploying AI models also poses real risks for our environment. The massive data centers that are critical for AI development requires substantial amounts of electricity, putting stress on the grid and potentially raising costs for consumers. These data centers also generate significant heat. Cooling them requires huge volumes of water, often in regions already facing droughts because of climate change, and some data centers have onsite backup diesel generators, which can cause respiratory and cardiovascular issues and can increase the risk of cancer for the surrounding community. The truth is, we know too little about both the environmental costs and benefits of AI. Mr. Smith, do you agree that it would be helpful for the government to conduct a comprehensive study on environmental impact of artificial intelligence? ¶
Mr. Smith 总的来说,是的。去年十二月刚完成了一项研究,我认为值得定期更新。 Generally, yes. One study was just completed last December and I think it is worth updating periodically. ¶
Senator Markey 你是否认为由政府召集包括产业界和学术界在内的利益相关方,来帮助更好地衡量 AI 的环境影响,会有帮助? Do you think it would be helpful for the government to convene stakeholders including from industry and academia to help better measure AI's environmental impact? ¶
Mr. Smith 我认为,和许多其他需要被衡量的东西一样,是的。我认为这里有一个可以发挥的作用。 I think as well as many other things that need to be measured. Yes, I think there is a role to be played. ¶
Senator Markey 奥特曼先生,你是否同意联邦政府应当帮助研究和衡量 AI 的环境影响? Mr. Altman, do you agree that the Federal Government should help with studying and measuring the environmental impact of AI? ¶
altman 我认为研究和衡量通常是一件好事。我确实认为,关于 AI 对环境的影响、以及相对的挑战和益处的讨论,已经变得有些失衡。我希望 AI——你知道的,我们长期以来一直在尝试应对气候环境挑战,但没有成功,或者说不够成功。我认为我们需要帮助。我认为 AI 能帮助我们做到这一点。我们已经提出、或者我们正在建设一个 10 吉瓦(gigawatt)的设施,而且我们还有另一个——【译注:碎片话轮,原文以破折号结尾,句意未完;按源件边界保留】 I think studying and measuring is usually a good thing. I do think that the conversation about the environmental impact on--of AI and the relative challenges and benefits has gotten somewhat out of whack. I am hopeful that AI--you know, we have been trying to address climate environmental challenges unsuccessfully or not successfully enough for a long time. I think we need help. I think AI can help us do that. We have proposed or we are in the process of building a 10-gigawatt facility and we have got another---- ¶
Senator Markey 我的问题是:联邦政府是否应当持续不断地研究 AI 的影响? My question is should the Federal Government be on an ongoing basis studying the impact of AI? ¶
altman 当然,而且我认为你们应该用 AI 来帮忙。 Sure, and I think you should use AI to help. ¶
Senator Markey 所以本届国会提出了《人工智能环境影响法案》,来研究 AI 的正面与负面后果。 随着技术继续发展、模型变得更高效、随着我们建起基础设施,我们需要做这件事。 是的,AI 也许能找到治愈癌症的办法。也许能。但 AI 也可能助推一场气候灾难。这同样是真的。 所以我们需要把这两件事都摆在桌面上,特别是在特朗普政府无视这样一个事实的时候:去年美国所有新增装机发电容量中 94% 是风能、太阳能和电池,而特朗普说过他要摧毁让这一切继续下去的所有激励。 这是你们必须发声的事情,确保他不会那么做。所以我期待在这一点上与各位合作。 现在我想转到 AI 对弱势群体的影响。毕竟,我们谈的不只是用人工智能写邮件或规划购物清单。 我们谈的是用来计算一个家庭按揭、筛选一个人的求职申请、决定一位老人医疗照护的技术。 在这些情境中使用时,绝对关键的是:由 AI 驱动的算法不能带有偏见与歧视。所以我们从一个简单的问题开始。 史密斯先生,算法会有偏见并造成歧视吗? So that is why this Congress introduced the ``Artificial Intelligence Environmental Impact Act'' to study both the positive and negative consequences of AI. As the technology continues to develop, as models become more efficient, and as we build out the infrastructure, we need to do it. Yes, AI might find--may find a cure for cancer. It may, but AI also could help to contribute to a climate disaster. That is also equally true. So we need to just keep both of those things right on the table, especially as the Trump administration is ignoring the fact that last year 94 percent of all new installed electrical generation capacity in the United States was wind, solar, and battery, and Trump has said he is going to destroy all incentives for continuation of that. That is something you have to weigh in on to make sure he does not do that. So I look forward to working with you on that. Now I want to turn to AI's impact on disadvantaged communities. After all, we are not just talking about using artificial intelligence to write e-mails or plan grocery lists. We are talking about technology used to calculate a family's mortgage, screen an individual's job application, and determine a senior's medical care. When used in these situations it is absolutely essential that AI-powered algorithms are free from bias and discrimination. So let us start with a simple question. Mr. Smith, can algorithms be biased and cause discrimination? ¶
Mr. Smith 会,这正是我们要做测试以避免那种结果的原因。 They can, which is why we test to avoid that outcome. ¶
Senator Markey 同样的问题,奥特曼先生。算法会有偏见并造成歧视吗? Same question, Mr. Altman. Can algorithms be biased and cause discrimination? ¶
altman 当然。 Of course. ¶
Senator Markey 当然。奥特曼先生,OpenAI 是否努力防范 ChatGPT 中的这类偏见与歧视? Of course. Mr. Altman, does OpenAI work to guard against such bias and discrimination in ChatGPT? ¶
altman 当然。 Of course. ¶
Senator Markey 当然。所以我很高兴听到这一点,因为你最近说过,政府不应对 AI 实施隐私监管,而应当——引用原话——「在问题出现时非常迅速地作出反应」;而我对这种做法深感担忧。 我们不需要等着看那些测试与训练不足的算法是否会伤害边缘化群体。人工智能已经在加剧我们社会中普遍存在的偏见与歧视。有偏见、有歧视的算法,意味着黑人和棕色人种家庭更不容易获得按揭贷款。 它意味着残障人士更不容易被推荐到某个职位空缺,也意味着女性更不容易获得高等教育的奖学金。 这些是此刻正在发生的真实伤害。应对伴随 AI 快速开发与部署而来的这些既有问题,是国会的职责;这也是我作为《AI 民权法案》提出者而感到自豪的原因——该法案将确保公司在开发和部署算法之前审查并消除其中的偏见与歧视。 这必须同步发生,并且当公司的算法对边缘化人群造成伤害时,将追究其责任。 我将为确保 AI 不代表「加速我国的不平等」而奋斗。我们在现实世界中拥有的一切保护,都应当搬到虚拟世界,因为同样的歧视——再说一次,女性、黑人、棕色人种、残障社群、LGBTQ 社群——都会转移到线上,而我们必须在最初就把防范这种偏见的保护措施建进去;否则那些同样的歧视做法会立刻迁移过去。而产业界的责任,就是与国会合作,确保我们把这些保护写进法律。 谢谢您,主席先生。 **The Chairman.** 谢谢。 彼得斯(Peters)参议员。 Of course. So I am glad to hear that because you recently stated that the government should not implement privacy regulations on AI but instead, quote, ``respond very quickly as the problems emerge,'' and I am very deeply worried about that approach. We do not need to wait and see if poorly tested and trained algorithms will harm marginalized communities. Artificial intelligence is already supercharging the bias and discrimination prevalent in our society. Biased and discriminatory algorithms mean black and brown families are less likely to obtain a mortgage. It means people with disabilities are less likely to be recommended for a job opening and it means women are less likely to receive scholarships for higher education. These are real harms that are happening right now. It is Congress' job to address these existing problems that come with the rapid development and deployment of AI and it is why I am the proud author of the ``AI Civil Rights Act'' which would ensure that companies review and eliminate bias and discrimination in their algorithms before developing and deploying them. It has to happen simultaneously, and it will hold companies accountable when their algorithms cause harms against marginalized population. I will be fighting to ensure AI does not stand for accelerating inequality in our Nation. All of the protections we have in the real world should be moved to the virtual world because the same discrimination--again, women, black, brown, communities with disabilities, LGBTQ community--are going to move online and we have to build in the protections against that bias right up front, because otherwise those same discriminatory practices will just migrate immediately and the responsibility of the industry will be to work with Congress to make sure we put those protections on the books. Thank you, Mr. Chairman. The Chairman. Thank you. Senator Peters. STATEMENT OF HON. GARY PETERS, U.S. SENATOR FROM MICHIGAN ¶
Senator Peters ### 尊敬的加里·彼得斯(Gary Peters)陈述 ### 密歇根州联邦参议员 谢谢您,主席先生;也感谢我们所有的证人。感谢各位到场。 这是一个极其重要的话题,我们感谢你们的专业见解。 当我们着眼于确保美国成为 AI 的世界领导者时,当然,我们一直在谈供应链、基础设施以及所有那些方面。 但我想特别聚焦的一个领域是劳动力和人,确保我们有人才在那里。这就是我起草《AI 服务奖学金法案》和《AI 培训法案》的原因。 这两项都在 2022 年签署成为法律。今年早些时候,我又提出了《AI 与关键技术劳动力框架法案》,沿着这些方向继续努力;我很乐意与各位每一位合作,研究为确保我们拥有受过训练、能够利用这项了不起技术的劳动力所需的其他立法。 我要特别向密歇根大学致意——它实际上成为世界上第一所为全体学生提供生成式 AI 工具的大学,为他们进入未来的劳动力市场做准备。所以我想稍微谈谈劳动力。 奥特曼先生,去年你在我办公室会面、我们有一次很好的交谈,你说多达 70% 的工作岗位可能被 AI 消除,你也承认这可能导致的社会扰动。 如果那正在发生,我们就必须为此做准备。我们不会挡在这些不可思议的机会前面;但如果这确实将要发生,我们就必须相当深入地思考它将如何被管理,确保每个人都能从 AI 中受益,而不只是少数被选中的人受益。 所以请跟我谈谈,你认为你们行业的领袖可以如何帮助减轻失业,或者应对那种——正如你去年所描述的——重大社会扰动? Thank you, Mr. Chairman, and thanks to all our witnesses. Thank you for being here. It is an incredibly important topic and we appreciate your expertise. As we are looking at making sure that the United States is the world leader in AI, certainly, we have been talking about supply chains and infrastructure and all of those aspects. But one area that I want to particularly focus on is workforce and people to make sure that we have the talent there. That is why I authored the ``AI Scholarship for Service Act'' and the ``AI Training Act''. Both of those were signed into law in 2022. Earlier this year I introduced my ``AI and Critical Technology Workforce Framework Act'' to continue the effort along those lines, and love to work with each of you as we look at other legislation necessary to make sure we have got the workforce trained to take advantage of this amazing technology. I do want to do a shout out to the University of Michigan that actually became the first university in the world to provide generative AI tools for their entire student body to prepare them for the workforce of tomorrow. So I want to talk a little bit about the workforce. Mr. Altman, when met last year in my office and had a great conversation you said that upwards of 70 percent of jobs could be eliminated by AI and you acknowledge the possible social disruption of this. If that is happening we have to prepare for it. We are not going to stand in the way of the incredible opportunities here but if this is, indeed, going to occur, we have got to be thinking pretty deeply about how that will be managed and make sure that everybody can benefit from AI, not just a select few that benefit. So talk to me about how you believe leaders in your industry can help mitigate job losses or deal with what could-- as you described it last year, major social disruption? ¶
altman 我认为这次与之前的技术革命不同的是潜在的速度。技术革命长期以来一直影响着就业和经济。有些工作消失了。有些新工作被创造出来。许多工作只是变得更高效,人们能够做更多、赚更多钱、创造更多,这很棒。在一段时间里,社会能够适应大量的工作变化,你可以看看过去两个世纪,看看那发生了多少变化。我不知道。我不认为有人确切知道这会有多快,但感觉它可能相当快。我认为我们能做的最重要的事情之一、或者最重要的事情之一,就是尽早把工具交到人们手中。我们有一个原则,叫作「迭代式部署」(iterative deployment)。我们希望人们在技术被开发出来的过程中就逐渐习惯它。自我们第一次发布产品以来,我们已经这样做近五年了。随着社会和这项技术共同演进,把优秀、强大的工具交到很多人手中,让他们弄清楚他们将用这些工具做什么、为彼此创造什么,并在这个我们拥有的新构建模块和社会的脚手架之上,想出并向社会提供价值——我认为,作为 OpenAI、作为我们的行业,这是我们为帮助平稳过渡所能做的最好的事情。 The thing that I think is different this time than previous technological revolutions is the potential speed. Technological revolutions have impacted jobs and the economy for a long time. Some jobs go away. Some new jobs get created. Many jobs just get more efficient and people are able to do more and earn more money and create more and that is great. Over some period of time society can adapt to a huge amount of job change, and you can look at the last couple of centuries and see how much that has happened. I do not know. I do not think anyone knows exactly how fast this is going to go, but it feels like it could be pretty fast. The most important thing or one of the most important things, I think, we can do is to put tools in the hands of people early. We have a principle that we call ``iterative deployment''. We want people to be getting used to this technology as it is developed. We have been doing this now for almost five years since our first product launch. As society and this technology co-evolve putting great, capable tools in the hands of a lot of people and letting them figure out the new things that they are going to do and create for each other and come up with and provide sort of value back to the world on top of this new building block we have and the sort of scaffolding of society that is, I think, the best thing we can do as OpenAI and as our industry to be--sort of help smooth this transition. ¶
Senator Peters 我们的想法是要走到这样一个境地:AI 不是在取代工作,而是在增强工作;人们更有生产力,去做一些我们大概连想象都想象不到人们会做的事。如果回看一百年前,我们有一些工作岗位,没有人—— The idea we want to get to the point where AI is not displacing work but actually enhancing work, that people are more productive and doing things that we probably cannot even imagine what people will do. If we would look a hundred years ago we have jobs that no one---- ¶
altman 你无法想象,我也不认为我们能想象这一另一端的工作是什么样子。但即使你看看今天编程领域正在发生的事——我会选这个,因为它多少是我的背景、也是我心之所系——在 2025 年 5 月,作为一名程序员、作为一名有效率的程序员意味着什么,与我上次在这里的 2023 年 5 月截然不同。这些工具真的改变了程序员有能力做到的事,改变了世界将获得的代码和软件的数量。这也不是说人们不再雇用软件工程师了。他们以不同的方式工作,而且生产力高得多。 You cannot imagine, and I do not think we can imagine the jobs on the other side of this. But even if you look today at what is happening with programming, which I will pick because it is sort of my background and near and dear to my heart, what it means to be a programmer and an effective programmer in May 2025 is very different than what it meant last time I was here in May 2023. These tools have really changed what a programmer is capable of, the amount of code and software that the world is going to get. And it is not like people do not hire software engineers anymore. They work in a different way and they are way more productive. ¶
Senator Peters 对。对。 苏博士,我们当然谈了很多开源 AI,但大部分讨论是关于软件的。然而,让技术开放、能够协同工作,在每一个层面都重要——如你所知,从为设备提供动力的芯片,到在后台运行的服务器。 所以我问你的问题是:在硬件层面而非软件层面,开放标准和系统互操作性有什么好处?它对创新、国家安全以及供应链韧性又意味着什么? Right. Right. Dr. Su, we certainly talk a lot about open source AI but most of the conversation has been about software. However, making technology open and able to work together matters at every level, as you know, from chips that power the devices to the servers that are running behind the scenes. So my question for you is, what are the benefits of open standards and system interoperability at the hardware level, not the software level, and what are the implications for innovation, national security, as well as resilience in the supply chain? ¶
Ms. Su 谢谢您的问题,参议员。 我认为,在硬件、软件和应用层面拥有一个开放生态,有着极其多的优势。 道理是——您知道——没有任何一个组织或群体拥有全部好点子;因此,让生态协同运作,使你能够在每一层选择最佳方案,并且在一大批参与方之间实现优化,这是好事。 我认为从安全角度看也非常好:确保——您知道,再说一次——有很多选择,这样我们就不依赖单一生态。所以,您知道,我们在开放标准和开放生态上会继续保持非常前瞻的思路。 Thank you for the question, Senator. I think there are an incredible number of advantages to having an open ecosystem at the hardware and the software and the application level. The idea is, you know, there is no one organization or one group that has all the good ideas and so enabling the ecosystem to work together so that you can choose the best solution at every level and then also optimization across a broad set of constituents is a good thing. I think it is also very good from a security standpoint to ensure that, you know, again, there are many choices so that we are not dependent on a single ecosystem. So, you know, we continue to be very forward thinking in open standards as well as open ecosystems. ¶
Senator Peters 那么你们的模式是开放模式。我理解英伟达(Nvidia)是封闭模式。那——优势与劣势各是什么?我们应当考虑什么? So your model is an open model. I understand Nvidia is a closed model. Is there--what are the advantages and disadvantages? What should we be thinking about? ¶
Ms. Su 我认为开放模式的主要优势——也是我们非常支持的一点——是这样一种理念:创新可以来自许多不同的方面;您知道,无论是硬件创新,也就是不同的芯片,还是把这一切组合起来的系统创新。 而且,您知道,我们的目标是确保我们始终拥有最好的中之最好,而做到这一点有很多不同的方式、很多不同的参与方可以贡献;这就是为什么我们在开放生态上非常积极进取。 I think the major advantage in an open model, and that is something that we very much support, is the idea that we can have innovation come from many different parties and, you know, whether that is hardware innovation so on the different chips or that is system innovation on putting all these things together. And, you know, our goal is to make sure that we always have the best of the best and there are many different ways--many different parties that can contribute to that and that is why we are very forward leaning in terms of open ecosystems. ¶
Senator Peters 很好。谢谢。谢谢您,主席先生。 **The Chairman.** 谢谢。 费特曼(Fetterman)参议员。 Great. Thank you. Thank you, Mr. Chairman. The Chairman. Thank you. Senator Fetterman. STATEMENT OF HON. JOHN FETTERMAN, U.S. SENATOR FROM PENNSYLVANIA ¶
Senator Fetterman ### 尊敬的约翰·费特曼(John Fetterman)陈述 ### 宾夕法尼亚州联邦参议员 谢谢您,主席先生。你好。 史密斯先生,我是能源的坚定支持者。对我来说,能源安全就是国家安全;当然,您知道,可再生能源关乎这一点。但当然,其他东西也一样——化石能源。而且当然也包括核能。核能很重要。 现在还有那种能源转型。我的关注点还在于,我要确保宾夕法尼亚州的电费缴纳者在这一切过程中不会被打得太重。 《华盛顿邮报》报道说,数据中心不断增长的电力需求将推高居民电费,也许高达 20%。 对我来说,这确实是个顾虑,对宾夕法尼亚的家庭当然也是。现在,数据中心在建设期间有重要的就业岗位,做那些事情,那当然是好事。 但我想它们不是长期的。而那些电价——那些费率可能持续更久。 还有,我一直在非常关注重启三里岛(TMI)的计划。我是说,我有个人的故事:1979 年那次熔毁时,我得抓上我的仓鼠撤离。 你可能会以为我反核,但并不是——实际上我非常支持核能,因为那是能源结构中的重要部分。如果你真的想应对气候变化,在我看来,你不能背对核能。 但我知道那是核电——微软的数据中心,所以现在——我真的很感谢这一点。但我要说的是:如果你能承诺那份购电协议不会推高宾夕法尼亚家庭的电价。 Thank you, Mr. Chairman. Hello. Mr. Smith, I am a big supporter of energy. For me energy security is national security and, of course, you know, renewables is about that. But, of course, other things as well, too--fossil. But also that also includes nuclear, of course. Nuclear is important. And now then there is that kind of energy transition. My focus is also that I want to make sure that rate payers in Pennsylvania really are not hit too hard for throughout all of this. Now, the Washington Post reported that increasing electricity demand for the data centers is going to raise up residential power bills, perhaps, as much by 20 percent. Now, to me, that is really a concern for me and certainly for Pennsylvania families. Now, the data center, you know, has important jobs during construction and doing those things and that is a great thing, of course. But they are not, I guess, long term. But the rate--those rates might last longer for that. And now, I have been very--tracking the plan to reopen TMI. I mean, I had my own personal story is I had to grab my hamster and evacuate, you know, in that--during the meltdown in 1979. You might consume--you might assume that I was anti-nuclear and that is not--it is a--I actually am very supportive of nuclear because that is an important part of the stack. If you really want to have--address climate change you cannot turn your back on nuclear, in my opinion. But I know that is the power nuke--Microsoft's data center so now--and I really appreciate that. But if I am saying, now, if you are able to commit that the power purchase agreement, you know, it is not going to raise electricity for Pennsylvania families. ¶
Mr. Smith 不,我认为你提出了一个关键点。我们在建设这些数据中心时遵循两条原则。 第一,我们会投资向电网注入与我们用电量相等的电力,这样我们就不会去抽取一个受制约的供应。 第二,我们会以这样的方式管理这一切:确保我们的活动不会推高社区的电价。 所以我刚才描述过:如果电网需要做改进——而往往需要——我们会去找公用事业委员会。我们会提出调整向我们收取的费率,好让我们为那项改进付费。 我只是认为这是现实;因为我认为你强调了关键的一点:建设期间有很多岗位,之后也有岗位,但没那么多了。 如果我们实际上是在向邻里征税——因为我们的到来而让每个人多付电费——那我们很快就会把好感耗尽。我们明白。我们知道我们必须是这个社区里良好而负责任的成员。 No, I think you raise a critical point. We have two principles that we follow when we are constructing these data centers. Number one, we will invest to bring onto the grid an amount of electricity that equals the amount of electricity that we will use so that we are not tapping a constricted supply. Number two, we will manage all of this in a way that ensures that our activity does not raise the price of electricity to the community. And so I was describing earlier how if there is improvements that need to be made to the grid, as there often are, we will go to the utility commission. We will propose a change in the rate that we are charged so that we can pay for that improvement. I just think it is a fact of life because I think you highlight something critical. There are a lot of jobs when the construction takes place. There are jobs afterwards but they are not as many. One will wear out the welcome quickly if we tax, in effect, the neighborhood by asking everyone to pay more for their electricity because we have arrived. We get it. We know we have to be a good and responsible member of the neighborhood. ¶
Senator Fetterman 现在,您知道,当参议员的好处之一是——至少对我来说——我有机会见到那些职业或事业比我更了不起的人。 而且,奥特曼先生,这将——我要把这算作一个亮点。 最近,我知道你所做的工作,你真的是推动 AI 的人之一,而现在这是一个机会。见到你我很兴奋。 而现在,人们——您知道,人们问我,比如说,如果你要谈 AI,那现在我可以问你——我是说, literally 的专家。 您知道,有些人担心 AI 或者别的什么,而我就说,那「奇点」呢,您知道,那类人。 如果你能谈谈那个,请讲。 Now, you know, one of the perks of being a senator is that--for me, anyway, I get an opportunity to meet people that have much more impressive kinds of jobs or careers that I have led. And, now, Mr. Altman, now, this is going to--I am going to count this as a highlight. Recently, like, I know the work that you have done you are really one of the people that are moving AI and now it is an opportunity. I was excited to meet you. And now, people--you know, people ask me it is, like, if you are going to talk about AI and now I get to ask you, I mean, like, the literal--the expert. You know, some people are worried about AI or whatever and I am, like, you know, what about the singularity so, you know, the people like that. If you would address that, please. ¶
altman 谢谢您,参议员,感谢您友善的话语,也感谢您在更多场合让连帽衫变得被接受。我很高兴看到这一点。我对进步的速度感到无比兴奋,但我也很谨慎,而且我会说,比如,我不知道——我觉得在它面前我很渺小之类的。我认为这超出了我们所有人目前完全理解它走向的范围。我相信,这属于人类有史以来将产生的最大——也许仍在努力成为最大的技术革命之列,我感到很荣幸能在这里。我对将要发生的事感到好奇和感兴趣。但我确实认为事情将发生相当重大的变化。我认为人类有一种奇妙的适应能力,那些看起来惊人的东西会很快变成新常态。我们将想出如何用这些工具去做我们以前永远做不到的事情,我认为那将非常非凡。但这些工具将能够做到我们还无法完全理解的事情,有些人称之为——你知道的,随着这些工具开始帮助我们创造下一个和未来的迭代版本,有些人称之为奇点(singularity)。有些人称之为起飞(take off)。不管它是什么,感觉就像是人类历史的一个新时代,我认为我们能亲历其中、并且能让它成为一件美好的事情,这非常令人兴奋。但我们必须以谦逊和某种谨慎来对待它。 Thank you, Senator, for the kind words and for normalizing hoodies in more spaces. I love to see that. I am incredibly excited about the rate of progress but I also am cautious, and I would say, like, I do not know--I feel small next to it or something. I think this is beyond something that we all fully yet understand where it is going to go. This is, I believe, among the biggest--maybe still trying to be the biggest technological revolutions humanity will have ever produced and I feel privileged to be here. I feel curious and interested in what is going to happen. But I do think things are going to change quite substantially. I think humans have a wonderful ability to adapt and things that seem amazing will become the new normal very quickly. We will figure out how to use these tools to just do things we could never do before and I think it will be quite extraordinary. But these are going to be tools that are capable of things that we cannot quite wrap our heads around, and some people call that--you know, as these tools start helping us to create next and future iterations some people call that singularity. Some people call that the take off. Whatever it is, it feels like a sort of new era of human history and I think it is tremendously exciting that we get to live through that and we can make it a wonderful thing. But we have got to approach it with humility and some caution. ¶
Mr. Fetterman 我是说,我刚才——对我来说,这一直是——我也有机会向很多爱迪生式的人物提问。 你们大家共同参与的那些事,将改造我们的社会;人们回望 50、60 年前,会看到发生了什么。所以对我来说——交给主席。谢谢。 **The Chairman.** 谢谢,费特曼参议员。 克洛布查尔参议员。 I mean, I just did--for me, it has been--I get a chance to ask questions to a lot of Edisons as well, too. The kinds of things that you are all collectively involved are going to transform our society, and people will look back 50, 60 years ago and see what has happened. So to me, over to the Chairman. Thank you. The Chairman. Thank you, Senator Fetterman. Senator Klobuchar. ¶
Senator Klobuchar 谢谢。费特曼参议员,好想法。谢谢。 你们几位在这儿坐得太久了,教皇都已经选出来了。 〔笑声。〕 Thank you. Good thought, Senator Fetterman. Thank you. So you guys have been sitting here so long that the Pope has been chosen. [Laughter.] ¶
Senator Klobuchar 我们还不知道是谁。 **The Chairman.** 恭喜你,艾米。 We do not know who. The Chairman. Congratulations, Amy. ¶
Senator Klobuchar 白烟已经升起来了。 **The Chairman.** 恭喜。 〔笑声。〕 The white smoke has come up. The Chairman. Congratulations. [Laughter.] ¶
Senator Klobuchar 不用客气。那样大概行不通。 但无论如何,那是——我离开去处理别的事,又回来了,因为我还有一个问题想问,它跟整个深度伪造问题有关,因为布莱克本参议员、库恩斯参议员、蒂利斯参议员和我在这上面非常努力,而且他们——布莱克本和库恩斯是那项法案的牵头人。 但我们最近看到了深度伪造视频:阿尔·罗克(Al Roker)在推销一种高血压疗法;一个布拉德·皮特(Brad Pitt)的深度伪造,躺在病床上要钱。索尼音乐已与各平台合作,下架了 7.5 万多首含未经授权深度伪造的歌曲,包括哈里·斯泰尔斯(Harry Styles)和碧昂丝(Beyoncé)的声音。 我最近——我是说,这不只是名人。明尼苏达有一位获格莱美提名的艺术家。我和他谈过数字复制品的情况。所以这是一个真实的担忧,它某种程度上触及了我和沙茨参议员早些时候谈新闻法案时说的东西。 但我只是想让各位了解这项立法,因为在这件事上有过一些分歧,而我们现在已经建立起一个联盟,包括 YouTube 在内都支持它,还有美国唱片业协会、美国电影协会、演员工会-美国电视和广播艺人联合会(SAG-AFTRA)。所以这是件大事,我希望各位都会去看看。 但能否请你们评论一下?我先问你,史密斯先生,关于保护人们不被人未经许可用 AI 复制其肖像;即便你们各位都承诺会做,我们明显的担忧是,可能还有其他公司不会那么做——这就是为什么我认为,在我们审视这些防护栏时,对数字——人们的数字权利的保护应当是其中一部分。 史密斯先生。 You are welcome. Probably would not work. But in any case, it was--I left for some other things, came back because I had one more question that I wanted to ask and it is related to just the whole deep fake issue just because Senator Blackburn and Senator Coons and Senator Tillis and I worked on this really hard, and they are--Blackburn and Coons are in the lead of the bill. But we have recently seen deep fake videos of Al Roker promoting a cure for high blood pressure, a deep fake of Brad Pitt asking for money from a hospital bed. Sony Music has worked with platforms to remove more than 75,000 songs with unauthorized deep fakes including voices of Harry Styles and Beyonce. I recently--I mean, it is not just famous people. There is a Grammy-nominated artist from Minnesota. Talked to him about what is going on with digital replicas. So there is a real concern and it kind of gets at what Senator Schatz and I were talking about earlier with the news bill. But I just wanted to make you all aware of this legislation because there were some differences on this and now we have gotten a coalition, including YouTube, supporting it as well as the Recording Industry Association, Motion Picture Association, SAG-AFTRA. So it is a big deal and I am hoping it is something that you will all look at. But could you just comment? I would go to you, Mr. Smith, first about protecting people from having their likenesses replicated through AI without permission, and even if you all pledge to do it our obvious concern is that there will may be other companies that would not and that is why I think as we look at what these guardrails are the protection of digital-- people's digital rights should be part of this. Mr. Smith. ¶
Mr. Smith 是的。不,我认为你指出这一点是对的。它已成为一个日益受关注的领域。您知道,去年总统大选期间,两边的竞选团队、两个政党都担心深度伪造可能被制造出来。 我们与两边的竞选团队和两个政党都合作应对这件事。我们看到它被以——真的,我会称之为滥用的方式使用,包括针对名人等等。 我认为这始于一种能力:识别某个东西是由 AI 生成的,而不是真实的、比如说,照片或视频影像;而我们确实发现,AI 在这件事上的能力比人眼和人的判断强得多。 我认为应当有某些防护栏,其中一些我们可以自愿适用。我们一直在整个行业这样做。 OpenAI 和微软去年都参与其中;有些用途大概应当被视为越线,因而应当是非法的,我认为这正是你所描述的那类倡议可以发挥特别重要作用的地方。 Yes. No, I think you are right to point to it. It has become a growing area of concern. You know, during the Presidential election last year both campaigns, both political parties, were concerned about the potential for deep fakes to be created. We worked with both campaigns and both parties to address that. We see it being used in, really, ways that I would call abusive including of celebrities and the like. I think it starts with an ability to identify when something has been created by AI and is not a genuine, say, photographic or video image, and we do find that AI is much more capable at doing that than, say, the human eye and human judgment. I think it is right that there be certain guardrails and some of these we can apply voluntarily. We have been doing that across the industry. OpenAI and Microsoft were both part of that last year, and there are certain uses that probably should be considered across the line and, therefore, should be unlawful, and I think that is where the kinds of initiatives that you are describing have a particularly important role to play. ¶
Senator Klobuchar 那你能看看那项立法吗? And could you look at that legislation? ¶
Mr. Smith 绝对可以。 Absolutely. ¶
Senator Klobuchar 很感谢。奥特曼先生,同样的问题,同样的内容。 Appreciate it. Mr. Altman, just the same question, same thing. ¶
altman 当然,我们很乐意审视这项立法。我认为这是一个大问题,而且是一个很快就要到来的问题。我不相信——我认为有几个方面着手。你可以谈论生成内容的 AI、分发内容的平台、下架(takedowns)如何运作、我们如何教育社会、以及我们如何内置鲁棒性以预期这将会发生。我不相信有可能阻止内容的生成。我认为开源、开放权重(open weight)模型总体上是一件好事,是我们需要追求的东西。但这确实意味着将会有大量这类模型四处流通、能够做这种事。至于大规模分发,我认为有可能设置更多护栏,这似乎很重要,但我不想忽视社会教育那一块。我认为每一项新技术,几乎总是伴随着某种新的骗局出现。我们越早让人们理解这些、对它们保持警惕、把它当作一件将要到来、然后正在发生的事情来谈论,我认为越好。人们正在非常快地理解内容可以是 AI 生成的,并在他们自己的头脑中建立新的防御机制。但尽管如此,你知道的,如果你接到一个电话,听起来完全像一个你认识的人,而且他们很慌张、需要帮助,或者如果你看到一段视频——就像你说的那些视频,这以一种非常深的心理方式击中我们,我认为我们需要建立社会韧性,因为这就要来了。 Of course, we would be happy to look at the legislation. I think this is a big issue and it is one coming quickly. I do not believe--I think there are a few areas to attack it. You can talk about AI that generates content, platforms that distribute it, how takedowns work, how we educate society, and how we build in robustness to expect this is going to happen. I do not believe it will be possible to stop the generation of the content. I think open source, open weight models are a great thing on the whole and something we need to pursue. But it does mean that there is going to be just a lot of these models floating around that can do this. The mass distribution, I think it is possible to put some more guardrails in place and that seems important but I do not want to neglect the sort of societal education piece. I think with every new technology there is some sort of-almost always some sort of new scams that come. The sooner we can get people to understand these, be on the lookout for them, talk about this as a thing that is coming and then a thing that is happening I think the better. People are very quickly understanding that content can be AI generated and building new kinds of defenses in their own minds about it. But still, you know, if you get a call and it sounds exactly like someone you know and they are panicked and they need help, or if you see a video that--like the videos you talked about, this, like, gets at us in a very deep psychological way and I think we need to build societal resilience because this is coming. ¶
Senator Klobuchar 嗯哼。它正在到来,但我们可以——总得有些办法保护人们的隐私权—— Mmm-hmm. It is coming, but we can there--there has got to be some ways to protect people's privacy rights---- ¶
altman 我们当然应该做一切——一定的。 We should do everything--for sure. ¶
altman 绝对应该。我们应该拥有所有这些。坏行为者仍然不总是遵守法律,所以我认为我们需要一个额外的盾牌,或者在任何我们能拥有它的时候。但是是的,我们绝对应该保护那一点。 Absolutely. We should have all of that. Bad actors still do not always follow the laws and so I think we need an additional shield or whenever we can have them. But yes, we should absolutely protect that. ¶
Senator Klobuchar 好。期待与你在这件事上合作。谢谢。 **The Chairman.** 那我得说,克洛布查尔参议员关于伪造和 AI 伪造的问题让我有点内疚,因为我确实发过一张 AI 生成的图片,把费特曼参议员画成格陵兰的教皇。所以我确实干了这事,尽管它可能不算伪造。它可能是真的。 All right. Look forward to working with you on it. Thank you. The Chairman. So I have to say Senator Klobuchar's question about fakes and AI fakes made me feel guilty because I did, in fact, tweet out an AI-generated picture of Senator Fetterman as the Pope of Greenland. So I am guilty of doing so, although it may not be a fake. It may be a real thing. ¶
Senator Klobuchar 好。哦,哇,法律允许戏仿。戏仿是允许的。那和我说的不一样,但费特曼参议员应该回应一下。 OK. Oh, whoa, parody is allowed under the law. Parody is allowed. That is different than what I am talking about but Senator Fetterman should respond. ¶
Senator Fetterman 或者那到底是不是 AI。 Or if it was AI. ¶
Senator Klobuchar 我知道。 **The Chairman.** 它可能——其实那张照片拍得不错。 好。我还有几个问题,然后就收尾。 奥特曼先生,你见过的 ChatGPT 最令人惊讶的用途是什么?你看到哪些应用是令人惊讶的? I know. The Chairman. It may be--it is a good shot, actually. All right. I have a few more questions and then we will wrap up. Mr. Altman, what has been the most surprising use for ChatGPT you have seen? What are applications that you are seeing that are surprising? ¶
altman 人们每天给 ChatGPT 发送数十亿条消息,所以他们用它做各种极具创造力的事情。我讲一个个人的故事,正如前面提到的,我最近有了一个新生儿。显然人们以前也这么做,但我不知道人们以前是如何在没有 ChatGPT 的情况下弄清楚如何照顾新生儿的。那一直是一个真正的救命稻草。主席:所以我给你讲一个我以前给你讲过的故事,但我十几岁的女儿几个月前给我发了这条长长的、详细的短信,里面充满感情,而且写得真的很好,我实际上还评论了。我说,哇,这写得真好。她说,哦,我用 ChatGPT 写的。比如,等等,你在给你爸爸发短信,你不——这就是新一代的某种特点,它如此无缝地融入生活,以至于她在发一封电子邮件、或者在做任何事情时,她甚至——甚至毫不犹豫地想到去找 ChatGPT 来捕捉她的想法。 People message ChatGPT billions of times per day so they use it for all sorts of incredibly creative things. I will tell one personal story, which as mentioned earlier I recently had a newborn. Clearly, people did it but I do not know how people figured out how to take care of newborns without ChatGPT. That has been a real lifesaver. The Chairman. So I will tell you a story that I have told you before but my teenage daughter several months ago sent me this long, detailed text, and it was emotional and it was it was really well written and I actually commented. I am, like, wow, this is really well written. She said, oh, I used ChatGPT to write it. Like, wait, you are texting your dad and you do not--it is something about the new generation that it is so seamlessly integrated into life that she is sending an e-mail, she is doing whatever, and she does not even--does not even hesitate to think about going to ChatGPT to capture her thoughts. ¶
altman 我对此心情复杂。[笑声] 主席:嗯,用这个应用,然后告诉我你的感受。 I have complicated feelings about that. [Laughter.] The Chairman. Well, use the app and then tell me what your feelings are. ¶
altman 好的。主席:Google 刚刚透露,他们在 Safari 上的搜索流量有史以来首次下降。 OK. The Chairman. Google just revealed that their search traffic on Safari declined for the first time ever. ¶
altman 它没有给我寄圣诞卡。主席:ChatGPT 会取代 Google 成为主要搜索引擎吗?如果是,什么时候? It did not send me a Christmas card. The Chairman. Will ChatGPT replace Google as the primary search engine, and if so, when? ¶
altman 可能不会。我是说,我认为人们如今使用搜索引擎的某些用例,肯定用像 ChatGPT 这样的服务做得更好,但 Google 是一个凶猛的竞争对手。他们有一支非常强的 AI 团队、大量基础设施、一个受到很好保护的生意,而且他们在把 AI 融入搜索方面正取得很大进展。主席:好的。那么,这是一个我花了很多时间与技术领域的商界领袖、CEO 们、AI 界谈论的问题,而我问的这个问题得到的答案各不相同——我很好奇你们四位怎么说——DeepSeek 有多重要?它是来自中国的一个重大的、震撼性的、令人震惊的进展吗?还是没那么重要?还是介于两者之间,接下来会发生什么?让我们四位各自谈谈。 Probably not. I mean, I think some use cases that people use search engines for today are definitely better done on a service like ChatGPT, but Google is like a ferocious competitor. They have a very strong AI team, a lot of infrastructure, a very well-protected business, and they are making great progress putting AI into their search. The Chairman. All right. So a question that I have spent a lot of time talking to business leaders, CEOs in the tech space, AI, and one question that I have asked that I get different answers on--and I am curious what the four of you say--how big a deal was DeepSeek? Is it a major, seismic, shocking development from China? Is it not that big a deal? Is it somewhere in between and what is coming next? And let us each of the four of you. ¶
