altman STATEMENT OF SAMUEL ALTMAN, CHIEF EXECUTIVE OFFICER, OPENAI, SAN FRANCISCO, CALIFORNIA(塞缪尔·奥特曼,OpenAI 首席执行官,加利福尼亚州旧金山) 谢谢。谢谢您,布卢门撒尔主席(Chairman Blumenthal),【译注:碎片话轮,原文以逗号结尾,句意未完;按源件话轮边界保留,不补写】 Thank you. Thank you, Chairman Blumenthal, ¶
altman 是的。我认为这是个很好的主意。我认为公司应当自行公布——你知道,「这是我们发布模型前的测试结果,这是它的弱点所在,这是它的强项所在。」但对此的独立审计也非常重要。这些模型正随时间变得更准确。你知道,我想我们比任何人都更响亮地说过:这项技术仍处于早期阶段。它确实仍会犯错。我们发现用户其实相当老练,明白错误在哪里、可能在哪里,明白自己需要负责核实模型的说法,会去查证。我担心的是,随着模型越来越好,用户围绕它的独立判别思考可能越来越少。但我认为,在这类对话中,用户的能力常常被我们低估。我认为许多披露——如果你用过 ChatGPT 就会看到——关于模型不准确之处的披露也很重要。我期待这样一个世界:公司随模型一并公布其行为方式、不准确之处的信息,独立机构或公司也提供这类信息。我认为这是个很好的主意。 Yes. I think that's a great idea. I think that companies should put their own sort of, you know, ``Here are the results of our tests of our model before we release it, here's where it has weaknesses, here's where it has strengths.'' But also, independent audits for that are very important. These models are getting more accurate over time. You know, as we have, I think, said as loudly as anyone, this technology is in its early stages. It definitely still makes mistakes. We find that people, that users are pretty sophisticated and understand where the mistakes are or are likely to be, that they need to be responsible for verifying what the models say, that they go off and check it. I worry that, as the models get better and better, the users can have sort of less and less of their own discriminating thought process around it. But I think users are more capable than we often give them credit for in conversations like this. I think a lot of disclosures--which, if you've used ChatGPT, you'll see--about the inaccuracies of the model are also important. And I'm excited for a world where companies publish, with the models, information about how they behave, where the inaccuracies are, and independent agencies or companies provide that, as well. I think it's a great idea. ¶
altman 与所有技术革命一样,我预计就业会受到重大影响,但这一影响具体是什么样子,很难预测。如果我们回到上一次技术革命的另一端,谈论那一端存在的工作——你知道,你可以回去读读这方面的书。当时人们就是这么说的。这很难。我相信,在这一革命的另一端会有好得多的工作,而今天的工作会变得更好。我认为重要的是——首先,我认为重要的是把 GPT-4 理解并视为一种工具,而不是一个生物(creature),这一点很容易混淆。它是一种人们对其使用方式拥有很大控制权的工具。其次,GPT-4 及类似系统擅长做的是任务(tasks),不是职业(jobs)。所以你已经看到,人们用 GPT-4 帮自己完成任务,从而更高效地做自己的工作。我想,GPT-4 会把一些工作完全自动化掉,也会创造我们相信会好得多的新工作。再说一次,我对技术史的理解是一场漫长的技术革命,而不是一堆不同革命的拼凑。但这一直在持续发生。随着我们的生活质量提高,随着我们创造的机器与工具帮助我们过上更好的生活,我们所做之事的门槛也随之提高,人类的能力、我们花时间追求的东西,会转向更雄心勃勃、更令人满足的事业。所以,就业会受到影响。我们努力对此非常坦诚。我认为这需要产业界与政府之间的伙伴关系,但主要靠政府采取行动,来弄清我们想如何缓解这一影响。但我对未来工作会有多好非常乐观。 Like with all technological revolutions, I expect there to be significant impact on jobs, but exactly what that impact looks like is very difficult to predict. If we went back to the other side of a previous technological revolution, talking about the jobs that exist on the other side--you know, you can go back and read books of this. It's what people said at the time. It's difficult. I believe that there will be far greater jobs on the other side of this and that the jobs of today will get better. I think it's important--first of all, I think it's important to understand and think about GPT-4 as a tool, not a creature, which is easy to get confused. And it's a tool that people have a great deal of control over, in how they use it. And, second, GPT-4 and other systems like it are good at doing tasks, not jobs. And so you see already people that are using GPT-4 to do their job much more efficiently by helping them with tasks. Now, GPT-4 will, I think, entirely automate away some jobs, and it will create new ones that we believe will be much better. Again, my understanding of the history of technology is one long technological revolution, not a bunch of different ones put together. But this has been continually happening. As our quality of life raises and as machines and tools that we create can help us live better lives, the bar raises for what we do, and our human ability and what we spend our time going after goes after more ambitious, more satisfying projects. So, there will be an impact on jobs. We try to be very clear about that. And I think it will require partnership between the industry and government, but mostly action by government, to figure out how we want to mitigate that. But I'm very optimistic about how great the jobs of the future will be. ¶
altman 是的。听着,我们一直努力非常清楚地说明这里风险的量级。我认为就业、工作,以及我们将如何支配自己的时间,真的事关重大。我同意,当我们拥有非常强大的系统时,格局会改变。我想我只是更乐观:我们人类有不可思议的创造力,有了更好的工具就会找到新的事情做,而且这会持续发生。我最深的恐惧是,我们——我们这个领域、这项技术、这个行业——给世界造成重大伤害。我认为这可能以很多不同方式发生。这正是我们创办这家公司的原因。这也是我今天来到这里、以及我们过去来到这里、得以与各位共度一些时光的重要原因。我认为如果这项技术出了问题,可能出很大的问题,我们想对此直言不讳。我们想与政府合作防止这种情况发生。但我们努力非常清醒地看待下行情况是什么样,以及为缓解它我们必须做的工作。 Yes. Look, we have tried to be very clear about the magnitude of the risks here. I think jobs and employment and what we're all going to do with our time really matters. I agree that when we get to very powerful systems, the landscape will change. I think I'm just more optimistic that we are incredibly creative, and we find new things to do with better tools, and that will keep happening. My worst fears are that we cause significant--we, the field, the technology, the industry cause significant harm to the world. I think that could happen a lot of different ways. It's why we started the company. It's a big part of why I'm here today and why we've been here in the past and we've been able to spend some time with you. I think if this technology goes wrong, it can go quite wrong, and we want to be vocal about that. We want to work with the Government to prevent that from happening. But we try to be very clear-eyed about what the downside case is and the work that we have to do to mitigate that. ¶
altman 是的。谢谢霍利参议员(Senator Hawley)的提问。这是我最担忧的领域之一:这些模型更一般的操纵、说服能力,提供某种一对一的、你知道,交互式虚假信息的能力。我认为这像是您所谈问题的更宽泛版本,但鉴于我们明年将面临选举,而这些模型正变得越来越好,我认为这是一个重大关切领域。我认为公司可以自愿采取很多政策,我很乐意谈谈我们在那方面的做法。我确实认为在这个议题上,一些监管会是相当明智的。早些时候有人提到——这是我们真心赞同的:人们需要知道自己在与 AI 对话,需要知道他们看到的内容可能是生成的、也可能不是。我认为把这一点讲清楚是一件大好事。我认为我们还需要规则与指南,规定提供可能具备您所说这类能力的模型的公司,在披露方面应达到什么要求。所以,我对此感到紧张。我认为人们能够相当快地适应。很久以前 Photoshop 刚出现时,你知道,有一段时间人们真的会被 PS 过的图像骗到,然后很快就形成了「图像可能经过 Photoshop 处理」的认知。这也会像那样,但猛烈得多(原文 on steroids)。而其交互性、真正对人类建模与预测的能力,正如您所谈到的,我认为将需要公司做正确的事、监管与公众教育三者结合。 Yes. Thank you, Senator Hawley, for the question. It's one of my areas of greatest concern: the more general ability of these models to manipulate, to persuade, to provide sort of one-on-one, you know, interactive disinformation. I think that's like a broader version of what you're talking about, but given that we're going to face an election next year and these models are getting better, I think this is a significant area of concern. I think there's a lot of policies that companies can voluntarily adopt, and I'm happy to talk about what we do there. I do think some regulation would be quite wise on this topic. Someone mentioned earlier--it's something we really agree with. People need to know if they're talking to an AI, if content that they're looking at might be generated or might not. I think it's a great thing to do, is to make that clear. I think we also will need rules, guidelines about what's expected in terms of disclosure from a company providing a model that could have these sorts of abilities that you talk about. So, I'm nervous about it. I think people are able to adapt quite quickly. When Photoshop came onto the scene a long time ago, you know, for a while people were really quite fooled by Photoshopped images and then pretty quickly developed an understanding that images might be Photoshopped. This will be like that, but on steroids. And the interactivity, the ability to really model, predict humans well, as you talked about, I think is going to require a combination of companies doing the right thing, regulation, and public education. ¶
altman 是的,我们应当对此感到担忧。说清楚一点,OpenAI 并不——你知道,我们没有基于广告的商业模式,所以我们并不试图建立用户的这类画像。我们并不试图让他们用得更多。事实上,我们倒希望他们用得少一点,因为我们的 GPU 不够用。但我认为其他公司已经在——而且未来肯定会——用 AI 模型做出非常好的广告预测,预测用户会喜欢什么。我认为这在很多方面已经在发生。 Yes, we should be concerned about that. To be clear, OpenAI does not--you know, we don't have an ad-based business model, so we're not trying to build up these profiles of our users. We're not trying to get them to use it more. Actually, we'd love it if they'd use it less, because we don't have enough GPUs. But I think other companies are already--and certainly will, in the future, use AI models to create, you know, very good ad predictions of what a user will like. I think that's already happening, in many ways. ¶
altman 谢谢您的提问,参议员。我还不知道这里的正确答案究竟是什么。我很乐意与您合作把它弄清楚。我确实认为,对于一项非常新的技术,我们需要一个新框架。当然,像我们这样的公司对我们投放到世界上的工具负有很大责任,但工具的使用者也有责任,还有那些在其之上构建、介于它们与最终消费者之间的人。我们想如何确立那里的责任框架,是一个极其重要的问题,我们很乐意一起合作。 Thank you for the question, Senator. I don't know yet exactly what the right answer here is. I'd love to collaborate with you to figure it out. I do think, for a very new technology, we need a new framework. Certainly, companies like ours bear a lot of responsibility for the tools that we put out in the world, but tool users do, as well, and also people that will build on top of it, between them and the end consumer. And how we want to come up with a liability framework there is a super important question, and we'd love to work together. ¶
altman 我也不满意。【译注:原文 Me, either. 为省略回应,承接德宾主席(Chair Durbin)「看着在线平台,我不确定我对结果感到满意」,意为「我也不满意」;按语境补出省略成分,未添加新信息】 Me, either. ¶
altman 我能简短插一句吗? Can I weigh in just briefly? ¶
altman 我想呼应支持马库斯先生(Mr. Marcus)所说的。我认为美国应当在这里引领、率先行动,但要有效,我们确实需要某种全球性的东西。正如您提到的,这可能发生在任何地方。有先例可循。我知道呼吁这样的东西听起来很天真,听起来真的很难。但有先例。我们以前通过国际原子能机构(IAEA)做到过。我们也谈过为其他技术这样做。鉴于制造这些模型所需的条件——芯片供应链、数量有限的高端 GPU、美国对这些公司的影响力——我认为存在可行路径,让美国设定一些其他国家需要协作并参与其中的国际标准,尽管这听起来表面上像个不切实际的主意。我认为这对世界会是件大好事。 I want to echo support for what Mr. Marcus said. I think the U.S. should lead here and do things first, but to be effective, we do need something global. As you mentioned, this can happen everywhere. There is precedent. I know it sounds naive to call for something like this, and it sounds really hard. There is precedent. We've done it before with the IAEA. We've talked about doing it for other technologies. Given what it takes to make these models, the chip supply chain, the sort of limited number of competitive GPUs, the power the U.S. has over these companies, I think there are paths to the U.S. setting some international standards that other countries would need to collaborate with and be part of that are actually workable, even though it sounds on its face like an impractical idea. And I think it would be great for the world. ¶