altman 不是什么大事。关于 DeepSeek 有两点。一是他们做了一个很好的开源模型,另一是他们做了一个消费者应用,它第一次短暂地超过了 ChatGPT,成为下载量最大的 AI 工具,也许是下载量最大的应用。将会有很多好的开源模型,显然,有非常有才华的人在 DeepSeek 工作、在做很棒的研究。所以我预计会有更多优秀的模型出现。希望我们、以及我们的一些同道也会发布优秀的模型。至于消费者应用,我认为如果一个——如果 DeepSeek 的消费者应用看起来要击败 ChatGPT 和我们美国同道的应用、成为人们使用的默认 AI 系统,那会很糟糕。但就我们目前看来,事情并非如此。 Not a huge deal. There are two things about DeepSeek. One is that they made a good open source model and the other is that they made a consumer app that, for the first time, briefly surpassed ChatGPT as the most downloaded AI tool, maybe the most downloaded app overall. There are going to be a lot of good open source models and, clearly, there are incredibly talented people working at DeepSeek doing great research. So I would expect more great models to come. Hopefully, also us and some of our colleagues will put out great models too. On the consumer app, I think if a--if the DeepSeek consumer app looked like it was going to beat ChatGPT and our American colleagues' apps as sort of the default AI systems that people use that would be bad. But that does not currently look to us like what is happening. ¶
Ms. Su 我要说,它介于两者之间,克鲁兹主席。当你想想我们学到了什么,我们学到的是——您知道——做事有很多不同的方式。 所以我们美国有很多极其有创新精神的人。美国的模型显然是目前为止最好的。然而,当你受到约束时,做事就有别的办法,我认为我们在这个过程中学到了一些东西。 我认为 DeepSeek 的开源性质,大概是影响最大的一点之一,就开源型模型和开放生态能做到多少而言。 但显然,美国在领先;正如我们所说,我们需要继续加速创新和采用——正如您以这个问题开场时所说。 I would say it is somewhere in between, Chairman Cruz. When you think about what we have learned, what we learned is, you know, there are different ways of doing things. So we have lots of incredibly innovative people in the United States. American models are, clearly, the best by far. However, when you have constraints that are placed there are other ways of doing things and I think we learned a few things in the process. I think the open source nature of DeepSeek was one of the things that probably was most impactful in just terms of how much can be done in an open source type of model and open ecosystem. But, clearly, the United States is leading and we need to continue, as we have said, to accelerate innovation and adoption as you started this hearing with. ¶
Mr. Intrator 我认为 DeepSeek 做了很多事。它做的其中一件事,是把中国 AI 能力的幽灵,抬到了一个比此前关注它的人广泛得多的受众面前,对吧;于是你看到那种震动在金融市场里回响。 你看到一种广泛基础的反应,突然间人人都知道 DeepSeek 是什么,也知道中国并非理论上在参与 AI 主导权竞赛,而是实际上是一个非常可畏的竞争者。 所以,您知道,它在某些方面是给更广泛人群的一记发令枪,也许让更广泛的意识明白:这不是既成事实,我们将必须作为美国共同协作,把我们的解决方案向前推进。所以我认为那将是我们会看到的持久影响之一。 I think DeepSeek did a lot of things. One of the things that it did was it sort of raised the specter of China's AI capability to a much broader audience than was perhaps focused on it prior to that, right, and so you saw that kind of reverberate through the financial markets. You saw, like, a broad-based reaction and suddenly everyone knows what DeepSeek is and the fact that China is not theoretically in the race for AI dominance but actually is very much a formidable competitor. And so, you know, it was a starting gun in some ways for the broader population and kind of maybe the broader consciousness of the fact that that this is not a fait accompli and that we are going to have to work as America together to kind of propel our solutions forward. And so I think that was one of the lasting impacts that we will see from that. ¶
Mr. Smith 我要像莉萨那样说,它介于两者之间。它并不令人震惊。我是说,它只是我们在中国跟踪的众多初创公司之一,我们此前就认为它有潜力在这个领域创新。 我确实认为有一个非常有意思也很重要的点:约束会以其他方式激励创新。而我认为关于 DeepSeek 的一个有趣事实是:在他们——比如说——200 多名员工中(发布这些模型时他们就是这个规模),几乎所有员工按设计都是大学毕业四年以内。 他们想招的是不会把传统的做事方式带进工作的人。 **The Chairman.** 所以是孩子们在接管世界? I would say like Lisa that it was somewhere in between. It was not shocking. I mean, it was one of a number of startups that we were following in China that we saw as having the potential to be innovative in this space. I do think there is a really interesting and important point that constraints encourage innovation in other ways and I just think one of the interesting facts about DeepSeek is that of their, say, 200 or more employees--that was the their size when they released these models--almost all of their employees by design were 4 years or less out of university. They wanted to hire people that would not bring to their work traditional ways of doing things. The Chairman. So the kids are taking over the world? ¶
Mr. Smith 每一代都这样。 〔笑声。〕 **The Chairman.** 与此相关——史密斯先生,你说完了吗? 与此相关,我们一开始谈到 AI 扩散规则被撤销,我很高兴。我认为那是一条糟糕的规则。我认为它过于复杂。我认为它给我们的一些贸易伙伴强加了不公平的限制,所以我很高兴总统正在撤销它。 但这并不必然意味着就不该有任何限制;关于 AI 扩散该有怎样的规则,存在各种各样的看法。 英伟达主张我们想让美国芯片到处都是,甚至在中国。另一些人主张至少应当限制最先进的处理器。 我很好奇——你们四位各自认为,如果要有东西取代 AI 扩散规则,那条规则应该是什么? 奥特曼先生,从你开始。 They do every generation. [Laughter.] The Chairman. Related to that--were you finished with that, Mr. Smith? Related to that, we talked at the outset about the AI diffusion rule being rescinded, which I am glad. I think it was a bad rule. I think it was overly complex. I think it put on a number of our trading partners unfair restrictions and so I am glad the President is rescinding it. That does not necessarily mean that there should be no restrictions and there are a variety of views on whether--what the rules should be concerning AI diffusion. Nvidia has argued that we want American chips everywhere, even in China. Others have argued that we want to restrict at least the most advanced processors. I am curious--each of the four of you what do you think the rule should be if anything is to replace the AI diffusion rule? And, Mr. Altman, we will start with you. ¶
altman 我也很高兴看到那项被撤销。我同意将需要一些约束。但我认为,如果我们的——如果那种思维模式是赢得扩散(winning diffusion)而不是阻止扩散,那在方向上似乎是对的。这并不意味着没有护栏。这并不意味着我们说,比如,我们要去某个其他国家建一个比美国更大的数据中心。我们的意图是在美国建设我们最大、最好的数据中心。在美国做训练。在这里造模型。把我们的核心研究放在这里。但我们确实想和世界各地的伙伴一起建设推理中心,我们一直在与政府在这方面合作。我认为那会是好事。在这一点上,影响力来自于人们采用美国的产品和服务、贯穿整个技术栈,也许最明显的是他们用 ChatGPT 而不是 DeepSeek,但也包括他们使用美国的芯片、美国的数据中心技术,以及微软所做的一切了不起的事情,那对我们是一场胜利,我认为我们应该拥抱那一点,但要确保,你知道的,最关键的东西——这些将极具影响力的模型的创造——那仍然应该在这里发生。 I also was glad to see that rescinded. I agree there will need to be some constraints. But I think if our--if the sort of mental model is winning diffusion instead of stopping diffusion that directionally seems right. That does not mean there is no guardrails. It does not mean we say, like, we are going to go build a bigger data center in some other country than the U.S. Our intention is to build our biggest and best data centers in the U.S. Do training in the U.S. Build models here. Have our core research here. But then we do want to build inference centers with our partners around the world and we have been working with the U.S. Government on that. I think that will be good. To this point that influence comes from people adopting U.S. products and services up and down the stack, maybe most obviously if they are using ChatGPT versus DeepSeek but also if they are using U.S. chips and U.S. data center technology and all of the amazing stuff Microsoft does, that is a win for us, and I think we should embrace that but make sure that, you know, the most critical stuff--the creation of these models that will be so impactful--that should still happen here. ¶
Ms. Su 我认为我们完全同意这样一个概念:某些限制是必要的。这既是 AI 扩散问题,也是国家安全问题。 既然如此,我们也很高兴看到它被撤销;我们把这视为一个真正去简化的机会,对吧。 归根结底,您知道,我们谈过需要推动我们技术和生态的广泛采用。您知道,简单、易于适用、真正让我们的盟友既能保护我们的技术、又能用上美国所能提供的最好东西的规则,我认为是我们在前进方向上的一个好起点;而且,您知道,再说一次,这是一个我认为魔鬼在细节里、需要大量平衡的领域。 所以从产业界的角度,您知道,我们的职责是戴上更宽广的帽子,与政府和国会携手合作,您知道,提出我们最好的建议,好让这项政策在向前推进时也有一定的稳定性。 **The Chairman.** 因特拉托先生。 I think we would totally agree with the concept that some restrictions are necessary. This is a matter of national security as much as it is about AI diffusion. That being the case, we were happy to see the rescinding as well and we view this as an opportunity to really simplify, right. At the end of the day, you know, we have talked about the need to drive widespread adoption of our technology and our ecosystem. You know, simple rules that can be easily applied that really allow our allies to protect our technology while still utilizing the best that the United States has to offer I think is a good start in terms of where we are going and, you know, again, this is an area where I think the devil is in the details and it requires a lot of balance. And so from an industry standpoint, you know, it is our job to put on the broader hat and work hand-in-hand with the administration and Congress to, you know, make our best recommendations so that it is a policy that has some stability as we go forward as well. The Chairman. Mr. Intrator. ¶
Mr. Intrator 我要呼应萨姆和莉萨所说的。但是,您知道,国家安全是最重要的;一旦你处理好了围绕国家安全的限制,那么在此之外与监管者一道搭建一套监管框架的机会就非常合理;而那条扩散规则没有给我们足够充分参与的机会,让我们觉得自己能得出一个在当下堪称最优的结果。 **The Chairman.** 史密斯先生。 So I will echo what Sam and Lisa said. But, you know, national security is paramount, and then once you have addressed the limitations around national security the opportunity to work with regulators to put together a regulatory framework beyond that makes a lot of sense, and the diffusion rule did not allow us that opportunity to participate fully enough to feel like we were going to come away with what would be an optimal outcome at this point. The Chairman. Mr. Smith. ¶
Mr. Smith 我认为我们都讨论过正确的配方。简化,取消那些削弱信心、削弱获取美国技术的「第二层级」数量限制;但同时要让这个国家最先进的 GPU 也能出口到由可信提供方运营、符合某些安全标准的数据中心。 这意味着既要有实体安全标准,也要有网络安全标准;要防范芯片被转运,也要防范某些用途,这意味着两件事。 一是要有管制措施,确保——比如说——解放军、中国军队无法在数据中心获取和使用这些先进模型或先进芯片,无论那个数据中心在哪个国家;二是有些有害用途是应当被禁止、被排除的,比如用模型制造下一场大流行病、生物武器、核武器。 我认为正在形成一种可以保留、可以向前推进、并且取得恰当平衡的做法。 **The Chairman.** 好。给各位最后一个问题。你们是否支持给各州设定一个 10 年的「学习期」,在此期间不得发布全面的 AI 监管;或者支持某种形式的联邦优先立法,为 AI 开发者和部署者创造公平的竞争环境? I think we have all discussed the right recipe. Simplify, eliminate these tier two quantitative restrictions that undermine confidence and access to American technology, but enable even the most advanced GPUs the country has to be exported to data centers that are run by a trusted provider, that meet certain security standards. That means both physical and cybersecurity standards. That there is protection against diversion of the chips and there are precautions against certain uses, and that means two things. One is that there are controls in place to ensure that, say, the PLA--the Chinese military--is not accessing and using these advanced models or advanced chips in a data center regardless of the country that it is in, and there are certain harmful uses that one should want to prohibit and preclude like using a model to create the next pandemic, a biological weapon, a nuclear weapon. And I think that there is an approach that is coming together that can be retained and can move forward and that strikes the right balance. The Chairman. OK. Final question for each of you. Would you support a 10-year learning period on states issuing comprehensive AI regulation or some form of Federal preemption to create an even playing field for AI developers and deployers? ¶
altman 我不确定「10 年学习期」是什么意思,但我认为有一种专注于轻量接触、在公平竞争环境上的联邦方式,在我看来很棒。 I am not sure what a 10-year learning period means, but I think having one Federal approach focused on light touch on an even playing field sounds great to me. ¶
Ms. Su 一个协调一致的联邦路径,加上——您知道——真正深思熟虑的监管,会非常、非常受欢迎。 Aligned Federal approach with, you know, really thoughtful regulation would be very, very much appreciated. ¶
Mr. Intrator 我同意两位同事的看法。 I agree with both my colleagues. ¶
Mr. Smith 是的;我认为这显然建立在您和格雷厄姆参议员去年发表的那篇评论文章之上;我认为给这个国家时间——您那个类比、那个例子,是说这在互联网上奏效过。 有很多细节需要敲定,但让联邦政府有能力去引领,特别是在产品安全和发布前审查这类领域,会有助于这个行业成长。 **The Chairman.** 好,我要感谢每一位证人。这是一场非常有意思的听证会。它很有信息量。这些问题很重要。 你们看到了两党双方对这个话题的巨大兴趣;所以我感谢——各位都很忙、都在做很多事,我感谢你们今天到场。 参议员们可在 5 月 15 日(星期四)下班前提交书面问题,证人可在 5 月 29 日(星期四)当天结束前回答这些问题。 就这样,今天的听证会到此结束。委员会休会。 〔听证会于下午 1:13 休会。〕 Yes, I think that builds, obviously, on the op- ed that you and Senator Graham published last year and I think giving the country time--your analogy, your example, was this worked for the internet. There is a lot of details that need to be hammered out, but giving the Federal Government the ability to lead, especially in the areas around product safety and pre-release reviews and the like, would help this industry grow. The Chairman. Well, I want to thank each of the witnesses. This was a very interesting hearing. It was informative. These issues matter. You saw a great deal of interest on both sides of the aisle in this topic and so I appreciate--each of you are very busy and doing a lot of things and I appreciate your being here today. Senators will have until the close of business on Thursday, May 15, to submit questions for the record and the witnesses will have until the end of the day on Thursday, May 29 to respond to those questions. And with that, that concludes today's hearing. The Committee stands adjourned. [Whereupon, at 1:13 p.m., the hearing was adjourned.] ¶
美国参议院商务、科学与交通委员会听证会「赢得 AI 竞赛:强化美国在计算与创新方面的能力」,2025 年 5 月 8 日 Hearing Before the U.S. Senate Committee on Commerce, Science, and Transportation “Winning the AI Race: Strengthening U.S. Capabilities in Computing and Innovation” May 8, 2025 ¶
OpenAI 联合创始人兼首席执行官 Sam Altman 的证词 Testimony of Sam Altman, Co-Founder and CEO, OpenAI ¶
感谢克鲁兹(Cruz)主席、坎特威尔(Cantwell)首席民主党议员以及委员会各位委员。 Thank you, Chairman Cruz, Ranking Member Cantwell, and Members of the Committee. ¶
我是 Sam Altman,OpenAI 的首席执行官。很荣幸能再次来到参议院,分享我们对 AI 现状与未来走向的看法。 I’m Sam Altman, Chief Executive Officer of OpenAI. It is an honor to return to the Senate and share our view of where AI is today and where we see it going. ¶
OpenAI 不是一家普通的公司,也永远不会是。 OpenAI is not a normal company and never will be. ¶
我们的使命是确保通用人工智能——AGI——造福全人类。AGI 是一个界定得并不严格的术语,但大体而言,我们指的是一个能够在许多领域以人类水平处理越来越复杂问题的系统。十多年前我们创立 OpenAI 时,我们围坐在厨房的餐桌旁面面相觑,不知从何做起。那时的 AI 还是研究者的小众工具,而非面向大众的技术。 Our mission is to ensure that artificial general intelligence – AGI – benefits all of humanity. AGI is a weakly defined term, but generally speaking we mean it to be a system that can tackle increasingly complex problems, at human level, in many fields. When we formed OpenAI more than 10 years ago, we stared at each other around a kitchen table, wondering how to get started. AI then was a niche tool for researchers, not the general public. ¶
2016 年,克鲁兹主席召开了首场 AI 听证会,我的联合创始人 Greg Brockman 作证说,AGI 模型距离我们大概还有 10 到 100 年。如今,AI 科学已经取得了巨大进步,我们现在有信心,我们将在特朗普(Trump)总统任内达到那个里程碑。 In 2016, Chairman Cruz convened his first AI hearing, and my co-founder, Greg Brockman, testified that AGI models were probably between 10 and 100 years away. Today, the science of AI has advanced so significantly that we are now confident that we’ll reach that milestone during President Trump’s time in office. ¶
纵观历史,人类一直打造工具来扩展我们的能力——而我们认为,AGI 将是有史以来创造出的最强大的工具。它将让人们能够为彼此创造出不可思议的东西,并改善他们的生活质量。 Throughout history, people have crafted tools to scale our abilities—and we believe AGI will be the most powerful tool ever created. It will enable people to build incredible things for each other and improve their quality of life. ¶
但 AGI 的全部潜力,只有在安全的前提下才能实现。那些让 AGI 能够支撑科学突破、加速人类进步的能力,同样会制造出新的风险领域。这正是我们不懈开展红队测试(red-team)、并在透明度方面引领行业的原因。 But AGI’s full potential won’t be realized unless it's safe. The same capabilities that will enable AGI to support scientific breakthroughs and accelerate human progress will also create new risk areas. That’s why we red-team relentlessly and lead the industry in transparency. ¶
最终,我相信好处会以数量级之差超过坏处,而且 AGI 将帮助我们进入我称之为「智能时代」(Intelligence Age)的纪元——在那个时代,每个人的生活都可以比今天任何一个人的生活更美好。 Ultimately, I believe the good will outweigh the bad by orders of magnitude, and that AGI will help bring us into what I call the Intelligence Age – an era when everyone’s lives can be better than anyone’s life today. ¶
这个未来的光明程度几乎难以想象,但前提是我们必须采取具体步骤,确保建立在自由与透明等民主价值观之上的、美国主导版本的 AI,战胜威权版本的 AI。 This future can be almost unimaginably bright, but only if we take concrete steps to ensure that an American-led version of AI, built on democratic values like freedom and transparency, prevails over an authoritarian one. ¶
局势的赌注再高不过了——而国会的判断是对的:美国必须走在最前面。在 OpenAI,我们坚定走民主 AI 这条路;ChatGPT 如今每周被超过 5 亿人使用,用来创造、发现、实现曾经遥不可及的突破——对此我们深感荣幸、心怀谦卑。 The stakes could not be higher—and Congress is right that the United States must lead the way. At OpenAI, we’re committed to the path of democratic AI, and we are humbled that ChatGPT is being used by more than 500 million people each week to create, discover, and achieve breakthroughs that were once out of reach. ¶
美国是一个创新者的国度,我们想要极大增强人们运用我们的技术改善自己生活的能力。 America is a nation of innovators, and we want to supercharge people’s ability to use our technology to make their lives better. ¶
我们想要开源能力非常强的模型。 We want to open source very capable models. ¶
我们想要在用户如何使用我们的工具这件事上给予他们极大的自由,并让他们能够个性化 ChatGPT,以最好地满足自己的需求。 We want to give our users a great deal of freedom in how they use our tools, and let them personalize ChatGPT to best meet their needs. ¶
我们想要为全世界构建一个「大脑」,并让人们使用起来极其简单,同时辅以防止伤害的常识性限制。 We want to build a brain for the world and make it super easy for people to use it, with common-sense restrictions to prevent harm. ¶
而事实是,AI 已经在让世界变得更好。 And the truth is that AI is already changing the world for the better. ¶
美国各国家实验室的科学家——包括橡树岭国家实验室(Oak Ridge National Laboratory)、洛斯阿拉莫斯国家实验室(Los Alamos National Laboratory)、阿贡国家实验室(Argonne National Laboratory)、普林斯顿等离子体物理实验室(Princeton Plasma Physics Laboratory)和太平洋西北国家实验室(Pacific Northwest National Laboratory)——正在使用我们的推理模型,加速能源等领域的突破。 Scientists at the US National Laboratories – including Oak Ridge National Laboratory, Los Alamos National Laboratory, Argonne National Laboratory, the Princeton Plasma Physics Laboratory, and the Pacific Northwest National Laboratory – are using our reasoning models to accelerate breakthroughs in areas like energy. ¶
在宾夕法尼亚州,ChatGPT 正在帮助州政府雇员更快地完成行政事务,从而腾出更多时间来改善公共服务的交付。 In Pennsylvania, ChatGPT is helping state employees do administrative tasks more quickly, freeing up more time to improve the delivery of public services. ¶
而在德克萨斯、北卡罗来纳、加利福尼亚等州,各大学正在把 ChatGPT 交到学生和教育者手中,以培养一支为 AI 做好准备的劳动力队伍。 And universities in states like Texas, North Carolina, and California are putting ChatGPT in the hands of students and educators to build an AI-ready workforce. ¶
AI 对于确保今天的学生为明天的工作做好准备至关重要。在美国,超过三分之一的大学年龄段年轻人在使用我们的模型,主要用于学习和辅导。在全球范围内,大多数 ChatGPT 用户的年龄在 35 岁以下。 AI will be vitally important to ensuring that today’s students are ready for tomorrow’s jobs. In the US, more than one-third of college-aged young people use our models, mainly for learning and tutoring. Around the world, most ChatGPT users are under age 35. ¶
我们为能免费提供一项已经为如此多人做了如此多事情的技术感到自豪,但 AI 最大的收益还在后头。 We’re proud to offer free access to a technology that is doing so much for so many people, but AI’s biggest gains are still to come. ¶
我们在 OpenAI 的工作表明,随着 AI 的进步,进展会不断加速,并且变得越来越负担得起,这体现在以下三条缩放原理(scaling principles)中: Our work at OpenAI suggests that as AI advances, progress accelerates and becomes increasingly affordable, as reflected in these three scaling principles: ¶
对 AI 投入更多,将会持续让它变得更好、更有能力。一个 AI 模型的智能,大致等于用于训练和运行它的资源的对数。直到最近,扩展方面的进展还主要来自训练算力和数据,但我们已经展示了如何 Investing more in AI will continue to make it better and more capable. The intelligence of an AI model roughly equals the log of the resources used to train and run it. Until recently, scaling progress has primarily come from training compute and data, but we have shown how ¶
让智能同样能从推理算力中扩展出来。用于预测这些收益的缩放定律(scaling laws)在跨越许多个数量级的范围内都极其精确。由此可以推知,进一步的投入将带来进一步的收益,以及给社会带来进一步的福祉:我们相信,线性增长的智能所带来的社会经济价值,本质上是超指数级的。 to make intelligence scale from inference compute, as well. The scaling laws that predict these gains are incredibly precise over many orders of magnitude. It follows that further investment will lead to further gains, and further benefits to society: We believe that the socioeconomic value of linearly increasing intelligence is super-exponential in nature. ¶
使用某一给定水平 AI 能力的成本,每 12 个月下降约 10 倍,而更低的价格会带来多得多的使用量。我们在 2023 年初的 GPT-4 与 2024 年中的 GPT-4o 之间的 token 成本变化中看到了这一点:在这段时间里,每 token 的价格下降了约 150 倍。摩尔定律预言,微芯片上的晶体管数量大约每两年翻一番;而使用 AI 的成本下降得甚至比这更剧烈。 The cost to use a given level of AI capability falls by about 10x every 12 months, and lower prices lead to much more use. We saw this in the change in token cost between GPT- 4 in early 2023 and GPT-4o in mid-2024, where the price per token dropped about 150x in that time period. Moore’s Law predicted that the number of transistors on a microchip would double roughly every two years; the decrease in the cost of using AI is even more dramatic. ¶
改进一个 AI 模型所需的时间在不断缩短。换句话说,AI 模型正在以越来越快的速度追赶人类智能。计算机要在某一个给定的基准测试上击败人类,所需的典型时间已经从该基准问世后的 20 年,缩短到 5 年,再到如今的 1 到 2 年——而且我们看不到有什么理由会让这些进步在不久的将来停下来。 The amount of time it takes to improve an AI model keeps decreasing. Put another way, AI models are catching up with human intelligence at an increasing rate. The typical time it takes for a computer to beat humans at a given benchmark has fallen from 20 years after the benchmark was introduced, to five years, and now to one to two years—and we see no reason why those advances will stop in the near future. ¶
那么,这在实践中意味着什么? So what does that mean practically? ¶
我相信未来三年我们会看到许多重大进展,下面是一些例子。 I believe we’ll see many major advances over the next three years, but here are some examples. ¶
在 2025 年,我们将发布能够处理复杂软件工程的 AI 驱动工具,以及能够处理现实世界任务——比如预约医生、帮忙经营企业——的 AI agent。这些 agent 将是超级助理,能够与各个行业的员工、所有专科的医生以及每个研究领域的科学家协作。 In 2025, we will release AI-powered tools that can handle sophisticated software engineering, and AI agents that can handle real-world tasks like making doctor’s appointments and helping to run a business. These agents will be super assistants who can collaborate with workers in every industry, doctors in all specialties, and scientists in every field of research. ¶
在 2026 年,AI 可能通过设计实验来应对美国在气候、健康和国家安全方面最棘手的挑战,从而解锁新一轮科学突破。 In 2026, AI may unlock a new wave of scientific breakthroughs by designing experiments to tackle America’s toughest challenges in climate, health, and national security. ¶
而在 2027 年,AI 驱动的机器人技术可能把 AI 带来的生产力提升推进到物理世界,处理例行任务,让人们把更多时间花在自己喜欢的工作和活动上。 And in 2027, AI-powered robotics could push AI-driven productivity gains into the physical world, handling routine tasks so people can spend more time on the work and activities they enjoy. ¶
随着 AI 系统变得更有能力,人们会更加想要使用它们。满足这种需求,需要更多的芯片、训练数据、能源和超级计算机。 As AI systems become more capable, people will want to use them even more. Meeting that demand requires more chips, training data, energy, and supercomputers. ¶
基础设施即命运(Infrastructure is destiny),而我们需要多得多的基础设施。 Infrastructure is destiny, and we need a lot more of it. ¶
今年早些时候,我与特朗普总统以及 Oracle(甲骨文)和 SoftBank(软银)的首席执行官们一起,宣布了星际之门项目(Stargate Project)——一项对美国 AI 基础设施的 5000 亿美元投资。 Earlier this year I joined President Trump and the CEOs of Oracle and SoftBank to announce the Stargate Project, a $500 billion dollar investment in American AI infrastructure. ¶
自星际之门启动以来,世界各国政府纷纷询问,如何把 AI 基础设施引入他们的国家,以及我们如何能确保民主的 AI 系统成为全球标准。 Since launching Stargate, governments around the world have asked about bringing AI infrastructure to their countries and how we can ensure that democratic AI systems become the global standard. ¶
作为回应,我们正在提供一种新型伙伴关系——OpenAI for Countries(OpenAI 面向各国)——来帮助这些国家建设自己的数据中心容量,以及 AI 创业公司和开发者生态。作为交换,这些国家将投资于星际之门项目——从而投资于美国主导的 AI 领导地位的延续,以及民主 AI 的全球化、不断成长的网络效应。 In response, we’re offering a new kind of partnership — OpenAI for Countries — to help these countries build up their data center capacity and ecosystems of AI start-ups and developers. In exchange, these countries would invest in the Stargate Project – and thus in continued US-led AI leadership and a global, growing network effect for democratic AI. ¶
最后请允许我以一段个人经历作结:我在圣路易斯长大,来自一个关系紧密又争强好胜的家庭,我们会玩「20 个问题」(20 Questions)的游戏来猜晚餐后的甜点是什么。我八岁那年,父母给我买了一台 Mac LC II。这台电脑是我人生中一条实实在在的分界线:一段是我拥有电脑之前的时光,另一段是之后的时光。我相信,AI 将对全国各地的孩子们——包括我自己的孩子——起到类似的塑造人生的作用。 To close on a personal note, I grew up in St. Louis, part of a close-knit and competitive family that played 20 Questions to guess what we were having for dessert. When I was eight, my parents bought me a Mac LC II. The computer was a literal dividing line in my life. There was the time before I had a computer, and there has been the time after. I believe that AI will play a similarly formative role for kids across the country, including my own. ¶
我要感谢克鲁兹主席、坎特威尔首席民主党议员以及本委员会各位委员在 AI 议题上持续的领导力。我很珍惜今天作证的机会,并期待回答各位的提问。 I want to thank Chairman Cruz, Ranking Member Cantwell, and the members of this Committee for your continued leadership on AI. I appreciate the opportunity to testify today and look forward to answering your questions. ¶
altman ## 1. 尊敬的罗杰·威克(Roger Wicker)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:AI 技术的政治与伦理决策 **背景:** xAI 由埃隆·马斯克于 2023 年 3 月 9 日创立,开发与 OpenAI 类似的技术;后者由马斯克于 2015 年与萨姆·奥特曼等人共同创办。Grok 是 xAI 的旗舰产品,运行于社交媒体平台 X(原 Twitter)之上。马斯克及其开发者已表示,有意测量并可能修改嵌入 AI 系统中的政治与伦理偏好。研究显示,像 OpenAI 的 ChatGPT 这样广受欢迎的 AI 模型往往表现出特定的意识形态倾向,尤其偏向环境保护,并表达左倾、自由意志主义的观点。 **问题:** 奥特曼先生,特朗普总统说过,为维持美国在 AI 领域的领导地位,我们必须开发不受意识形态偏见或人为社会议程影响的系统。你曾说过,你预期 AI 在达到超人级通用智能之前,就会先具备超人级的说服能力。根据 xAI 与 Scale AI 顾问丹·亨德里克斯(Dan Hendrycks)最近发表的研究,AI 系统在其价值体系中表现出显著左翼偏见。应当怎么做来防止 AI 的超人级说服?AI 的超人级说服应当被禁止吗?你们正在做什么来防止 OpenAI 系统中的超人级说服? **答复:** 我们的工具赋予学习、创造与创新的自由。我们正在加速知识、创造力与自由表达。我们的系统有稳健的防护栏。我们对这些防护栏如何运作保持透明,并努力确保我们在赋能创造力、保护每个人使用 AI 的自由。我们的模型被专门设计为不「带有议程」,这一点在我们公开的文档(例如描述模型运作方式的《模型规范》Model Spec)中已有说明;AI 助手的目标是协助人类,而不是塑造人类。 Response to Written Question Submitted by Hon. Roger Wicker to Sam Altman Political and Ethical Decisions by AI Technology Background: xAI was founded by Elon Musk on March 9, 2023, and develops technology similar to OpenAI, which Musk helped found alongside Sam Altman and others in 2015. Grok is xAI's flagship product, which runs on X (formerly Twitter), the social media platform. Musk, along with its developers, has expressed an intent to measure and potentially modify the political and ethical preferences embedded in AI systems. Studies have shown that popular AI models like OpenAI's ChatGPT tend to exhibit specific ideological leanings, particularly favoring environmental protection and expressing left-leaning, libertarian viewpoints. Question. Mr. Altman, President Trump has said that to maintain U.S. leadership in AI, we must develop systems that are free from ideological bias or engineered social agendas. You have said that you expect AI to be capable of superhuman persuasion well before it is superhuman at general intelligence. According to research recently published by xAI and Scale AI advisor Dan Hendrycks, AI systems exhibit significant left-wing biases in their value systems. What should be done to prevent superhuman persuasion by AI? Should superhuman persuasion by AI be banned? What are you doing to prevent superhuman persuasion in OpenAI's systems? Answer. Our tools enable the freedom to learn, the freedom to create, and the freedom to innovate. We are accelerating knowledge, creativity, and free expression. Our systems have robust guardrails. We are transparent about how those guardrails work and we work hard to make sure we are enabling creativity and protecting everyone's freedom to use AI. Our models are specifically designed not to ``have an agenda'' which is outlined in our public documentation, such as the Model Spec, describing how our models work; the goal of an AI assistant is to assist humanity, not to shape it. ______ ¶
altman ## 2. 尊敬的玛莎·布莱克本(Marsha Blackburn)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) **问题:** 美国企业的实力长期依赖知识产权的执行。最近,OpenAI 公开主张,除非允许美国的 AI 公司宽泛地主张对受版权保护内容的合理使用,否则美国将在 AI 领域失去对中国的竞争优势。〔1〕具体而言,OpenAI 最近的一封意见函称「毫无疑问,中国的人工智能开发者将享有不受限制的数据获取——包括受版权保护的数据——从而改进他们的模型」。这一主张与长期以来把知识产权保护视为创新驱动力和抵御外国竞争之保障的美国原则形成鲜明对比。美国对维护财产权和法治的承诺,是其技术发展领导地位的核心。你是否建议美国采取一种更接近中国的知识产权方针? **答复:** 美国的知识产权法支撑了一代又一代美国技术领导地位,从个人电脑到商业互联网。美国的创意专业人士受益于我国法律给予创作者的强力保护;美国的技术创新者也受益于既有的、长期存在的知识产权原则,例如合理使用——它允许新技术以转化性方式与受版权保护的作品互动。这一平衡的框架成就了美国 AI 的成功,美国应当坚守这套一直如此有效地服务我国的既有法律框架。 Response to Written Questions Submitted by Hon. Marsha Blackburn to Sam Altman Question. The strength of American businesses has long depended on the enforcement of intellectual property (IP) rights. Recently, OpenAI has publicly argued that unless AI companies in the U.S. are permitted to broadly claim fair use of copyrighted content, the Nation will lose its competitive advantage to China in the AI sector.[1] Specifically, a recent comment letter by OpenAI stated ``there is little doubt that the PRC's AI developers will enjoy unfettered access to data--including copyrighted data--that will improve their models.'' This assertion stands in stark contrast to the long-standing American principles that prioritize IP protection as a driver of innovation and a safeguard against foreign competition. The United States' commitment to upholding property rights and the rule of law has been central to its leadership in technological development. Would you suggest that the U.S. adopt an approach to intellectual property rights more akin to that of China? Answer. America's intellectual property laws have underpinned generations of American technology leadership, from the personal computer to the commercial internet. America's creative professionals have benefitted from the strong protections our laws give to creators, and America's technology innovators have benefitted from existing and longstanding IP doctrines such as fair use, which permits new technologies to interact with copyrighted works in transformative ways. This balanced framework has enabled the success of American AI, and the United States should stand by the existing American legal framework that has served our country so well. ______ ¶