altman 谢谢您,参议员。这是我们非常感兴趣的领域。我想说,首先,我们认为创作者应当有权控制他们的创作成果如何被使用,以及在其将该成果发布到世界之后所发生的事情。其次,我认为我们需要用这项新技术找到新的方式,让创作者能够获胜、成功、拥有充满活力的生活。我乐观地认为,这将呈现它——【译注:碎片话轮,原文以破折号结尾,句意未完;按源件话轮边界保留,不补写】 Thank you, Senator. This is an area of great interest to us. I would say, first of all, we think that creators deserve control over how their creations are used and what happens sort of beyond the point of them releasing it into the world. Second, I think that we need to figure out new ways with this new technology that creators can win, succeed, have a vibrant life. And I'm optimistic that this will present it---- ¶
altman 这正是我接下来要说的。 That's exactly what I was going to say. ¶
altman 我们目前正在与艺术家合作,包括视觉艺术家、音乐家,以弄清楚人们想要什么。不幸的是,有非常多不同的意见,而在某个时候,我们将——【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】 We're working with artists now, visual artists, musicians, to figure out what people want. There's a lot of different opinions, unfortunately, and at some point, we'll have---- ¶
altman 我不熟悉 SoundExchange(声音交换公司)。 I'm not familiar with SoundExchange. ¶
altman 抱歉,请您再说一遍。 I'm sorry. ¶
altman 好的。 Okay. ¶
altman 所以,首先,Jukebox 不是我们提供的一款产品。那是一次研究性发布,但它不是——你知道的——不像 ChatGPT 或 DALL·E。【译注:原文 DALL·E 在采集件中显示为乱码「DALL␁E」,按通行写法还原为 DALL·E,标记待核】 So, first of all, Jukebox is not a product we offer. That was a research release, but it's not--you know, unlike ChatGPT or DALLE. ¶
altman 哦,我明白了。是的。当然。 Oh, I understand. Yes. For sure. ¶
altman 我手头没有 Jukebox 作为研究性发布的具体数字。我可以随后跟您的办公室跟进,但 Jukebox 并不是一个受到很多关注或使用的东西。它被发布出来是为了展示某种事情是可能的。 I don't know the numbers on Jukebox on the top of my head as a research release. I can follow up with your office, but Jukebox is not something that gets much attention or usage. It was put out to show that something's possible. ¶
altman 是的。我们绝对在这件事上有所投入。再次重申我先前的观点,我们认为内容创作者、内容所有者需要从这项技术中受益。确切的经济模式是什么——我们仍在艺术家和内容所有者讨论他们想要什么。我认为有很多方式可以实现这一点。但非常清楚的是,无论法律如何,正确的做法是确保人们从这项新技术中获得重大的上升收益。我们相信它确实会带来这种收益。但内容所有者的肖像——人们完全应当控制其使用方式,并从中受益。 Yes. We are absolutely engaged on that. Again, to reiterate my earlier point, we think that content creators, content owners need to benefit from this technology. Exactly what the economic model is--we're still talking to artists and content owners about what they want. I think there's a lot of ways this can happen. But very clearly, no matter what the law is, the right thing to do is to make sure people get significant upside benefit from this new technology. And we believe that it's really going to deliver that. But the content owners' likenesses--people totally deserve control over how that's used and to benefit from it. ¶
altman 对此有几点想法。第一,我们认为人们应当能够说:「我不希望我的个人数据被用于训练。」我觉得这就像是——【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】 So, a few thoughts about this. Number one, we think that people should be able to say, ``I don't want my personal data trained on.'' I think that's like this---- ¶
altman 是的。我认为强有力的隐私——【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】 Yes. I think strong privacy---- ¶
altman 好的。 Okay. ¶
altman 谢谢您,参议员。我们此前稍微谈过这个。我们对它可能对选举产生的影响相当担忧。我认为这是一个希望整个行业和政府能够迅速合作的领域。有很多方法,我会谈一些我们所做的事情,但在此之前,我认为用社交媒体的框架来套是有诱惑力的,但这不是社交媒体。这是不同的东西。因此,我们需要的应对也是不同的。你知道的,这是一个用户在使用的工具,用来比以前更高效地生成内容。他们可以修改它,可以测试它的准确性。如果他们不喜欢,可以拿到另一个版本,但随后它仍然会通过社交媒体或其他方式传播。比如,ChatGPT 是一种,你知道的,单人体验,你只是在使用它。所以我认为,在我们思考该做什么时,理解这一点很重要。在那里有很多我们能够——而且确实在——做的事情。有些东西是模型拒绝生成的。我们有政策。重要的是,我们也有监测。因此,在规模上,我们能够检测到某人在生成大量这类推文,即便生成一条推文是可以的。 Thank you, Senator. We talked about this a little bit earlier. We are quite concerned about the impact this can have on elections. I think this is an area where hopefully the entire industry and the Government can work together quickly. There's many approaches, and I'll talk about some of the things we do, but before that, I think it's tempting to use the frame of social media, but this is not social media. This is different. And so the response that we need is different. You know, this is a tool that a user is using to help generate content more efficiently than before. They can change it, they can test the accuracy of it. If they don't like it, they can get another version, but it still then spreads through social media or other ways. Like, ChatGPT is a, you know, single-player experience where you're just using this. And so I think, as we think about what to do, that's important to understand. There's a lot that we can--and do--do, there. There's things that the model refuses to generate. We have policies. We also, importantly, have monitoring. So, at scale, we can detect someone generating a lot of those tweets, even if generating one tweet is okay. ¶