altman ## 3. 尊敬的玛丽亚·坎特韦尔(Maria Cantwell)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:AI 标准 美国与世界上最先进的民主国家一道推动 AI 标准的发展,这为我们提供了在全球舞台上为 AI 设定「道路规则」的机会。 **问题 1.** 在回应我关于 NIST 标准的问题时,你表示 NIST 标准并非必要,但可能有助于提升我们的竞争力。能否解释你如何看待 NIST 标准有助于美国的竞争力? **答复:** 我们知道,在 5G 竞赛中,中国在补贴和支持标准制定机构中的企业方面发挥了积极作用。他们在表决机构中的影响力帮助设定了 5G 的道路规则。类似的动态正在 AI 标准上发生,NIST 可以在支持美国企业应对这些程序方面发挥作用,特别是对非传统技术企业而言。 为了让 AI 造福世界,技术标准将有助于各国在美国技术之上构建、并推广民主的 AI。这正是我们的「OpenAI for Countries」倡议所要实现的目标。 **问题 2.** 你希望 NIST 制定并推广哪些标准,以提升美国的竞争力? **答复:** 随着模型能力增强,推进评估科学与衡量这些能力的指标将变得重要。NIST 的测量科学举措(基准、试验台、密码学验证、AI 评估)可以帮助推进这些评估方法的发展,并给产业界一套围绕安全与安保的共同语汇。 另外,各国正在提出前沿系统风险管理的框架。NIST 可以在协调跨辖区的这些框架方面发挥作用,帮助美国产品进入新市场。 最后,智能体将对互联网上任务执行与沟通的方式构成新挑战。NIST 可以帮助推动并建立基于共识的产业技术标准来应对这一挑战。 ### 主题:AI 安全 我们看到深度伪造和其他 AI 内容激增,威胁着普通人在媒体中辨别真相的能力。而这只是 AI 领域中提出复杂安全问题的其中一个方面。美国 AI 安全研究所在确保 AI 系统被负责任地开发、最先进的模型得到充分测试方面扮演关键角色。这对建立信任、促进更广泛采用至关重要。 **问题 3.** 你是否支持美国 AI 安全研究所的工作? **答复:** OpenAI 与美国 AI 安全研究所建立了建设性的伙伴关系,聚焦于两用 AI 能力所带来的国家安全风险。在我们看来,这是一个良好的模式,说明联邦政府与私营部门之间的自愿伙伴关系如何能够保护美国国家安全、增强我们的经济竞争力。 ### 主题:科学领域的公共投资 政府对基础科学的投入一直是美国在技术与创新上取得成功的支柱。如果美国想要在竞争中胜过外国对手,就不能削减对国家科学基金会、国家标准与技术研究院(NIST)、能源部各实验室,或支撑 AI 劳动力与生态的 STEM 教育项目的拨款。AI 领域的领导地位需要持续的公共投资,而不是缺乏数据支撑的、考虑不周的削减。 **问题 4.** 贵公司在人工智能开发中如何受益于国家科学基金会、NIST 或能源部各实验室,或与之合作? **答复:** OpenAI 非常珍视与美国公共机构合作的机会,特别是与能源部国家实验室(包括洛斯阿拉莫斯国家实验室)合作,推进人工智能在科学发现中的安全有益使用。 我们合作的亮点包括: - **在洛斯阿拉莫斯国家实验室(LANL)安全部署 OpenAI 模型:** 在一项史无前例的伙伴关系中,我们使我们的模型得以部署在 Venado 超级计算集群上,支持安全政府环境内的高保障科学研究。 - **在国家实验室生态内更广泛的参与:** 我们与能源部实验室网络各处的科学家建立了合作,包括举办一场「1000+ 科学家 Jam Session」,九个国家实验室的 1,500 多名研究人员探讨了 AI 如何加速研究。这些合作带来相互学习:我们的模型因真实的科学反馈而改进,实验室科学家则获得前沿工具的实操经验。 - **关于面向未来项目的持续讨论:** 我们继续与各实验室在使命一致的领域接触,例如能源研究、生物科学和材料发现,共同目标是负责任地驾驭 AI,以支持美国的科学与技术领导地位。 **问题 5.** 削减 NSF 拨款将如何影响你们的劳动力和人才搜寻? **答复:** OpenAI 受益于、并深深珍视由国家科学基金会(NSF)等美国机构所培育的强大科学生态。NSF 资助的项目在支持塑造人工智能未来的研究者、学生与发现方面发挥着重要作用。 一个健康的学术研究环境: - 扩大了具备 AI 能力的人才库,包括许多我们有幸聘用的研究者。 - 强化了基础科学,而其中很大一部分支撑着机器学习及邻近领域的进步。 - 支持更广泛的创新,确保 AI 的发展既受益于、也贡献于更广大的科学事业。 我们致力于与政府和学术伙伴合作,确保美国在人才与创新两方面都保持全球领导者地位。 **问题 6.** 削减联邦对大学科学与研究的拨款,将对美国在 AI 领域的竞争力产生什么影响? **答复:** 美国研究型大学的实力长期以来是先进技术(包括 AI)国家竞争力的关键驱动力。对大学研究的公共投资,在促成基础发现与人才发展两方面都发挥着独特作用。 就 AI 而言: - 许多核心创新都出自受联邦拨款支持的大学实验室。 - AI 模型本身正越来越多地被用于支持科学发现,放大了研究投资的价值。 对大学研究的持续支持,确保美国保持在科学与技术进步的最前沿。我们对有机会继续与学术界和联邦伙伴一道为这一共同使命作贡献充满热情。 ### 主题:能源需求与聚变能研发 为 AI 数据中心供电的电力需求增长之快令人震惊。据一些估计,到 2030 年,全球数据中心的电力需求预计将增长一倍以上,超过 945 太瓦时(TWh)。这将给电网和能源供应商带来压力。 **问题 7.** 贵公司有什么计划来满足 AI 的能源需求?你们正在对非化石能源(例如聚变)作哪些投资? **答复:** 我们预计 AI 的能源需求将激励对电网基础设施的大规模新投资,并推动能源技术的创新。AI 的进步——包括推理模型——对科学发现具有重大前景,包括在充裕、可负担的能源解决方案领域。事实上,我们现有的、为国家实验室部署推理模型的伙伴关系就包括劳伦斯利弗莫尔实验室,那里的科学家在世界上首次演示了聚变点火。我们也将与业内其他人一道,继续努力寻找新方法,确保我们的技术尽可能高效,包括在能耗方面。即便我们持续看到有前景的研究与创新,我们同样专注于在研究与部署中优化使用现有算力。 **问题 8.** 就聚变能而言,随着聚变技术持续发展,政府可以如何与私营部门合作以实现其规模化? **答复:** 缩放定律是清楚的。美国在 AI 的领导地位是能源、数据和芯片的函数。政府对聚变研究与试点项目的支持可以至关重要,因为企业正设法找到可行路径,为其持续的科学进展、并最终实现规模化开发筹集所需资本。随着政府继续支持私营实体及我国国家实验室的聚变研究,我们希望我们与国家实验室的合作、以及与一系列裂变和聚变公司的早期探索,能早日表明 AI 在推进能源充裕方面可以扮演的角色。 Response to Written Questions Submitted by Hon. Maria Cantwell to Sam Altman AI Standards The U.S. driving development of AI standards alongside the most advanced democracies in the world offers us an opportunity to set the ``rules of the road'' for AI on the global stage. Question 1. In response to my question regarding NIST standards, you stated that NIST standards would not be necessary, but that they could be helpful in improving our competitiveness. Can you explain how you view NIST standards as helping the United States' competitiveness? Answer. We know that in the race for 5G China took an active role in subsidizing and supporting companies in standard setting bodies. Their influence in voting bodies helped set the rules of the road for 5G. Similar dynamics are occurring with AI standards and NIST can play a role in supporting American companies to navigate these processes, especially for non-legacy technology companies. In order for AI to benefit the world, technical standards would help countries build on U.S. technology and promote democratic AI. This is what our initiative, OpenAI for Countries, aims to achieve. Question 2. What standards would you like to see NIST develop and promote to improve U.S. competitiveness? Answer. As models become more capable, it will be important to further the science of evaluations and metrics for measurement of such capabilities. NIST's measurement science initiatives (benchmarks, test beds, cryptographic validation, AI evaluations) can help further the development of these methods of evaluation and give industry a common vocabulary around safety and security. Separately, various countries are proposing frameworks for risk management of frontier systems. NIST could play a role in harmonizing those across jurisdictions to help American products get to new markets. Finally, agents will present a new challenge to how tasks and communication are conducted on the internet. NIST could help drive and establish consensus based industry technical standards to address this challenge. AI Safety We are seeing a proliferation of deepfakes and other AI content that threatens the average person's ability to discern truth in media. And that's just one area in the field of AI that presents complicated safety questions. The U.S. AI Safety Institute plays a critical role in ensuring that AI systems are developed responsibly and that the most advanced models are fully tested. This is crucial for building trust and promoting wider adoption. Question 3. Do you support the work of the U.S. AI Safety Institute? Answer. OpenAI has had a constructive partnership with the U.S. AI Safety Institute, focused on national security risks posed by dual-use AI capabilities. In our view, this is a good model for how a voluntary partnership between the Federal government and private sector can protect American national security and strengthen our economic competitiveness. Public Investment in Science Government investment in fundamental science has been the backbone of American success in technology and innovation. If the United States wants to outcompete foreign adversaries, it cannot defund the National Science Foundation, National Institute of Standards and Technology (NIST), Department of Energy labs, or STEM education programs that power the AI workforce and ecosystem. Leadership in AI requires sustained public investment, not ill-conceived cuts that are not data driven. Question 4. How has your company benefited from or collaborated with the National Science Foundation, NIST or the Department of Energy Labs in artificial intelligence development? Answer. OpenAI has greatly valued the opportunity to collaborate with U.S. public institutions, particularly the Department of Energy National Laboratories, including Los Alamos National Laboratory, in advancing the safe and beneficial use of artificial intelligence for scientific discovery. Highlights of our collaboration include: Secure deployment of OpenAI models at Los Alamos National Laboratory (LANL): In a first-of-its-kind partnership, we enabled the deployment of our models on the Venado supercomputing cluster, supporting high-assurance scientific research within secure government environments. Wider engagement across the national lab ecosystem: We have partnered with scientists from across the DOE laboratory network, including hosting a ``1,000+ Scientists Jam Session'' where over 1,500 researchers at nine national labs explored how AI can accelerate research. These collaborations provide mutual learning: our models improve with real scientific feedback, and lab scientists gain hands-on experience with frontier tools. Ongoing discussions on future-focused projects: We continue to engage with the labs on mission-aligned areas such as energy research, bioscience, and materials discovery, with the shared goal of responsibly harnessing AI to support U.S. scientific and technological leadership. Question 5. How will cuts to NSF funding impact your workforce and search for talent? Answer. OpenAI benefits from and deeply values the strong scientific ecosystem fostered by U.S. institutions such as the National Science Foundation (NSF). NSF-funded programs play an important role in supporting the researchers, students, and discoveries that shape the future of artificial intelligence. A healthy academic research environment: Expands the pool of AI-ready talent, including many of the researchers we are proud to have hired. Strengthens foundational science, much of which underpins progress in machine learning and adjacent fields. Supports broader innovation, ensuring that developments in AI benefit from and contribute to the wider scientific enterprise. We are committed to working with government and academic partners to ensure the U.S. remains a global leader in both talent and innovation. Question 6. What impact will cuts to Federal funding for science and research at universities have on U.S. competitiveness in AI? Answer. The strength of America's research universities has long been a key driver of national competitiveness in advanced technologies, including AI. Public investment in university research plays a unique role in enabling both fundamental discovery and talent development. In the context of AI: Many core innovations have emerged from university labs supported by Federal grants. AI models themselves are increasingly being used to support scientific discovery, amplifying the value of research investment. Continued support for university-based research ensures that the U.S. remains at the forefront of scientific and technological progress. We are enthusiastic about the opportunity to continue contributing to this shared mission alongside academic and Federal partners. Energy Needs and R&D for Fusion Energy The growing demand for electricity to power AI data centers is staggering. By some estimates, global electricity demand from data centers is projected to more than double by 2030 exceeding 945 terawatt-hours (TWh). It will strain electric grids and energy providers. Question 7. What plan does your company have to meet energy needs for AI, and what investments are you making into non-fossil fuel sources of energy such as fusion? Answer. We anticipate that AI's energy needs will incentivize substantial new investment in grid infrastructure and drive innovation in energy technologies. Advances in AI, including reasoning models, hold significant promise for scientific discovery, including in the field of abundant, affordable energy solutions. Indeed, our existing partnership to deploy reasoning models for use by the National Labs includes Lawrence Livermore, whose scientists were first in the world to demonstrate fusion ignition. Alongside others within the industry, we also will continue our work to find new ways to ensure our technology is as efficient as possible, including when it comes to energy consumption. Even as we continue to see promising research and innovation, we also remain focused on optimal use of available computing power, both in research and deployment. Question 8. With respect to fusion energy, how can the government partner with the private sector to scale fusion technology as it continues to develop? Answer. The scaling laws are clear. American AI leadership is a function of energy, data, and chips. Government support for fusion research and pilots can be crucial, as companies look to identify viable paths to raising the capital they need for continued scientific progress and ultimately development at scale. As the government continues to support fusion research by private entities and at our national labs, we hope that our work with the National Labs as well as early explorations with a range of fission and fusion companies provides early indication of the role that AI can play in advancing energy abundance. ______ ¶
altman ## 4. 尊敬的艾米·克洛布查尔(Amy Klobuchar)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:工程师人才培养 你在证词中说,到今年年底,OpenAI「将发布能够处理复杂软件工程任务的 AI 驱动工具」。我担心,如果 AI 取代初级工程师,我们将无法培养出资深工程师。 **问题:** 你们将如何确保 OpenAI 培养未来成功所需的人才? **答复:** 我们非常注重培养和训练软件工程人才——包括通过我们的技术在提振各层级美国教育、扩大我国中小学及经济各部门计算机科学能力方面可以发挥的作用。我们的 AI 工具与既有软件专业知识高度互补,能大幅提升小企业和大企业的能力。一个例子是,我们的工具可以减少软件工程师的重复性手工工作,让这些工程师专注于更复杂的任务,从而强化他们在批判性思维、创造力和问题解决等重要领域的技能。 Response to Written Question Submitted by Hon. Amy Klobuchar to Sam Altman Topic: Workforce Development of Engineers In your testimony you said that by the end of this year Open AI ``will release AI powered tools that can handle sophisticated software engineering.'' I'm concerned we won't be able to grow senior engineers if AI replaces junior engineers. Question. How will you ensure Open AI grows the talent needed for future success? Answer. We are very focused on fostering and training software engineering talent--including through the role our technology can play in bolstering American education at all levels and expanding computer science capabilities across our schools and all sectors of the economy. Our AI tools are highly complementary to existing software know-how and can dramatically increase the capabilities of small and large businesses. One example is how our tools can reduce a software engineer's manual work and allow those engineers to focus on more complex tasks, thereby strengthening their skills in important areas like critical thinking, creativity, and problem-solving. ______ ¶
altman ## 5. 尊敬的布莱恩·沙茨(Brian Schatz)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:工作的未来 **问题 1.** 贵公司如何利用你们所赋能的自动化来扩大生产率,同时不把劳动者落在后面? **答复:** 我们的 AI 工具在 OpenAI 内部有广泛用途,例如帮助人们完成开发代码、与客户互动、分析数据等任务。我们把先进 AI 视为帮助个人和企业获得新能力、增进创新、提升生产率的方式。随着 AI 工具能力不断增强,我们认为在各层级强化教育很重要,从小学一直到职业中期的劳动者。每个人都应当有机会从 AI 中受益。 **问题 2.** 一旦 OpenAI 在投资上「产生了数量级」的回报,你是否认为联邦政府有责任确保这些好处在美国人之间公平分配? **答复:** OpenAI 是一家相对年轻的公司,我们持续大规模投资以开发新的、能力更强的 AI 模型,并在实现我们使命所需的基础设施上投入巨大。我们在现阶段并不是一家盈利公司,不像那些拥有庞大而稳定利润率、更成熟的技术公司。我们认为,联邦政府应当确保美国人在 AI 进步时拥有获取并受益于 AI 的自由。 **问题 3.** OpenAI 的「利润上限」结构最初部分是为了减轻劳动力自动化的伤害,即用 AI 的超额利润去支持那些失去工作的人。然而,根据你们的企业重组公告,你们现在打算取消那个上限。在新结构下,你们是否仍打算支持可能被取代的劳动者? **答复:** 本月早些时候,我们重申了对 OpenAI 非营利实体掌控该组织的承诺。此前存在的营利性子公司——最初结构为「利润上限」有限责任公司——将转制为公益公司(PBC),这个新实体将继续由非营利实体控制。PBC 的使命将与非营利实体相同,即确保 AGI 造福全人类。新结构将使我们能够增强吸引资本、人才和资源的能力,同时保留我们创立时的使命:让我们的服务广泛惠及全人类。我们相信,我们实现安全且有益的先进 AI 的使命,将帮助个人和企业获得新能力、增进创新、提升生产率。 **问题 4.** OpenAI 是否会承诺,不仅为数据质量制定明确标准,也为 AI 训练数据供应链上的劳动保护和负责任做法制定明确标准? **答复:** OpenAI 在透明性方面有良好记录,包括我们的模型如何构建、我们如何确保其安全并被设计为防止各类潜在伤害。随着技术推进,我们将继续与公共和私营部门合作,提供洞见与理解。我们还公开了我们的供应商行为准则。 ### 主题:企业重组 2017 年,你说过:「这就是我们为什么是非营利:我们永远不想为了股东利益做决策。我们唯一想对其负责的,是全人类。」在此前向参议院的证词中,你解释了 OpenAI 结构中确保其忠于慈善使命的具体保障措施。2025 年 5 月 5 日,OpenAI 宣布将把其营利性业务转为公益公司(PBC),但非营利实体将保留控制权。 **问题 5.** 有哪些机制防止使命漂移,并确保 PBC 的行动与 OpenAI 的根本目标一致? **答复:** 非营利实体将控制 PBC,并成为其大股东。非营利实体与 PBC 将拥有同一使命——确保 AGI 造福全人类。我们已在这份近期声明中说明了我们的计划,包括我们对使命的持续承诺。其他 AI 实验室,如 Anthropic 和 xAI,也是 PBC;其他以使命为驱动的公司,如巴塔哥尼亚(Patagonia),也是如此。 **问题 6.** OpenAI 将用什么标准、指标或基准来评估其行动是否服务于公共利益? **答复:** 我们将继续对模型如何运作保持透明,包括发布「模型规范」,概述我们的模型如何设计、训练中纳入了哪些安全防护栏。我们还维护一套公开的「防范框架」,详述我们如何评估和缓解潜在的 AI 危害。 此外,我们的公司结构从根本上就是为了服务公共利益而设计的。作为公益公司(PBC),PBC 董事会将负有信义义务,维护我们章程中列明的公共利益目标。这些目标与 OpenAI 非营利实体的目标一致——事实上完全相同。这意味着服务公共利益不仅是一条指导原则,而是嵌入我们治理模式的法律义务。 通过把技术透明、以安全为核心的指标、以及使命一致的公司结构结合起来,我们确保我们的行动始终牢牢聚焦于推进公共利益。 **问题 7.** OpenAI 是否会承诺定期公开披露其运营、决策过程和 AI 发展情况? **答复:** 我们在这些问题上有长期的透明与参与记录,包括已发表的深度研究、系统卡、安全规范与测试信息、持续的研究项目与学术合作,以及关于我们的模型和安全工作如何开发与实施的披露。 **问题 8.** 你是否仍然同意:股东的利益与公众的利益并不相同,而且未必总与美国的安全利益一致? **答复:** 我们的使命是确保 AGI 造福全人类,这一点不会改变。我们相信,由美国及志同道合国家引领、以对自由和民主原则的承诺为锚的民主 AI,是同时确保我们的使命与美国安全利益的最佳方式。正如我们在 5 月早些时候宣布的,OpenAI 创立时是非营利组织,今天由该非营利实体监督和控制。今后,它将继续由该非营利实体控制。我们的营利性有限责任公司自 2019 年起一直隶属于该非营利实体,现将转制为公益公司(PBC)——一种以使命为驱动、必须同时考虑股东利益与使命的公司结构。非营利实体将控制 PBC,并成为其大股东,使非营利实体拥有更好的资源去支持诸多公益。我们的使命不变,PBC 将拥有同一使命。 **问题 9.** OpenAI 将如何管理利润驱动的投资者预期与其非营利使命之间的潜在张力? **答复:** 我们的使命不变——确保 AGI 造福全人类。非营利实体与 PBC 共享同一使命。 ### 主题:安全与安保 **问题 10.** 你是否同意,董事会的核心职责之一,是确保 OpenAI 的模型在发布前经过彻底测试,以确保它们不会伤害公众? **答复:** 我们进行广泛的安全测试,并说明我们的模型如何设计、以及我们为防范潜在风险所做的工作。这一流程在我们的网站上有所列明,并在大量公开文档中有所讨论和详述。去年,董事会成立了安全与安保委员会,这是一个独立的监督委员会,聚焦模型的安全与安保。该委员会听取公司领导层关于安全评估的汇报,并对重大模型发布行使监督。 我们还与美国 AI 安全研究所保持着富有成效的伙伴关系,开展聚焦国家安全风险的 AI 研究、测试与评估。 **问题 11.** 你是否同意,应当鼓励员工就美国安全所受威胁提出关切? **答复:** OpenAI 相信,开放的沟通对成功的工作环境至关重要,所有员工都应当能够无所畏惧地提出关切。我们一直努力培育这样一种文化:人们感到有责任提出潜在的安全关切,并着手解决它们。这体现在我们的《提出关切》政策中,该政策向所有 OpenAI 员工提供。 **问题 12.** OpenAI 是否维持其在拜登政府《自愿 AI 承诺》下所作的承诺? **答复:** OpenAI 继续优先考虑 AI 技术安全、可靠、透明的开发与使用。 **问题 12a.** 如果是,OpenAI 在其新的 PBC 结构下是否会继续维持这些承诺? **答复:** 无论公司结构如何,OpenAI 都将始终致力于 AI 技术安全、可靠、透明的开发与使用。 **问题 13.** OpenAI 是否维持其在《前沿 AI 安全承诺》下所作的承诺? **答复:** OpenAI 仍致力于履行 2024 年 5 月首尔 AI 峰会上作出的自愿性《前沿 AI 安全承诺》。在 2025 年 2 月巴黎 AI 行动峰会之前,我们发布了一份更新,说明我们在首尔及此前各次 AI 峰会上所作自愿承诺的进展,见此处:https://cdn.openai.com/global-affairs/paris-summit-update-on-voluntary-commitments-20250207.pdf **问题 13a.** 如果是,OpenAI 在其 PBC 结构下是否会继续维持这些承诺? **答复:** 我们将继续优先考虑 AI 技术安全、可靠、透明的开发与使用。 ### 主题:加速科学研究 **问题 14.** OpenAI 如何扩大伙伴关系或工具开发,以加速生物技术及其他科学研究的突破?为负责任地扩大这些努力,你们需要联邦机构提供什么额外支持? **答复:** 我们对我们的 AI 模型能够加速科学进步感到无比兴奋。我们已与国家实验室建立了强有力的伙伴关系以推进这些努力。我们将继续鼓励公共部门在从事科学与健康工作的各机构中采用 AI 工具。我们也强烈鼓励把 AI 纳入各层级教育的努力,这将使更多人有能力驾驭 AI 去实现科学突破。 **问题 15.** OpenAI 的模型越来越多地生成用于科学推理的内容。你认为需要哪些评估协议或基准,来确保 AI 在科学领域生成的结果稳健、可复现、最新且可信? **答复:** 不同领域对此有不同做法,但总的来说,稳健的科学应当是可复现、经同行评议的。这有助于确保科学工作的有效性、重要性与原创性,也有助于提升各领域研究的整体质量。 ### 主题:阿拉伯联合酋长国(UAE) **问题 16.** 请说明你为什么认为阿联酋是扩大 AI 合作的可信伙伴。 **答复:** 阿联酋正力争成为把 AI 融入其产业与社会的世界领导者。他们花了多年制定计划,把政府、投资者、私营部门和教育机构汇聚到这一目标上。我们认为,在美国企业推进自身计划时,美国公司不应当袖手旁观。 当我们在「OpenAI for Countries」倡议下与各国合作时,我们确保民主价值塑造 AI 的未来。为安全地做到这一点,我们需要聪明的出口管制,在创新与安全之间取得平衡,同时让各国围绕言论自由、防范监控等权利形成一致。我们正在与联邦政府合作,确保我们的国际伙伴关系符合最高标准的安全与合规要求。 **问题 17.** 阿联酋的哪些行动会让你重新考虑对其 AI 合作与协作的支持? **答复:** 我们与外国政府的伙伴关系,只有在遵循美国政府设定的安全与合规标准时才能成功。外国政府也将被要求遵守我们的使用政策。 ### 主题:能源消耗与取胜的成本 **问题 18.** 你是否支持特朗普总统为 AI 数据中心扩大煤电的努力? **答复:** 我们支持确保 AI 基础设施由可负担、可靠、可扩展的能源供电的努力——这些目标对美国的经济竞争力和就业创造至关重要。如何最好地实现这一点、如何在多种能源之间搭配,是政策制定者和更广泛能源部门的事情。 **问题 19.** 你们如何应对新基础设施建设在短期内带来的成本,包括用水、污染和气候影响方面? **答复:** 我们非常严肃地对待 AI 基础设施项目所带来的影响。我们打算积极与联邦、州和地方官员以及社区接触,负责任地管理用水和气候影响。我们的目标是确保我们的 AI 基础设施项目不仅支持技术进步,也对可持续性和生活质量作出积极贡献。 **问题 20.** 你们有什么计划来采购清洁能源,并公开报告公司的排放? **答复:** 我们预计近期的基础设施项目将包含相当比例的清洁能源成分,包括核能、太阳能发电、风能和规模化电池储能,并将遵守所有适用的环境与报告要求。我们也在积极探索如何支持先进清洁能源技术(包括先进核能与聚变)加速落地的时间表。 Response to Written Questions Submitted by Hon. Brian Schatz to Sam Altman Future of Work Question 1. How is your company taking advantage of the automation you're empowering to scale productivity without leaving workers behind? Answer. Our AI tools are used in a wide range of ways at OpenAI, such as helping people with tasks like developing code, interacting with customers, and analyzing data. We see advanced AI as a way to help people and businesses gain new capabilities, increase innovation, and uplift productivity. As AI tools become increasingly capable, we think it's important to bolster education at all levels, from elementary schools through mid-career workers. Everyone should have the opportunity to benefit from AI. Question 2. Once OpenAI has ``generated orders of magnitude'' of returns on investments, do you believe the Federal government has any responsibility to make sure those benefits are distributed equitably across all Americans? Answer. OpenAI is a relatively young company and we continue to invest at scale to develop new and more capable AI models, and we invest heavily in the infrastructure required to achieve our mission. We are not a profitable company at this stage, unlike more mature technology firms that have large, sustained profit margins. We believe the Federal government should ensure Americans have the freedom to access and benefit from AI as it advances. Question 3. OpenAI's capped profit structure was originally designed, in part, to mitigate the harms of workforce automation by using excess profits from AI to support those who lost their jobs. However, per your corporate restructuring announcement, you now intend to remove that cap. Do you still intend to support potential displaced workers under this new structure? Answer. Earlier this month we reaffirmed our commitment to the OpenAI nonprofit having control over the organization. The previously existing for-profit subsidiary--originally structured as a ``capped- profit'' LLC--will be converted into a Public Benefit Corporation (PBC), and this new entity will remain under the control of the nonprofit. The PBC's mission will be the same as the nonprofit's, which is to ensure AGI benefits all of humanity. The new structure will allow us to strengthen our ability to attract capital, talent, and resources, while preserving our founding mission to make our services broadly available to all of humanity. We believe our mission of achieving safe and beneficial advanced AI will help people and businesses gain new capabilities, increase innovation, and uplift productivity. Question 4. Will OpenAI commit to developing clear standards not only for data quality, but for labor protections and responsible practices across the AI training data supply chain? Answer. OpenAI has a strong track record of transparency around how our models are built and how we ensure they're safe and designed to prevent a wide range of potential harms. We will continue to work with the public and private sectors to provide insights and understanding as our technology advances. We also have made public our supplier code of conduct. Corporate Restructuring In 2017, you said ``That's why we're a nonprofit: we don't ever want to be making decisions to benefit shareholders. The only people we want to be accountable to is humanity as a whole.'' In your previous testimony before the Senate, you explained the specific safeguards in OpenAI's structure that ensure it remains true to its charitable mission. On May 5, 2025, OpenAI announced that it would transition its for-profit operations to a Public Benefit Corporation (PBC), but that the nonprofit would retain control. Question 5. What mechanisms are in place to prevent mission drift and ensure that the PBC's actions align with OpenAI's foundational goals? Answer. The nonprofit will control and be a large shareholder of the PBC. Both the nonprofit and the PBC will have the same mission--to ensure that AGI benefits all of humanity. We have described our plans, including our ongoing commitment to our mission, in this recent statement. Other AI labs, like Anthropic and xAI, are also PBCs, as are other purpose driven companies like Patagonia. Question 6. What criteria, metrics, or benchmarks will OpenAI use to evaluate whether its actions serve the public interest? Answer. We will continue to be transparent about how our models work, including publishing a ``model spec'' that outlines how our models are designed and the safety guardrails incorporated into training. We also maintain a public preparedness framework, which details how we evaluate and mitigate potential AI harms. In addition, our corporate structure is fundamentally designed to serve the public interest. As a Public Benefit Corporation (PBC), the PBC board of directors will hold a fiduciary duty to uphold the public benefit objectives outlined in our charter. These objectives are aligned with--and in fact identical to--those of the OpenAI nonprofit. This means that serving the public interest is not just a guiding principle, but a legal obligation embedded in our governance model. By combining technical transparency, safety-focused metrics, and a mission-aligned corporate structure, we ensure that our actions remain squarely focused on advancing the public good. Question 7. Will OpenAI commit to regular public disclosures about its operations, decision-making processes, and AI developments? Answer. We have a long track record of transparency and engagement on these issues, including published, in-depth research, system cards, safety specifications and testing information, ongoing research programs and academic partnerships, and disclosures about how our models and safety work are developed and implemented. Question 8. Do you still agree that the interests of your shareholders are not the same as the interests of the public, and might not always be aligned with America's security interests? Answer. Our mission is to ensure that AGI benefits all of humanity, and that will not change. We believe that advancing democratic AI, led by the U.S. and like-minded nations and anchored by a commitment to freedom and democratic principles, is the best way to ensure both our mission and America's security interests. As we announced earlier in May, OpenAI was founded as a nonprofit, and is today overseen and controlled by that nonprofit. Going forward, it will continue to be controlled by that nonprofit. Our for-profit LLC, which has been under the nonprofit since 2019, will transition to a Public Benefit Corporation (PBC)--a purpose-driven company structure that has to consider the interests of both shareholders and the mission. The nonprofit will control and also be a large shareholder of the PBC, giving the nonprofit better resources to support many benefits. Our mission remains the same, and the PBC will have the same mission. Question 9. How will OpenAI manage potential tensions between profit-driven investor expectations and its nonprofit mission? Answer. Our mission remains the same--to ensure that AGI benefits all of humanity. The nonprofit and PBC share the same mission. Safety and Security Question 10. Do you agree that one of the board's core responsibilities is to ensure OpenAI's models are thoroughly tested before their release to ensure they won't harm the public? Answer. We conduct extensive safety testing and outline both how our models are designed and the work we do to safeguard against potential risks. This process is outlined on our website and is discussed and detailed in extensive public documentation. Last year, the Board formed the Safety and Security Committee, an independent oversight committee focused on model safety and security. The Committee is briefed by company leadership on safety evaluations and exercises oversight over major model releases. We also maintain a productive partnership with the U.S. AI Safety Institute that enables AI research, testing, and evaluation focused on national security risks. Question 11. Do you agree that employees should be encouraged to raise concerns about threats to the security of the United States? Answer. OpenAI believes that open communication is essential to a successful work environment and that all employees should feel free to raise issues of concern without fear of reprisal. We have worked to foster a culture where people feel a responsibility to raise potential safety concerns and work to address them. This is encompassed in our Raising Concerns policy, which is made available to all OpenAI employees. Question 12. Does OpenAI maintain the commitments it made under the Biden Administration Voluntary AI Commitments? Answer. OpenAI continues to prioritize the safe, secure, and transparent development and use of AI technology. Question 12a. If so, will OpenAI continue to maintain these commitments under its new PBC structure? Answer. Regardless of our corporate structure, OpenAI will always be committed to the safe, secure, and transparent development and use of AI technology. Question 13. Does OpenAI maintain the commitments it made under the Frontier AI Safety Commitments? Answer. OpenAI remains committed to fulfilling the voluntary Frontier AI Safety Commitments made at the Seoul AI Summit in May 2024. In advance of the Paris AI Action Summit in February 2025, we published an update showing our progress on the voluntary commitments made at Seoul and at previous AI summits, available here. [https:// cdn.openai.com/global-affairs/paris-summit-update-on-voluntary- commitments-20250207.pdf] Question 13a. If so, will OpenAI continue to maintain these commitments under its PBC structure? Answer. We will continue to prioritize the safe, secure, and transparent development and use of AI technology. Accelerating Scientific Research Question 14. How is OpenAI expanding partnerships or tool development to accelerate breakthroughs in biotechnology and other scientific research, and what additional support do you need from Federal agencies to scale these efforts responsibly? Answer. We are incredibly excited about how our AI models can accelerate scientific progress. We have established a strong partnership with the National Laboratories to further these efforts. We would continue to encourage public sector uptake of AI tools across institutions that work on science and health. And we strongly encourage efforts to incorporate AI into education at all levels, which will equip many more people to harness AI for scientific breakthroughs. Question 15. OpenAI models increasingly generate content used in scientific inference. What evaluation protocols or benchmarks do you believe are necessary to ensure that AI-generated results in scientific fields are robust, reproducible, up-to-date, and trustworthy? Answer. Different fields have different approaches to this, but generally speaking, robust science should be reproducible and peer reviewed. This helps ensure the validity, significance, and originality of scientific work and helps improve the overall quality of research across various fields. United Arab Emirates (UAE) Question 16. Please share why you believe the UAE is a trustworthy partner for increased collaboration on AI. Answer. The UAE is aiming to be a world leader in integrating AI into their industry and society. They have spent years developing plans that bring together government, investor, private sector and educational institutions towards this end. We believe it's important that American companies don't sit on the sidelines as they proceed with their plans. As we partner with countries under our OpenAI for Countries initiative we are ensuring that democratic values shape the future of AI. To do this securely, we need smart export controls that balance innovation and safety, while aligning nations around rights like free expression and safeguards against surveillance. We are working with the Federal government to ensure our international partnerships meet the highest standards of security and compliance. Question 17. What actions by the UAE would cause you to reconsider support for cooperation and collaboration on AI? Answer. Our partnerships with foreign governments will only succeed by following the standards of security and compliance set forth by the United States government. Foreign governments will also be required to comply with our usage policies. Energy Consumption and Cost of Winning Question 18. Do you support President Trump's efforts to expand coal power for AI data centers? Answer. We support efforts to ensure AI infrastructure is powered by energy that is affordable, reliable, and scalable--goals that are critical to U.S. economic competitiveness and job creation. How best to achieve that, across a mix of energy sources, is a matter for policymakers and the broader energy sector. Question 19. How are you addressing the costs associated with new infrastructure development in the short-term, including in terms of water consumption, pollution, and climate impacts? Answer. We take very seriously the impacts associated with AI infrastructure projects. We intend to actively engage with federal, state, and local officials, as well as communities, to responsibly manage water usage and climate impacts. Our aim is to ensure that our AI infrastructure projects not only support technological advancement but also contribute positively to sustainability and quality of life. Question 20. What plans do you have to source clean energy and to publicly report your companies' emissions? Answer. We expect our near-term infrastructure projects to incorporate significant clean energy components, including nuclear energy, solar generation, wind and large-scale battery storage, as well as to comply with all applicable environmental and reporting requirements. We are also actively exploring how we can support accelerated timelines for advanced clean energy technologies, including advanced nuclear and fusion. ______ ¶
altman ## 6. 尊敬的爱德华·马基(Edward Markey)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:数据中心建设的综合影响 **问题 1.** 在规划数据中心建设时,贵公司是否会进行「从摇篮到坟墓」的基础设施研究,涵盖数据中心运营寿命期间及之后对野生动物、社区和污染的影响?如果是,你们从这些研究中学到了什么?如果不是,为什么? **答复:** 在我们持续的选址过程中,我们依据所有适用的联邦和州环境法律评估选址,并利用行业经验、顾问和最佳实践。我们与甲骨文(Oracle)、微软等基础设施伙伴密切合作,迄今为止这些活动由它们负责。 ### 主题:备用发电 **问题 2.** 贵公司是否在任何设施使用备用柴油发电机? **问题 2a.** 如果是,请提供每个使用柴油发电机的设施清单,以及发电机的位置、数量和类型。 **问题 2b.** 如果是,贵公司是否考虑过使用电池储能技术作为柴油发电机的替代方案?请解释你们的决策过程。 **答复:** 我们目前不使用备用柴油发电机。在我们的选址与设计过程中,我们会考虑柴油发电机的替代方案,包括电池储能系统(BESS)技术。我们有电池备用电源。 ### 主题:能源结构 **问题 3.** 贵公司是否使用任何现场或同址能源发电来为数据中心供电? **问题 3a.** 如果是,请详细说明有多少电力来自现场和同址发电。 **问题 3b.** 如果是,请列出所有用于为贵公司数据中心供电的现场和同址能源(例如可再生、核能、水电、燃气轮机等)。 **答复:** 在开发我们的 AI 基础设施时,我们将同时考虑现场发电和公用事业供电,包括同址能源。 **问题 4.** 贵公司如何确保本地电费缴纳者不必为满足数据中心能源需求所需的新能源基础设施(例如输电线路)付费? **答复:** 我们打算与公用事业公司、监管者和地方政府积极协作,确保与我们的数据中心相关的基础设施成本不会不公平地加重本地电费缴纳者的负担,为所有能源消费者维持公平与可负担性。 ### 主题:AI 模型训练中的能耗 **问题 5.** 你们估计训练 GPT-4.1 以及任何新 AI 模型用了多少吉瓦时(GWh)能源? **答复:** 我们正在积极努力降低 AI 开发各阶段的能耗,从初始训练到实际推理。我们持续投资于硬件与软件效率的改进,并与研究者、政策制定者和行业伙伴密切合作,以建立更强的行业标准和可信、开放的基准,例如 ML Energy Initiative 和 Epoch AI 所开发的那些。 在探索提升效率的新方式时,我们始终致力于审慎、负责任地使用算力资源。像模型蒸馏这样的创新——让我们能够训练大型、全面的模型,随后再精炼为面向特定应用的更小、极高效的版本——是这一战略的关键部分。随着算法与 AI 技术持续演进,我们预期能效与资源节约(包括用水减少)都会持续改善。 **问题 6.** 数据中心降温需要大量水。当水量处于危急低位时,贵公司是否会继续抽取用于建筑降温的水?是否有应急方案,以避免对供水造成进一步压力、避免从家庭、农业或小企业那里占用有限资源? **答复:** 我们目前正在评估确保可持续用水的策略,包括在选址中、通过先进冷却技术、以及通过优先考虑社区用水需求的运营调整与应急方案,从而确保负责任的资源管理,并尽量减少对当地农业、家庭和小企业的潜在影响。 ### 主题:政府伙伴关系 **问题 7.** 贵公司专门为公共部门提供 AI 产品,这些产品现已在联邦各机构以及州和地方政府中使用。鉴于政府使用 AI 所涉风险尤为突出——包括剥夺权利或服务获取资格、以及关于政府福利与项目的虚假或错误信息——你们采取了哪些额外步骤来确保这些工具在政府情境下安全有效? **答复:** 我们向消费者、企业和政府提供先进 AI 工具。每一种用例都不同,但每一种都需要安全且有益的 AI。我们在网站上列明了我们的安全与安保工作,并发布了「模型规范」,详细说明我们在模型开发中如何工程化地实现安全。 **问题 8.** 你们有什么协议与政府机构合作,在伤害或错误发生时予以纠正? **答复:** 我们征求客户和伙伴(包括政府机构)的反馈。我们不决定特定客户将如何使用我们的模型,但努力确保这些模型能力强大且部署安全。我们还通过「信任与透明」报告资源,提供一条安全的伤害报告途径。 ### 主题:商业伙伴关系 **问题 9.** 当你把 AI 系统提供给其他用户/部署者时,你们会就哪些类型的问题达成一致? **答复:** 我们为各类产品和工具(包括 API)制定了使用政策;企业通常以 API 为基础构建新产品和服务。我们的使用政策是公开的。 **问题 10.** 你们向那些其他方/部署者提供什么信息? **答复:** 我们分享关于模型如何训练、如何被设计为安全、以及我们的使用政策的公开信息。 **问题 11.** 贵公司采取什么步骤来降低这样一种可能性:下游用户以可能伤害他人、或违反你们使用条款的方式使用你们的服务? **答复:** 我们结合使用自动化系统、人工审核和用户举报来发现并评估可能违反我们政策的使用行为。违规可能导致对内容或客户账户采取措施,例如警告、分享限制,或取消纳入我们服务的资格。 ### 主题:AI 幻觉 近期报道表明,随着技术能力增强,生成式 AI 的幻觉正在变得更糟。在某些情形下,幻觉可能造成重大伤害,例如在评估求职申请时,或者更危险地,在国家安全情境中。 **问题 12.** 你是否同意,我们需要防护栏来确保 AI 工具不被以可能伤害他人的方式使用或误用? **答复:** 是的,我们有旨在防范潜在伤害的使用政策。 **问题 13.** 贵公司正在采取什么步骤来应对这个问题? **答复:** 我们有旨在防范潜在伤害的使用政策。 Response to Written Questions Submitted by Hon. Edward Markey to Sam Altman Comprehensive Impacts of Data Center Construction Question 1. When planning for data center construction, does your company conduct a cradle-to-grave infrastructure study that includes wildlife, community, and pollution impacts during and beyond the operational lifespan of a data center? If yes, what have you learned from those studies? If no, why not? Answer. In our ongoing site selection process, we are assessing sites in accordance with all applicable Federal and state environmental laws, as well as utilizing industry experiences, consultants, and best practices. We work closely with our infrastructure partners such as Oracle and Microsoft which have been responsible for these activities to date. Backup Energy Generation Question 2. Does your company use backup diesel generators at any facilities? Question 2a. If yes, please provide a list of each facility where diesel generators are being used, along with the location, quantity, and type of generators. Question 2b. If yes, did your company consider the use of battery storage technology as an alternative to diesel generators? Please explain your decision process. Answer. We are not currently utilizing backup diesel generators. In our site selection and design process, we will consider alternatives to diesel generators, including BESS technologies. We have battery backup power. Energy Mix Question 3. Does your company utilize any on-site or colocated energy generation to power your data centers? Question 3a. If yes, please provide detail how much power comes from on-site and colocated energy generation. Question 3b. If yes, please list all on-site and colocated energy sources (e.g., renewable, nuclear, hydropower, gas-powered turbines, etc.) that are being utilized to power your data centers. Answer. In developing our AI infrastructure, we will consider both on-site and utility power, including co-located energy sources. Question 4. How does your company ensure local ratepayers are not responsible for paying the cost of new energy infrastructure, such as transmission lines, needed to meet the data center's energy demand? Answer. We intend to actively collaborate with utilities, regulators, and local governments to ensure that infrastructure costs associated with our data centers do not unfairly burden local ratepayers, maintaining fairness and affordability for all energy consumers Energy Consumption in AI Model Training Question 5. How many GWh of energy do you estimate was used to train GPT-4.1 and any new AI models? Answer. We are actively working to reduce energy consumption across all stages of AI development, from initial training to practical inference. We invest continuously in improvements to both hardware and software efficiency, collaborating closely with researchers, policymakers, and industry partners to establish stronger industry standards and credible, open benchmarks, such as those developed by ML Energy Initiative and Epoch AI. As we explore new ways to enhance efficiency, we remain committed to thoughtful and responsible use of computing resources. Innovations like model distillation, which allows us to train large, comprehensive models that can subsequently be refined into smaller, highly efficient versions for targeted applications, are a key part of this strategy. As algorithms and AI technologies continue evolving, we anticipate ongoing improvements in both energy efficiency and resource conservation, including reductions in water usage. Question 6. Data centers require vast amounts of water for cooling. When water is at critically low levels, does your company continue to pull water for building cooling? Does it have a contingency for operating as to not put further stress on the water supply and potentially take limited resources from households, agriculture, or small businesses? Answer. We are currently evaluating strategies to ensure sustainable water use, including in site selection, through advanced cooling technologies, and via operational adjustments and contingency plans that prioritize community water needs, ensuring responsible resource management and minimizing potential impacts on local agriculture, households, and small businesses. Government Partnerships Question 7. Your company offers AI products specifically for the public sector, which are now used across Federal agencies and state and local governments. Given the especially heightened risks related to governments' use of AI--including the denial of rights or access to services and false or incorrect information about government benefits and programs--what additional steps have you taken to ensure that these tools are safe and effective to use in the context of government? Answer. We provide advanced AI tools to consumers, businesses, and governments. Each of these use cases is different, but each requires safe and beneficial AI. We have outlined our safety and security work on our website and have published a ``model spec'' which provides extensive details about how we engineer safety in model development. Question 8. What protocols do you have in place to work with government agencies to rectify any harms or errors when they occur? Answer. We solicit feedback from customers and partners, including government agencies. We don't decide how our models will be used by particular customers but work to ensure these models are highly capable and safe to deploy. We also provide a secure way to report harms through our Trust & Transparency reporting resources. Business Partnerships Question 9. When you make your AI systems available to other users/ deployers, what are the types of issues you agree on? Answer. We have usage policies in place for our various products and tools, including our API, which businesses typically use as a base to build new products and services. Our usage policies are publicly available. Question 10. What information do you provide to those other parties/deployers? Answer. We share publicly available information about how the models are trained, how they are designed to be safe, and our usage policies. Question 11. What step does your company take any steps to reduce the likelihood that, for instance, those downstream users don't use your services in ways that could harm people or that violate your terms of use? Answer. We use a combination of automated systems, human review, and user reports to find and assess uses that potentially violate our policies. Violations can lead to actions against the content or a customer account, such as warnings, sharing restrictions, or ineligibility for inclusion in our services. AI Hallucinations Recent reporting suggests that generative AI hallucinations are getting worse as the technologies become more powerful. Hallucinations can lead to great harm in certain scenarios, such as when assessing job applications, or even more dangerously in the national security context. Question 12. Do you agree that we need guardrails to ensure that AI tools are not used or misused in ways that could cause harm to people? Answer. Yes and we have usage policies that are designed to protect against potential harms. Question 13. What steps is your company taking to address this issue? Answer. We have usage policies that are designed to protect against potential harms. ______ ¶
altman ## 7. 尊敬的塔米·鲍德温(Tammy Baldwin)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) **问题:** 你曾公开谈到,你认为 AI 可以用来推进医学研究、改善患者结局。然而,我担心在开发这类模型时敏感个人信息会遭到泄露,也担心系统可能固化偏见。能否谈谈你设想 AI 在医疗照护中扮演什么角色,以及需要哪些保护措施来保障患者隐私? **答复:** OpenAI 已在 AI 全生命周期(从模型开发到模型部署)中开发并实施了一系列隐私保障措施。我们在训练阶段做出大量努力,以限制个人信息的存在,并降低模型输出敏感个人信息的可能性;我们还向用户提供信息与工具,以回应其隐私关切。我们坚决承诺保障我们从用户处获得或关于用户的一切信息的安全。这些保障措施在我们的隐私政策中有所列明。 我们已经看到我们的工具被用于加速科学研究、帮助对抗疾病。我们在医学研究和临床照护方面都有合作。我们已列明与医疗照护提供方合作的方式,包括面向患者的隐私与安全设置。这对确保 AI 工具在医疗照护中被安全使用至关重要——在医疗领域,AI 是一个改变游戏规则的东西。 Response to Written Question Submitted by Hon. Tammy Baldwin to Sam Altman Question. You have spoken publicly on how you believe AI can be used to advance medical research and improve patient outcomes. However, I am concerned about sensitive personal information being compromised in the development of such models, and how systems may intrench bias. Can you share what role you envision AI having in health care and what protections need to be in place to safeguard patient privacy? Answer. OpenAI has developed and implemented a range of privacy safeguards across the AI lifecycle, from model development to model deployment. We undertake extensive efforts at the training stages to limit the presence of personal information and decrease the likelihood that our models output sensitive personal information, and we provide users information and tools to address their privacy concerns. We are strongly committed to keeping secure any information we obtain from users or about users. These safeguards are outlined in our privacy policy. We are already seeing our tools being used to accelerate scientific research and help fight disease. We have partnerships across medical research and clinical care. We have outlined our approach to working with healthcare providers, including privacy and security settings for patients. This is critical to ensuring AI tools are used safely in healthcare, where AI is a game changer. ______ ¶
altman ## 8. 尊敬的杰基·罗森(Jacky Rosen)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) ### 主题:模型安全 **问题 1.** 确保我们使用的 AI 模型不成为又一个网络安全漏洞,至关重要。针对大型 AI 模型或高风险模型、以及其训练所依托的基础设施的自愿性网络安全标准,是否有助于建立信任? **答复:** 我们致力于持续提升网络安全实践,既从安全角度(保护我们的用户、系统和知识产权),也从安全(safety)视角(为可能造成伤害的两用前沿能力,例如网络能力,作准备)。 为确保我们模型与系统的安全,我们正在利用自身的 AI 技术来扩展网络防御、保护我们的用户、系统和知识产权;与第三方网络安全专家合作,通过贴近实战的红队测试严格检验我们的网络防御;并与伙伴一道,努力采用零信任架构、硬件支持的安全方案等业界领先的安全实践。 为确保我们模型的安全(safety),我们依照《防范框架》严格评估模型能力,并发布大量文档,说明我们内建于模型的保障措施以及我们为确保安全与安保所采用的方法。我们与网络安全专家密切合作,对我们的模型开展严格的第三方评估,包括与专注应对网络风险的政府机构合作。最后,我们持续监测并阻断恶意行为者利用我们技术的企图。 ### 主题:AI 测试与评估(T&E)人才 确保美国继续在 AI 竞赛中领先世界的一个关键因素,是确保我们开发的 AI 是最好的、因而也是最可信的。验证模型输出是建立信任的重要一步。然而,眼下美国既没有标准,也没有受过训练的人才来评估 AI 模型,以确立我们可以信任模型输出。 **问题 2.** 国会应当考虑什么来激励并壮大 AI 测试与评估人才队伍? **答复:** 我们支持国会采取行动,进一步拓展教育机会,以壮大未来的 AI 劳动力。此外,OpenAI 与美国 AI 安全研究所保持着富有成效的伙伴关系,开展聚焦国家安全风险的 AI 研究、测试与评估。这一伙伴关系依赖政府一侧技术精湛的专家。更广泛地说,随着 AI 工具在政府各部门被采用,在联邦劳动力中提升 AI 技能与素养将很重要,以确保有一支有能力的队伍来进行 AI 开发与实施,并开展你所描述的那类 AI 安全工程与测试。 **问题 3.** 国会如何支持对 AI 测试与评估采取更多跨学科的方法?例如,我们如何确保一个用于医疗照护场景的模型,既经过模型技术专家的评估,也经过其所部署的医疗照护场景专家的评估? **答复:** 不同的 AI 用例将需要不同的方法。我们与金融服务、医疗照护等受监管行业广泛合作,确保我们的工具符合其监管要求。这对确保 AI 在这些行业被采用、并满足安全与安保要求非常重要。 Response to Written Questions Submitted by Hon. Jacky Rosen to Sam Altman Model Security Question 1. It's essential we ensure the AI models we use do not become another cybersecurity vulnerability. Would voluntary cybersecurity standards for large AI models or high-risk models and the infrastructure they were trained on be helpful in establishing trust? Answer. We are committed to continuously upleveling our cybersecurity practices, both from a security perspective (protecting our users, systems, and intellectual property) and from a safety lens (preparing for dual-use frontier capabilities, such as cyber, that have the potential to cause harm). To ensure the security of our models and systems, we are leveraging our own AI technology to scale our cyber defenses and protect our users, systems, and intellectual property; partnering with third-party cybersecurity experts to rigorously test our cyber defenses through realistic red-teaming; and working to adopt industry-leading security practices such as zero-trust architectures and hardware-backed security solutions, together with our partners. To ensure the safety of our models, we rigorously evaluate model capabilities consistent with our Preparedness Framework, and publish extensive documentation about how safeguards we have built into our models and approaches we use to ensure safety and security. We work closely with cybersecurity experts to conduct rigorous third-party assessments of our models, including with government agencies that are focused on combatting cyber risks. Lastly, we continuously monitor and disrupt attempts by malicious actors to exploit our technology. AI T&E Workforce A key factor in ensuring the U.S. continues to lead the world in the AI race is by ensuring the AI we develop is the best and therefore the most trustworthy. Validating model outputs is an important step in establishing trust. Right now, however, the U.S. has neither the standards nor the trained workforce to evaluate AI models to establish that we can trust model outputs. Question 2. What should Congress consider to incentivize and grow the AI test and evaluation workforce? Answer. We support congressional action to further educational opportunities to grow the AI workforce of the future. Additionally, OpenAI maintains a productive partnership with the U.S. AI Safety Institute that enables AI research, testing, and evaluation focused on national security risks. This partnership relies on skilled technical experts on the government side. More broadly, as AI tools are adopted across government, it will be important to promote AI skills and literacy across the Federal workforce, in order to ensure a capable workforce for AI development and implementation, and for the type of AI safety engineering and testing that you describe. Question 3. How can Congress support more interdisciplinary approaches to testing and evaluating AI? For example, how do we ensure a model being used in a healthcare setting has been evaluated both by experts in the model technology, but also experts in the healthcare setting in which it will be deployed? Answer. Different AI use cases will require different approaches. We have worked extensively with regulated industries like financial services and healthcare to ensure our tools comply with their regulatory requirements. This is very important to ensuring AI is adopted across these sectors and that safety and security requirements are met. ______ ¶
altman ## 9. 尊敬的莉萨·布朗特·罗切斯特(Lisa Blunt Rochester)提交的书面问题 → 萨姆·奥尔特曼(Sam Altman) 奥特曼先生,AI 正日益融入我们关键的经济与社会基础设施。麦肯锡(McKinsey)称,从企业用例看,AI 的长期机会可能是约 4.4 万亿美元的额外生产率增长潜力。 但供应商锁定可能是个真实问题:某家 AI 供应商大幅落后于竞争,给客户留下一个远为逊色的产品,这可能威胁该 AI 产品所服务的关键行业。 **问题:** 对于在金融、医疗等关键行业运营的 AI 产品,你们是否有任何计划或策略来缓解锁定,以防可能伤害这些关键行业及其从业者的潜在锁定效应? **答复:** AI 是一个竞争非常激烈的领域,很多公司都在开发先进的模型能力和新产品。最先进水平在快速推进。企业通常同时使用多个模型、多个开发者,从而降低被某一特定供应商锁定的风险。我们认为,重要的是消费者和企业能够选择最好的 AI 模型,以及运行这些模型的最好云基础设施。 Response to Written Question Submitted by Hon. Lisa Blunt Rochester to Sam Altman Mr. Altman, AI is becoming integrated into our critical economic and societal infrastructure, with McKinsey stating that long-term AI opportunity could be about $4.4 trillion in added productivity growth potential from corporate use cases. But vendor lock in could be a real issue, where an AI vendor dramatically falls behind the competition and leaves its client with a vastly inferior product, which could threaten key industries the AI product operates in. Question. Do you have any plans or strategies regarding mitigating lock-in for your AI products operating in critical sectors, like the financial and medical sectors, to prevent potential lock-in effects that might harm these critical sectors and the folks therein? Answer. AI is a highly competitive space, with lots of companies developing advanced model capabilities and new products. The state of the art is advancing rapidly. Firms typically work with multiple models and multiple developers, reducing the risk of lock-in to a particular provider. We think it's important that consumers and businesses are able to choose the best AI models and the best cloud infrastructure on which to run those models. ______ ¶
Ms. Su ## 10. 尊敬的托德·杨(Todd Young)提交的书面问题 → 苏姿丰博士(Dr. Lisa Su) ### 主题:AI 公众认知与教育 要赢下扩散竞赛,不仅需要为技术与应用更广泛地进入一般商业流程提供路径,还需要加强公众对 AI 的认知与教育。 **问题:** 苏博士,你愿意就此谈谈吗?特别是就培养一支能够解决更先进科学研发挑战的劳动力而言。如果这支劳动力不在本国,它会去哪里,后果又是什么? **答复:** 美国在人工智能领域的领导地位最终建立在人才之上。我们长期竞争力唯一的、最大的决定因素,是我们能否吸引、培养并留住那些将设计、建造和治理这些系统的人。今天,在三大类别上都存在关键缺口: - **先进 AI 研究与系统工程。** 能够设计和扩展前沿 AI 模型的博士级人才严重短缺,这包括机器学习、算法优化和系统协同设计方面的专长。该领域许多顶尖研究者正被海外机构积极招揽——并留住——往往有国家主导战略的支持。 - **应用 AI 与部署人才。** 我们也缺少足够的工程师,能够安全、负责任地把 AI 整合进医疗照护、制造、国家安全和物流等关键行业。这些是跨学科角色,要求同时精通 AI 及其所应用的领域。如果我们想让 AI 创新转化为广泛的经济收益,弥合这一缺口至关重要。 - **技术基础设施与硬件专门化。** AI 领导地位要求在底层硬件栈上领先——从芯片架构、互连设计到先进封装与电源管理。美国长期短缺半导体工程师、固件开发者和熟练技术员。这些岗位对让 AI 模型高性能、高效率、可扩展至关重要。 弥合这些缺口所需的,不只是研究生奖学金或大学经费。它需要一种协调一致的国家路径——从 K-12 的 STEM 教育,到社区大学的通路,到高技能移民的签证政策,再到让学生和劳动者真正接触前沿 AI 开发的公私伙伴关系。 如果这支劳动力不在美国,它就会在别的国家,包括战略竞争对手那里,对他们有利、对我们不利。这也包括能够吸引并留住经过适当审查的外国人才——他们中许多人希望在这里学习、工作和生活,为我们的技术领导地位作贡献。 如果我们想在 AI 上引领世界,我们就必须在人才上引领世界。这是根基——而眼下,我们在太多领域都在追赶。 Response to Written Question Submitted by Hon. Todd Young to Dr. Lisa Su AI Public Awareness and Education Winning the diffusion race not only requires providing a pathway for greater adoption of the technology and its applications into the general stream of commerce but also bolstering our public awareness and education of AI. Question. Dr. Su, would you like to comment on this, especially as it relates to building a workforce capable of solving more advanced scientific R&D challenges? If the workforce isn't here at home, where will it go and what are the consequences of that? Answer. U.S. leadership in artificial intelligence is ultimately based on talent. The single greatest determinant of our long-term competitiveness is whether we can attract, train, and retain the people who will design, build, and govern these systems. Today, there are critical gaps across three main categories: Advanced AI research and systems engineering. There is a significant shortfall of PhD-level talent capable of designing and scaling frontier AI models. This includes expertise in machine learning, algorithm optimization, and systems co- design. Many of the leading researchers in this field are being aggressively recruited--and retained--by overseas institutions, often with the backing of state-directed strategies. Applied AI and deployment talent. We also lack enough engineers who can safely and responsibly integrate AI into critical sectors like healthcare, manufacturing, national security, and logistics. These are cross-disciplinary roles that require fluency in both AI and the domain in which it's being applied. Bridging that gap is essential if we want AI innovation to translate into broad-based economic benefit. Technical infrastructure and hardware specialization. AI leadership requires leading in the underlying hardware stack-- everything from chip architecture and interconnect design to advanced packaging and power management. The U.S. has a chronic shortage of semiconductor engineers, firmware developers, and skilled technicians. These roles are vital for making AI models performant, efficient, and scalable. Addressing these gaps requires more than just graduate fellowships or university funding. A coordinated, national approach--from STEM education in K-12 to community college pathways, visa policy for high- skill immigrants, and public-private partnerships that give students and workers real exposure to cutting-edge AI development. If the workforce is not in the United States, it will be in other countries, including strategic competitors, to their benefit and our detriment. This includes being able to attract and retain properly vetted foreign talent, many of whom seek to study, work and live here to contribute to our technology leadership. If we want to lead the world in AI, we must lead the world in talent. That is the foundation--and right now, we are playing catch-up in too many areas. ______ ¶
Ms. Su ## 11. 尊敬的玛丽亚·坎特韦尔(Maria Cantwell)提交的书面问题 → 苏姿丰博士 ### 主题:AI 标准 美国与世界上最先进的民主国家一道推动 AI 标准的发展,这为我们提供了在全球舞台上为 AI 设定「道路规则」的机会。 **问题 1.** NIST 标准如何有助于美国的竞争力? **答复:** NIST 的《风险管理框架》,包括它与其他 AI 框架和国际标准的比较,有助于美国竞争力,因为它是通过透明的公开程序制定的,从而提升了客户信心。NIST 已在网络安全和 AI 等领域成功运用这一框架,AMD 也一直倚重它。 **问题 2.** 你希望 NIST 制定或推广哪些标准,以提升美国的竞争力? **答复:** 有些标准(如芯片设计)最好由产业联盟来推动;而保护国家安全的标准,如网络安全和硅来源可溯(silicon provenance),最好由 NIST 与产业界合作推广,以提升美国竞争力。 **问题 3.** 贵公司在人工智能开发中如何受益于国家科学基金会、NIST 或能源部各实验室,或与之合作? **答复:** 通过与能源部十多年的伙伴关系,AMD 现在为世界上两台最快的超级计算机提供动力:2021 年在橡树岭国家实验室投入运行的「前沿」(Frontier),以及去年年底在劳伦斯利弗莫尔国家实验室投入运行的「埃尔卡皮坦」(El Capitan)。这些系统是美国国家安全与科学领导地位的关键基础设施,支撑着药物发现、医学研究、气候研究、高超声速飞行乃至训练未来几代更强 AI 模型的最新进展。 **问题 4.** 削减 NSF 拨款将如何影响你们的劳动力和人才搜寻? **答复:** NSF 对美国高校科学与工程研究和教育的资助,为美国半导体劳动力管道提供了形成性的训练场。我们支持继续为这些目的提供资金,以维持美国的 AI 领导地位。 **问题 5.** 削减联邦对大学科学与研究的拨款,将对美国在 AI 领域的竞争力产生什么影响? **答复:** 美国在 AI 的竞争力建立在五个优先事项上:(1)加速美国芯片与系统创新,以保持在 AI 算力基础设施上的领先——美国大学在其中发挥了关键作用;(2)开放生态,让大学能直接贡献创新;(3)为先进半导体制造与封装提供美国研究经费;(4)以研究投资驱动美国 AI 人才发展;(5)网络技术研究,以强化美国防御能力。联邦对美国大学中与这些优先事项相关研究的资助,对维持我们的领先至关重要。若可用于这些优先事项的联邦资助减少,私营部门将需要提供额外资金与支持。 **问题 6.** 就你们的供应链而言,你最担心什么? **答复:** 作为一家无晶圆厂(fabless)公司,AMD 专注于确保其所设计的产品能够顺利生产并交付客户。因此,AMD 支持让这些努力得以继续的贸易政策,包括不损害当前提升美国半导体制造能力工作的政策。我们必须为先进半导体制造与封装建立强健的国内供应链。AI 领导地位取决于建造完整、集成系统的能力。这意味着我们要在最先进节点的晶圆制造、下一代封装技术,以及把这一切整合起来所需的先进系统能力上,都拥有国内能力。这是一个强有力公私伙伴关系至关重要的领域。整个半导体行业在一点上是一致的:需要携手并与政府合作,大幅扩大美国本土的芯片生产与先进封装能力。 **问题 7.** 关税带来的更高成本和潜在供应链中断,将如何影响你们建设 AI 基础设施的计划? **答复:** AMD 支持把领先制造带回美国的努力。我们应当清醒地认识到,建设领先的半导体制造将是一个以年计的项目。AMD 对台积电新宣布在美追加 1000 亿美元投资感到鼓舞。把 100% 的半导体制造放在一个地方并不可行。作为一个行业,我们需要找到在美投资、同时构建供应链韧性的方式。 **问题 8.** 你预期全国各地 AI 基础设施的建设或其他工作会出现延误吗?如果是,这些影响可能在哪里最严重? **答复:** 这个问题最好由在美国建设 AI 基础设施的各方来回答。 Response to Written Questions Submitted by Hon. Maria Cantwell to Dr. Lisa Su AI Standards The U.S. driving development of AI standards alongside the most advanced democracies in the world offers us an opportunity to set the ``rules of the road'' for AI on the global stage. Question 1. How do NIST standards help the United States' competitiveness? Answer. NIST's Risk Management Framework, including its comparisons with other AI frameworks and international standards, supports U.S. competitiveness because it was developed through a transparent public process which improves customer confidence. NIST has successfully used this framework in areas such as cybersecurity and AI, which AMD has relied upon. Question 2. What standards would you like to see NIST develop or promote to improve U.S. competitiveness? Answer. While some standards, like chip design, are best driven by industry consortium, standards that protect national security, such as cybersecurity and silicon providence, are best when NIST partners with industry to promote U.S. competitiveness. Question 3. How has your company benefited from or collaborated with the National Science Foundation, NIST or the Department of Energy Labs in artificial intelligence development? Answer. Through more than a decade of partnership with the Department of Energy, AMD now powers the world's two fastest supercomputers: Frontier, which went into operation at Oak Ridge National Labs in 2021, and El Capitan, which went into operation at Lawrence Livermore National Labs late last year. These systems are critical infrastructure for U.S. national security and scientific leadership, including the latest advances in drug discovery, medical research, climate research, hypersonic flight, and even training future generations of more capable Al models. Question 4. How will cuts to NSF funding impact your workforce and search for talent? Answer. NSF funding for science and engineering research and education at U.S. colleges and universities provide formative training ground for the American semiconductor workforce pipeline. We support continued funding for these purposes to maintain U.S. AI leadership. Question 5. What impact will cuts to Federal funding for science and research at universities have on U.S. competitiveness in AI? Answer. U.S. competitiveness in AI is based on five priorities: (1) accelerating U.S. chip and system innovations to keep our leadership in Al compute infrastructure where U.S. universities have played a key role; (2) open ecosystems that enable universities to directly contribute innovation; (3) U.S. research funding for advanced semiconductor manufacturing and packaging; (4) research investment do drive U.S. AI talent development; and (5) research in cyber techniques to strengthen U.S. defense capabilities. Federal funding for research related to these priorities at U.S. universities has been critical to maintaining our lead. To the extent less Federal funding is available for research related to these priorities at U.S. universities, the private sector will need to provide additional funding and support. Question 6. What are you most concerned with when it comes to your supply chains? Answer. As a fabless company, AMD focuses on ensuring that the products it designs can be readily produced and delivered to customers. AMD consequently supports trade policies that allow those efforts to continue, including policies that do not impair the work underway to enhance semiconductor manufacturing capabilities in the United States. We must build a robust domestic supply chain for advanced semiconductor manufacturing and packaging. AI leadership depends on the ability to build complete, integrated systems. That means ensuring we have domestic capabilities in both wafer manufacturing at the most advanced nodes and next-generation packaging technologies, as well as the advanced system capabilities needed to bring it all together. This is an area where strong public-private partnerships are critical. The entire semiconductor industry is aligned on the need to work together and partner with the government to significantly scale U.S. chip production and advanced packaging capabilities here at home. Question 7. How will the higher costs from tariffs and potential supply chain disruptions impact your plans for building AI infrastructure? Answer. AMD supports efforts to bring leading-edge manufacturing back to the U.S. We should be clear-eyed about the fact that building leading-edge semiconductor manufacturing will be a years-long project. AMD is encouraged by the new announcement by TSMC of an increase of $100 billion investment in the U.S. It is not viable to have 100 percent of semiconductor manufacturing in one location. As an industry, we need to find ways to invest in the U.S. while building resiliency in our supply chain. Question 8. Do you anticipate any delays in construction or other work on AI infrastructure around the country? If so, where might these impacts hit the hardest? Answer. This is a question best answered by parties building AI infrastructure in the U.S. ______ ¶
Ms. Su ## 12. 尊敬的艾米·克洛布查尔(Amy Klobuchar)提交的书面问题 → 苏姿丰博士 ### 主题:芯片供应链 两党共同的《芯片与科学法案》是对本土半导体研发与制造的一次里程碑式投资,我坚决支持。然而,设计与制造只是供应链的一部分。例如,据美国外交关系委员会(Council on Foreign Relations)称,芯片通常被送往东南亚进行组装、测试和封装。 **问题 1.** 在芯片供应链中,还有哪些关键环节是我们需要着力以增强美国经济韧性的? **答复:** 我们感谢对强化美国半导体能力的持续关注。展望未来,任何未来的立法都提供了一个机会,去处理那些对我国国家安全和经济安全至关重要、但在以往立法中可能权重不足的领域。我提出三项关键优先事项。 - **第一,美国在芯片设计上的领导地位必须被视为国家优先事项。** 设计是技术差异化发生的地方——性能、效率与能力都在这里被定义。其他国家正在这个领域大举投资,而不只是在制造上。一个前瞻性的政策框架应当包括有针对性的税收激励和可退还的研发抵免,以确保美国成为先进半导体架构、软件工具链和 AI 加速器创新更具吸引力的环境。 - **第二,先进封装应当获得远为更大的战略重视。** 计算的未来日益依赖芯粒(chiplet)架构与三维集成。然而今天,正如你所说,大部分高端先进封装产能位于亚洲。这对商业应用和国家安全应用都是日益增长的风险。对本土封装研发枢纽与供应链准备度的定向投资,可以在一个美国仍有时间领先的领域创造出实质能力——但窗口正在关闭。 - **第三,人才必须始终处于任何长期战略的中心。** 当前半导体工程师、软硬件协同设计专家和封装技术员的短缺是一项结构性挑战。我们鼓励联邦继续并扩大对劳动力发展项目的支持——从学徒制模式、社区大学,到与产业需求直接对接的大学高级研究伙伴关系。 简而言之,我们认为未来的行动应当校准到战略杠杆最高的领域——设计、封装和人才——并且其结构应当鼓励广泛的私营部门投资,同时不扭曲市场。 我们感谢委员会的领导,并期待携手制定政策,确保美国在未来几十年保持半导体领导地位。 Response to Written Questions Submitted by Hon. Amy Klobuchar to Dr. Lisa Su Topic: Chips Supply Chain The bipartisan CHIPS and Science Act was a landmark investment in domestic semiconductor R&D and manufacturing that I strongly support. Yet design and manufacturing are only part of the supply chain. For example, according to the Council on Foreign Relations, chips are typically sent to Southeast Asia for assembly, testing, and packaging. Question 1. What are additional key points in the chip supply chain that we need to focus on to bolster U.S. economic resiliency? Answer. We appreciate the continued focus on strengthening America's semiconductor capabilities. Looking ahead, any future legislation offers an opportunity to address areas that are essential to our national and economic security but may have been underweighted in prior legislation. I'd offer three key priorities. First, U.S. leadership in chip design must be treated as a national priority. Design is where technological differentiation happens--where performance, efficiency, and capability are defined. Other nations are investing aggressively in this space, not just in manufacturing. A forward-looking policy framework should include targeted tax incentives and refundable R&D credits to ensure the U.S. becomes a more attractive environment for advanced semiconductor architecture, software tooling, and AI accelerator innovation. Second, advanced packaging deserves far greater strategic emphasis. The future of computing increasingly depends on chiplet architectures and 3D integration. Yet today, as you note, most of the high-end advanced packaging capacity resides in Asia. This is a growing risk for both commercial and national security applications. Focused investments in domestic packaging R&D hubs and supply chain readiness can create meaningful capability in an area where the U.S. still has time to lead--but the window is closing. Third, talent must remain at the center of any long-term strategy. The current shortage of semiconductor engineers, software-hardware co-design specialists, and packaging technicians is a structural challenge. We encourage continued and expanded Federal support for workforce development programs--from apprenticeship-style models and community colleges to advanced university research partnerships that align directly with industry needs. In short, we believe future action should be calibrated toward areas where strategic leverage is highest--design, packaging, and talent--and should be structured in a way that encourages broad private-sector investment without distorting the market. We thank the Committee for its leadership and look forward to working together on policies that secure U.S. semiconductor leadership for the decades ahead. ______ ¶
Ms. Su ## 13. 尊敬的布莱恩·沙茨(Brian Schatz)提交的书面问题 → 苏姿丰博士 ### 主题:保障美国 AI 系统的安全 **问题 1.** 请详细说明你们已采用、或承诺采用哪些实体安全与网络安全标准,以防止外国对手窃取知识产权和进行网络破坏? **答复:** 美国企业与制造商有着长期伙伴关系,这些制造商以最高敏感度对待芯片技术知识产权,以确保客户继续使用它们进行制造。 就 AMD 产品制造之后的实体安全而言,我们非常严肃地对待出口管制法下的义务。作为芯片制造商,我们实施严格的「了解你的客户」(KYC)程序,并与政府伙伴密切合作以确保合规。 有效的执法最终取决于产业与政府之间强有力的伙伴关系、清晰的法规,以及贯穿整个供应链的严格监督。 **问题 2.** 你们在硬件安全研发上作了哪些投资(包括了解你的客户流程),以防止产品落入不当之手? **答复:** AMD 维护一套严格的出口管制合规计划,该计划依据美国工业与安全局(BIS)的指引制定。每一笔订单都必须通过出口合规审查。我们关于全球贸易合规有完善的管理政策,定期审查和更新,并分发给所有员工。我们计划的每一个要素都对应美国及国际法规的特定要求,以确保我们的出口符合自身义务。我们的合规系统还利用软件管控来阻断被禁止的销售,例如向政府制裁名单销售,以及反映对某些地点和特定最终用途出口禁令的管控。 此外,我们利用最新技术,并对合规工作采取基于风险的方法。我们的目标是了解我们的客户;我们也通过销售点记录收集关于间接客户的信息。在考虑与任何一方开展业务之前,我们会就制裁问题进行筛查。我们部署了领先的第三方资源——包括美国政府所信任的同一批资源——来审查潜在客户。这些平台持续研究开源数据,看是否能产出新的情报,我们用其最新发现刷新我们的系统。如果出现警示信号,我们的团队会在采取任何进一步步骤之前进行强化尽职调查。 在 AMD,合规是全企业范围的努力。所有员工每年都要接受强制性合规培训,以跟上其在出口管制方面的职责。我们还有一支由全球专家组成的核心 GTC 团队,负责我们的出口管制政策、流程与程序。该团队分布在六个国家,共同拥有数十年在全球范围内驾驭出口合规法规、并管理和执行出口合规政策的经验。 **问题 3.** 这些流程是行业标准吗?我们如何确保你们的竞争对手也执行这些流程? **答复:** 如上所述,AMD 力求拥有同类最佳的合规计划。鼓励所有公司都这样做,并确保政府有足够资源以一致、透明的方式监督合规并适用法规,将增强现行管制的有效性。 ### 主题:中国部署 AI **问题 4.** 你认为中华人民共和国部署 AI 系统所构成的最大国家安全威胁是什么? **答复:** AMD 五十多年来一直是一家自豪的美国公司。我们希望美国繁荣、创新并领先。这对我们公司有利,对我们的员工有利——不仅在美国,也在世界各地。作为一家公司,AMD 并不处于识别或排列国家安全威胁优先次序的位置。 **问题 5.** 对于应对这些威胁,你有什么建议? **答复:** 出口管制法规描述了这些威胁,并据此施加限制。如上所述,AMD 广泛的合规计划旨在遵守这些限制、应对已识别的威胁并防止转运,同时与政府伙伴密切合作以确保合规。 ### 主题:加速科学研究 **问题 6.** AMD 与美国国家实验室的合作如何深化,以确保各学科科学家都能获得下一代算力? **答复:** AMD 在超级计算领域的工作体现了与国家实验室合作的好处。如上所述,通过与能源部十多年的伙伴关系,AMD 现在为世界上两台最快的超级计算机提供动力:2021 年在橡树岭国家实验室投入运行的「前沿」,以及去年年底在劳伦斯利弗莫尔国家实验室投入运行的「埃尔卡皮坦」。这些系统是美国国家安全与科学领导地位的关键基础设施,支撑着药物发现、医学研究、气候研究、高超声速飞行乃至训练未来几代更强 AI 模型的最新进展。同样地,AMD 随时准备在下一代算力上继续并加强与美国国家实验室的合作。 Response to Written Questions Submitted by Hon. Brian Schatz to Dr. Lisa Su Securing American AI Systems Question 1. Could you please detail what physical security and cybersecurity standards you have adopted, or will commit to adopting, to prevent IP theft and cyber disruptions by foreign adversaries? Answer. U.S. companies have long-standing partnerships with manufacturers who treat chip technology IP with the utmost sensitivity to ensure customers continue using them for manufacturing. For the physical security of AMD's products after manufacturing, we take our obligations under export control laws very seriously. As chip manufacturers, we implement strict Know Your Customer procedures and work closely with government partners to ensure compliance. Effective enforcement ultimately depends on a strong partnership between industry and government, clear regulations, and rigorous oversight across the supply chain. Question 2. What investments, including know-your-customer processes, are you making in hardware security research and development to prevent your products from falling into the wrong hands? Answer. AMD maintains a rigorous export control compliance program, which was developed in accordance with guidance from the U.S. Bureau of Industry and Security (``BIS''). Every single order must pass an export compliance review. We have well-established management policies on global trade compliance, which we regularly review and update, and distribute to all our employees. Each element of our program addresses specific requirements associated with U.S. and international regulations to make sure we export our products consistent with our obligations. Our compliance system also leverages software controls to block prohibited sales, such as to government sanctions lists, as well as controls reflecting bans against exports to certain locations and for specified end-uses. Additionally, we leverage the latest technology and take a risk- based approach to our compliance efforts. Our goal is to know our customers, and we also gather information on indirect customers with point-of-sale records. Before we consider doing business with any party, we screen them for sanctions issues. We have deployed leading third-party resources--including the same ones trusted by the U.S. government--to scrutinize potential customers. These platforms continually study open-source data to see if it yields new intelligence, and we refresh our systems with their up-to-date findings. If red flags arise, our team conducts enhanced due diligence before taking any further steps. At AMD, compliance is an enterprise-wide effort. All of our employees undertake mandatory compliance training on a yearly basis in order to stay current with their export control responsibilities. We also have a core GTC team of specialists around the world who are responsible for our export control policies, processes, and procedures. This team is located in six countries and collectively has decades of experience navigating export compliance regulations--and managing and executing export compliance policies--on a global scale. Question 3. Are these processes industry standards? How do we ensure that your competitors also exercise them? Answer. AMD strives to have a best-in-class compliance program, as described above. Encouraging all companies to do the same and ensuring the government has sufficient resources to monitor compliance and apply the regulations in a consistent and transparent manner, will enhance the effectiveness of the current controls. PRC Deployment of AI Question 4. What do you believe is the greatest national security threat posed by the People's Republic of China's deployment of AI systems? Answer. AMD is a proud American company for over five decades. We want the U.S. to thrive, innovate and lead. That is good for us as a company and good for our employees, not just in the U.S. but around the world. As a company, AMD is not in a position to identify or prioritize national security threats. Question 5. What are your recommendations for addressing these threats? Answer. The export control regulations describe the threats and impose restrictions, accordingly. As noted above, AMD's extensive compliance program is designed to comply with those restrictions, address the identified threats and prevent diversion, working closely with government partners to ensure compliance. Accelerating Scientific Research Question 6. How can AMD's collaboration with U.S. national labs be deepened to ensure scientists across disciplines can access next- generation compute? Answer. AMD's work in supercomputing illustrates the benefits of collaboration with the national labs. As mentioned above, through more than a decade of partnership with the Department of Energy, AMD now powers the world's two fastest supercomputers: Frontier, which went into operation at Oak Ridge National Labs in 2021, and El Capitan, which went into operation at Lawrence Livermore National Labs late last year. These systems are critical infrastructure for U.S. national security and scientific leadership, including the latest advances in drug discovery, medical research, climate research, hypersonic flight, and even training future generations of more capable Al models. In a similar vein, AMD stands ready to continue and enhance its collaboration with U.S. national labs in next-generation compute. ______ ¶