altman 是的。我希望我们正在创造的工具能够帮助新闻机构做得更好。我认为拥有一个充满活力的国家媒体至关重要。而且,姑且这么说,互联网的第一轮在这方面并不好。 Yes. Like, it is my hope that tools like what we're creating can help news organizations do better. I think having a vibrant national media is critically important. And, let's call it, round one of the internet has not been great for that. ¶
altman 当然。 For sure. ¶
altman 当然。 For sure. ¶
altman 嗯——【译注:碎片话轮,原文仅「Well----」,按源件边界保留】 Well---- ¶
altman 是的。再说一次,我们的模型——你知道的,当前版本的 GPT-4 于 2021 年结束训练。它不是查找近期新闻的好方式,我认为它也不是一项能很好地做外链的服务,尽管也许通过我们的插件,这是可能的。如果有我们能够做的事情来帮助地方新闻,我们当然愿意。再次强调,我认为这至关重要。 Yes. Again, our model--you know, the current version of GPT-4 ended training in 2021. It's not a good way to find recent news, and I don't think it's a service that can do a great job of linking out, although maybe with our plugins, it's possible. If there are things that we can do to help local news, we would certainly like to. Again, I think it's critically important. ¶
altman 是的。我们主张我们需要共同努力,找到一种全新的方法。我不认为第 230 条(Section 230)甚至是正确的框架。 Yes. We're claiming we need to work together to find a totally new approach. I don't think Section 230 is even the right framework. ¶
altman 这超出了我的法律专业范围。 That is beyond my area of legal expertise. ¶
altman 不是因为那个,没有。 Not for that, no. ¶
altman 是的,OpenAI 被起诉过。 Yes, OpenAI gets sued. ¶
altman 是的,我们以前被起诉过。 Yes, we've gotten sued before. ¶
altman 我是说,它们大多是非常琐碎的事情,就像我认为发生在任何公司身上的那样。 I mean, they've mostly been, like, pretty frivolous things, like I think happens to any company. ¶
altman 是的。我认为公司需要承担明确的责任。 Yes. I think there needs to be clear responsibility by the companies. ¶
altman 不,我不认为我们在说任何类似那样的话。 No, I don't think we're saying anything like that. ¶
altman 是的。我们一直在呼吁这个。我们认为任何——【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】 Yes. We've been calling for this. We think any---- ¶
altman 是的。我们会对此非常热情。 Yes. We'd be enthusiastic about that. ¶
altman 参议员,我们已经清楚说明了我们认为的好处是什么,我想你能从用户身上看到他们有多享受、从中获得了多少价值。但我们也清楚说明了弊端是什么。 Senator, we've been clear about what we think the upsides are, and I think you can see from users how much they enjoy and how much value they're getting out of it. But we've also been clear about what the downsides are. ¶
altman 所以这就是为什么我们认为我们需要一个机构。 And so that's why we think we need an agency. ¶
altman 这是一项重大的新技术。 It's a major new technology. ¶
altman 我们认为它会是——【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】 We think it'll be---- ¶
altman 这就是为什么我们同意你。 That's why we're agreeing with you. ¶
altman 是的。 Yes. ¶
altman 那应该是——【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】 That should be---- ¶
altman 显然,那应该是某个机构权力的一部分。 Clearly, that should be part of what an agency can do. ¶
altman 正确。 Correct. ¶
altman 嗯,那就是为什么我认为它不一定非得是一个世界性组织,但必须有某种——这里有很多选项。必须有某种标准,某种控制体系——【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】 Well, that's why I think it doesn't necessarily have to be a world organization, but there has to be some sort of--and there's a lot of options here. There has to be some sort of standard, some sort of set of controls---- ¶
altman 那些确实具有全球效力的。 That do have global effect. ¶
altman 我有 1 分钟? I've got 1 minute? ¶
altman 好吧。这对 1 分钟来说是个难题。这远远超出了我的专业领域,但我——【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】 All right. That's a tough question for 1 minute. This is very far out of my area of expertise, but I---- ¶
altman 我认为我们不应该允许那样。 I think we shouldn't allow that. ¶
altman 当然可以。 Sure. ¶
altman 谢谢您,参议员。这是个很好的问题。我想通过谈谈我们为什么还要部署来切入:也就是说,我们为什么把这些系统投放到世界上。有一个显而易见的答案,就是有好处,人们把它用于各种美好的事情并获得巨大价值,这让我们高兴。但我们这样做的很大一部分原因是,我们相信迭代式部署(iterative deployment),让人们、让我们的制度、让你们所有人有时间去适应这项技术,去理解它,去发现它的局限和益处,去了解我们需要围绕它制定什么法规,去弄清楚确保它安全需要什么——这真的很重要。暗中构建一套超级强大的 AI 系统,然后一次性把它扔到世界上,我认为这不会有好结果。所以,我们策略的很大一部分是,在这些系统还相对弱、还远不完美的时候,设法让人们获得使用它们的经验,与现实接触,并弄清楚我们需要做什么来让它更安全、更好。这是我在新技术和产品史上所见过的、通往一个很好结果的唯一途径。因此,与世界的这种互动非常重要。当然,在我们把某样东西投放出去之前,它需要满足一道安全门槛。同样,我们在 GPT-4 完成训练之后,又花了远超 6 个月的时间,经历所有这些不同的事项,并决定标准应该是什么,才把东西投放出去,试图找出我们知道的危害以及应对方法。让我们感到欣慰的一件事是,甚至我们一些最严厉的批评者在看过 GPT-4 后都说:「哇,OpenAI 在……上取得了巨大进展」——【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】 Thank you, Senator. It's a great question. I'd like to frame it by talking about why we deploy at all: like, why we put these systems out into the world. There's the obvious answer about there's benefits and people are using it for all sorts of wonderful things and getting great value, and that makes us happy. But a big part of why we do it is that we believe that iterative deployment and giving people and our institutions and you all time to come to grips with this technology, to understand it, to find its limitations and benefits, the regulations we need around it, what it takes to make it safe--that's really important. Going off to build a super powerful AI system in secret and then dropping it on the world all at once, I think would not go well. So, a big part of our strategy is, while these systems are still relatively weak and deeply imperfect, to find ways to get people to have experience with them, to have contact with reality, and to figure out what we need to do to make it safer and better. And that is the only way that I've seen in the history of new technology and products of this magnitude to get to a very good outcome. And so that interaction with the world is very important. Now, of course, before we put something out, it needs to meet a bar of safety. And again, we spent well over 6 months with GPT-4, after we finished training it, going through all of these different things and deciding also what the standards were going to be, before we put something out there, trying to find the harms that we knew about it and how to address those. One of the things that's been gratifying to us is even some of our biggest critics have looked at GPT-4 and said, ``Wow, OpenAI made huge progress on''---- ¶