Ms. Su ## 14. 尊敬的杰基·罗森(Jacky Rosen)提交的书面问题 → 苏姿丰博士 ### 主题:AI 生态中的竞争 在美国建设强大 AI 产业的同时,我们也必须着力建立一个有竞争力的产业,这至关重要。 **问题 1.** 国会应当考虑什么来促进 AI 生态及整个技术栈中的竞争,例如互操作性要求,或保障研究者获得算力? **答复:** 在 AMD,我们坚信开放生态与互操作性是驱动创新、增强安全、培育强健竞争的基础——在快速演进的 AI 领域尤其如此。 开放生态让各种规模的开发者、研究者和公司能够协作、在共享标准之上构建,并加速突破。通过把我们的平台设计为支持开源软件和行业标准接口,我们降低了创新者的进入门槛,使更广泛、更多元的参与者能够作贡献。这种获取的民主化带来更快的迭代周期和更丰富的创新管道,因为没有任何一家公司——无论多大——能垄断好主意。 互操作性确保来自不同厂商的技术能够无缝协同。这不仅给客户更多选择,也防止供应商锁定、鼓励健康的市场竞争。从国家竞争力的角度看,它通过创造韧性、冗余和灵活性,强化了整个 AI 供应链。 关键在于,开放性也促进透明与安全。开放标准和社区驱动的软件让漏洞能被更快发现和解决。这种协作性的审视,使系统比在封闭孤岛中建造的更为安全。 简而言之,互操作性与开放生态不只是良好的工程实践——它们是战略必需。它们催化创新、增强信任,并确保美国保持在 AI 全球技术领导地位的最前沿。 在我们思考如何培育一个更具包容性和竞争力的 AI 生态时,必须强调:这不仅是创新问题——它也是国家安全的必需。 开放生态与互操作性尤其有助于确保我们不会把风险、能力或决策权集中到少数公司或平台手中。在一个 AI 系统将越来越多地用于支持国防、关键基础设施和情报行动的世界里,生态的韧性与多样性同原始能力一样重要。 ### 主题:AI 技能 **问题 2.** 你们两位在证词中都提到数字技能的重要性。能否谈谈,如果我们不借助国会授权的联邦项目——例如《数字公平法》下设立的那些明确旨在帮助美国人建立数字技能(如教老年人、小企业和退伍军人使用 AI)的项目——这会如何损害美国与中国竞争的能力? **答复:** AMD 认识到数字技能对美国在国际舞台上竞争并领先的重要性。人才必须始终处于任何长期战略的中心,才能成功。当前半导体工程师、软硬件协同设计专家和封装技术员的短缺是一项结构性挑战。我们鼓励联邦继续并扩大对劳动力发展项目的支持——从学徒制模式、社区大学,到与产业需求直接对接的大学高级研究伙伴关系。 我们必须投资于人才,并确保我们在 STEM 教育和劳动力培训上的国家战略。私营部门当然可以做得更多,包括扩大大学伙伴关系、投资再培训项目、培养成功所需的跨学科人才。我们应当激励企业加大其最关键的 AI 研发努力。我们应当让美国成为世界上对 AI 人才绝对最好的地方。 为此,我们认识到教育在塑造明日领导者和创新者方面具有巨大力量,我们热忱于提升数字素养、让学生为 21 世纪劳动力的要求做好准备。这就是为什么我们与学校、教育工作者和本地非营利组织合作,为 AMD 学习实验室配备基于 AMD 处理器的设备,帮助赋能教师、激励学生投身 STEM 教育。 作为 AMD 致力于技能发展和培养明日领导者的更多例子:AMD 支持安·理查兹青年女性领导者学校(Ann Richards School for Young Women Leaders),这是得克萨斯州奥斯汀一所独特的公立学校,提供跳出常规的教育策略和增益机会,纳入真实世界的手把手项目,让学生做好准备、有能力以大创意应对大问题。高中生完成女性历来代表性不足的 STEM 领域的「大学到职业」通路。在大学层面,AMD 大学计划为教授和讲师提供免费软件许可、硬件捐赠和教育资源,支持数字设计、嵌入式系统、计算机科学与 AI 的课堂教学。 在课堂之外,AMD 还于 2023 年底通过「雇用我们的英雄」(Hiring Our Heroes, HOH)企业奖学金项目迎来了首批退伍军人,壮大了我们的劳动力。HOH 项目由美国商会开发,让即将转业的军人、退伍军人、军属和军人照护者沉浸到平民劳动力中,通过积累其经验、提高其在业内就业的可能性来创造经济机会。 通过私营部门的倡议(例如 AMD 的这些做法)以及联邦项目的支持,我们可以为美国在国际舞台上竞争并成功铺就数字技能素养之路。 Response to Written Questions Submitted by Hon. Jacky Rosen to Dr. Lisa Su Competition Across the AI Ecosystem It's essential that as we build a strong AI industry in the U.S., we also focus on establishing a competitive one. Question 1. What should Congress consider to promote competition in the AI ecosystem and across the tech stack, such as interoperability requirements or securing access to computing power for researchers? Answer. At AMD, we strongly believe that open ecosystems and interoperability are foundational to driving innovation, enhancing security, and fostering robust competition--especially in the fast- evolving AI sector. Open ecosystems allow developers, researchers, and companies of all sizes to collaborate, build upon shared standards, and accelerate breakthroughs. By designing our platforms to support open-source software and industry-standard interfaces, we lower the barriers to entry for innovators, enabling a broader and more diverse set of participants to contribute. This democratization of access leads to faster iteration cycles and richer innovation pipeline because no one company, however large, has a monopoly on good ideas. Interoperability ensures that technologies from different vendors can work together seamlessly. This not only gives customers more choice, but it also prevents vendor lock-in and encourages healthy market competition. From a national competitiveness standpoint, it strengthens the entire AI supply chain by creating resilience, redundancy, and flexibility. Critically, openness also promotes transparency and security. Open standards and community-driven software allow vulnerabilities to be identified and resolved more quickly. This collaborative scrutiny results in more secure systems than those built in closed silos. In short, interoperability and open ecosystems are not just good engineering practices--they are strategic imperatives. They catalyze innovation, enhance trust, and ensure that the U.S. remains at the forefront of global technology leadership in AI. As we consider how to foster a more inclusive and competitive AI ecosystem, it's essential to emphasize that this is not just an innovation issue--it's also a national security imperative. Open ecosystems and interoperability are particularly useful to help ensure that we are not concentrating risk, capability, or decision-making power in the hands of a few companies or platforms. In a world where AI systems will increasingly be used to support defense, critical infrastructure, and intelligence operations, resilience and diversity in the ecosystem are just as important as raw capability. AI Skills Question 2. Both of you mentioned the importance of digital skills in your testimony. Can you discuss how it might hurt the U.S.'s ability to compete with China if we don't leverage congressionally-mandated Federal programs like those created under the Digital Equity Act, which were explicitly designed to help Americans build digital skills, like teaching seniors, small businesses, and veterans how to use AI? Answer. AMD recognizes the importance of digital skills for the U.S. to compete and lead on the international stage. Talent must remain at the center of any long-term strategy to succeed. The current shortage of semiconductor engineers, software-hardware co-design specialists, and packaging technicians is a structural challenge. We encourage continued and expanded Federal support for workforce development programs--from apprenticeship-style models and community colleges to advanced university research partnerships that align directly with industry needs. We must invest in talent and ensure our national strategy for STEM education and workforce training. The private sector can certainly do more, including expanding university partnerships, investing in reskilling programs, and developing the cross-disciplinary talent required for success. We should incentivize companies to increase their most critical Al R&D efforts. We should make America the absolute best place for Al talent in the world. To that end, we recognize education holds immense power in shaping the leaders and innovators of tomorrow, and we are passionate about enhancing digital literacy and preparing students for the demands of the 21st century workforce. This is why we partner with schools, educators, and local nonprofit organizations to outfit AMD Learning Labs with AMD processor-based equipment, helping empower teachers and inspire students to pursue STEM education. As additional examples of AMD's commitment to skills development and training the leaders of tomorrow, AMD supports the Ann Richards School for Young Women Leaders, a unique public school in Austin, Texas, which provides out-of-the-box education strategies and enrichment opportunities that incorporate real-world, hands-on projects that prepare and equip students to tackle big problems with big ideas. High school students complete a college-to-career pathway in STEM fields where women are historically underrepresented. In the university context, AMD University Program offers professors and lecturers free software licenses, hardware donations, and educational resources to support classroom teaching in digital design, embedded systems, computer science and AI. Beyond the classroom, AMD also welcomed our first cohort of veterans in late 2023 through the Hiring Our Heroes (HOH) Corporate Fellowship Program, strengthening our workforce. The HOH Program, developed by the U.S. Chamber of Commerce, immerses transitioning service members, veterans, military spouses, and military caregivers in the civilian workforce, creating economic opportunity by building their experience and increasing the possibility of a job in the industry. Through private sector initiatives, such as those at AMD, and the support of Federal programs, we can pave the way for the digital skills literacy for the U.S. to compete and succeed on the international stage. ______ ¶
Ms. Su ## 15. 尊敬的约翰·费特曼(John Fetterman)提交的书面问题 → 苏姿丰博士 **问题 1.** 苏博士,数据中心使用的能源实在太多——劳伦斯伯克利国家实验室的美国官方数据预计,到 2028 年它们将用掉全美电力的 12%。〔1〕当我的共和党同事忙于推翻重要的能效标准时,美国产业在能效和可再生能源上引领,比以往任何时候都更关键。硬件制造商可以在提升数据中心能效方面扮演什么角色? > 〔1〕https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/ **答复:** 我们同意,随着 AI 系统能力增强,对算力的需求——以及为其供电的能源需求——正在快速上升。没有深思熟虑的规划,我们就有让这种压力不公平地落到共享同一电网的普通美国人身上的风险。 我们应对这一能源关切最重要的方式之一是在芯片层面。提升每瓦算力——也就是我们从每一单位能源中获得多少有用的工作——至关重要。作为一家芯片设计公司,我们把这视为我们使命的核心部分。每一代 AI 硬件都必须比上一代显著更节能。这不仅工程上是好事——经济上、能源政策上也是好事。 我们还需要更聪明地安排 AI 工作负载在哪里、以何种方式运行。更高效的芯片,加上智能的系统设计,可以大幅降低训练和部署这些模型所需的电力——在大规模场景下尤其如此。 与此同时,产业需要与公用事业公司和政策制定者合作,确保增长的成本不会转嫁给家庭。这意味着投资清洁能源、现代化电网基础设施,并在财务上和运营上成为解决方案的一部分。 AI 的领导地位不应当以可负担性或可持续性为代价。如果我们在效率上领先——从芯片层一直往上——我们就能以更少的能源交付更多创新,并确保 AI 的好处被广泛分享。 **问题 2.** 我很欣赏你在证词中提到技术与流程效率。具体而言,AMD 在提升其数据中心芯片的能效方面正在做什么? **答复:** AMD 高度专注于提升其数据中心芯片的能效。AMD 对能效设计采取整体性方法,在构成芯片设计的众多复杂架构杠杆之间寻求平衡,把计算与内存通过芯粒架构紧密集成,并纳入先进封装、软件分区和新型互连。我们在所有产品上的首要目标之一,是在平衡能耗的同时榨取尽可能高的性能。 此外,AMD 不仅在提升自身解决方案的效率,也与伙伴及更大的生态合作,优化 AI 管道的几乎每一个环节。对其处理单元、以及连接芯片、系统和机架的各种连接技术的优化,都将有助于提升效率,同时还包括模型量化、软件改进和算法微调。AMD 优化能效的整体性方法,意味着持续处理虚拟 AI 链条上的每一环,以最大化每瓦性能。 这是一个重要的考量,因为它意味着产品刚上市时的功率与能源需求,通常会在产品生命周期内改善。AMD 逐年取得显著的效率提升,使用 AMD 技术建造的超级计算机在 GREEN500 榜单上名列前茅。曾有一度,橡树岭国家实验室由 AMD 提供动力的 Frontier TDS(测试与开发系统)实际上登上了 GREEN500 榜首。GREEN500 按能效对 TOP500 榜单上的超级计算机进行排名。 可能取得显著效率提升的关键领域之一与数据移动有关。最大的 AI 模型需要海量数据。当比特从 GPU 或加速器芯片内部微小的寄存器文件,移动到缓存内存、再到高带宽内存(HBM)、再到 CPU 等等,能耗呈指数级增长。让尽可能多的数据尽可能靠近加速器,对最大化能效至关重要。这就是为什么 AMD 在每一代 Instinct 加速器上都持续增加缓存与内存容量,也是公司持续探索优化数据处理方式的原因。 如果我们看今天典型的大规模 AI 系统,运行该系统所需总功率的大约一半由 GPU 的高带宽内存(HBM)消耗,但另一半由 CPU、纵向与横向扩展网络,以及冷却和其他数据中心设施开销等各类项目构成。AMD 的目标是最大化系统级性能,同时最小化总功耗——不只是其芯片的功耗,也包括数据中心里芯片周围一切东西的功耗。 尽管今天 AI 所需的巨量算力资源是一个隐忧,AMD 正在努力最大化其平台的效率,并达成相关的效率目标。 Response to Written Questions Submitted by Hon. John Fetterman to Dr. Lisa Su Question 1. Dr. Su, data centers use far too much energy--U.S. government data from the Lawrence Berkley National Lab expects that they'll use up to 12 percent of all U.S. electricity by 2028.\1\ While my Republican colleagues have been busy overturning important energy efficiency standards, it's more critical than ever for American industry to lead on energy efficiency and renewable energy. What role can hardware manufacturers play in improving data center energy efficiency? --------------------------------------------------------------------------- \1\ https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report- evaluates-increase-in-electricity-demand-from-data-centers/ --------------------------------------------------------------------------- Answer. We agree that, as AI systems grow more capable, the demand for compute--and the energy to power them--is rising quickly. Without thoughtful planning, we risk that pressure falling unfairly on everyday Americans who share the grid. One of the most important ways that we can address this energy concern is at the chip level. Improving compute per watt--that is, how much useful work we get out of each unit of energy--is essential. As a chip design company, we see this as a central part of our mission. Every generation of AI hardware must be significantly more energy efficient than the last. That's not just good engineering--it's good economics and good energy policy. We also need to be smarter about where and how AI workloads are run. More efficient chips, paired with intelligent system design, can dramatically reduce the power needed to train and deploy these models-- especially at scale. At the same time, industry needs to partner with utilities and policymakers to ensure the costs of growth are not passed on to households. That means investing in clean energy, modernizing grid infrastructure, and being part of the solution--financially and operationally. AI leadership shouldn't come at the cost of affordability or sustainability. If we lead on efficiency--from the chip level all the way up--we can deliver more innovation with less energy, and make sure the benefits of AI are shared broadly. Question 2. I appreciate that you referenced technological and process efficiency in your testimony. Specifically, what is AMD doing to improve the energy efficiency of its chips for data center use? Answer. AMD is highly focused on improving the energy efficiency of its chips for data center use. AMD takes a holistic approach to energy efficient design, balancing advancements across the many complex architectural levers that make up chip design, incorporating tight integration of compute and memory with chiplet architectures, advanced packaging, software partitions, and new interconnects. One of our primary goals across all of our products is to extract as much performance as possible while balancing energy use. In addition, AMD is not only working on improving the efficiency of its own solutions, but working with partners and the larger ecosystem to optimize virtually every aspect of the AI pipeline. Optimization of its processing units and myriad of connectivity technologies, which link chips, systems and racks, will all help enhance efficiency, along with quantizing models, improving software, and tweaking algorithms. AMD's holistic approach to optimizing power efficiency means continually addressing every link in the virtual AI chain to maximize performance-per-watt. This is an important consideration because it means the power and energy requirements of products when they initially hit the market, typically improve over the lifetime of the product. AMD has made significant efficiency gains year over year, and supercomputers built using AMD technologies have earned top rankings on the GREEN500. At one point, the AMD-powered Frontier TDS (test and development system) at Oak Ridge National Labs actually topped the GREEN500 list. The GREEN500 ranks supercomputers from the TOP500 list, in terms of energy efficiency. One of the key areas where significant efficiency gains are possible relate to data movement. The largest AI models require huge amounts of data. As bits move from the tiny register files inside GPUs or accelerator chips, to cache memory, out to High Bandwidth Memory, and to the CPU, and so on, energy consumption grows exponentially. Keeping as much data as close to the accelerator as possible is paramount to maximizing energy efficiency. It's why, for example, AMD continues to increase the amount of cache and memory on its Instinct accelerators generation-on-generation, and why the company continually explores ways to optimize how the data is actually processed. If we look at typical, large-scale AI system today, roughly half the total power required to run the system is consumed by the GPU's high bandwidth memory (HBM), but the other half is comprised of CPU, scale-up and scale-out networking, and various things like cooling and other data center facility overhead. AMD's goal is to maximize system- level performance, while also minimizing total power consumption, not just from its chips, but from everything around them in the data center as well. While the significant amounts of compute resources required for AI today are a concern, AMD is working hard to maximize the efficiency of its platforms and meet its related efficiency goals. ______ ¶
Mr. Intrator ## 16. 尊敬的玛丽亚·坎特韦尔(Maria Cantwell)提交的书面问题 → 迈克尔·因特拉托(Michael Intrator) ### 主题:AI 标准 美国与世界上最先进的民主国家一道推动 AI 标准的发展,这为我们提供了在全球舞台上为 AI 设定「道路规则」的机会。 **问题 1.** NIST 标准如何有助于美国的竞争力? **答复:** NIST 特别出版物通过广泛征求利益相关方意见制定,为美国公司提供具体、可适配的指引,用于构建开发和部署 AI 技术的框架。例如,《NIST 网络安全框架》通过提供共同语汇和信息安全与网络风险管理的自愿性指南,帮助美国企业。 **问题 2.** 你希望 NIST 制定并推广哪些标准,以提升美国的竞争力? **答复:** 通过提供技术基础、为产业主导的标准制定确立共同语汇,NIST 可以帮助确保美国的安全做法塑造全球基准。 ### 主题:科学领域的公共投资 政府对基础研究的投入,对美国在科学发现、创新方面的进展以及美国的竞争力作出了重大贡献。对美国研究综合体的持续投入,对维持美国的经济实力与国家安全至关重要。AI 也是如此,因为我们需要持续发展高性能计算能力,以及为未来的数据中心供电所需的下一代能源技术。 **问题 3.** 贵公司在人工智能开发中如何受益于国家科学基金会、NIST 或能源部各实验室,或与之合作? **答复:** 截至 2025 年 5 月,据其所知,CoreWeave 尚未与能源部各实验室、国家科学基金会或 NIST 进行正式合作。不过,我们确实期待有机会就能源部关于在联邦土地上开发数据中心的努力与之合作。 **问题 4.** 削减 NSF 拨款将如何影响你们的劳动力和人才搜寻? **答复:** CoreWeave 支持 NSF 在劳动力发展方面的重要作用,即为美国劳动者提供新兴技术培训。例如,NSF 的「新兴与新颖技术体验式学习」项目,旨在为有兴趣进入人工智能、半导体、先进制造等关键技术职业路径的人提供体验式学习机会。这类项目增强美国竞争力,并确保美国有一支支撑 AI 基础设施与 AI 开发的人才管道。 一支有技能、受过训练的劳动力,对 AI 数据中心的稳定与扩张至关重要——这些中心依赖专业的数据中心技术员、网络与电气工程师、网络安全专业人员和项目经理。CoreWeave 支持培训本土劳动力的努力,这些技能劳动者是满足不断增长的 AI 需求、加速 AI 创新所必需的。 **问题 5.** 削减联邦对大学科学与研究的拨款,将对美国在 AI 领域的竞争力产生什么影响? **答复:** 科学与研究型大学在维持 AI 全球领导地位方面扮演重要角色。研究者和学生在尖端项目上与技术产业协作,塑造 AI 创新的未来图景。美国在 AI 竞赛中一个独特的优势,是我们有能力支持大学里的研究者和学生,并在学术界、产业界和政府之间建立有效的伙伴关系。 CoreWeave 坚定致力于支持 AI 研究、科学与教育,以及公私协作伙伴关系。 例如,CoreWeave 很自豪能与微软、普林斯顿大学和新泽西州经济发展局一道,成为新泽西 AI 中心(New Jersey AI Hub)的创始伙伴。创始伙伴将共同投入超过 7200 万美元,支持该中心的长期成功;该中心聚焦研发工作、AI 在若干行业部门的应用,以及 AI 劳动力发展与教育。 ### 主题:关税影响 高关税和过度扩张的出口管制——尤其那些未与私营部门及美国盟友充分协调的——有可能扰乱供应链、抬高美国公司的成本。这使建设数据中心、芯片、电厂和电网现代化等 AI 基础设施变得更加昂贵。 **问题 6.** 就你们的供应链而言,你最担心什么? **答复:** CoreWeave 在驱动 AI 的设备上投入数十亿美元。像任何购买长期资本资产的企业一样,CoreWeave 依赖成本与政策的确定性来规划业务、在美国进行关键投资。这使我们能够为客户提供可预测的定价。 我们最关心的事情之一,是确保供应链中的战略投资稳定性与价格可预测性。我们认为,重要的是继续与志同道合、可靠的国际战略伙伴(包括芯片制造商、原始设备制造商和软件供应商)保持强健的供应链关系,以继续在美国扩大 AI 数据中心运营,并在全球 AI 竞赛中领先。 **问题 7.** 关税带来的更高成本和潜在供应链中断,将如何影响你们建设 AI 基础设施的计划? **答复:** 关键部件(如先进半导体和网络设备)全球供应链的波动,可能抬高成本或打乱部署时间表,从而不利于美国公司快速扩展满足领先企业和 AI 实验室要求所需的 AI 能力。 获取驱动 AI 工作负载所需的高性能部件,需要管理复杂的全球供应链、维持强健的供应链关系。与领先的全球供应商和战略伙伴持续接触,对确保美国 AI 基础设施的持续运行、扩张与快速部署,以及维护美国竞争力,至关重要。 我们聚焦于一种供应链战略:对我们继续开发 AI 基础设施所需的关键部件,保持强健而有韧性的获取渠道。我们希望与贸易伙伴正在进行的讨论,能提供美国公司所需的清晰度与确定性,以便进行大规模资本投入,以当下所需的规模与紧迫性建设 AI 基础设施。 **问题 8.** 你预期全国各地 AI 基础设施的建设或其他工作会出现延误吗?如果是,这些影响可能在哪里最严重? **答复:** 不,我们不预期 AI 基础设施的建设或机器(例如 GPU、存储、网络)方面的工作会出现延误。CoreWeave 确实会遇到一些挑战,我们会把它们纳入规划和建设流程。获取公用事业电力时,由于需求高企和关键部件供应有限,开关设备和大型变压器可能交期很长。获取数据中心设备及钢铁、混凝土等基础材料,需要与数据中心运营商和供应链供应商密切协作。 劳动力约束可能构成另一项挑战,因为数据中心基础设施需求很高,而熟练技工和技术员的人才池有限。即便在熟悉数据中心区划与许可的市场,政府管理这些流程的行政能力也可能成为限制因素。 Response to Written Questions Submitted by Hon. Maria Cantwell to Michael Intrator AI Standards The U.S. driving development of AI standards alongside the most advanced democracies in the world offers us an opportunity to set the ``rules of the road'' for AI on the global stage. Question 1. How do NIST standards help the United States' competitiveness? Answer. NIST special publications, developed through broad stakeholder consultation, offer American companies specific and adaptable guidance for creating frameworks to develop and deploy AI technologies. For example, the NIST Cybersecurity Framework assists U.S. firms by providing a common vocabulary and voluntary guidelines for information security and cyber risk management. Question 2. What standards would you like to see NIST develop and promote to improve U.S. competitiveness? Answer. By providing the technical foundation and establishing common vocabulary for industry-led standards development, NIST can help ensure that U.S. security approaches shape global benchmarks. Public Investment in Science Government investment in basic research has contributed significantly to U.S. advances in scientific discovery, innovation and U.S. competitiveness. Continued investment in the U.S. research complex is essential to maintaining U.S. economic strength and national security. This is also the case with AI given the need to continue to develop our high performance computing capabilities and next generation energy technologies required to power the data centers of the future. Question 3. How has your company benefited from or collaborated with the National Science Foundation, NIST or the Department of Energy Labs in artificial intelligence development? Answer. As of May 2025, to the best of its knowledge CoreWeave has not formally collaborated with the Department of Energy Labs, the National Science Foundation, or NIST. However, we do look forward to opportunities to collaborate with the Department of Energy on its efforts regarding potential data center development on Federal lands. Question 4. How will cuts to NSF funding impact your workforce and search for talent? Answer. CoreWeave supports NSF's important role in workforce development by training U.S. workers in emerging technologies. For example, NSF's Experiential Learning in Emerging and Novel Technologies program aims to provide experiential learning opportunities for individuals interested in career pathways in key technologies such as artificial intelligence, semiconductors, advanced manufacturing, and more. These types of programs bolster U.S. competitiveness and ensure that the U.S. has a pipeline of talent to support AI infrastructure and AI development. A skilled and trained workforce is vital for the stability and expansion of AI data centers--which rely on specialized data center technicians, network and electrical engineers, cybersecurity professionals, and project managers. CoreWeave supports efforts to train the domestic workforce comprised of the skilled workers required to meet the growing AI demand and to accelerate AI innovation. Question 5. What impact will cuts to Federal funding for science and research at universities have on U.S. competitiveness in AI? Answer. Science and research universities play an important role in maintaining global leadership in AI. Researchers and students collaborate with the technology industry on cutting-edge projects to shape the future landscape of AI innovation. A unique American advantage in the AI race is our ability to support our researchers and students at universities, as well as to create effective partnerships between academia, industry, and government. CoreWeave is deeply committed to supporting AI research, science, and education, and public-private collaborative partnerships. For example, CoreWeave is proud to be a founding partner of the New Jersey AI Hub, along with Microsoft, Princeton University, and the New Jersey Economic Development Authority. The founding partners will collectively invest over $72 million to support the long-term success of the NJ AI Hub, which focuses on research and development efforts, applications of AI in several industry sectors, and AI workforce development and education. Tariff Impacts High tariffs and overreaching export controls, especially that are not well coordinated with the private sector and U.S. allies, have the potential to disrupt supply chains and raise costs for U.S. companies. That makes building AI infrastructure like data centers, chips, power plants, and grid modernization more expensive. Question 6. What are you most concerned with when it comes to your supply chains? Answer. CoreWeave invests billions of dollars in the equipment which powers AI. Like any business which purchases long-lived capital assets, CoreWeave relies on cost and policy certainty to plan its business and make critical investments in the U.S. This enables us to provide predictable pricing to our customers. One of our top concerns is ensuring strategic investment stability and predictable prices in the supply chain. We believe it's important to continue to maintain robust supply chain relationships with like- minded and reliable international strategic partners, including chipmakers, original equipment manufacturers, and software providers, to continue to scale AI data center operations in the U.S. and lead in the global AI race. Question 7. How will the higher costs from tariffs and potential supply chain disruptions impact your plans for building AI infrastructure? Answer. Volatility in the global supply chain for critical components, such as advanced semiconductors and networking equipment, can raise costs or disrupt deployment timelines, adversely affecting American companies' ability to rapidly scale the AI capabilities necessary to meet the requirements of leading enterprise companies and AI labs. Acquiring the necessary high-performance components to power AI workloads requires managing a complex global supply chain and maintaining robust supply chain relationships. Continued engagement with leading global suppliers and strategic partners is vital to ensuring the continued operation, expansion, and rapid deployment of U.S. AI infrastructure and to uphold U.S. competitiveness. We are focused on a supply chain strategy that maintains robust, resilient access to the critical components we need to continue to develop AI infrastructure. We hope the ongoing discussions with trading partners will provide the clarity and certainty American companies require to make the large scale capital necessary to build the AI infrastructure at the scale and urgency this moment requires. Question 8. Do you anticipate any delays in construction or other work on AI infrastructure around the country? If so, where might these impacts hit the hardest? Answer. No, we do not anticipate delays in construction or machine (e.g., GPU, storage, networking) work on AI infrastructure. CoreWeave does encounter challenges that we factor into our planning and construction processes. Acquiring utility power can involve long lead times for switch gear and large transformers due to high demand and the availability of critical components. Access to both data center equipment and underlying materials, such as steel or concrete, requires close collaboration with data center operators and supply chain vendors. Labor constraints can pose another challenge, given the high demand for data center infrastructure and the limited pool of skilled tradesmen and technicians. Even in markets experienced with data center zoning and permitting, government administrative capacity to manage these processes can be a limiting factor. ______ ¶
Mr. Intrator ## 17. 尊敬的艾米·克洛布查尔(Amy Klobuchar)提交的书面问题 → 迈克尔·因特拉托 ### 主题:云基础设施中的竞争 联邦贸易委员会(FTC)提出关切:初创企业进入人工智能市场面临重大障碍,因为大型科技公司更容易获得海量数据、具备必要专长的人才以及算力资源。 **问题:** 能否谈谈确保有创新精神的 AI 初创企业能与大型成熟科技公司同场兴旺的重要性?作为 AI 基础设施市场相对较新的进入者,哪些因素对确保一个有竞争力的生态很重要? **答复:** CoreWeave 的基础设施是唯一一个专为 AI 与机器学习(ML)工作负载、以最高性能与效率为目标而从头构建的云平台。CoreWeave 迄今的成功表明,尖端创新、快速执行和无与伦比的性能,使公司得以与在位平台竞争。 我们近期的增长表明,市场可以回报那些以性能实现差异化的公司——为客户的具体需求交付专门化的解决方案。通过我们专有的软件能力,我们使客户在既有基础设施云环境中,相对于其他 AI 产品获得显著更高的系统整体性能和更有利的正常运行时间,并在规模上释放速度。 确保所有行业利益相关方公平竞争的政策,有助于确保技术优劣与创新决定成败。 Response to Written Question Submitted by Hon. Amy Klobuchar to Michael Intrator Topic: Competition in Cloud Infrastructure The Federal Trade Commission has raised concerns that startups face significant barriers to entry into artificial intelligence (AI) markets because large technology firms have more access to troves of data, individuals with the necessary expertise, and computational resources. Question. Can you describe the importance of ensuring that innovative AI startups are able to thrive in markets alongside large, established tech companies? As a relatively new entrant into the AI infrastructure market, what factors are important to ensure a competitive ecosystem? Answer. CoreWeave's infrastructure is the only cloud platform that was purpose-built for AI and Machine Learning (ML) workloads at the maximum performance and efficiency. CoreWeave's success to date demonstrates that cutting-edge innovation, rapid execution, and unparalleled performance has enabled the company to compete with incumbent platforms. Our recent growth shows that the market can reward companies that differentiate based on performance, delivering specialized solutions to customers' specific needs. Through our proprietary software capabilities, we enable our customers to achieve substantially higher total system performance and more favorable uptime relative to other AI offerings within existing infrastructure cloud environments and unlock speed at scale. Policies that ensure a level playing field for all industry stakeholders help ensure that technical merit and innovation determine success. ______ ¶
Mr. Intrator ## 18. 尊敬的布莱恩·沙茨(Brian Schatz)提交的书面问题 → 迈克尔·因特拉托 ### 主题:能源消耗与取胜的成本 **问题 1.** 你是否支持特朗普总统为 AI 数据中心扩大煤电的努力? **答复:** 在 CoreWeave,我们相信为 AI 数据中心供电有丰富的解决方案。我们公司致力于推进由现代、高效能源技术驱动的 AI 基础设施。虽然我们理解满足 AI 数据中心快速增长的电力需求所面临的挑战,但我们认为未来在于一种多元化的能源组合,优先考虑可再生能源、核能,以及采用最高效、低排放技术的化石能源。 **问题 2.** 你们如何应对新基础设施建设在短期内带来的成本,包括用水、污染和气候影响方面? **答复:** CoreWeave 优先考虑在增长与高效、负责任地使用资源之间取得平衡。关于用水,CoreWeave 往往部署闭环风冷基础设施和液冷技术,与传统数据中心相比可大幅减少甚至消除用水。我们运营所在的许多数据中心,为我们的运营提供非排放和低排放能源。 **问题 3.** 你们有什么计划来采购清洁能源,并公开报告公司的排放(如果尚未报告)? **答复:** CoreWeave 目前所在的若干数据中心使用清洁能源、拥有现场可再生能源和/或购买可再生能源证书(REC)。随着我们扩展到新的数据中心,我们正积极采购清洁能源为我们的服务供电。 为履行加州法律 SB 253 下的合规义务,CoreWeave 将于 2026 年报告其范围一和范围二温室气体排放。 **问题 4.** 你认为当前的激励机制是在推动还是忽视能源创新的机会? **答复:** 过去和当前的激励机制在 21 世纪帮助压低了各类能源的成本,其中包括研发与税收政策。 为维持并强化美国在 AI 领域的地位,需要持续的政策来确保充足、可靠、可负担的清洁能源供应。CoreWeave 支持那些能促成能源技术平衡、多元组合的政策,包括先进核能、碳捕集与封存,以及可再生能源的持续进步。国家还需要实现电网现代化并提升能效。 ### 主题:中国部署 AI **问题 5.** 你认为中华人民共和国部署 AI 系统所构成的最大国家安全威胁是什么? **答复:** 围绕中华人民共和国(PRC)部署 AI 系统的国家安全关切是相互关联的。其中包括把 AI 整合进其长期的军民融合战略,形成两用能力,从而推进其经济和国防目标。与此同时,中国力图通过把其 AI 生态确立为众多国家的基础性要素,来扩大其全球 AI 影响力,增强其塑造全球 AI 开发与部署规则的能力。 **问题 6.** 对于应对这些威胁,你有什么建议? **答复:** 为维持其在 AI 的领导地位并有效应对国家安全威胁,美国必须扩大其 AI 生态,并在 AI 技术栈的每一层培育创新。精心校准出口管制与贸易协定以应对安全关切,将限制中国获取关键 AI 技术,削弱其全球 AI 影响力。让美国 AI 生态向那些承诺遵守安全与技术框架的盟友和志同道合国家扩散,也将应对这一威胁。强化美国的 AI 领导地位,将确保民主价值与安全利益塑造全球 AI 开发的规则。 ### 主题:加速科学研究 **问题 7.** 你认为把先进云计算能力提供给学术和国家实验室研究者存在哪些障碍? **答复:** 我们处于有利位置来服务学术和国家实验室研究者,因为我们提供的产品正是为这类用户设计的。由于私营部门和 AI 实验室的需求日益庞大,我们在寻找可用容量方面可能面临挑战。 **问题 7a.** 国会可以如何帮助负责任地减少这些障碍? **答复:** 国会可以考虑设立机会与资源,促成私营部门与联邦研究中心之间的伙伴关系。一个专门汇聚各研究机构算力需求的资源池或联合体,可以创造更大的采购议价能力,并减少当前构成约束的不确定性。 与此相关,为先进云计算实施简化、标准化的采购工具,可以减少行政障碍,使之更好地兼顾研究时间表与产业商业模式。 **问题 8.** 是否应当有专门的联邦渠道或采购通路,确保美国联邦研究者能够以科学需求的节奏获得 AI 算力? **答复:** 算力分配目前通过与私营公司的巨额采购合同来决定。我们寻找并优先争取这类合同,因为其规模大、期限长,有助于我们降低财务运营风险。这导致我们把期限较短、金额较小的合同排在后面。联邦政府可以考虑建立机制,汇聚学术与国家实验室机构的需求,确保它们能获得维持美国 AI 领导地位所需的算力。 Response to Written Questions Submitted by Hon. Brian Schatz to Michael Intrator Energy Consumption and Cost of Winning Question 1. Do you support President Trump's efforts to expand coal power for AI data centers? Answer. At CoreWeave, we believe that there are abundant solutions for powering AI data centers. Our company is committed to advancing AI infrastructure powered by modern, efficient energy technologies. While we understand the challenges of meeting the rapidly growing electricity demands of AI data centers, we believe the future lies in a diversified energy mix that prioritizes renewables, nuclear power, and fossil energy sources using the most efficient, low emitting technologies. Question 2. How are you addressing the costs associated with new infrastructure development in the short-term, including in terms of water consumption, pollution, and climate impacts? Answer. CoreWeave prioritizes balancing growth with also being efficient and responsible with resources. Regarding water consumption, CoreWeave often deploys closed-loop, air-cooled infrastructure and liquid cooling technologies that drastically reduce or eliminate water use compared to traditional data centers. Many of the data centers where we operate provide non and low-emitting energy sources for our operations. Question 3. What plans do you have to source clean energy and to publicly report your companies' emissions, if you are not already? Answer. CoreWeave is currently in several data centers that use clean energy, have on-site renewables and/or buy Renewable Energy Certificates. As we expand into new data centers, we are actively engaged in procuring clean energy to power our services. To meet its compliance obligations under California law SB 253, CoreWeave will report its Scope 1 and Scope 2 greenhouse gas emissions in 2026. Question 4. Do you believe current incentive structures drive or ignore the opportunity for energy innovation? Answer. Past and current incentive structures have helped to drive down the cost of all energy sources during the 21st century. These include research and development and tax policies. To maintain and strengthen the U.S. position in the AI-continued policies are required to ensure adequate supplies of clean, reliable supplies of affordable energy CoreWeave supports policies which will enable a balanced diverse portfolio of energy technologies including advanced nuclear, carbon capture and sequestration, and continued advances in renewable energy. The nation also needs to modernize the grid and improve energy efficiency. PRC Deployment of AI Question 5. What do you believe is the greatest national security threat posed by the People's Republic of China's deployment of AI systems? Answer. The national security concerns surrounding the People's Republic of China's (PRC) deployment of AI systems are interconnected. These include the integration of AI into their long-standing civil- military fusion strategy, resulting in dual-use capabilities that further their economic and defense goals. At the same time, the PRC aims to expand its global AI influence by establishing its AI ecosystem as a foundational element in numerous countries, increasing its capacity to shape the rules governing AI development and deployment worldwide. Question 6. What are your recommendations for addressing these threats? Answer. To maintain its leadership in AI and effectively address national security threats, the U.S. must scale its AI ecosystem and foster innovation across all layers of the AI stack. Carefully calibrating export controls and trade agreements to address security concerns will limit the PRC's access to vital AI technologies and diminish its global AI influence. Enabling the diffusion of the American AI ecosystem to allies and like-minded nations that commit to security and technology frameworks will also address this threat. Strengthening American AI leadership will ensure that democratic values and security interests shape the rules governing global AI development. Accelerating Scientific Research Question 7. What barriers do you see to making advanced cloud computing capacity available to academic and national lab researchers? Answer. We are well-positioned to serve academic and national lab researchers because we offer products designed with those users in mind. We may face challenges in finding available capacity due to increasingly large demand from the private sector and AI labs. Question 7a. How could Congress help to responsibly reduce those barriers? Answer. Congress could consider establishing opportunities and resources to enable partnerships between the private sector and Federal research centers. A dedicated resource or consortium that aggregates computing demand across research institutions could create more procurement power and reduce uncertainty that currently acts as a constraint. Relatedly, implementing streamlined, standardized procurement vehicles for advanced cloud computing could reduce administrative barriers that better align with both research timelines and industry business models. Question 8. Should there be dedicated Federal channels or procurement pathways to ensure U.S. Federal researchers can access AI compute at the pace of scientific need? Answer. Compute allocation is presently dictated through large procurement contracts with private sector companies. We seek out and prioritize these contracts as they help us derisk our financial operations due to their large size and length. This leads us to deprioritize shorter-term and smaller-value contracts. The Federal government could consider mechanisms to aggregate demand across academic and national lab institutions to ensure they can access the compute needed to maintain the U.S. leadership in AI. ______ ¶
Mr. Intrator ## 19. 尊敬的爱德华·马基(Edward Markey)提交的书面问题 → 迈克尔·因特拉托 ### 主题:数据中心建设的综合影响 **问题 1.** 在规划数据中心建设时,贵公司是否会进行「从摇篮到坟墓」的基础设施研究,涵盖数据中心运营寿命期间及之后对野生动物、社区和污染的影响?如果是,你们从这些研究中学到了什么?如果不是,为什么? **答复:** CoreWeave 正在制定一套更正式的选址流程,以确保我们跟踪气候风险与韧性,并审视更广泛的生态风险。虽然我们目前并不对每个设施都开展正式的「从摇篮到坟墓」生命周期分析,但我们的方法整合了严格的环境尽职调查、利益相关方参与和适应性最佳实践,以尽量减少生态与社区影响。 我们选择以适应性再利用为重点的场地,往往改造既有建筑而非开发绿地,从而减少对土地利用的影响。此外,我们专注于尽可能高效地使用现有数据中心,以便在更小空间里容纳更多服务器,从而减少总体建筑面积需求。 ### 主题:备用发电 **问题 2.** 贵公司是否在任何设施使用备用柴油发电机? **答复:** 我们的数据中心提供方在我们运营所在的数据中心使用备用柴油发电机。这些备用发电机是归我们数据中心提供方所有的资产。它们向我们提供的服务,包括为数据中心提供对公用电源的冗余。 **问题 2a.** 如果是,请提供每个使用柴油发电机的设施清单,以及发电机的位置、数量和类型。 **答复:** 我们的数据中心提供方使用备用发电机,这些资产归提供方所有。 **问题 2b.** 如果是,贵公司是否考虑过使用电池储能技术作为柴油发电机的替代方案?请解释你们的决策过程。 **答复:** 我们目前就在这样做。我们的数据中心配备了不间断电源基础设施,可在断电后为数据中心覆盖最初 4–6 分钟。 CoreWeave 正在探索把电池储能纳入备用发电的方式,作为保持竞争优势、聚焦网络韧性的一部分。备用电池储能将使我们能为 CoreWeave 的运营和整个电网提供更大的韧性。 我们把电池储能视为一种补充性技术,可以进一步增强我们的微电网能力、管理可再生能源的间歇性,并为关键任务工作负载提供快速响应的备用电源。 ### 主题:能源结构 **问题 3.** 贵公司是否使用任何现场或同址能源发电来为数据中心供电? **答复:** 不使用。 **问题 3a.** 如果是,请详细说明有多少电力来自现场和同址发电。 **问题 3b.** 如果是,请列出所有用于为贵公司数据中心供电的现场和同址能源(例如可再生、核能、水电、燃气轮机等)。 **问题 4.** 贵公司如何确保本地电费缴纳者不必为满足数据中心能源需求所需的新能源基础设施(例如输电线路)付费? **答复:** 在评估新场地时,我们正在评估我们的能源成本结构。至少对于我们今年晚些时候的一项大型扩建,CoreWeave 支付的费率将包含附加费,用于为本地电费缴纳者提供费率保护。我们理解存在这样一种担忧:既有电费缴纳者将承担数据中心电力需求增长的成本。CoreWeave 承诺为其电力需求所需成本支付公平份额。 ### 主题:AI 模型训练中的能耗 **问题 5.** 数据中心降温需要大量水。当水量处于危急低位时,贵公司是否会继续抽取用于建筑降温的水?是否有应急方案,以避免对供水造成进一步压力、并可能从家庭占用有限资源? **答复:** CoreWeave 力求减少用水量,我们对数据中心降温有不同的做法。只要可能,我们就设置风冷冷水机组,这是一种闭环系统。系统注满后,所需的水量有限。每个数据中心都会经过评估,以权衡用电与用水的环境与气候成本效益。 Response to Written Questions Submitted by Hon. Edward Markey to Michael Intrator Comprehensive Impacts of Data Center Construction Question 1. When planning for data center construction, does your company conduct a cradle-to-grave infrastructure study that includes wildlife, community, and pollution impacts during and beyond the operational lifespan of a data center? If yes, what have you learned from those studies? If no, why not? Answer. CoreWeave is developing a more formal site selection process that will ensure we are tracking climate risk and resiliency, as well as examine broader ecosystem risks. While we do not currently conduct formal cradle-to-grave lifecycle analyses for every facility, our approach integrates rigorous environmental due diligence, stakeholder engagement, and adaptive best practices to minimize ecological and community impacts. We select sites with adaptive reuse as the focus and we often retrofit existing structures rather than developing greenfield sites, allowing us to have less impact on land use. In addition, we focus on using our existing data centers as efficiently as possible so that we can fit more servers into less space, therefore reducing the need for overall square footage. Backup Energy Generation Question 2. Does your company use backup diesel generators at any facilities? Answer. Our data center providers use backup diesel generators in the data centers in which we operate. These backup generators are assets that belong to our data center providers. Their service to us includes redundancy to the utility power source for the data center. Question 2a. If yes, please provide a list of each facility where diesel generators are being used, along with the location, quantity, and type of generators. Answer. Our data center providers use backup generators, and these assets belong to the providers. Question 2b. If yes, did your company consider the use of battery storage technology as an alternative to diesel generators? Please explain your decision process. Answer. We do this currently. Our data centers have uninterruptible power supply infrastructure in place that would cover the immediate 4-6 minutes of a power outage to the data center. CoreWeave is exploring ways to incorporate battery storage into our back-up generation as part of maintaining our competitive edge with a focus on network resiliency. Backup battery storage would allow us to have greater resiliency for CoreWeave's operations and the grid overall. We see battery storage as a complementary technology that can further strengthen our microgrid capabilities, manage renewable intermittency, and provide fast-responding backup for mission-critical workloads. Energy Mix Question 3. Does your company utilize any on-site or colocated energy generation to power your data centers? Answer. No. Question 3a. If yes, please provide detail how much power comes from on-site and colocated energy generation. Question 3b. If yes, please list all on-site and colocated energy sources (e.g., renewable, nuclear, hydropower, gas-powered turbines, etc.) that are being utilized to power your data centers. Question 4. How does your company ensure local ratepayers are not responsible for paying the cost of new energy infrastructure, such as transmission lines, needed to meet the data center's energy demand? Answer. As we review new sites, we are assessing the cost structure of our energy. For at least one of our large expansions coming later this year, the rate CoreWeave pays will include additional surcharges in order to fund rate protection for local ratepayers. We understand that there is a concern that incumbent ratepayers will bear the cost of increased data center power demand. CoreWeave is committed to paying its fair share of the costs required to meet its power demand. Energy Consumption in AI Model Training Question 5. Data centers require vast amounts of water for cooling. When water is at critically low levels, does your company continue to pull water for building cooling? Does it have a contingency for operating as to not put further stress on the water supply and potentially take limited resources from households? Answer. CoreWeave aims to reduce the amount of water it uses and we have a different approach to cooling data centers. Whenever possible, we set up air cooled chillers, which are a closed loop system. Once the system is filled, there is a limited amount of water required. Each data center is evaluated to consider the environmental and climate cost-benefit analysis of power usage versus water usage. ______ ¶
Mr. Intrator ## 20. 尊敬的杰基·罗森(Jacky Rosen)提交的书面问题 → 迈克尔·因特拉托 ### 主题:AI 生态中的竞争 在美国建设强大 AI 产业的同时,我们也必须着力建立一个有竞争力的产业,这至关重要。 **问题 1.** 国会应当考虑什么来促进 AI 生态及整个技术栈中的竞争,例如互操作性要求,或保障研究者获得算力? **答复:** 竞争的活力对美国持续保持 AI 领导地位至关重要。确保所有行业利益相关方公平竞争的政策,将更广泛地有益于美国的活力与创新。为此,重要的是简化政府认证流程、确保出口许可审查等既有要求的高效管理,并确保较新的进入者能够有效竞标政府合同。关键在于确保技术优劣与创新决定成败。 此外,开放标准是规模化运营我们云平台的重要因素。CoreWeave 目前参与若干界定开放标准与硬件互操作性的行业组织。如果没有这些行业组织界定硬件互操作性,CoreWeave 就无法以客户所需的规模和速度部署他们所要求的异构硬件组合。开放标准可以增强供应链韧性、加速创新,同时降低不同及新进入者参与的障碍。 ### 主题:模型安全 **问题 2.** 确保我们使用的 AI 模型不成为又一个网络安全漏洞,至关重要。针对大型 AI 模型或高风险模型、以及其训练所依托的基础设施的自愿性网络安全标准,是否有助于建立信任? **答复:** 是的,针对 AI 基础设施的自愿性网络安全标准,有助于建立信任、为利益相关方确立共同语汇,特别是通过产业主导的倡议。 Response to Written Questions Submitted by Hon. Jacky Rosen to Michael Intrator Competition Across the AI Ecosystem It's essential that as we build a strong AI industry in the U.S., we also focus on establishing a competitive one. Question 1. What should Congress consider to promote competition in the AI ecosystem and across the tech stack, such as interoperability requirements or securing access to computing power for researchers? Answer. Competitive dynamism is essential for continued American AI leadership. Policies that ensure a level playing field for all industry stakeholders will benefit American dynamism and innovation more broadly. In this vein, it is important to streamline government certification processes, ensure the efficient administration of existing requirements like export license reviews, and ensure that newer entrants can effectively bid for government contracts. The key is ensuring that technical merit and innovation determine success. In addition, open standards are an important factor for operating our cloud at scale. CoreWeave currently participates in industry groups that define open standards and hardware interoperability. Without these industry groups defining hardware interoperability, CoreWeave would not be able to deploy the heterogeneous mix of hardware that our customers demand at the scale and speed that they need. Open standards can enhance supply chain resiliency and accelerate innovation, while also reducing barriers for different and new stakeholders to participate. Model Security Question 2. It's essential we ensure the AI models we use do not become another cybersecurity vulnerability. Would voluntary cybersecurity standards for large AI models or high-risk models and the infrastructure they were trained on be helpful in establishing trust? Answer. Yes, voluntary cybersecurity standards for AI infrastructure can be helpful in establishing trust and establishing common vocabulary for stakeholders, particularly through industry-led initiatives. ______ ¶
Mr. Smith ## 21. 尊敬的罗杰·威克(Roger Wicker)提交的书面问题 → 布拉德·史密斯(Brad Smith) ### 主题:光纤电缆供应链 **背景:** 尽管本土制造有所扩张,美国仍严重依赖从泰国进口光纤电缆,并对来自中国的部件有特别依赖。2024 年,中国有 3 亿公里的过剩光纤产能,压低了全球价格。 **问题:** 史密斯先生,当我们谈 AI 基础设施时,往往聚焦算力需求以及为 AI 数据中心供电的能源缺口。但也存在关键的基础设施需求。本委员会常从宽带重要性的角度考虑光纤基础设施,但它对 AI、对我们保持在 AI 开发上领先的能力同样必不可少。能否谈谈把 AI 数据中心彼此连接起来的重要性,以及这为什么重要? **答复:** AI 工作负载需要大量算力,而光纤网络在为进出数据中心的高速、低时延实时数据传输方面扮演关键角色。超大规模云服务商往往拥有大量光纤传输基础设施,其中一部分自有,一部分通过与电信服务商签约使用。例如,微软建设了一张光纤 AI 广域网,以每秒 400 太字节的光纤骨干连接其数据中心版图,容量是传统数据中心所获容量的十倍。这张专网对微软云的运作至关重要,因为数据要在数据中心之间流动。 把 AI 数据中心彼此连接起来,对提升 AI 运营的性能、可靠性与可扩展性至关重要。互联的中心使分布式处理成为可能,优化资源利用,同时促进负载均衡,确保没有单一数据中心被压垮。实现冗余与故障切换机制,也能增强灾难恢复、确保连续运行。公共网络同样扮演关键角色,必须足够强健,才能可靠地把数据送入和送出数据中心网络。 Response to Written Question Submitted by Hon. Roger Wicker to Brad Smith Fiber Optic Cable Supply Chain Background: Despite some domestic manufacturing expansion, the U.S. remains heavily dependent on foreign imports of fiber optic cable from Thailand, with a particular dependency on components from China. In 2024, China had 300 million km of excess fiber capacity, which has depressed global prices. Question. Mr. Smith, when we talk about AI infrastructure, we often focus on compute needs and the gaps in sufficient energy to power AI data centers. But there's also a critical need for infrastructure. This committee often considers fiber infrastructure in terms of its importance to broadband, but it is also essential for AI and our ability to stay ahead in AI development. Can you describe the importance of connecting AI data centers to one another and why this matters? Answer. AI workloads require significant computational power and fiber networks play a crucial role in delivering high-speed, low- latency connections for real-time data transfer in and out of datacenters. Hyperscalers often have significant fiber optic transport infrastructure, some of which they own and some of which they contract with telecom service providers to use. For example, Microsoft has constructed a fiber-optic AI wide-area network connecting its data center footprint with a 400 terabyte per second fiber-optic backbone, capacity that is ten times what was enabled for traditional data centers. This private network is critical to the functioning of the Microsoft cloud as it moves data between datacenters. Connecting AI data centers to one another is vital for enhancing performance, reliability, and scalability for AI operations. Interconnected centers enable distributed processing and optimize resource utilization while facilitating load balancing to ensure no single datacenter is overwhelmed. Enabling redundancy and failover mechanisms also enhances disaster recovery and ensures continuous operation. Public networks play a crucial role too and must be sufficiently robust to reliably carry data to and from the data center network. ______ ¶
Mr. Smith ## 22. 尊敬的托德·杨(Todd Young)提交的书面问题 → 布拉德·史密斯 ### 主题:联邦对科学研究的投资 美国在技术上的全球领导地位并非偶然——它建立在数十年强劲、持续的联邦基础研究投入之上。 **问题 1.** 史密斯先生,能否谈谈为什么这类来自联邦政府的根基性支持在今天依然必不可少,以及它如何帮助美国成为全球创新引擎,尤其是在 AI 与量子等新兴领域? **答复:** 过去 80 年,美国以其科学与技术实力引领世界,产出了具有变革性的产品与能力。要在竞争中胜过中国这样的国家,我们必须大幅提升研发投入,并确保我们拥有专注于新兴技术的熟练研究者和科学家。专家预测,未来十年中国将持续向 AI、先进制造、清洁能源、量子计算和半导体等下一代技术投入大量资源。美国今天需要的那种自觉,与我们在 20 世纪初所拥有的一样——当时有一次刻意的努力,把产业界、政府和学术界汇聚起来推动科学进步。 AI 技术正处于一个拐点。恰恰在这个阶段——当科学突破即将规模化之时——公共投资最为重要。削减基础研究、特别是 AI 领域的基础研究,有可能延迟甚至使美国在这个现代 AI 时刻的发展轨迹脱轨。没有可预测的长期投资,美国将在科学领导地位和关键技术的商业化两方面落后。 ### 主题:AI 公众认知与教育 要赢下扩散竞赛,不仅需要为技术与应用更广泛地进入一般商业流程提供路径,还需要加强公众对 AI 的认知与教育。 **问题 2.** 史密斯先生,能否谈谈商务部长如何采取全政府的方式,来促进更高的 AI 素养认知、扩大 STEM 机会,以培养我们下一代的劳动力? **答复:** AI 如同所有新技术一样,将扰动经济、取代部分岗位。然而,我们相信 AI 将帮助降低许多职业的进入门槛、取代机械重复的任务,并为人类创造力建立在 AI 工具之上的基础。AI 将创造新的经济机会,让创业者创办新企业、创造新就业。我们已经在微软内部和整个经济中看到其中一些好处。领英(LinkedIn)最近一份报告强调的研究显示,使用 AI 的企业在创新与创造力上、甚至在扩大劳动力方面都看到了好处。 各种年龄和背景的美国人都需要 AI 技能,才能在这个新的工作世界中竞争。对多数人而言,一个关键机会是培养 AI 流畅度,使他们能在工作中使用 AI,就像今天使用笔记本电脑、智能手机、软件应用和互联网一样。 美国商务部包括 NIST、NTIA、MBDA 和人口普查局等关键机构,每一个在经济发展、数据治理和创新中都扮演独特角色。AI 素养可以通过与各部门使命对接而整合进来——例如,NIST 可以支持基于标准的 AI 教育,NTIA 可以把 AI 培训嵌入其数字项目。近期的立法和国家战略也呼吁该部牵头开展公众认知宣传,并通过图书馆、非营利组织和劳动力项目资助 AI 学习倡议。 ### 主题:AI 与量子 我们知道 AI 和量子计算都是战略前沿,但越来越清楚的是,我们最需要的突破可能来自这些技术如何相互作用。 **问题 3.** 史密斯先生,从你的视角看,美国应当如何思考 AI 与量子的融合——不只是作为两个彼此独立的优先事项,而是作为在基础技术上超越中国的统一战略的一部分? **答复:** 人工智能与量子计算是两个战略技术前沿——它们的融合有望开启前所未有的突破。在我们看来,AI 和量子计算都是「基础技术」,将驱动从医疗到国防的各行业创新。我们相信,把 AI 与量子计算结合起来存在强大的协同效应。例如,量子计算可以极大地加速 AI。某些让经典计算机吃不消的计算,在量子机器上可以指数级更快地运行。这暗示了量子在给机器学习优化、大规模数据分析、复杂模拟等 AI 任务「加涡轮」方面的潜力。具体来说,未来的量子协处理器或许能大幅加速处理高维数据的聚类算法,或使 AI 能分析经典计算机无法胜任的组合场景。反过来,AI 也能帮助量子进步。AI 算法协助设计更好的量子电路和纠错方法,本质上是用 AI 去克服量子工程难题。AI 还可以生成合成数据或启发式方法,引导量子算法开发——本质上,用 AI 去探索量子计算机应当处理哪些问题、以及如何处理。通过共同开发这些技术,我们开启新的能力——从更精准的药物发现(例如 AI 识别候选物、量子评估分子相互作用),到气候建模和材料科学的突破。 驾驭这股双重动能不仅是科学上的必需,也是战略上的必需。鉴于中国在人工智能和量子技术两方面协调一致、国家主导的推进,我们认为美国必须在研究、政策、产业和人才等所有维度上作出大胆、全面的投入。这包括对其融合的定向支持、跨学科协作,以及赋能私营部门创新。 Response to Written Questions Submitted by Hon. Todd Young to Brad Smith Federal Investment in Science Research America's global leadership in technology didn't happen by accident--it's been built on decades of strong, sustained Federal investment in basic research. Question 1. Mr. Smith, Can you speak to why that kind of foundational support from the Federal government remains essential today, and how it has helped position the United States as a global engine of innovation, particularly in emerging fields like AI and quantum? Answer. For the last 80 years, the United States has led the world with its scientific and technological prowess, resulting in transformative products and capabilities. To outcompete nations like China, we must significantly boost our R&D investments and ensure we have skilled researchers and scientists focused on emerging technologies. Experts predict China will continue to invest substantial resources into next-generation technologies such as AI, advanced manufacturing, clean energy, quantum computing, and semiconductors over the next decade. The United States needs the same level of intentionality today that we had in the early 20th century--where a deliberate effort brought together industry, government and academia to propel scientific advancement. AI technology is at an inflection point. It is precisely at this stage--when scientific breakthroughs are on the cusp of scaling--that public investment matters most. Cuts to basic research, particularly in AI, risk delaying or even derailing the U.S. trajectory in this modern AI moment. Without predictable, long-term investment, the United States will fall behind in both scientific leadership and the commercialization of critical technologies. AI Public Awareness and Education Winning the diffusion race not only requires providing a pathway for greater adoption of the technology and its applications into the general stream of commerce but also bolstering our public awareness and education of AI. Question 2. Mr. Smith, can you speak to how the Secretary of Commerce can have a whole of government approach to fostering greater awareness for AI literacy and growing STEM opportunities to create the next generation of our workforce? Answer. AI, like all new technologies, will disrupt the economy and displace some jobs. However, we believe AI will help lower the barriers to entry for many professions, replace rote tasks, and create a foundation for human creativity that builds on AI tools. AI will create new economic opportunities, allowing entrepreneurs to start new businesses and create new jobs. We are already seeing some of these benefits both at Microsoft and across the economy. A recent LinkedIn report highlighted research on how businesses using AI are seeing the benefits in innovation and creativity, and even in expanding their workforce. Americans of all ages and backgrounds need AI skills to complete in this new world of work. A key opportunity for most people will be to develop an AI fluency that will enable them to use AI in their jobs, much as they use laptops, smartphones, software applications, and the Internet today. The U.S. Department of Commerce includes key agencies like NIST, NTIA, MBDA, and the Census Bureau, each playing a unique role in economic development, data governance, and innovation. AI literacy can be integrated across these departments by aligning with their missions--for example, NIST can support standards-based AI education, and NTIA can embed AI training in their digital programs. Recent legislation and national strategies also call on the Department to lead public awareness campaigns and fund AI learning initiatives through libraries, nonprofits, and workforce programs. AI and Quantum We know AI and quantum computing are both strategic frontiers, but what's becoming clearer is that the breakthroughs we'll need most may come from how these technologies interact. Question 3. Mr. Smith, from your vantage point, how should the United States be thinking about the convergence of AI and quantum--not just as two separate priorities, but as part of a unified strategy to outpace China in foundational technologies? Answer. Artificial Intelligence and quantum computing are two strategic frontiers of technology--and their convergence is poised to unlock unprecedented breakthroughs. In our view, both AI and quantum computing are ``foundational technologies'' that will drive innovation across industries, from healthcare to defense. We believe there is a powerful synergy in combining AI with quantum computing. For example, quantum computing can vastly accelerate AI. Certain computations that choke classical computers can run exponentially faster on quantum machines. This hints at quantum's potential to turbocharge AI tasks like machine learning optimization, large-scale data analysis, and complex simulations. In practical terms, a future quantum co-processor might, for example, dramatically speed up a clustering algorithm working on high-dimensional data or enable AI to analyze combinatorial scenarios that are infeasible for classical computers. Conversely, AI can help quantum progress. AI algorithms assist in designing better quantum circuits and error-correction methods, essentially using AI to overcome quantum engineering challenges. AI can also generate synthetic data or heuristics to guide quantum algorithm development--essentially, using AI to explore which problems a quantum computer should tackle and how. By co-developing these technologies, we unlock new capabilities-- from more accurate drug discovery (e.g., AI identifying candidates, quantum evaluating molecular interactions) to climate modeling and materials science breakthroughs. Harnessing this dual momentum is not just a scientific imperative, it is a strategic necessity. Given China's coordinated, state-led advancements in both artificial intelligence and quantum technologies, we believe the United States must pursue bold, comprehensive investment across all dimensions: research, policy, industry, and workforce. This includes targeted support for their convergence, cross-disciplinary collaboration, and enabling private-sector innovation. ______ ¶
Mr. Smith ## 23. 尊敬的玛丽亚·坎特韦尔(Maria Cantwell)提交的书面问题 → 布拉德·史密斯(Brad Smith) ### 主题:AI 标准 美国若与世界上最先进的民主国家一道推动 AI 标准的制定,就有机会在全球舞台上为 AI 设定「道路规则」。 **问题 1.** NIST 的标准如何有助于美国的竞争力? **答复:** 由产业主导的自愿性标准,通过为 AI 系统的质量、安全性和互操作性建立全球公认的基准,对驱动美国竞争力起着关键作用。这些标准确保美国 AI 技术被视为可靠、可信且性能优异——这对国内和国际的采用都至关重要。NIST 在政府、产业界与公民社会之间的召集职能,鼓励美国实体积极参与前期标准化材料的制定,而这些材料最终将影响全球 AI 格局。NIST 的优先事项——作为本届政府优先事项的反映——也向产业界发出重要信号,鼓励各方围绕共同价值观与全球贸易实现战略对齐。 **问题 2.** 你希望 NIST 制定或推广哪些标准,以提升美国的竞争力? **答复:** 今天,即便在产业界内部,也缺乏共识性的参照基准,用来识别最可能出现的 AI 风险,以及如何可靠地评估和缓解这些风险。微软正与同业共同制定共识性最佳实践,但仍有扩大和加速这些工作的空间。应当发挥 NIST 在测试、评估、验证与确认(TEVV)以及计量科学方面的专长,为开发者和部署者提供结构与指引,说明如何最好地开展评估、以在加速采用的同时把风险降到最低。例如,NIST 可以制定「对评估本身进行评估」的实践,确保评估在科学上有效;也可以汇编评估清单,供政府和产业界用来评估特定风险或能力。 ### 主题:AI 出口与出口管制 美国需要一套强有力的国家技术出口战略,以及其他美国出口品的战略。联盟创造市场,而与盟友接触对于出口管制的有效协调与执行不可或缺。 **问题 3.** 美国政府——具体说是美国商务部——支持美国公司出口 AI 技术的最佳方式是什么?我们应当把重点放在哪里,以确保我们的外国对手不会抢先一步? **答复:** 国会和政府都在通过阻止对手获取先进技术部件来保护我们的国家安全方面扮演关键角色。政策必须保持平衡,使美国公司既能蓬勃发展,又能为该技术确立全球标准。 拟议中的「AI 扩散」规则,提出了一套要求建设 AI 数据中心的企业遵守的、相当完备的安全标准。我们认为,这些定性护栏为美国政策指出了合理方向,尤其是在涉及国家安全关切时。然而,我们不同意对各国进行「分级」,也不同意在缺乏通用许可的情况下施加某些量化限制——因为我们认为,任意施加限制会削弱国际社会对获取美国关键 AI 技术的信心,从而限制美国企业在全球竞争的能力。 **问题 4.** 你提到,世界各国只有在对美国 AI 有信任时才会使用它。美国政府如何与私营部门合作建立这种信任? **答复:** 商业规划与投资决策在很大程度上依赖双方之间可预测且可信的关系。同样地,不确定性会使决策与执行变得复杂并拖慢节奏。清晰而一致的政策,使像微软这样的公司能够对创新、劳动力发展和基础设施建设作出长期承诺,无论是在海外还是在国内。联邦政府有机会在自身的采购规则、以及对 AI 产品与服务的采用和使用上发挥引领作用。 ### 主题:隐私 数据是任何 AI 系统技术栈中的基础要素。AI 的进步将推动对数据的需求增长,既用于训练模型,也用于为模型提供事实依据。这进一步凸显了就消费者数据的收集与使用划定明确红线的必要性。划定这些红线的最佳方式,是制定一部强有力、全面性的联邦数据隐私法。 **问题 5.** 你是否同意,围绕消费者隐私划定明确红线将促进人工智能创新? **答复:** 同意。提高隐私保护水平将促进创新,并通过增强消费者信心、确保数据在仍可供有益用途使用的同时得到妥善保护,来惠及美国经济。这两个方面对 AI 发展都至关重要。消费者对其个人信息的收集与处理依然深感担忧。¹ 他们表示不清楚这些信息流向了哪里,并对社交媒体和 AI 带来的隐私丧失及其对儿童的影响表示忧虑。² 他们也对政府解决问题的能力缺乏信心:71% 的人表示,他们并不指望社交媒体公司会被约束,尽管有 77% 的人表示希望政府采取更多行动。³ 这种担忧跨越政治光谱:68% 的共和党人和 78% 的民主党人认为应当对个人信息施加更多监管。⁴ 从某些指标看,问题实际上还在恶化。 **脚注:** 1. 《美国人如何看待数据隐私》,Colleen McClain、Michelle Faverio、Monica Anderson、Eugenie Park,皮尤研究中心,2023 年 10 月 18 日。见:https://www.pewresearch.org/internet/2023/10/18/how-americans-view-data-privacy/ 2. 同上。 3. 同上。 4. 同上。 表示自己对隐私法「了解很少/完全不了解」的成年人比例,已从 2019 年的 63% 上升到 2023 年的 72%。⁵ 这一切都不应令人意外。消费者被各种关于个人信息丢失或被滥用可能如何伤害他们的信息所淹没。无论是通过数据泄露通知、对日益精密的身份盗用骗局的恐惧,⁶ 雇主对个人信息的滥用,⁷ 还是诸如健身应用泄露美军秘密基地位置这类出人意料的负面后果,⁸ 个人信息的利用与滥用都可能渗入消费者生活的方方面面。 **脚注:** 5. 同上。 6. 《威瑞森(Verizon)2024 年数据泄露调查报告》,见:2024 Data Breach Investigations Report | Verizon 7. Michelle Boorstein、Marisa Iati、Annys Shin,《美国天主教高层官员在手机数据被用于追踪其 Grindr 使用及同性酒吧行踪后辞职》,《华盛顿邮报》(2021 年 7 月 21 日),见:https://www.washingtonpost.com/religion/2021/07/20/bishop-misconduct-resign-burrill/ 8. Jeremy Hsu,《Strava 热力图与秘密的终结》,《连线》(Wired,2018 年 1 月 29 日),见:Strava Data Heat Maps Expose Military Base Locations Around the World | WIRED 这种消费者信心的缺失正在直接影响经济增长。世界贸易组织与经合组织(OECD)最近的一份报告发现,如果所有国家都采用隐私保障措施,全球出口将增长 3.6%,全球 GDP 将增长 1.77%。⁹ 而缺乏隐私保护将使全球 GDP 下降近 1%、全球进口下降超过 2%,其中高收入经济体受到的影响最大。¹⁰ 美国隐私格局的碎片化——例如目前美国各州拼凑而成的州法体系——也有可能对美国经济造成负面影响。一项行业估算认为,州法拼凑体系的成本在十年内将超过 1 万亿美元,其中 2,000 亿美元的负担落在小企业身上。¹¹ 鉴于这些重要收益与迫近的潜在成本,通过联邦综合性隐私立法对于促进 AI 创新至关重要。 **脚注:** 9. OECD/WTO(2025),《数据监管的经济影响:平衡开放与信任》,OECD 出版,巴黎,https://doi.org/10.1787/aa285504-en。 10. 同上,第 39 页。 11. 信息技术与创新基金会(ITIF),《州隐私法拼凑体系迫近的成本》(2022 年 1 月),见:https://itif.org/publications/2022/01/24/50-state-patchwork-privacy-laws-could-cost-1-trillion-more-single-federal/ **问题 6.** 微软是否会与我合作,推动联邦综合性隐私立法以划定这些明确红线? **答复:** 会。微软将与你以及国会任何其他成员合作,推动联邦综合性隐私立法。我们自 2005 年起就一直在倡导此事,因为我们认为这对建立消费者信任、推动美国经济增长至关重要。¹² **脚注:** 12. 《微软倡导联邦综合性隐私立法》,微软新闻,2005 年 11 月 3 日,见:https://news.microsoft.com/2005/11/03/microsoft-advocates-comprehensive-federal-privacy-legislation/ ### 主题:能源需求与聚变能源研发 为 AI 数据中心供电的电力需求增长之快令人震惊。据一些估算,到 2030 年全球数据中心用电需求预计将增长一倍以上,超过 945 太瓦时(TWh)。这将对电网和能源供应商造成压力。 **问题 7.** 贵公司有什么计划来满足 AI 的能源需求?你们正在对聚变等非化石能源作哪些投资? **答复:** 在考虑任何电力来源时,我们都寻求可靠、可规模化且具成本效益、并能按与我们数据中心增长相匹配的快速时间表开发出来的方案。为抵销我们产生碳排放的用电,微软已签订购电协议,在全球新增 34 吉瓦可再生能源装机。2024 年,我们签订了一份购电协议,从星座能源(Constellation)的鹤清洁能源中心(Crane Clean Energy Center)新增 835 兆瓦无碳核电,该机组预计于 2028 年投入运行,将为 PJM 互联电网提供额外容量。 2023 年,我们宣布与 Helion Energy 达成协议,从其正在开发的首座聚变发电厂采购电力。我们对聚变能源未来能在电网结构中发挥重要作用持乐观态度。 **问题 8.** 就聚变能源而言,政府如何与私营部门合作,在聚变技术持续发展的过程中实现规模化? **答复:** 要实现聚变突破并使聚变能源规模化,政府与私营部门的伙伴关系不可或缺,用以降低风险、加速创新,并为聚变商业化奠定基础。政府可以支持成本分担项目,并针对聚变采用清晰、可适应的专门监管框架,以加快部署。把公共资源与私营创新对齐,有助于让聚变能源更快、更可负担地进入市场。 ### 主题:AI 安全 我们看到深度伪造及其他 AI 内容大量涌现,威胁到普通人在媒体中辨别真相的能力。而这只是 AI 领域中提出复杂安全问题的其中一个方面。美国 AI 安全研究所在确保 AI 系统负责任地开发、确保最先进模型得到充分测试方面扮演关键角色。这对建立信任、促进更广泛的采用至关重要。 **问题 9.** 你是否支持美国 AI 安全研究所的工作? **答复:** 支持。我们支持美国 AI 标准与创新中心(原美国 AI 安全研究所)的工作。他们在促进人工智能可靠发展、建立严格安全标准、推动跨部门协作以及应对潜在风险方面的努力,对负责任的创新至关重要。我们认同他们通过可靠性、透明度和问责原则确保 AI 造福社会的承诺。 ### 主题:AI 劳动力 美国需要一支熟练的劳动力来建设和维护 AI 基础设施,包括电工、管道工、木匠和工程师。劳动力短缺已经在拖慢数据中心建设,而我们却没有培养或留住这些人才的全国性战略。 **问题 10.** 你认为国会和政府应当做些什么来支持 AI 教育、培训和劳动力发展? **答复:** AI 如同所有新技术一样,将扰动经济并取代部分岗位,我们知道这引起了许多人和整个劳动力的担忧。我们相信,AI 将创造新的机会,其收益将超过前方的许多挑战。AI 将帮助降低许多职业的进入门槛、取代机械重复的任务,并为建立在 AI 工具之上的人类创造力奠定基础。AI 将创造新的经济机会,让创业者创办新企业、创造新就业。我们已经在微软内部和整个经济中看到其中一些好处。领英(LinkedIn)最近一份报告强调了关于使用 AI 的企业如何在创新与创造力、甚至在扩大劳动力方面获益的研究。我们鼓励国会投资于聚焦重大 AI 基础设施计划的联邦学徒项目,设立在职培训的联邦项目,并支持重新授权《劳动力创新与机会法》(Workforce Innovation and Opportunity Act)。 **问题 11.** 贵公司及其他企业面临本届政府移民政策的哪些挑战?你对高技能移民所受影响有何关切? **答复:** 美国面临严重的人才瓶颈。我们的大学培养了世界上一些最有才华的工程师、科学家和创业者,但过时的移民政策往往迫使他们在毕业不久后就离开美国。像我们这样的公司在签证和绿卡流程上面临漫长延误与持续的不确定性,这使得招聘和留住所需人才变得困难。在人工智能、先进工程和量子计算等高需求领域,对专门技能的需求远远超过国内可用供给。 这并非要取代美国劳动者,而是通过以全球人才补充美国劳动力来促成创新。要建设未来的技术,我们需要与来自世界各地最优秀的头脑合作。美国公司为全球客户服务,让来自世界各地聪慧的人才在我们的团队中工作,有助于我们打造真正满足全球受众需求的产品。 为保持全球竞争力,美国需要一个能够吸引并留住顶尖人才的移民体系。这意味着为高技能人才——尤其是美国大学的毕业生——创造快速、可靠的通道,并解决来自等待时间极长国家的劳动者的绿卡积压问题。如果我们想在 AI 及其他新兴技术上引领世界,就必须继续做吸引世界最优秀、最聪明人才的磁石,并确保他们有机会在美国建设自己的未来。 Response to Written Questions Submitted by Hon. Maria Cantwell to Brad Smith AI Standards The U.S. driving development of AI standards alongside the most advanced democracies in the world offers us an opportunity to set the ``rules of the road'' for AI on the global stage. Question 1. How do NIST standards help the United States' competitiveness? Answer. Industry-led voluntary standards are instrumental in driving U.S. competitiveness by establishing globally recognized benchmarks for quality, safety, and interoperability in AI systems. These standards ensure that American AI technologies are seen as reliable, trustworthy, and high performing, which is critical for adoption both domestically and internationally. NIST's convening function across government, industry, and civil society encourages active participation by U.S. entities in the development of pre- standardization materials--which will ultimately influence the global AI landscape. NIST's priorities--as a reflection of the Administration's priorities--also provide an important signal to industry, encouraging strategic alignment on shared values and global trade. Question 2. What standards would you like to see NIST develop or promote to improve U.S. competitiveness? Answer. Today, there is a dearth of consensus reference points-- even across industry--that identify the most likely AI risks and how to evaluate and mitigate them reliably. Microsoft is working with industry peers to develop consensus best practices, but there are opportunities to expand and accelerate these efforts. NIST's expertise in test, evaluation, validation, and verification (TEVV) and measurement science should be leveraged to provide structure and guidance to developers and deployers on how to best develop and use evaluations to accelerate adoption while minimizing risks. For example, NIST could develop practices for evaluating evaluations to ensure they are scientifically valid and could assemble panels of evaluations for use by the government and industry to evaluate specific risks or capabilities. AI Exports and Export Controls The U.S. needs a strong national export strategy for technology and other U.S. exports. Alliances create markets and engaging our allies is essential to effective coordination and implementation of export controls. Question 3. What is the best way for the U.S. Government-- specifically, the U.S. Department of Commerce--to support U.S. companies in exporting AI technologies? Where should we target to make sure our foreign adversaries do not get there first? Answer. Congress and the Administration both have key roles in protecting our national security by preventing adversaries from obtaining advanced technology components. Policies must be balanced so that American companies can thrive and set the global standard for the technology. The proposed ``AI Diffusion'' rule, sets forth a robust set of security standards expected of companies building AI datacenters. We believe these qualitative guardrails point in a sensible direction for U.S. policy especially as it relates to national security concerns. However, we disagree with the ``tiering'' of countries as well as certain quantitative restricts--absent a universal license--as we think arbitrary restrictions undermine international confidence in access to critical American AI technology and therefore restrict the ability of U.S. firms to compete globally. Question 4. You mentioned that countries around the world will only use American AI if they can trust it. How can the U.S. government partner with the private sector to build that trust? Answer. Business planning and investment decisions rely heavily on predictable and trusted relationships on both sides. By the same token, uncertainty complicates and slows decision making and execution. Clear and consistent policies enable companies like Microsoft to make long- term commitments to innovation, workforce development, and infrastructure investment both abroad and domestically. The Federal government has an opportunity to lead in their own procurement rules and the adoption and utilization of AI products and services. Privacy Data is a foundational element in the tech stack for any AI system. Advances in AI will spur an increase in demand for data, both to train and ground AI models. This enhances the need for bright lines related to consumer data collection and usage. The best way to set these bright lines is through a strong, comprehensive Federal data privacy law. Question 5. Do you agree that bright lines around consumer privacy will spur innovation in artificial intelligence? Answer. Yes, raising privacy protections will spur innovation and benefit the U.S. economy by bolstering consumer confidence and ensuring that data will be appropriately protected while still being available for beneficial uses. Both components are critical to AI development. Consumers remain deeply concerned about the collection and processing of their personal information.\1\ They report confusion about where it is going and cite concern over a loss of privacy from social media, AI, and impact on children.\2\ They also report little faith in government's ability to solve the problem, with 71 percent saying that they do not expect social media companies to be reined in even as 77 percent say they would like more government action.\3\ Concern is shared across the political spectrum with 68 percent of Republicans and 78 percent of Democrats saying there should be more regulation over personal information.\4\ By some measures the problem is actually getting worse. --------------------------------------------------------------------------- \1\ How Americans View Data Privacy, Colleen McClain, Michelle Faverio, Monica Anderson and Eugenie Park, Pew Research Center, October 18, 2023. See: https://www.pewresearch.org/internet/2023/10/18/how- americans-view-data-privacy/ \2\ Id. \3\ Id. \4\ Id. --------------------------------------------------------------------------- The share of adults reporting that they understand privacy laws very little/not at all has risen from 63 percent in 2019 to 72 percent in 2023.\5\ None of this should be a surprise. Consumers are deluged by the ways that lost or misused personal information can harm them. Whether it's through data breach notifications and fears over increasingly sophisticated identity theft scams,\6\ misuse of personal information by employers,\7\ or unexpected negative outcomes such as a fitness app revealing the location of secret U.S. military bases,\8\ the use and misuse of personal information can spill into every facet of a consumer's life. --------------------------------------------------------------------------- \5\ Id. \6\ Verizon 2024 Data Breach Investigations Report, see: 2024 Data Breach Investigations Report | Verizon \7\ Michelle Boorstein, Marisa Iati, and Annys Shin, ``Top U.S. Catholic Church official resigns after cellphone data used to track him on Grindr and to gay bars,'' Wash. Post (July 21, 2021), see: https:// www.washingtonpost.com/religion/2021/07/20/bishop-misconduct-resign- burrill/ . \8\ Jeremy Hsu, ``The Strava Heat Map and the End of Secrets'', Wired (Jan 29, 2018), see: Strava Data Heat Maps Expose Military Base Locations Around the World | WIRED --------------------------------------------------------------------------- This lack of consumer confidence is directly impacting economic growth. A recent report by the World Trade Organization and the OECD found that if all nations adopted privacy safeguards, global exports would increase by 3.6 percent and global GDP would grow by 1.77 percent.\9\ An absence of privacy protections would cause global GDP to fall by almost 1 percent and global imports by over 2 percent with the largest impact on high-income economies.\10\ A fragmented U.S. privacy landscape--such as the state patchwork currently in place in the US-- also risks negatively impacting the U.S. economy. One industry estimate puts the cost of a patchwork of state laws at over $1 trillion over 10 years, with $200 billion of that burden falling on small businesses.\11\ Given these important benefits and looming potential costs, passage of Federal comprehensive privacy legislation is critical for spurring AI innovation. --------------------------------------------------------------------------- \9\ OECD/WTO (2025), Economic Implications of Data Regulation: Balancing Openness and Trust, OECD Publishing, Paris, https://doi.org/ 10.1787/aa285504-en. \10\ Id pg. 39. \11\ ITIF, ``The Looming Cost of a Patchwork of State Privacy Laws,'' (January 2022) see: https://itif.org/publications/2022/01/24/ 50-state-patchwork-privacy-laws-could-cost-1-trillion-more-single- federal/ Question 6. Will Microsoft work with me on Federal comprehensive privacy legislation to set those bright lines? Answer. Yes, Microsoft will work with you and any other members of congress on Federal comprehensive privacy legislation. We've been advocating for it since 2005, because we believe it is critical to building consumer trust and driving economic growth in the United States.