altman 我正要——【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 was just about---- ¶
altman 说到那里。 To get there. ¶
altman 好吧。不好意思。是的。我认为事先赋予模型价值观是极其重要的一组(事情)。你知道,RLHF 是做同一件事的另一种方式。但不管怎样,通过合成数据或人工生成的数据,你在说:「这里是价值观。这是我希望你体现的东西」,或者「这是社会所允许的一切的宽泛边界,然后在这个范围内,作为用户的你来选择,比方说你想要这边这套价值体系还是那边那套价值体系。」我们认为这非常重要。有多种技术路径,但我们需要给政策制定者和整个世界提供工具,让他们能够说:「这就是价值观,去实现它们。」 All right. Sorry about that. Yes. I think giving the models values up front is an extremely important set. You know, RLHF is another way of doing that same thing. But somehow or other, with synthetic data or human-generated data, you're saying, ``Here are the values. Here's what I want you to reflect,'' or ``Here are the wide bounds of everything that society will allow, and then within there, you pick, as the user, you know, if you want value system over here or value system over there.'' We think that's very important. There's multiple technical approaches, but we need to give policymakers and the world as a whole the tools to say, ``Here's the values, and implement them.'' ¶
altman 谢谢您,参议员。第一,我会成立一个新机构,对任何超出一定能力规模的努力发放许可,并可以吊销该许可、并确保其遵守安全标准。第二,我会制定一套安全标准,聚焦于你在第三个假设中所说的危险能力评估。我们过去用过的一个例子,是看一个模型能否自我复制并自我泄露(self-exfiltrate)到外部世界。我们可以给你办公室提供一份很长的其他清单,列出我们认为在那里重要的事情,但具体来说,是一个模型在部署到世界之前必须通过的具体测试。然后,第三,我会要求独立审计,不只是来自公司或该机构,而是由专家来判断该模型是否遵守这些既定的安全阈值、以及在这些方面的问题 X 或 Y 上的表现百分比。 Thank you, Senator. Number one, I would form a new agency that licenses any effort above a certain scale of capabilities and can take that license away and ensure compliance with safety standards. Number two, I would create a set of safety standards focused on what you said in your third hypothesis, as the dangerous capability evaluations. One example that we've used in the past is looking to see if a model can self-replicate and self-exfiltrate into the wild. We can give your office a long other list of the things that we think are important there, but specific tests that a model has to pass before it can be deployed into the world. And then, third, I would require independent audits, so not just from the company or the agency, but experts who can say the model is or isn't in compliance with these stated safety thresholds and these percentages of performance on question X or Y. ¶
altman 我们会这样做的。 We will do that. ¶
altman 我爱我现在的工作。[笑声] I love my current job. [Laughter.] ¶
altman 我们很乐意给您推荐合适的人选,是的。 We'd be happy to send you recommendations for people out there, yes. ¶
altman 不。我的薪水够付医保。我在 OpenAI 没有股权。 No. I'm paid enough for health insurance. I have no equity in OpenAI. ¶
altman 是的。 Yes. ¶
altman 我做这件事是因为我热爱它。 I'm doing this because I love it. ¶
altman 是的。我举几个例子。一个是关于暴力内容。另一个是关于鼓励自残的内容。还有一个是成人内容。并不是说我们认为成人内容本身有害,但有些东西可能与之相关,而我们无法足够可靠地区分,所以我们一概拒绝。 Yes. I'll give a few examples. One would be about violent content. Another would be about content that's encouraging self-harm. Another is adult content. Not that we think adult content is inherently harmful, but there's things that could be associated with that that we cannot reliably enough differentiate, so we refuse all of it. ¶
altman 要谈你所说的前一部分,有一些事情,除了「这条内容该不该被生成?」之外,我认为也同样重要。所以,你提到的那张图片是被生成的——我认为一个好的政策是,在所有场合都明确标注生成的图片是生成的。这样的话,我们仍然让那张图片留在那里,但至少要求人们说明这是一张生成的图片。 To touch on the first part of what you said, there are things besides, you know, ``Should this content be generated or not?'' that I think are also important. So, that image that you mentioned was generated--I think it'd be a great policy to say generated images need to be made clear in all contexts that they were generated. And, you know, then we still have the image out there, but we're at least requiring people to say this was a generated image. ¶
altman 我认为许可制度发挥作用的地方,不在于这些模型今天能做什么,因为,正如你指出的,你不需要一个新的许可机构来做到那一点。但当我们迈向——你知道的,这可能需要很长时间,我不确定——当我们迈向人工通用智能,以及它将带来的影响和这项技术的力量时,我认为我们需要像对待其他非常强大的技术一样认真对待它。这就是我个人认为我们需要这样一种制度的原因。 Where I think the licensing scheme comes in is not for what these models are capable of today, because, as you pointed out, you don't need a new licensing agency to do that. But as we head--and, you know, this may take a long time. I'm not sure. As we head towards artificial general intelligence and the impact that will have and the power of that technology, I think we need to treat that as seriously as we treat other very powerful technologies. And that's why I personally think we need such a scheme. ¶