\12\ --------------------------------------------------------------------------- \12\ ``Microsoft Advocates Comprehensive Federal Privacy Legislation,'' Microsoft News, November 3, 2005, see: https:// news.microsoft.com/2005/11/03/microsoft-advocates-comprehensive- federal-privacy-legislation/. --------------------------------------------------------------------------- Energy Needs and R&D for Fusion Energy The growing demand for electricity to power AI data centers is staggering. By some estimates, global electricity demand from data centers is projected to more than double by 2030 exceeding 945 terawatt-hours (TWh). It will strain electric grids and energy providers. Question 7. What plan does your company have to meet energy needs for AI, and what investments are you making into non-fossil fuel sources of energy such as fusion? Answer. As we consider any power sources, we look for reliable, scalable and cost-effective options that can be developed in rapid timelines aligned to our datacenter growth. To match our carbon emitting electricity use, Microsoft has entered into power purchase agreements to add 34 GW of renewable energy capacity globally. In 2024, we entered into a power purchase agreement to add 835 MW of carbon- free, nuclear energy from Constellation's Crane Clean Energy Center that is anticipated to be put into service by 2028, which will provide additional capacity to the PJM interconnect. In 2023, we announced an agreement with Helion Energy to procure power from its first fusion power plant that is under development. We are optimistic that fusion energy can play an important part of the grid mix in the future. Question 8. With respect to fusion energy, how can the government partner with the private sector to scale fusion technology as it continues to develop? Answer. To achieve fusion breakthrough and scale fusion energy, government and private sector partnerships are essential in order to reduce risk, accelerate innovation, and build the foundation for fusion commercialization. The government can support cost-sharing programs and adopt clear, adaptive regulatory frameworks specific to fusion to accelerate deployment. Aligning public resources with private innovation can help bring fusion energy to market faster and more affordably. AI Safety We are seeing a proliferation of deepfakes and other AI content that threatens the average person's ability to discern truth in media. And that's just one area in the field of AI that presents complicated safety questions. The U.S. AI Safety Institute plays a critical role in ensuring that AI systems are developed responsibly and that the most advanced models are fully tested. This is crucial for building trust and promoting wider adoption. Question 9. Do you support the work of the U.S. AI Safety Institute? Answer. Yes, we support the work of the U.S. Center for AI Standards and Innovation, formerly known as the U.S. AI Safety Institute. Their efforts to promote reliable development of artificial intelligence, establish rigorous safety standards, foster cross-sector collaboration, and address potential risks are crucial for responsible innovation. We share their commitment to ensuring AI benefits society through principles of reliability, transparency, and accountability. AI Workforce The U.S. needs a skilled workforce to build and maintain AI infrastructure, including electricians, pipefitters, carpenters, and engineers. Labor shortages are already slowing down data center construction, yet there is no national strategy to train or retain this talent. Question 10. What do you think Congress and the administration should be doing to support AI education, training, and workforce development? Answer. AI, like all new technologies, will disrupt the economy and displace some jobs, and we which we know causes concern for many people and the workforce at large. We believe AI will create new opportunities that will outweigh many of the challenges ahead. AI will help lower the barriers to entry for many professions, replace rote tasks, and create a foundation for human creativity that builds on AI tools. AI will create new economic opportunities, allowing entrepreneurs to start new businesses and create new jobs. We are already seeing some of these benefits both at Microsoft and across the economy. A recent LinkedIn report highlighted research on how businesses using AI are seeing the benefits in innovation and creativity, and even in expanding their workforce. We encourage congress to invest in Federal apprenticeships focused on major AI infrastructure initiatives, establish Federal programs for on-the-job training and support the reauthorization of the Workforce Innovation and Opportunity Act. Question 11. What challenges do your companies and others face from the administration's immigration policies, and what concerns do you have about impacts on high-skilled immigration? Answer. The United States faces a critical talent bottleneck. Our universities educate some of the world's most talented engineers, scientists, and entrepreneurs, but outdated immigration policies often force them to leave the country shortly after graduation. Companies like ours face long delays and constant uncertainty in the visa and green card process, which makes it difficult to hire and retain the talent we need. In high-demand fields like artificial intelligence, advanced engineering, and quantum computing, the need for specialized expertise far exceeds the available domestic supply. This isn't about replacing American workers. It's about enabling innovation by complementing the U.S. workforce with global talent. To build the technologies of the future, we need to work with the best minds from around the world. American companies serve customers across the globe, and having a smart people from around the world work on our teams helps us create products that truly meet the needs of a global audience. To remain globally competitive, the United States needs an immigration system that attracts and retains top talent. That means creating fast, reliable pathways for highly skilled individuals-- especially graduates of U.S. universities--and addressing the green card backlog for workers from countries facing extreme wait times. If we want to lead the world in AI and other emerging technologies, we must continue to be a magnet for the world's best and brightest--and ensure they have the opportunity to build a future in the United States. ______ ¶
Mr. Smith ## 24. 尊敬的布莱恩·沙茨(Brian Schatz)提交的书面问题 → 布拉德·史密斯 ### 主题:工作的未来 **问题 1.** 你对工作的未来有何愿景?在近期和长期,未来的有价值岗位是什么? **答复:** 我们相信 AI 将极大地影响工作的未来——而且大体上是往好的方向。同时,现在也是政府参与这些讨论的重要时刻。今年年初,我就这一问题发表了一篇博文。领英最近也发布了一份报告《AI 与全球经济》,考察 AI 已经在如何影响劳动力和整个经济。 **问题 2.** 贵公司如何利用你们所赋能的自动化来提升生产率,同时不让劳动者掉队? **答复:** 我们设计技术是为了增强人的能力——让人们在工作中的每一个小时都能取得更大的成效。这意味着投资于覆盖面广的技能培训倡议,包括学徒项目与劳动力伙伴关系,使每一位劳动者——不只是科技行业的人——都能从 AI 中受益。我们鼓励员工和合作伙伴试用 Copilot 这类工具,为自身工作找到最佳用例。我们还建立了反馈机制,以确保持续改进,并让劳动者体验始终处于我们创新的中心。 ### 主题:中国对 AI 的部署 **问题 3.** 你认为中华人民共和国部署 AI 系统所带来的最大国家安全威胁是什么? **答复:** 中华人民共和国部署 AI 系统带来一系列国家安全风险。例如,中国技术向第三国的扩散可能损害全球网络安全与信息完整性。通过先进的 AI 驱动监控、网络间谍活动和虚假信息行动,中国可能利用关键基础设施的漏洞、操纵公众舆论,并侵蚀人们对民主制度的信任。 **问题 4.** 你有哪些应对这些威胁的建议? **答复:** 微软主张以全面、多层次的方式来应对中国带来的国家安全风险。为应对 AI 驱动的监控与网络间谍活动,微软建议部署零信任架构、增强终端检测与响应能力,并投资于 AI 驱动的威胁情报,以侦测和缓解高级持续性威胁。加强公私伙伴关系对确保实时信息共享、协调应对网络入侵不可或缺。 微软也认识到包括中国在内的民族国家行为体所带来的风险——它们运用网络手段实施影响力行动,针对美国境内的关键机构。微软调动了「AI 向善实验室」(AI for Good Lab)与微软威胁分析中心(MTAC)团队的数据科学与技术能力,评估这些风险,包括这些行为体是否在这些行动中使用了 AI。在适当情况下,团队会借助微软数字犯罪部门的专长,投入资源并把对 AI 驱动犯罪活动的早期侦测投入实战,并通过主动提起民事诉讼以阻断和威慑此类活动、以及通过威胁情报项目和数据共享,作出相应回应。 此外,微软致力于推进信息完整性,并认为引入内容凭证是这方面的重要驱动力。我们是内容来源与真实性联盟(C2PA)的创始成员。为实现透明度、支持信息完整性并赋能用户,我们正在多款产品中采用 C2PA 的「内容凭证」开放标准。例如,在领英上,含有「内容完整性」技术的媒体已被自动标注,用户开始在含 C2PA 元数据的图片和视频上看到「Cr」图标。 除改进防御的技术方案之外,微软也强调政治解决方案对这些威胁的重要性——各国政府应在对中国恶意网络活动的归因、曝光和威慑方面发挥更主动、更协调的作用。这包括及时的公开归因、外交压力,以及在适当情况下的法律行动。对国家支持的网络攻击给予明确后果,对改变成本收益计算、强化反对数字侵略的国际规范至关重要。 对于以关键基础设施为目标、意在为未来突发事件「预置」破坏性或摧毁性攻击的网络入侵,这一点尤其成立。此类行动使平民面临重大风险,与传统的间谍活动有实质区别。预置网络攻击应被认定为国际法所禁止的「武力」威胁,必须通过跨领域的充分后果加以威慑。 最后,确保美国 AI 技术保持全球领先、并确保监管与政策不会不必要地阻碍美国 AI 的全球扩散,是对抗中国的关键一环。 ### 主题:加速科学研究 **问题 5.** 微软的「AI 助力健康」和「AI 助力地球」项目已支持了数百个学术与非营利研究项目。你认为政府应如何更好地构建合作结构,以利用 AI 加速科学发现? **答复:** *见下方答复。 > 译注:原文此处即为「\*See answer below」,指向紧随其后的第 6 问答复,属原文件的交叉引用,非脱漏。 **问题 6.** 你一直推动环境与生物医学数据集的汇聚和标准化。产业界和政府还应共同采取哪些进一步措施,以确保研究者能够获得结构良好、可互操作且已标注的数据集,用于 AI 驱动的发现? **答复:** 政府可以通过支持让数据更易被发现、获取和使用的倡议,在扩大数据集可及性和加速科学发现方面发挥关键作用。发起一场全国性开放数据运动,将使国家档案馆、国会图书馆、史密森学会及其他政府机构有能力把非机密、非敏感数据数字化、整理并共享出来用于 AI 训练。此外,在非营利、学术和文化机构内部创建并扩展开放数据公地,可以进一步让有价值数据集的使用民主化。解决限制科研获取的付费墙、设立专项资金以解锁闭源期刊,也将大大有助于让关键知识可用于创新。 政府还应优先制定并执行清晰的数据共享与标注准则。需要更强的激励与监督机制,确保由联邦资助的研究数据按照资助协议及时、安全地共享。支持数据集采用元数据与来源标准,可以提升其效用与可靠性,在 AI 模型生成合成数据的时代尤其如此。通过采取这些步骤,政府可以培育一个强健的生态,使数据成为科学进步与技术发展的支柱。 ### 主题:能源消耗与「赢下竞赛」的代价 **问题 7.** 你是否支持特朗普总统为 AI 数据中心扩大煤电的努力? **答复:** 可靠、有韧性且具成本效益的电力供应,对美国经济增长不可或缺。总统近期关于煤电的行政令,凸显了满足能源部门需求增长的紧迫性,包括在建设新型能源选项的过程中,把煤电作为通向未来无碳发电的桥梁。在微软,当我们考虑任何潜在的专用电源时,我们寻找的是与我们对于可靠、具成本效益且可持续电力的需求相契合的发电选项。我们也投资于下一代能源供应技术,以支持加速创新和成本下降。在实践中,这意味着我们使用并投资于多样化的发电技术组合,以契合我们对可靠且具成本效益电力的需求。这些技术还必须契合我们到 2030 年实现碳负排放的承诺。在可靠性、成本效益与可持续性所有这些优先事项上,我们都看不到扩大使用煤炭来满足数据中心能源需求的商业理由。相反,我们认为有越来越强的理由在系统中纳入更多可再生能源、核能、净零和净负排放发电技术,以及多样化的储能技术。尽管这些新技术中有许多在未来提供了有前景的选项,我们也认识到,在壮大国家能源基础设施的过程中,有时确实需要在近期建设其他选项。 **问题 8.** 你如何应对短期内与新基础设施建设相关的成本,包括用水、污染和气候影响方面? **答复:** 当为服务我们的负荷而需要改进电网时,我们与当地公用事业公司合作,确保我们为服务本站点所需的改进付费。从决定建设数据中心的那一刻起,我们就审慎考虑 AI 的资源强度。我们正致力于推进低碳材料并创建全球市场,以帮助推动各行业的可持续性。在运营数据中心时,我们优化能效与水效,包括公布一种制冷零耗水的新数据中心设计。我们也越来越多地采用循环方式,以实现到 2030 年零废弃的目标。我们的做法是先着力防止产生废弃物,再聚焦于再利用和材料回收。这包括对建筑与拆除废弃物进行再利用和回收、把运营废弃物分流出去,以及推进云端硬件的循环利用与包装循环。 Response to Written Questions Submitted by Hon. Brian Schatz to Brad Smith Future of Work Question 1. What is your vision of the future of work and what are the valuable jobs of the future, in the near-term and long-term? Answer. We believe that AI is going to impact the future of work tremendously--largely for the better. Additionally, this is an important moment for the government to be engaged in these conversations. At the beginning of the year, I published a blog post connected to this question. LinkedIn also recently published a report, AI and the Global Economy, examining how AI is already impacting the workforce and the economy at large. Question 2. How is your company taking advantage of the automation you're empowering to scale productivity without leaving workers behind? Answer. We design our technologies to augment human capability-- empowering people to achieve greater impact with each hour they spend at work. That means investing in broad-based skilling initiatives, including apprenticeships and workforce partnerships, so every worker, not just those in tech, can benefit from AI. We encourage employees and partners to experiment with tools like Copilot to find the best use cases for their work. And we've built feedback mechanisms to ensure continuous improvement and keep worker experience at the center of our innovation. PRC Deployment of AI Question 3. What do you believe is the greatest national security threat posed by the People's Republic of China's deployment of AI systems? Answer. The People's Republic of China's deployment of AI systems presents a range of national security risks. For example, the spread of Chinese technology to third countries may undermine global cybersecurity and information integrity. Through advanced AI-driven surveillance, cyber-espionage, and disinformation campaigns, China could exploit vulnerabilities in critical infrastructure, manipulate public opinion, and erode trust in democratic institutions. Question 4. What are your recommendations for addressing these threats? Answer. Microsoft advocates a comprehensive, multi-layered approach to counter the national security risks posed by China. To address AI- driven surveillance and cyber-espionage, Microsoft recommends deploying zero-trust architectures, enhancing endpoint detection and response capabilities, and investing in AI-powered threat intelligence to detect and mitigate advanced persistent threats. Strengthening public-private partnerships is essential to ensure real-time information sharing and coordinated responses to cyber intrusions. Microsoft also recognizes the risks posed by nation-state actors including China employing cyber enabled influence operations targeting critical institutions within America. Microsoft harnessed the data science and technical capabilities of our AI for Good Lab and Microsoft Threat Analysis Center (MTAC) teams to assess these risks including whether or not these actors were utilizing AI in these operations. When appropriate, the team calls on the expertise of Microsoft's Digital Crimes Unit to invest in and operationalize the early detection of AI- powered criminal activity and respond fittingly, through the filing of affirmative civil actions to disrupt and deter that activity and through threat intelligence programs and data sharing. In addition, Microsoft is committed to advancing information integrity and believes that including content credentials is an important driver for this. We are a founding member of the Coalition for Content Provenance and Authenticity (C2PA). To achieve transparency, support information integrity, and empower our users, we are leveraging C2PA's ``content credentials'' open standard across several products. For example, content containing the ``Content Integrity'' technology has been automatically labelled on LinkedIn, with users beginning to see the ``Cr'' icon on images and videos that contain C2PA metadata. Beyond technological solutions that improve defenses, Microsoft also stresses the importance of political solutions to these threats-- governments taking a more proactive and coordinated role in attributing, exposing, and deterring Chinese malicious cyber activity. This includes timely public attribution, diplomatic pressure, and legal action where appropriate. Clear consequences for state-sponsored cyberattacks are essential to shift the cost-benefit calculus and reinforce international norms against digital aggression. This is especially true for cyber intrusions which target critical infrastructure in order to ``preposition'' for disruptive or destructive attacks in a future contingency. Such operations put civilians at significant risk and are meaningfully different from traditional espionage. Prepositioning cyberattacks should be recognized as a ``threat'' of force prohibited by international law that must be deterred via sufficient consequences across domains. Finally, ensuring that U.S. AI technology retains its global leadership and that regulations and policies do not unnecessarily hinder the global diffusion of U.S. AI is a key component to countering China. Accelerating Scientific Research Question 5. Microsoft's AI for Health and AI for Earth programs have supported hundreds of academic and nonprofit research projects. How would you suggest the government can better structure collaborations to accelerate scientific discovery using AI? Answer. *See answer below. Question 6. You've led efforts to aggregate and standardize environmental and biomedical datasets. What further steps should industry and government take together to ensure researchers have access to well-structured, interoperable, and annotated datasets for AI-driven discovery? Answer. The government can play a pivotal role in increasing access to datasets and accelerating scientific discovery by supporting initiatives that make data more discoverable, accessible, and usable. Launching a national open data campaign would empower institutions such as the National Archives, the Library of Congress, the Smithsonian, and other government agencies to digitize, organize, and share non- classified and non-sensitive data for AI training. Additionally, the creation and expansion of open data commons within non-profit, academic, and cultural institutions could further democratize access to valuable datasets. Addressing paywalls that restrict access to scientific research and establishing dedicated funds to unlock closed- access journals would also contribute significantly to the availability of critical knowledge for innovation. The government should also prioritize the development and enforcement of clear guidelines for data sharing and annotation. Stronger incentives and monitoring mechanisms are needed to ensure timely and secure sharing of federally funded research data, in alignment with funding agreements. Supporting the adoption of metadata and provenance standards for datasets can enhance their utility and reliability, particularly in the era of synthetic data generated by AI models. By taking these steps, the government can foster a robust ecosystem where data serves as the backbone for scientific advancement and technological progress. Energy Consumption and Cost of Winning Question 7. Do you support President Trump's efforts to expand coal power for AI data centers? Answer. The availability of reliable, resilient, and cost-effective electricity is essential for economic growth in the United States. The recent Executive Orders from the President on coal generation highlight the urgency to meet energy sector demand growth, including as a bridge to future carbon free generation as new energy options are built out. At Microsoft, when we consider any potential dedicated power source, we look for generation options that align with our need for reliable, cost-effective, and sustainable electricity. We also invest in the next generation of energy supply technologies to support accelerated innovation and cost-declines. In practice, this has meant that we use and invest in a diverse mix of electricity generation technologies that align with our needs for reliable and cost-effective electricity. These technologies also must align with our commitment to be carbon negative by 2030. Across all of these priorities--reliability, cost- effectiveness, and sustainability--we do not see a business case for expanded use of coal to meet the energy needs of our data centers. Rather, we see increasingly strong arguments for integrating more renewables, nuclear, net-zero, and net-negative electricity generation technologies into the system in addition to a diverse set of energy storage technologies. While many of these new technologies present promising options in the future, we also recognize that there will sometimes be a need to build other options in the near term as we grow our Nation's energy infrastructure. Question 8. How are you addressing the costs associated with new infrastructure development in the short-term, including in terms of water consumption, pollution, and climate impacts? Answer. When improvements to the grid are required to serve our load, we work with our local utility to ensure that we pay for the improvements required to serve our site. We are thoughtful about the resource intensity of AI from the moment that we decide to build a datacenter. We're working to advance low-carbon materials and create global markets to help advance sustainability across industries. When operating our datacenters, we optimize energy and water efficiency, including announcing a new datacenter design that consumes zero water for cooling. We also are increasingly adopting a circular approach to reach our target of zero waste by 2030. We target preventing waste first, then focus on reusing, and recovering materials. This includes reusing and recycling construction and demolition waste, diverting operational waste and advancing circular cloud hardware and packaging. ______ ¶
Mr. Smith ## 25. 尊敬的埃德·马基(Edward Markey)提交的书面问题 → 布拉德·史密斯 ### 主题:数据中心建设的综合影响 **问题 1.** 在规划数据中心建设时,贵公司是否开展「从摇篮到坟墓」的基础设施研究,涵盖数据中心运营期内及之后的野生动物、社区与污染影响?如果是,你们从这些研究中学到了什么?如果否,为什么不? **答复:** 数据中心是长期投资,其规划是一个多年期、资本密集的程序,需要在一系列因素之间取得一致才能成功。这些因素包括能源、水、光纤、土地适宜性、环境考量(含野生动物影响),以及一支可用的、训练有素的劳动力,以确保我们按客户期望的时间表部署数据中心。 从决定建设数据中心的那一刻起,我们就审慎考虑 AI 的资源强度。我们正致力于推进低碳材料并创建全球市场,以帮助推动各行业的可持续性。在运营数据中心时,我们优化能效与水效,包括公布一种制冷零耗水的新数据中心设计。 我们也越来越多地采用循环方式,以实现到 2030 年零废弃的目标。我们的做法是先着力防止产生废弃物,再聚焦于再利用和材料回收。这包括对建筑与拆除废弃物进行再利用和回收、把运营废弃物分流出去,以及推进云端硬件的循环利用与包装循环。 ### 主题:备用发电 **问题 2.** 贵公司是否在任何设施使用柴油备用发电机? **答复:** 是。数据中心的发电机(多数以柴油为燃料)在提供可靠备用电力方面起关键作用,使我们能够满足众多依赖我们服务的客户的需求,包括医院、急救人员和教育机构。在我们各数据中心站点,每台发电机的年使用时间都不超过数小时,多数情况下用于例行维护或电网停电期间的备用供电。 **问题 2a.** 如果是,请提供使用柴油发电机的每一处设施清单,并说明位置、数量和发电机类型。 **答复:** 我们在所有数据中心站点都使用发电机。我们数据中心站点的完整清单可在此查看:https://datacenters.microsoft.com/globe/explore/ 。我们的备用发电机使用非常节制。 **问题 2b.** 如果是,贵公司是否考虑过以电池储能技术替代柴油发电机?请说明你们的决策过程。 **答复:** 在考虑任何备用电源时,我们都寻求可靠、可规模化且具成本效益、能够按与我们数据中心增长相匹配的快速时间表开发和部署、并能在我们数据中心设施的约束条件内运行的方案,包括电池和其他技术(例如替代燃料)。此类备用发电的首要要求是:能够迅速提升出力以应对紧急事件,并且在站点进出受限的情况下,能够储存足够能量在停电期间为数据中心供电。 ### 主题:能源结构 **问题 3.** 贵公司是否利用任何现场或同址发电来为数据中心供电? **答复:** 没有。 **问题 3a.** 如果是,请详细说明有多少电力来自现场和同址发电。 **答复:** 不适用。 **问题 3b.** 如果是,请列出所有用于为数据中心供电的现场和同址能源来源(例如可再生能源、核能、水电、燃气轮机等)。 **答复:** 不适用。 **问题 4.** 贵公司如何确保当地电费缴纳者不必为满足数据中心能源需求所需的新能源基础设施(如输电线路)成本买单? **答复:** 当为服务我们的负荷而需要改进电网时,我们与当地公用事业公司合作,确保我们为服务本站点所需的改进付费。 ### 主题:AI 模型训练中的能耗 **问题 5.** 过去一年,你们估计用于训练新 AI 模型的能耗是多少吉瓦时(GWh)? **答复:** 迄今为止,我们多数基础设施站点同时为客户提供 AI 与传统云服务,以及客户所依赖的关键业务与通信需求。AI 在这些站点共享空间与资源,因此我们尚无法把往年的 AI 能耗单独剥离出来。在规模化部署这项新技术的过程中,我们持续在基础设施的每一个环节提升效率。我们正在对能耗作进一步分析。 **问题 6.** 数据中心制冷需要大量水。当水位处于危急低位时,贵公司是否仍继续取水用于建筑制冷?是否有应急预案,以避免对供水造成进一步压力、避免占用本已有限的家用、农业或小企业用水? **答复:** 就数据中心而言,用水需求取决于所采用的制冷技术。微软正在转向在缺水地区采用制冷零耗水的方案。不过,卫生间和休息室等基本安全与行政功能通常仍需接入供水和下水道。 ### 主题:与政府的伙伴关系 **问题 7.** 贵公司专门为公共部门提供 AI 产品,这些产品如今已在美国联邦机构以及州和地方政府中使用。考虑到政府使用 AI 所带来的风险格外高——包括剥夺权利或服务获取资格,以及关于政府福利与项目的虚假或错误信息——你们采取了哪些额外措施,确保这些工具在政府场景中安全有效? **答复:** 联邦政府在 AI 采用上的引领,对确立标准、让美国保持在创新前沿至关重要。通过把 AI 整合进自身运作并使用现有 AI 应用,政府可以加快公共服务交付、推动广泛采用、改进服务并提振产业信心。为有效做到这一点,我们已采取额外措施,确保在政府中使用的工具安全有效。例如: 微软设有专为公共部门设计的云实例,即其云服务的专门版本,旨在满足政府机构在安全、合规和运营方面的独特需求。例如,Microsoft Azure Government 和 Microsoft 365 Government 就是为处理敏感数据和受监管工作负载而构建,并已通过认证满足 FedRAMP High 标准。尽管底层技术往往与我们商用产品相同,这些实例的配置与治理方式则与公共部门要求保持一致。对受监管客户提供的 AI,在多数情况下与我们向商用客户提供的完全一致。我们的商用 Azure 云同样获得 FedRAMP High 授权,承载着一批今天把其作为主要云服务的州、地方和联邦机构。 **问题 8.** 你们有哪些协议安排,用于与政府机构合作,在出现损害或错误时予以纠正? **答复:** 微软提供一套分层框架,把技术、程序与伦理保障结合起来。我们认识到在公共部门部署 AI 意味着更高的责任,尤其是在可能影响权利、服务和福利获取的重大决策上。我们负责任 AI 的方法建立在问责(含人类监督与控制)、透明度、公平性以及可靠性与安全等原则之上。 例如,我们采用了以下做法: **人在回路监督:** 我们确保涉及资格或服务获取的关键决策包含人工审核与推翻机制。 **事件响应框架:** 我们与各机构密切合作,建立清晰的升级路径与补救规程,用于识别、报告和纠正与 AI 相关的错误或损害。 **偏见与公平性审计:** 我们的模型在部署前接受严格审计,并在运行中持续审计,以侦测和缓解偏见,在敏感用例中尤其如此。 **透明文档:** 我们提供模型卡、数据表和决策可追溯性,以支持可解释性与问责。 微软还提供负责任 AI 仪表板的访问权限,使各机构能够监测不同人口群体之间的公平性、准确性和错误率。我们致力于与政府伙伴持续协作,确保 AI 系统安全、公平并与公共价值一致。我们乐见有机会共同开发适合你们使命需求的治理框架。 ### 主题:商业伙伴关系 **问题 9.** 当你们把 AI 系统提供给其他用户/部署者时,你们就哪些类型的问题达成约定? **答复:** 微软与其客户(即其 AI 系统的部署者)密切合作,确保 AI 技术以负责任、安全且符合法律要求的方式使用。当微软向客户提供 AI 系统时,双方就负责任使用 AI 的关键问题达成约定,微软向客户提供大量信息与指引,并采取主动措施防止 AI 服务被滥用。这些内容已写入微软的使用条款、可接受使用政策以及《微软 AI 服务行为准则》,客户必须遵守。此外,部分被滥用风险较高的服务只能通过「有限访问计划」获得。 **问题 10.** 你们向那些其他当事方/部署者提供什么信息? **答复:** 微软向部署其 AI 系统的客户提供大量信息、工具与指引,使他们能够理解如何负责任且有效地使用该技术。这包括详细文档与透明度说明,解释 AI 技术如何运作、其能力与局限,以及如何取得最佳结果。我们还提供使用准则、最佳实践和负责任 AI 资源,指导部署者安全实施(例如关于人类监督、测试和公平性检查的建议)。 此外,我们直接在 AI 服务中内建内容过滤和滥用检测模型等安全功能,并提供持续支持(包括技术支持渠道和更新),帮助客户正确部署 AI。这有助于确保部署者理解系统,并拥有维护所约定原则的工具。 **问题 11.** 贵公司采取什么步骤来降低下述可能性:例如,下游用户不以可能伤害他人或违反你们使用条款的方式使用你们的服务? **答复:** 我们采取主动措施降低 AI 服务被下游滥用的风险。我们的服务在多个层面(模型、API 服务和应用)内建安全控制,可自动过滤或阻断有害内容。我们执行自身的使用条款,包括对试图绕过安全措施的当事方采取法律行动——例如那些刻意开发专门用于规避安全护栏、以生成攻击性和有害内容的工具的网络犯罪分子。此外,一年多前我们实施了安全政策:如果我们观察到民族国家行为体、网络犯罪分子或其他恶意行为体使用我们的 AI 工具或服务,我们将立即中断并禁用其访问,通知其可能正在使用的任何服务提供商,并与公众和利益相关方分享我们的经验,以改进 AI 生态(参见《在 AI 时代领先于威胁行为体》,微软安全博客)。进一步地,我们通过从新威胁中学习、更新安全系统来不断增强护栏,并与伙伴合作、为用户提供举报滥用行为的渠道。这些措施共同有助于确保:即便在部署之后,我们的 AI 服务也不会被以伤害他人或违反所约定条款的方式使用。 ### 主题:AI 幻觉 近期报道显示,随着技术能力增强,生成式 AI 的幻觉问题正在变得更糟。幻觉在某些场景中可能造成严重损害,例如在评估求职申请时,或者更危险地,在国家安全语境中。 **问题 12.** 你是否同意,我们需要护栏来确保 AI 工具不被以可能伤害他人的方式使用或滥用? **答复:** 同意。我们认为,审慎的、基于风险的护栏对于推进安全、可信和负责任的 AI 是必要的。 **问题 13.** 贵公司正在采取哪些步骤来应对这一问题? **答复:** 数年来,微软形成了一套结构化方法来负责任地发布生成式 AI 应用,其指导框架是「绘制、度量、管理」。在这一过程的每个阶段,我们都嵌入了由真实经验支撑的最佳实践、准则和工具。一个关键重点是应对幻觉风险,也称「无据性」,即 AI 模型生成看似合理但缺乏支撑的内容。作为这一综合方法的一部分,产品团队配备了集中式工具来评估输出无据的可能性,并获得针对其具体应用量身定制的设计模式与缓解策略支持。 这些风险缓解实践如何运作的一个关键例子,是 2023 年发布的 Copilot Studio——它利用生成式 AI,让没有编程或 AI 技能的客户也能构建自己的 copilot。该产品的一项关键风险就是「有据性」,与所有生成式应用一样,Copilot Studio 工程团队在部署前按照我们的治理框架对风险进行了绘制、度量和管理。通过调整元提示来改进有据性缓解措施,Copilot Studio 团队显著提升了域内查询的响应表现,把域内通过率从 88.6% 提高到 95.7%。这意味着当用户提出域内(即主题适切)的问题时,用 Copilot Studio 构建的 copilot 能够更准确地作答。这一改动还使答复率在实施后仅一周内显著提升 6 个百分点。换言之,改进后的有据性过滤也减少了 copilot 拒绝作答的查询数量,改善了整体用户体验。 团队还为输出引入了引注,使 copilot 用户能获得更多关于 AI 生成输出中所含信息来由的信息。通过修改安全系统消息并利用内容过滤,Copilot Studio 团队把引注准确率从 85% 提升到 90%。在「绘制、度量、管理」框架的指引和强健治理流程的支撑下,Copilot Studio 团队推出的体验让客户能够构建更安全、更可信的 copilot。 正如我们在生成式产品的平台层和应用层度量并管理 AI 风险一样,我们也赋能客户做同样的事。我们向客户提供功能,在使用有据文档的应用(如问答类 copilot 和文档摘要应用)中检测生成式 AI 输出里的无据陈述。有据性检测能够找出 AI 生成输出中的无据陈述,并允许客户实施缓解措施,例如触发对无据陈述的重写。 Response to Written Questions Submitted by Hon. Edward Markey to Brad Smith Comprehensive Impacts of Data Center Construction Question 1. When planning for data center construction, does your company conduct a cradle-to-grave infrastructure study that includes wildlife, community, and pollution impacts during and beyond the operational lifespan of a data center? If yes, what have you learned from those studies? If no, why not? Answer. Datacenters are long-term investments, and planning is a multi-year, capital intensive program that requires alignment across a range of factors to be successful. These factors include energy, water, fiber, land suitability, environmental considerations including wildlife impacts, and an available, trained workforce to ensure that we deploy the data center on the timelines our customers expect. We are thoughtful about resource intensity of AI from the moment that we decide to build a datacenter. We're working to advance low- carbon materials and create global markets to help advance sustainability across industries. When operating our datacenters, we optimize energy and water efficiency, including announcing a new datacenter design that consumes zero water for cooling. We also are increasingly adopting a circular approach to reach our target of zero waste by 2030. We target preventing waste first, then focus on reusing, and recovering materials. This includes reusing and recycling construction and demolition waste, diverting operational waste and advancing circular cloud hardware and packaging. Backup Energy Generation Question 2. Does your company use backup diesel generators at any facilities? Answer. Yes. Generators at datacenters, most often powered by diesel fuel, play a key role in delivering reliable backup power so we can meet the needs of the many customers that rely on our services, including hospitals, first responders, and educational institutions. Each of these generators is used for no more than a few hours a year or less at our datacenter sites, most often for routine maintenance or for backup power during a grid outage. Question 2a. If yes, please provide a list of each facility where diesel generators are being used, along with the location, quantity, and type of generators. Answer. We use generators at all of our datacenter sites. A complete list of our datacenter sites can be found here. [https:// datacenters.microsoft.com/globe/explore/] We operate our backup generators sparingly. Question 2b. If yes, did your company consider the use of battery storage technology as an alternative to diesel generators? Please explain your decision process. Answer. As we consider any back-up power sources, we look for reliable, scalable and cost-effective options that can be developed and deployed in rapid timelines aligned to our datacenter growth and operate within the constraints of our datacenter facilities, including batteries and other technologies (e.g., alternative fuels). The primary requirements for such backup generation are that they can quickly ramp to meet the emergency event and can store sufficient energy to supply the datacenter through an outage in case site access is limited. Energy Mix Question 3. Does your company utilize any on-site or colocated energy generation to power your data centers? Answer. No. Question 3a. If yes, please provide detail how much power comes from on-site and colocated energy generation. Answer. N/A Question 3b. If yes, please list all on-site and colocated energy sources (e.g., renewable, nuclear, hydropower, gas-powered turbines, etc.) that are being utilized to power your data centers. Answer. N/A Question 4. How does your company ensure local ratepayers are not responsible for paying the cost of new energy infrastructure, such as transmission lines, needed to meet the data center's energy demand? Answer. When improvements to the grid are required to serve our load, we work with our local utility to ensure that we pay for the improvements required to serve our site. Energy Consumption in AI Model Training Question 5. In the past year, how many GWh of energy do you estimate was used to train new AI models? Answer. To date, most of our infrastructure sites have been serving both AI and traditional cloud services for our customers and the critical business and communications requirements they rely on us for. AI has shared space and resources in those locations, so we are not yet able to specifically separate out AI energy use for prior years. We continue to drive efficiency into every part of this infrastructure as we deploy this new technology at scale. We are working on further analysis of energy use. Question 6. Data centers require vast amounts of water for cooling. When water is at critically low levels, does your company continue to pull water for building cooling? Does it have a contingency for operating as to not put further stress on the water supply and potentially take limited resources from households, agriculture, or small businesses? Answer. For the datacenter, water requirements will vary depending on the cooling technology used. Microsoft is moving to solutions that utilize zero water for cooling in water-stressed areas. However, water and sewer connections are typically required for basic safety and administrative functions, like restrooms and breakrooms. Government Partnerships Question 7. Your company offers AI products specifically for the public sector, which are now used across Federal agencies and state and local governments. Given the especially heightened risks related to governments' use of AI--including the denial of rights or access to services and false or incorrect information about government benefits and programs--what additional steps have you taken to ensure that these tools are safe and effective to use in the context of government? Answer. The Federal government's leadership in AI adoption is vital for setting standards and keeping the U.S. at the forefront of innovation. By integrating AI into its operations and using existing AI applications, it can speed up public service delivery, drive widespread adoption, improve services, and boost industry confidence. To do this effectively, we've taken additional steps to ensure the tools used in government are safe and effective. For example: Microsoft has cloud instances designed specifically for the public sector, meaning specialized versions of its cloud services designed to meet unique security, compliance, and operational needs of government agencies. For example, Microsoft Azure Government and Microsoft 365 Government are built to handle sensitive data and regulated workloads and are certified to meet the standards of FedRAMP High. While the underlying technology is often the same as in our commercial offerings, these instances are configured and governed to align with public sector requirements. Our AI for regulated customers is in most cases exactly the same as what we offer our commercial customers. Our commercial Azure cloud is also FedRAMP High authorized and hosts an array of State, Local, Federal agencies who utilize this as their primary cloud offering today. Question 8. What protocols do you have in place to work with government agencies to rectify any harms or errors when they occur? Answer. Microsoft offers a layered framework that combines technical, procedural, and ethical safeguards. We recognize the heightened responsibility that comes with deploying AI in the public sector, especially where consequential decisions can impact access to rights, services, and benefits. Our approach to Responsible AI is grounded in principles such as accountability (including human oversight and control), transparency, fairness, and reliability & safety. For example, we utilized the following techniques: Human-in-the-Loop Oversight: We ensure that critical decisions involving eligibility or access to services include a human review and override mechanisms. Incident Response Framework: We work closely with agencies to establish clear escalation paths and remediation protocols for identifying, reporting, and correcting AI-related errors or harms. Bias and Fairness Audits: Our models undergo rigorous pre- deployment and ongoing audits to detect and mitigate bias, especially in sensitive use cases. Transparent Documentation: We provide model cards, data sheets, and decision traceability to support explainability and accountability. Microsoft also provides access to our responsible AI dashboard, which enables agencies to monitor fairness, accuracy, and error rates across demographic groups. We are committed to continuous collaboration with government partners to ensure AI systems are safe, equitable, and aligned with public values. We welcome the opportunity to co-develop governance frameworks tailored to your mission needs. Business Partnerships Question 9. When you make your AI systems available to other users/ deployers, what are the types of issues you agree on? Answer. Microsoft works closely with its customers, the deployers of its AI systems, to ensure that AI technologies are used responsibly, safely, and in line with legal requirements. When Microsoft makes an AI system available to a customer, both parties agree on key issues of responsible AI use, Microsoft provides extensive information and guidance to the customer, and Microsoft takes proactive steps to prevent misuse of the AI services. These are codified in Microsoft's Terms of Use, Acceptable Use Policy, and the Microsoft AI Services Code of Conduct, to which customers must adhere. In addition, some services, which have a higher risk of misuse, are available only through a Limited Access Program. Question 10. What information do you provide to those other parties/deployers? Answer. Microsoft provides extensive information, tools, and