altman We think this is really important. One example is that we worked with the government of Iceland, which is a language with fewer speakers than many of the languages that are well represented on the internet, to ensure that their language was included in our model. And we've had many similar conversations, and I look forward to many similar partnerships with lower resource languages to get them into our models. GPT-4 is unlike previous models of ours, which were good at English and not very good at other languages--now pretty good at a large number of languages. You can go pretty far down the list, ranked by number of speakers, and still get good performance. But for these very small languages, we're excited about custom partnerships to include that language into our model run. And the part of the question you asked about values and making sure that cultures are included, we're equally focused on that, excited to work with people who have particular data sets and to work to collect a representative set of values from around the world, to draw these wide bounds of what the system can do. I also appreciate what you said about the benefits of these systems and wanting to make sure we get those to as wide of a group as possible. I think these systems will have lots of positive impact on a lot of people, but in particular, historically underrepresented groups in technology, people who have not had as much access to technology around the world. This technology seems like it can be a big lift up. 我们认为这非常重要。一个例子是,我们与冰岛政府合作——冰岛语这门语言的使用者比许多在互联网上有充分代表的语言都要少——以确保他们的语言被纳入我们的模型。我们进行过许多类似的对话,我期待与低资源语言开展更多类似的合作,把它们纳入我们的模型。GPT-4 不像我们以前的模型——以前的模型擅长英语而不太擅长其他语言——现在在大量语言上都相当不错。你可以按使用者数量排名的名单往下走得相当远,仍能获得良好的表现。但对于这些非常小的语言,我们对定制合作、把那种语言纳入我们的模型运行感到兴奋。至于你问题中关于价值观、以及确保各种文化被纳入的部分,我们同样予以关注,很高兴与拥有特定数据集的人合作,努力从世界各地收集一套有代表性的价值观,以划定系统能做什么的这些宽泛边界。我也感谢你所说的关于这些系统的益处,以及希望确保让尽可能广泛的群体获得这些益处。我认为这些系统会对很多人产生很多积极影响,但尤其是对历史上在技术领域代表性不足的群体,以及世界各地的许多人——那些此前没能充分接触技术的人。这项技术似乎能成为一次大的提升。 We think this is really important. One example is that we worked with the government of Iceland, which is a language with fewer speakers than many of the languages that are well represented on the internet, to ensure that their language was included in our model. And we've had many similar conversations, and I look forward to many similar partnerships with lower resource languages to get them into our models. GPT- 4 is unlike previous models of ours, which were good at English and not very good at other languages--now pretty good at a large number of languages. You can go pretty far down the list, ranked by number of speakers, and still get good performance. But for these very small languages, we're excited about custom partnerships to include that language into our model run. And the part of the question you asked about values and making sure that cultures are included, we're equally focused on that, excited to work with people who have particular data sets and to work to collect a representative set of values from around the world, to draw these wide bounds of what the system can do. I also appreciate what you said about the benefits of these systems and wanting to make sure we get those to as wide of a group as possible. I think these systems will have lots of positive impact on a lot of people, but in particular, historically underrepresented groups in technology, people who have not had as much access to technology around the world. This technology seems like it can be a big lift up. ¶
altman 是的。谢谢你的提问,Ossoff 参议员。我认为这极其重要。我认为这里有非常不同的层次,我认为重要的是,任何新方法、任何新法律都不要阻止较小公司、开源模型、以及从事较小规模工作的研究者进行创新。这是我们这个生态系统、也是美国的美好组成部分,我们不想拖慢它。那里可能仍然需要一些规则,但我认为我们可以在需要严格许可的系统这里划一条线。最容易的做法——我不确定它是不是最好的,但最容易的——是谈论投入这样一个模型的算力。所以,你知道的,我们可以定义一个算力阈值——它必须变化。它可能上升也可能下降;随着我们发现更高效的算法,它可能下降——规定「超过这个算力,你就进入这个制度」。我更偏好的——这更难做,但我认为更准确——是定义一些能力阈值,然后说:「一个能做 X、Y 和 Z 这类事情的模型——由你们来决定——现在就进入这个许可制度。」但能力较弱——你知道的,我们不想阻止我们的开源社区,不想阻止个人研究者,不想阻止新初创公司——的模型,可以在一个不同的框架下继续推进。 Yes. Thanks for asking, Senator Ossoff. I think it's super important. I think there are very different levels here, and I think it's important that any new approach, any new law does not stop the innovation from happening with smaller companies, open-source models, researchers that are doing work at a smaller scale. That's a wonderful part of this ecosystem and of America, we don't want to slow that down. There still may need to be some rules there, but I think we could draw a line at systems that need to be licensed in a very intense way. The easiest way to do it--I'm not sure if it's the best, but the easiest would be to talk about the amount of compute that goes into such a model. So, you know, we could define a threshold of compute--and it'll have to change. It could go up or down; down, as we discover more efficient algorithms, that says, ``Above this amount of compute, you are in this regime.'' What I would prefer--it's harder to do, but I think more accurate--is to define some capability thresholds and say, ``A model that can do things X, Y, and ''Z--up to you all to decide--``that's now in this licensing regime.'' But models that are less capable--you know, we don't want to stop our open-source community, we don't want to stop individual researchers, we don't want to stop new startups--can proceed, you know, with a different framework. ¶