guidance to customers deploying its AI systems, so they can understand how to use the technology responsibly and effectively. This includes detailed documentation and transparency notes that explain how AI technology works, its capabilities and limitations, and how to achieve the best results. We also provide usage guidelines, best practices, and responsible AI resources to guide deployers in safe implementation (e.g., recommendations for human oversight, testing, and fairness checks). Additionally, we build in safety features like content filtering and abuse detection models directly into our AI services and offer ongoing support, including technical support channels and updates, to help customers deploy AI correctly. This helps ensure deployers understand the system and have the tools to uphold the agreed principles. Question 11. What step does your company take any steps to reduce the likelihood that, for instance, those downstream users don't use your services in ways that could harm people or that violate your terms of use? Answer. We take active steps to reduce the risk of downstream misuse of our AI services. Our services have built-in safety controls at multiple levels (the model, the API service, and the application) to automatically filter or block harmful content. We enforce our terms of use, including by pursuing legal action against actors who try to bypass safety measures, such as cybercriminals who intentionally develop tools specifically designed to circumvent the safety guardrails to create offensive and harmful content. Additionally, over a year ago, we implemented security policies to include, if we observe a nation- state actor, cybercriminal or other malicious actor using our AI tools or services, we will disrupt and disable them immediately, notify any service providers they may be using and share our learnings with the public and stakeholders to improve the AI ecosystem. Staying ahead of threat actors in the age of AI | Microsoft Security Blog Furthermore, we continually strengthen our guardrails by learning from new threats and updating our safety systems, and we work with partners and provide channels for users to report abuses. These measures collectively help ensure that even after deployment, our AI services are used in ways that do not harm people or violate the agreed-upon terms. AI Hallucinations Recent reporting suggests that generative AI hallucinations are getting worse as the technologies become more powerful. Hallucinations can lead to great harm in certain scenarios, such as when assessing job applications, or even more dangerously in the national security context. Question 12. Do you agree that we need guardrails to ensure that AI tools are not used or misused in ways that could cause harm to people? Answer. Yes. We agree that thoughtful, risk-based guardrails are necessary for advancing safe, trustworthy, and responsible AI. Question 13. What steps is your company taking to address this issue? Answer. Over several years, Microsoft has developed a structured approach to responsibly releasing generative AI applications, guided by a ``map, measure, and manage'' framework. At each stage of this process, we've embedded best practices, guidelines, and tools informed by real-world experience. A key focus has been addressing the risk of hallucinations, also known as ungroundedness, where AI models generate plausible but unsupported content. As part of this comprehensive approach, product teams are equipped with centralized tools to evaluate the likelihood of ungrounded outputs and are supported with design patterns and mitigation strategies tailored to their specific applications. A key example of how these risk mitigation practices work is the 2023 release of Copilot Studio, which harnesses generative AI to enable customers without programming or AI skills to build their own copilots. One of the key risks for this product is groundedness, and, as with all generative applications, the Copilot Studio engineering team mapped, measured, and managed risks according to our governance framework prior to deployment. By improving groundedness mitigations through metaprompt adjustments, the Copilot Studio team significantly enhanced in-domain query responses, increasing the in-domain pass rate from 88.6 percent to 95.7 percent. This means that when a user submits a question that is in-domain--or topically appropriate--copilots built with Copilot Studio are able to respond more accurately. This change also resulted in a notable 6 percent increase in answer rate within just one week of implementation. In other words, the improved groundedness filtering also reduced the number of queries that copilots declined to respond to, improving the overall user experience. The team also introduced citations for outputs, so copilot users have more information about the source of information included in AI- generated outputs. By amending the safety system message and utilizing content filters, the Copilot Studio team improved citation accuracy from 85 percent to 90 percent. Following the map, measure, and manage framework and supported by robust governance processes, the Copilot Studio team launched an experience where customers can build safer and more trustworthy copilots. Just as we measure and manage AI risks across the platform and application layers of our generative products, we empower our customers to do the same. We offer features to our customers that detect ungrounded statements within generative AI outputs in applications using grounded documents, such as Q&A Copilots and document summarization applications. Groundedness detection finds ungrounded statements in AI-generated outputs and allows the customer to implement mitigations such as triggering rewrites of ungrounded statements. ______ ¶
Mr. Smith ## 26. 尊敬的塔米·鲍德温(Tammy Baldwin)提交的书面问题 → 布拉德·史密斯 **问题 1.** 我担任拨款委员会 LHHS 小组委员会的少数党首席成员,该小组委员会负责确保我们投资于下一代美国劳动者的教育。微软对芒特普莱森特(Mount Pleasant)数据中心的投资中,有一部分是与盖特威技术学院(Gateway Technical College)合作建设一所数据中心学院。能否谈谈微软为何决定投资于本地 STEM 教育? **答复:** 微软致力于本地劳动力发展。该计划旨在五年内培训超过 1,000 名学生,使他们具备从事 IT 与数据中心运营职业所需的技能。通过把基础设施投资与教育对齐,微软力求确保当地社区直接从所创造的经济机会中受益。这也有助于公司更宏大的目标——在有望迎来科技增长的地区,建设一支多元化、面向未来的劳动力队伍。 **问题 2.** 2023 年底,微软与美国劳工联合会—产业工会联合会(AFL-CIO)建立了此类合作中的首例伙伴关系,以确保劳动者在人工智能的开发和实施中拥有席位。我认为,技术以增强而非消灭岗位的方式进步至关重要。能否谈谈微软为何决定推进这一伙伴关系,以及你们从中收获了什么? **答复:** 微软与 AFL-CIO 合作,是为了确保劳动者在 AI 如何开发与部署上拥有发言权。这一伙伴关系聚焦于扩大 AI 教育的可及性、把劳工反馈纳入产品设计,以及塑造支持包容性经济机会的公共政策。它体现了微软对负责任 AI、以及对建设一个惠及所有人的工作未来的更广泛承诺。微软认识到了把劳动者声音直接嵌入 AI 技术部署的价值,并进一步确认了推行包容性 AI 技能战略的必要性。 Response to Written Questions Submitted by Hon. Tammy Baldwin to Brad Smith Question 1. I serve as Ranking Member of the LHHS subcommittee of the Appropriations Committee where we are tasked with ensuring we are investing in the education of the next generation of American workers. Part of Microsoft's investment in the Mount Pleasant data center includes partnering with Gateway Technical College to build a Data Center Academy. Can you share why Microsoft has decided to invest in local STEM education? Answer. Microsoft is committed to local workforce development. This initiative is designed to train over 1,000 students in 5 years, equipping them with the skills needed for careers in IT and data center operations. By aligning infrastructure investment with education, Microsoft aims to ensure that local communities benefit directly from the economic opportunities created. It also supports the company's broader goal of building a diverse, future-ready workforce in regions poised for tech growth. Question 2. At the end of 2023, Microsoft entered into a first of its kind partnership with AFL-CIO to ensure workers are at the table in the development and implementation of artificial intelligence. I believe it is critically important that technology advances in a way that enhances our workforce instead of eliminating jobs. Can you share why Microsoft decided to pursue this partnership and what you have gained from it? Answer. Microsoft partnered with the AFL-CIO to ensure that workers have a voice in how AI is developed and deployed. The partnership focuses on expanding access to AI education, incorporating labor feedback into product design, and shaping public policy that supports inclusive economic opportunity. It reflects Microsoft's broader commitment to responsible AI and to building a future of work that benefits everyone. Microsoft has learned the value of embedding worker voice directly into the deployment of AI technologies and reinforced the need for inclusive AI skilling strategies. ______ ¶
Mr. Smith ## 27. 尊敬的杰基·罗森(Jacky Rosen)提交的书面问题 → 布拉德·史密斯 ### 主题:对抗性 AI 网络安全专家之所以能够识别出 DeepSeek 平台上的安全缺口,原因在于它是一个开源模型。 **问题 1.** 国会如何保护美国人免受外国对手关联实体未来开发的模型之害——这些模型无论开源与否,都可能使用户数据面临风险? **答复:** *见下方答复。 > 译注:原文此处为「\*See the answer below」,指向紧随其后的第 2 问答复,属原文件的交叉引用,非脱漏。 **问题 2.** 用户——无论是联邦用户还是商业用户——是否有办法安全地使用敌对国家企业开发的 AI 模型? **答复:** 对这项技术的需求是明确存在的。随着越来越多的企业和个人寻求创新并开拓新市场,他们会希望考虑范围广泛的、最契合自身需求的 AI 模型类型。获取最新 AI 技术能够促成创新,但应在可信且安全的环境中使用。在处理敏感数据或用于高风险用例时,可采用红队测试、自动化评估和深入安全审查等额外措施,以发现并缓解模型及其所在系统中的潜在风险。按照共担责任的实践,我们鼓励客户仔细审阅模型与系统文档和透明度说明,采用内容安全过滤等平台防护措施,并针对自身具体用例开展自己的安全与保障评估。 ### 主题:模型安全 **问题 3.** 确保我们使用的 AI 模型不会成为又一个网络安全漏洞,这一点至关重要。针对大型 AI 模型或高风险模型及其训练所用基础设施的自愿性网络安全标准,是否有助于建立信任? **答复:** 针对 AI 模型、特别是前沿模型及其训练所用基础设施的自愿性、基于风险的网络安全标准,将是迈向建立信任的有价值一步。 以往几波技术浪潮已清楚表明:要推动新技术的创新与采用,用户必须既信任技术本身的运行表现,也信任该技术能够成功、安全且可靠地被使用。网络安全框架与标准为这种信任提供了基础。过去十年间,由 NIST 和 ISO/IEC 等组织制定的基于风险的网络安全标准,帮助公共和私营部门的利益相关方有效管理网络安全风险,无论其规模或成熟度如何。这些标准也支撑着合规与保证机制,带来经济与安全双重效益。 在白宫科技政策办公室(OSTP)就制定《AI 行动计划》征求意见(RFI)时,微软在公开评论中鼓励联邦政府避免制定重复政策,而应利用现有的基于风险、以结果为导向的网络安全标准,例如 NIST 网络安全框架和安全软件开发框架(SSDF)。这些标准提供了精简监管、促进一致性、推动创新与跨境合作所需的灵活性。 与规定性的网络安全要求不同,基于风险的标准具有灵活性和敏捷性,可以精简监管、推动一致性,并激励创新与跨境商务。从零开始制定新标准是资源密集型过程,联邦政府可以通过使用并发展现有的基于风险的标准节省大量资源,同时仍确保强健的网络安全措施到位。在评论中,我们还鼓励政府投资于 NIST,因为它们在为全球 AI 讨论提供信息、把美国视角带入 ISO/IEC 等国际标准化机构方面扮演关键角色。美国网络安全路径与国际标准化路径之间的对齐与互操作性,对跨境商务的增长和对美国 AI 的信任都很重要。 ### 主题:AI 技能 **问题 4.** 你在证词中提到了数字技能的重要性。如果我们不利用国会授权的联邦项目——例如《数字公平法》下设立的项目,它们明确旨在帮助美国人建立数字技能,比如教老年人、小企业和退伍军人使用 AI——这会如何损害美国与中国竞争的能力? **答复:** 各种年龄和背景的美国人都需要 AI 技能,才能在这个新的工作世界中竞争。对多数人而言,一个关键机会是培养 AI 流畅度,使他们能在工作中使用 AI,就像今天使用笔记本电脑、智能手机、软件应用和互联网一样。中国拥有一些世界上最有才华的计算机科学研究人员,在培养 STEM 毕业生方面也居世界前列。因此,美国——政府、私营部门和非营利组织——投资于 AI 与数字技能培训至关重要。例如,我们与美国社区学院协会合作推出了「社区学院 AI」项目,旨在通过提供有价值的资源与支持来赋能学生和教育者。该项目为教职员工提供适配各种技能水平的 AI 培训,确保对 AI 概念有全面理解;它将提供以 AI 为重点的课程,使学生具备满足地区劳动力需求的热门技能,同时提供 AI 技术以增强社区学院各部门的职能。通过帮助教育者把 AI 技能直接融入课堂,该项目确保学生做好充分准备,作为具备 AI 能力的专业人士进入劳动力市场,回应各行业对 AI 专长日益增长的需求。 ### 主题:AI 与频谱 人工智能有潜力大幅改进我们管理和利用频谱的方式,尤其是像 Wi-Fi 所用的那些非授权频段。 **问题 5.** AI 在增强动态频谱接入、减少干扰、优化拥塞环境中的性能方面,有哪些最有前景的应用方式?国会应当做些什么来支持这些努力? **答复:** 有一些公司正在运用机器学习与 AI 打造自动化工具,通过共享使用来更好地管理各频段的频谱利用。也有一些大学正在研究动态频谱共享所涉的一些更基础的问题。可以推测,其中部分研究正在探索如何应用 AI 工具。不过,微软并未参与这些工作。 **问题 6.** 公民宽带无线电服务(CBRS)模式已表明,联邦与非联邦用户之间的动态共享可以释放出有价值的频谱用于创新。AI 如何通过实现更敏捷、实时的频谱协调,来增强和扩展这类频谱共享框架?联邦政府应采取哪些步骤来加速开发 AI 驱动的频谱管理工具? **答复:** 尽管微软曾参与十多年来多轮 CBRS 程序,但近年来公司未再参与。 ### 主题:宽带与 Wi-Fi 的必要性 **问题 7.** Wi-Fi 是我们家庭、学校、医院和工作场所连接的基础。随着 AI 更深地嵌入各行业的应用——从诊断与患者监测到智能工厂和个性化教育——我们继续投资于全美宽带基础设施和强健的高容量 Wi-Fi 网络,以充分实现 AI 的经济与社会潜力,这有多关键? **答复:** 微软同意,支持 Wi-Fi 的设备在当今通信网络中扮演不可或缺的角色,未来也将继续如此。联邦通信委员会(FCC)决定在全部 5925–7125 兆赫(6 GHz)频段授权非授权低功率室内(LPI)设备和极低功率(VLP)设备,并在 6 GHz 频段的部分区段授权由自动频率协调(AFC)系统管控的更高功率标准功率(SP)设备——这使多个大带宽信道成为可能,为个人和企业用户接入宽带服务提供了大型入口,而这正是高容量、低时延 Wi-Fi 网络的前提条件。FCC 对 LPI、VLP 以及(由 AFC 管控的)SP 设备的授权,已为大规模私营部门投资奠定基础。在 AI 赋能设备方面,微软近期发布的若干 Copilot+ PC 机型就搭载了支持 320 兆赫超宽信道的 Wi-Fi 7 无线模块。我们正与客户一道,实时了解 AI 嵌入式系统与服务如何最好地利用高容量、低时延的 Wi-Fi 连接。 ### 主题:AI 测试与评估人才队伍 确保美国继续在 AI 竞赛中领先世界的一个关键因素,是确保我们开发的 AI 是最好的、从而是最可信的。验证模型输出是建立信任的重要一步。然而目前,美国既没有相应标准,也没有训练有素的劳动力来评估 AI 模型,以确立我们可以信任模型输出。 **问题 8.** 国会应考虑什么来激励和壮大 AI 测试与评估人才队伍? **答复:** 国会应考虑设立 AI 技能提升的行业协作体:由政府资助,把使用相同或相似 AI 技术的公司聚在一起,帮助更高效、更大规模地培训劳动者。这对那些可能没有条件建立 AI 培训项目的领域专家尤其有益。此外,我们建议倡导扩大、精简和推广第 127 条(Section 127),该条允许雇主为在职攻读学位的员工提供免税教育资助。在扩大或精简第 127 条的任何努力中,立法也可以突出或进一步激励雇主把这项福利用于 AI 技能提升。第 127 条的福利往往未被充分利用,通过强调它如何能满足 AI 技能需求,也许有机会进一步推广。 **问题 9.** 国会如何支持对 AI 开展测试与评估时采用更多跨学科方法?例如,我们如何确保用于医疗场景的模型既经过模型技术专家的评估,也经过该模型将要部署于其中的医疗场景专家的评估? **答复:** 微软一贯强调,对 AI 开展跨学科测试——特别是针对具体部署场景的测试——对安全可信的 AI 使用不可或缺。在我们自己的产品开发中,我们在模型层和应用层都开展压力测试,必要时开展红队测试。对模型做红队测试有助于识别模型可能如何被滥用、界定其能力范围并理解其局限。这些洞见不仅指导平台层评估与缓解措施的制定(供模型在应用中使用),也可用于为模型的未来版本提供参考。应用层的 AI 红队测试采取系统视角,基础模型只是其中一部分。通过纳入应用特有的缓解措施和安全系统,这有助于识别超出模型本身的失效。在 AI 产品开发全程中适时开展红队测试,可以揭示此前未知的风险、确认潜在风险是否在应用中实际发生,并为度量与风险管理提供依据。这一实践还有助于厘清 AI 应用能力与局限的范围、识别滥用可能性,并揭示需要进一步调查的方面。 国会可以通过支持并资助 NIST 来协助这些努力。NIST 目前正在考虑为相关工作流划定范围,以制定 AI 测试、评估、验证与确认的方法和度量。 Response to Written Questions Submitted by Hon. Jacky Rosen to Brad Smith Adversarial AI The reason cybersecurity experts were able to identify gaps in security in DeepSeek's platform was because it is an open-source model. Question 1. How can Congress protect Americans from future models developed by entities affiliated with foreign adversaries that may put users' data at risk, whether they're open source or not? Answer. *See the answer below Question 2. Are there ways users--whether Federal or commercial-- could safely use AI models developed by companies in adversarial nations? Answer. There is a clear demand for this technology. As more and more businesses and individuals look to innovate and explore new markets, they'll want to consider a wide range of AI model types that best fit their needs. Access to the latest AI technologies enables innovation, but they should be used in a trusted and safe environment. When handling sensitive data or used in high-risk use cases, additional measures, such as red teaming, automated assessments, and in-depth security reviews, can be used to discover and mitigate potential risks in models and the systems they're used in. Consistent with shared responsibility practices, we encourage customers to carefully review model and system documentation and transparency notes and adopt platform safeguards, such as content safety filters, and to conduct their own security and safety evaluations tailored for their specific use case. Model Security Question 3. It's essential we ensure the AI models we use do not become another cybersecurity vulnerability. Would voluntary cybersecurity standards for large AI models or high-risk models and the infrastructure they were trained on be helpful in establishing trust? Answer. Voluntary risk-based cybersecurity standards for AI models, particularly cutting-edge models and the infrastructure they're trained on--would be a valuable step toward building trust. Previous waves of technology have highlighted that to drive innovation and adoption of new technologies, users must both trust in how technology itself performs and trust that the technology can be used successfully, safely and securely. Cybersecurity frameworks and standards provide a foundation for that trust. For the past decade, risk-based cybersecurity standards developed by organizations such as NIST and ISO/IEC have helped public and private sector stakeholders manage cybersecurity risks effectively, regardless of their size or maturity. These standards also support compliance and assurance schemes that deliver both economic and security benefits. In our public comments to the White House Office of Science and Technology Policy (OSTP) in its Request for Information (RFI) for the Development of an AI Action Plan, Microsoft encouraged the Federal government to avoid developing duplicative policies by leveraging existing risk-based and outcome-focused cybersecurity standards, such as the NIST Cybersecurity Framework and the Secure Software Development Framework (SSDF). These standards offer the flexibility needed to streamline regulation, promote consistency, and foster innovation and cross-border collaboration. Unlike prescriptive cybersecurity requirements, risk-based standards offer flexibility and agility to streamline regulations, drive consistency, and incentivize innovation and cross-border commerce. Developing new standards from scratch is a resource-intensive process, and the Federal government can save substantial resources using and building existing risk-based standards while still ensuring robust cybersecurity measures are in place. In our comments, we also encourage the administration to invest in NIST, as they play a critical role in informing the global AI conversation, bringing U.S. perspectives to international standardization bodies like ISO/IEC. Alignment and interoperability between U.S. cybersecurity and international standardization approaches is important for the growth of cross-border commerce and trust in American AI. AI Skills Question 4. You mentioned the importance of digital skills in your testimony. Can you discuss how it might hurt the U.S.'s ability to compete with China if we don't leverage congressionally-mandated Federal programs like those created under the Digital Equity Act, which were explicitly designed to help Americans build digital skills, like teaching seniors, small businesses, and veterans how to use AI? Answer. Americans of all ages and backgrounds will need AI skills to compete in this new world of work. A key opportunity for most people will be to develop an AI fluency that will enable them to use AI in their jobs, much like they use laptops, smartphones, software applications, and the Internet today. China is home to some of the world's most talented computer science researchers. They also lead the world in graduating STEM students. Therefore, it is critical the United States, government, private sector, and non-profits, to invest in AI and digital skilling. For example, we launched the AI for Community Colleges program, in collaboration with the American Association of Community Colleges, designed to empower both students and educators by providing valuable resources and support. The program offers AI training for faculty and staff tailored to all skill levels, ensuring a comprehensive understanding of AI concepts. It will deliver AI-focused curriculum that equips students with in-demand skills to meet regional workforce needs, as well as AI technology to enhance various departmental functions at community colleges. By helping educators integrate AI skilling directly into their classrooms, the program ensures students are well-prepared to enter the workforce as AI-ready professionals, addressing the growing demand for AI expertise across industries. AI For Spectrum Artificial Intelligence has the potential to dramatically improve how we manage and utilize spectrum, particularly unlicensed bands like those used for Wi-Fi. Question 5. What are some of the most promising ways AI can be applied to enhance dynamic spectrum access, reduce interference, and optimize performance in congested environments--and what should Congress do to support those efforts? Answer. There are companies that are applying machine learning and AI to create automated tools to better manage spectrum utilization in various frequency bands through shared use. There are also universities examining some of the more fundamental questions associated with dynamic spectrum sharing. Presumably, some of this research is exploring how AI tools can be applied. However, Microsoft is not involved in these efforts. Question 6. The Citizens Broadband Radio Service (CBRS) model has demonstrated how dynamic sharing between Federal and non-federal users can unlock valuable spectrum for innovation. How can AI enhance and expand these types of spectrum sharing frameworks by enabling more agile, real-time spectrum coordination, and what steps should the Federal government take to accelerate the development of AI-powered spectrum management tools? Answer. Although Microsoft participated in multiple CBRS proceedings that go back over a decade, in recent years the company has not been engaged. Need for Broadband and Wi-Fi Question 7. Wi-Fi is the foundation of connectivity in our homes, schools, hospitals, and workplaces. As AI becomes more embedded in applications across sectors--from diagnostics and patient monitoring to smart factories and personalized education--how critical is it that we continue to invest in broadband infrastructure across the U.S. and robust, high-capacity Wi-Fi networks to realize AI's full economic and social potential? Answer. Microsoft agrees that Wi-Fi-enabled devices play an essential role in today's communications networks and will continue to do so in the future. The Federal Communication Commission's (FCC) decisions to authorize unlicensed Low Power Indoor (LPI) and Very Low Power (VLP) devices across the entire 5925-7125 Megahertz (6 GHz) band and higher-power Standard Power (SP) devices under control of an Automated Frequency Coordination (AFC) system over portions of the 6 GHz band, allow for multiple, large bandwidth channels that serve as large on-ramps for individuals and enterprise users to access broadband services, and which is a prerequisite for high-capacity low-latency Wi- Fi networks. FCC authorization of LPI, VLP, and SP (under control of an AFC) devices has set the stage for significant private sector investments. With respect to AI enabled devices, several of Microsoft's more recently released Copilot + PC computer models incorporate Wi-Fi 7 radios that feature 320 Megahertz ultra-wide channels. Along with our customers, we are learning in real-time how AI embedded systems and services can best leverage high-capacity low latency Wi-Fi connections. AI T&E Workforce A key factor in ensuring the U.S. continues to lead the world in the AI race is by ensuring the AI we develop is the best and therefore the most trustworthy. Validating model outputs is an important step in establishing trust. Right now, however, the U.S. has neither the standards nor the trained workforce to evaluate AI models to establish that we can trust model outputs. Question 8. What should Congress consider to incentivize and grow the AI test and evaluation workforce? Answer. Congress should consider creating AI upskilling sector- based collaboratives, where governments fund efforts that bring together companies that use the same or similar AI technologies to help train workers more efficiently and at scale, which would be especially beneficial to subject matter experts that may not otherwise have the infrastructure to establish AI training programs. Additionally, we recommend advocating for expanding, streamlining and promoting Section 127, which allows employers to provide tax-free educational assistance to employees pursuing education while working. As part of any effort to expand or streamline Section 127, legislation could also highlight or further incentivize employers using this benefit for AI upskilling. There may be an opportunity to further promote Section 127 benefits, which are often underutilized, by highlighting the ways in which it can address AI skilling needs. Question 9. How can Congress support more interdisciplinary approaches to testing and evaluating AI? For example, how do we ensure a model being used in a healthcare setting has been evaluated both by experts in the model technology, but also experts in the healthcare setting in which it will be deployed? Answer. Microsoft has consistently emphasized that cross- disciplinary testing of AI--specifically through testing specific to a deployment setting--is essential to secure and trustworthy AI use. During our own product development, we conduct stress testing, or red teaming if necessary, at both the model and the application layer. Red teaming the model helps to identify how a model can be misused, scope its capabilities, and understand its limitations. These insights not only guide the development of platform-level evaluations and mitigations for use of the model in applications but can also be used to inform future versions of the model. Application-level AI red teaming takes a system view, of which the base model is one part. This helps to identify failures beyond just the model, by including the application specific mitigations and safety system. Red teaming throughout AI product development, when appropriate, can surface previously unknown risks, confirm whether potential risks materialize in an application, and inform measurement and risk management. The practice also helps clarify the scope of an AI application's capabilities and limitations, identify potential for misuse, and surface areas to investigate further. Congress can assist these efforts by supporting and funding NIST, which is currently considering scoping for workstreams to develop methods and metrics for AI testing, evaluation, verification, and validation. ______ ¶
Mr. Smith ## 28. 尊敬的约翰·费特曼(John Fetterman)提交的书面问题 → 布拉德·史密斯 史密斯先生,我是可再生能源的坚定支持者,这其中也包括核能。每当谈到能源转型——正如我们在本次听证会上讨论的——我的重点一直很明确:确保宾夕法尼亚州的电费缴纳者不受损害。《华盛顿邮报》报道称,数据中心带来的电力需求增长正把居民电费推高 20%。¹³ 这是不可接受的。这些数据中心甚至没有提供长期稳定的岗位:新增岗位集中在建设阶段,而更高的电价却会永久持续。我一直在关注重启三里岛(Three Mile Island)以为微软数据中心能源需求供电的计划。我欣赏创新方案,但宾夕法尼亚州人民优先。 **脚注:** 13. https://www.washingtonpost.com/business/2024/11/01/ai-data-centers-electricity-bills-google-amazon/ **问题.** 你是否会向我承诺——如你在听证会上所做的那样——你与星座能源(Constellation)的购电协议不会提高宾州家庭的电价? **答复:** 微软与星座能源签订的购电协议将为 PJM 地区增加 835 兆瓦电力。通过签订这份购电协议,微软为鹤清洁能源中心核电机组所发的这部分电力担保了一个客户,并完全承担该设施所发电力与容量的成本。 Response to Written Question Submitted by Hon. John Fetterman to Brad Smith Mr. Smith, I'm a big supporter of renewable energy, and that includes nuclear. Whenever we talk about the energy transition, as we discussed at this hearing, my focus has been clear: making sure ratepayers in Pennsylvania aren't hurt. The Washington Post has reported that increasing electricity demand from data centers is jacking up residential power bills by 20 percent.\13\ That's unacceptable. These data centers don't even offer long lasting, stable jobs: the new jobs are in the construction phase, but the higher utility rates last forever. I've been tracking the plan to reopen Three Mile Island to power Microsoft's data center energy needs. I appreciate innovative plans, but Pennsylvanians come first. --------------------------------------------------------------------------- \13\ https://www.washingtonpost.com/business/2024/11/01/ai-data- centers-electricity-bills-google-amazon/ Question. Will you commit to me--as you did at the hearing--that your power purchase agreement with Constellation will not raise electricity rates for PA households? Answer. The power purchase agreement that Microsoft entered into with Constellation Energy will add 835 MW of electricity to the PJM region. By entering into this power purchase agreement, Microsoft is guaranteeing a customer for this power produced by the nuclear unit at the Crane Clean Energy Center and fully responsible for the costs of the energy and capacity from this facility. ______ ¶
Mr. Smith ## 29. 尊敬的莉萨·布朗特·罗切斯特(Lisa Blunt Rochester)提交的书面问题 → 布拉德·史密斯 ### 主题:网络安全与 AI 史密斯先生,我知道像贵公司这样的许多科技企业都把 AI 智能体视为 AI 进步的重要部分。 但 AI 智能体也可能包含关于其用户的敏感数据,比如一个订票智能体会包含支付信息和位置数据。这些数据必须在网络攻击面前保持安全。 此外,AI 智能体也有可能被网络犯罪分子用来比人类更快、规模大得多地策划攻击。 **问题 1.** 我知道微软对全球网络安全威胁有相当充分的了解。你如何评估 AI 网络安全的现状?本委员会应考虑采取哪些具体步骤来加强 AI 智能体的网络防护? **答复:** 本月早些时候,在微软年度开发者大会 Build 上,我们阐述了关于这样一个世界的愿景:AI 智能体跨用户、团队或组织作出决策并执行任务。要实现这一愿景,AI 智能体必须既有能力、又安全可靠。当今的威胁态势——从网络犯罪分子发动的前所未有规模的勒索软件攻击,到国家支持的行为体持续进行的网络间谍活动与攻击——已经要求我们作出强有力、协调一致的应对。AI 智能体,尤其是那些与敏感数据交互的智能体,带来了新的风险,必须通过可信的多方利益相关方伙伴关系主动应对。 AI 智能体可以自主运行、实时决策、与外部工具和数据交互,甚至与其他智能体协作。这有可能变革各行业,从优化电网到协调自动驾驶车队。已有超过 23 万家组织——包括《财富》500 强中的 90%——使用 Copilot Studio 构建了 AI 智能体和自动化流程。 这些强大的能力也带来了新的风险。AI 智能体可能通过提示或数据源被操纵,去执行有害操作——例如对员工做画像、制作有针对性的钓鱼邮件,或泄露敏感信息。在多智能体系统中,单个被攻陷的智能体就可能引发级联失效。例如,一台被入侵的仓储机器人可能扰乱整条供应链。攻击者可以在某份用于训练面向公众的智能体、或该智能体所访问的文档中藏入秘密指令。当某个安全的 AI 智能体随后与那个面向公众的智能体交互时,嵌入的指令就能诱使它绕过自身防护、泄露敏感数据。由于这些系统往往在没有直接人类监督的情况下运行,这些风险会被放大,也就提出了关于问责的重要问题。 我们乐见有机会就此议题与您进一步沟通,并希望看到委员会支持以下几点: - 强调强健的网络安全基础,例如面向所有 AI 系统(包括智能体)的「设计即安全、默认即安全」实践和零信任架构。 - 支持自愿性、基于风险、可适配、技术上有依据且与国际对齐的 AI 网络安全标准。 - 消除云采用的障碍,认识到云服务为 AI 智能体部署带来的安全、可扩展性和效率优势。 - 鼓励 Agent2Agent(A2A)和模型上下文协议(MCP)等开放、安全的协议,以实现智能体之间、智能体与工具之间安全且可互操作的协作。 - 投资于公私协作,以共享威胁情报、模拟对抗场景并演进最佳实践。 ### 主题:AI 与竞争 史密斯先生,AI 正日益融入我们的关键经济与社会基础设施。麦肯锡指出,从企业用例看,AI 的长期机会可能带来约 4.4 万亿美元的额外生产率增长潜力。 但供应商锁定可能是个真实问题:当某家 AI 供应商大幅落后于竞争,其客户就被困在远为逊色的产品上,这可能威胁该 AI 产品所服务的关键行业。 **问题 2.** 对于在金融、医疗等关键行业运营的 AI 产品,你是否有任何计划或策略来缓解锁定问题,以防止可能损害这些关键行业及其从业者的潜在锁定效应? **答复:** 今天的 AI 生态是开放的、不断演进的,并且日益去中心化。微软正努力确保它保持这种状态。 我们看到各行业创新涌动,新模型、新工具和新平台不断涌现。在金融服务和公共卫生领域尤其如此——AI 的潜力仍处在被发现和被应用的阶段。从欺诈检测到疾病监测,应用正在迅速扩展,客户可做的选择也在迅速增多。 在微软,我们致力于支持这种多样性。Azure 承载范围广泛的 AI 模型,包括来自 OpenAI、Meta、Mistral 以及开源社区的模型;我们最近把 xAI 的 Grok 3 和 Grok 3 Mini 加入了 Azure AI Foundry。客户也可以自带模型、对其进行微调,或使用预训练模型。他们可以在 Azure 上、本地部署环境或多云环境中运行模型。我们支持这种灵活性,因为我们认为它对信任与创新不可或缺。 我们也认识到,许多客户在受监管环境中运营。正因如此,我们投资于互操作性、可移植性和合规性——使他们能够以最小的摩擦与约束,按需迁移工作负载。 Response to Written Questions Submitted by Hon. Lisa Blunt Rochester to Brad Smith Cybersecurity and AI Mr. Smith, I know that many tech companies like yours see AI agents as a big part of AI advancement. But AI agents also can contain sensitive data about its users, like a flight agent containing payment information and location data. This data needs to be secure from cyberattacks. Also, there is potential for AI agents to be used by cybercriminals to orchestrate attacks more quickly and at a far larger scale than humans could. Question 1. I know that Microsoft has a good amount of visibility into global cybersecurity threats. How would you assess the current state of AI cybersecurity, and what concrete steps should this committee consider to strengthen the cyber-protection of AI agents? Answer. Earlier this month at Microsoft Build, our annual developer conference, we outlined our vision for a world where AI agents make decisions and perform tasks across users, teams, or organizations. To realize our vision, AI agents must be both capable and secure. Today's threat landscape--from the unprecedented volume of ransomware attacks by cybercriminals to the sustained cyberespionage and attacks from state-sponsored adversaries--already demands a strong, coordinated response. AI agents, especially those interacting with sensitive data, introduce new risks that must be addressed proactively and through trusted multistakeholder partnerships. AI agents can operate autonomously, make real-time decisions, interact with external tools and data, and even collaborate with other agents. This has the potential to transform industries, from optimizing energy grids to coordinating fleets of autonomous vehicles. More than 230,000 organizations--including 90 percent of the Fortune 500--have already used Copilot Studio to build AI agents and automations. These powerful capabilities also introduce new risks. AI agents can be manipulated through prompts or data sources to perform harmful actions--like profiling employees, crafting targeted phishing e-mails, or leaking sensitive information. In multi-agent systems, a single compromised agent can trigger cascading failures. For example, a hacked warehouse robot could disrupt an entire supply chain. An attacker could hide secret instructions in a document used for training of or accessed by a public-facing agent. When a secure AI agent later interacts with the public-facing agent, the embedded instructions can trick it into bypassing its safeguards and leak sensitive data. These risks are amplified because these systems often operate without direct human oversight, raising important questions about accountability. We would welcome the opportunity to engage with you further on this topic, and hope to see the Committee's support for the following: Emphasize robust cybersecurity foundations, such as secure- by-design-and default practices and Zero Trust Architecture for all AI systems--including agents. Support voluntary, risk-based AI cybersecurity standards that are adaptable, technically grounded, and internationally aligned. Remove barriers to cloud adoption, recognizing the security, scalability and efficiency benefits cloud services provide for AI agent deployment. Encourage open, secure protocols like Agent2Agent (A2A) and Model Context Protocol (MCP) to enable safe and interoperable agent-to-agent and agent-to-tool collaboration. Invest in public-private collaboration to share threat intelligence, simulate adversarial scenarios, and evolve best practices. AI and Competition Mr. Smith, AI is becoming integrated into our critical economic and societal infrastructure, with McKinsey stating that long-term AI opportunity could be about $4.4 trillion in added productivity growth potential from corporate use cases. But vendor lock in could be a real issue, where an AI vendor dramatically falls behind the competition and leaves its client with a vastly inferior product, which could threaten key industries the AI product operates in. Question 2. Do you have any plans or strategies regarding mitigating lock-in for your AI products operating in critical sectors, like the financial and medical sectors, to prevent potential lock-in effects that might harm these critical sectors and the folks therein? Answer. Today's AI ecosystem is open, evolving, and increasingly decentralized. Microsoft is working to ensure it stays that way. We are seeing a surge of innovation across sectors, with new models, tools, and platforms emerging regularly. This is especially true in financial services and public health, where AI's potential is still be discovered and applied. From fraud detection to disease surveillance, the applications are expanding rapidly. And so are the choices available to customers. At Microsoft, we are committed to supporting that diversity. Azure hosts a wide range of AI models, including those from OpenAI, Meta, Mistral, and open-source communities; we recently added xAI's Grok 3 and Grok 3 Mini into the Azure AI Foundry. Customers can also bring their own models, fine-tune them, or use pre-trained ones. They can run models in Azure, on-premises, or across multiple clouds. We support that flexibility because we believe it's essential to trust and innovation. We also recognize that many of our customers operate in regulated environments. That's why we invest in interoperability, portability, and compliance--so they can move workloads as needed, with minimal friction and constraint. [all] ¶