altman 我很想——与其当场即兴作答,不如随后给你办公室一个经过深思熟虑的——【译注:碎片话轮,原文以「-- --」断裂结尾,句意未完;按源件边界保留】 I would love, rather than to do that off the cuff, to follow up with your office with, like, a thoughtful-- -- ¶
altman 好吧。本着只是发表看法的心态,我认为,一个能够说服、操纵、影响一个人行为或其信念的模型——这会是一个不错的门槛。我认为,一个能够帮忙制造新型生物制剂(biological agents)的模型会是一个很好的门槛。诸如此类。 All right. In the spirit of just opining, I think a model that can persuade, manipulate, influence a person's behavior or a person's beliefs--that would be a good threshold. I think a model that could help create novel biological agents would be a great threshold. Things like that. ¶
altman 我认为我们并不确定地知道这个问题的答案,但可以说,它至少能在那里有一些影响。 I think we don't know the answer to that for sure, but let's say it can at least have some impact there. ¶
altman 是的。我认为非常重要的是,我们要继续理解这些是人类用来做出人类判断的工具,而我们不要剥夺人的判断。例如,我不认为人们应该基于一个 AI 系统的输出而被起诉。 Yes. I think it's very important that we continue to understand that these are tools that humans use to make human judgments and that we don't take away human judgment. I don't think that people should be prosecuted based off of the output of an AI system, for example. ¶
altman 我认为那会是好事。 I think it'd be good. ¶
altman 再说一次,这远远超出了我的专业领域。我认为有许许多多隐私专家,他们能比我更好地就一部法律需要什么发表意见。 Again, this is very far out of my area of expertise. I think there's many, many people that are privacy experts that could weigh in on what a law needs much better than I can. ¶
altman 我是说,我认为最低限度是,用户应当能够选择不让他们的个人数据被像我们这样的公司或社交媒体公司使用。删除你的数据应当很容易。我认为这些是——但从我经营一家 AI 公司的角度来看,我认为重要的是,如果你不希望你的数据被用于训练这些系统,你就有权这样做。 I mean, I think a minimum is that users should be able to sort of opt out from having their data used by companies like ours or the social media companies. It should be easy to delete your data. I think those are--but the thing that I think is important, from my perspective running an AI company, is that if you don't want your data used for training these systems, you have the right to do that. ¶
altman 我刚才说的是一件稍微不同的事,也就是某个人生成的数据、他们向我们系统提出的问题、他们在那里输入的东西,被用于训练。而公共网络上可获取的数据——即使我们不用它训练,模型当然也可以链接到它。所以那不是我所指的东西。我认为,你知道的,有一些方法——或者说应该有更多方法——让你的数据从公共网络上被移除,但当然,具备网页浏览能力的模型将能够搜索网络并链接到它。 So, I was speaking about something a little bit different, which is the data that someone generates, the questions they ask our system, things that they input there, training on that. Data that's on the public web, that's accessible--even if we don't train on that, the models can certainly link out to it. So, that was not what I was referring to. I think that, you know, there's ways to have your data or there should be more ways to have your data taken down from the public web, but certainly models with web-browsing capabilities will be able to search the web and link out to it. ¶
altman 是的。 Yes. ¶
altman 嗯,我认为两者都是。 Well, I think both. ¶
altman 我是纵深防御(defense in depth)的信奉者。我认为,对于一个已部署的模型能做什么、以及它实际做什么,都应该有限制。 I'm a believer in defense in depth. I think that there should be limits on what a deployed model is capable of and then what it actually does, too. ¶
altman 嗯,你必须——我是说,你必须使用产品得满 18 岁,或者 13 岁以上要得到父母许可。但我们理解,人们总是能绕过这些防护措施。所以我们试图做的就是设计一个安全的产品。有些决定,如果我们知道只有成年人在用,我们是会允许的,但因为我们知道儿童也会以某种方式使用它,我们就不在产品里允许这些。特别是,鉴于这些系统在教育中被如此广泛地使用,我们——希望意识到这一点正在发生。 Well, you have to be--I mean, you have to be 18 or up, or have your parent's permission at 13 and up, to use a product. But we understand that people get around those safeguards all the time. And so what we try to do is just design a safe product. And there are decisions that we make that we would allow if we knew only adults were using it, that we just don't allow in the product because we know children will use it some way or other, too. In particular, given how much these systems are being used in education, we, like, want to be aware that that's happening. ¶
altman 我们完全同意。我想我们时间到了,但如果能回应,我很乐意谈那个问题。 We couldn't agree more. I think we're out of time, but I'm happy to talk about that if I can respond. ¶
altman 好的。首先,我认为我们试图设计不追求最大化用户参与度(engagement)的系统。事实上,我们 GPU 非常短缺,人们越少使用我们的产品越好。但我们不是基于广告的模式。我们并不试图让人们越来越多地使用它。我认为这与广告支持的社交媒体是不同形态的。第二,这些系统确实有能力以明显的方式、也有以非常微妙的方式施加影响,我认为这对儿童安全尤其重要,但也会影响我们所有人。我们自己会做的一件事——无论有没有监管,但我认为监管方式对此也有好处——是关于这些系统的价值观如何被设定、以及这些系统如何回应那些可能造成影响的问题的要求。所以我们很乐意与你们合作。在重要性上完全同意。 Okay. First of all, I think we try to design systems that do not maximize for engagement. In fact, we're so short on GPUs, the less people use our products, the better. But we're not an advertising-based model. We're not trying to get people to use it more and more. And I think that's a different shape than ad-supported social media. Second, these systems do have the capability to influence in obvious and in very nuanced ways, and I think that's particularly important for the safety of children, but that will impact all of us. One of the things that we'll do ourselves, regulation or not, but I think a regulatory approach would be good for, also, is requirements about how the values of these systems are set and how these systems respond to questions that can cause influence. So, we'd love to partner with you. Couldn't agree more on the importance. ¶
altman 我们最初是一家非营利组织,真正专注于这项技术将如何被构建。当时,像 AGI 这样的事情甚至是否可能,都远远超出「奥弗顿之窗」(Overton Window)。这一点已经改变了很多。我们当时不知道规模会如此重要,但我们确实知道,我们希望在心中怀着人类最佳利益来构建它,并相信这项技术如果能按我们想要的方向发展,如果我们能做到 Marcus 教授提到的那些事情,就能深刻改变世界。我们希望成为一股推动走向积极——的力量。【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】 We started as a nonprofit, really focused on how this technology was going to be built. At the time, it was very outside the Overton Window that something like AGI was even possible. That's shifted a lot. We didn't know at the time how important scale was going to be, but we did know that we wanted to build this with humanity's best interests at heart and a belief that this technology could, if it goes the way we want, if we can do some of those things Professor Marcus mentioned, really deeply transform the world. And we wanted to be as much of a force for getting to a positive---- ¶
altman 我不会说永远不会。比如,我认为可能有些我们想向其提供服务的对象,而没有任何其他可行的模式,但我真的很喜欢基于订阅的模式。我们有 API 开发者付费给我们,也有 ChatGPT 消费者付费给我们。 I wouldn't say never. Like, I think there may be people that we want to offer services to, and there's no other model that works, but I really like having a subscription-based model. We have API developers pay us, and we have ChatGPT consumers pay us. ¶
altman 当然。 Sure. ¶
altman 我认为会有很多人开发模型。开源社区正在发生的事情令人惊叹,但会有数量相对较少的供应商能够制造真正处于最前沿(bleeding edge)的模型——【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 there will be many people that develop models. What's happening on the open-source community is amazing, but there will be a relatively small number of providers that can make models at the true leading edge---- ¶
altman 我认为其中既有益处也有危险。比如,就像我们谈论 AI 的所有危险时所说的,你真正需要密切留意的人数越少——在绝对的能力最前沿——这有好处,但我认为也需要有足够多的——而且会有,因为其中有如此巨大的价值——让消费者有选择,让我们有不同的想法。 I think there is benefits and danger to that. Like, as we were talking about all of the dangers with AI, the fewer of us that you really have to keep a careful eye on, on the absolute, like, bleeding edge of capabilities, there's benefits there, but I think there needs to be enough--and there will, because there's so much value--that consumers have choice, that we have different ideas. ¶
altman 我还想补充一件事。我认为非常重要的一点是,这些系统被对齐(aligned)到什么、什么样的价值观、那些边界是什么,这些应当以某种方式由整个社会、由各国政府来设定。因此,创建那个数据集、那个对齐数据集——它可能是一部 AI 宪法,无论是什么——它必须非常广泛地来自整个社会。 One more thing I wanted to add. One thing that I think is very important is that what these systems get aligned to, whose values, what those bounds are, that that is somehow set by society as a whole, by governments as a whole. And so creating that data set, the alignment data set--it could be an AI constitution, whatever it is--that has got to come very broadly from society. ¶
altman 谢谢您,参议员。第一,我认为美国必须继续引领。这件事发生在美国。我很自豪它发生在美国。 Thank you, Senator. One, I think America has got to continue to lead. This happened in America. I'm very proud that it happened in America. ¶
altman 我个人相信有一种方法可以两者兼顾。我认为在这件事上拥有全球视角很重要,因为无论这项技术在哪里被开发,它都会影响美国人和我们所有人。但我认为我们希望美国来引领。我们希望——【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 personally believe there's a way to do both. And I think it is important to have the global view on this, because this technology will impact Americans and all of us wherever it's developed. But I think we want America to lead. We want---- ¶
altman 嗯,这是一个。我是说,那是一个危险——【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】 Well, that's one. I mean, that is a peril---- ¶
altman 我认为那是个很棒的类比。 I think it's a great analogy. ¶
altman 其一是,我们不使用提交给我们 API 的任何数据进行训练。所以,如果你是我们的企业客户并提交数据,我们完全不会用它训练。我们确实会保留 30 天,仅出于信任与安全执行的目的。但那与用它训练不同。如果你使用 ChatGPT,你可以选择退出让我们用你的数据训练。你也可以删除你的对话历史或整个账户。 One is that we don't train on any data submitted to our API. So, if you're a business customer of ours and submit data, we don't train on it at all. We do retain it for 30 days, solely for the purpose of trust and safety enforcement. But that's different than training on it. If you use ChatGPT, you can opt out of us training on your data. You can also delete your conversation history or your whole account. ¶
altman 我当然同意那些是重要的观点。我想补充的是——Marcus 教授也谈到了这一点。我想补充的是,随着我们——我们今天大部分时间都花在当前的风险上,我认为这是恰当的,我很高兴我们这样做了。随着这些系统确实变得更有能力——我不确定那还有多远,但也许不是很远——我认为重要的是,我们也要花时间谈论我们将如何应对那些挑战。 I certainly agree that those are important points. I would add that--and Professor Marcus touched on this. I would add that as we--we spent most of the time today on current risks, and I think that's appropriate, and I'm very glad we have done it. As these systems do become more capable-and I'm not sure how far away that is, but maybe not super far--I think it's important that we also spend time talking about how we're going to confront those challenges. ¶
altman 你知道我有多在意。 You know how much I care. ¶
altman 谢谢。 Thank you. ¶
altman 所以,首先,在我们完成 GPT-4 的训练之后,我们等了 6 个多月才部署它。我们目前没有在训练将成为 GPT-5 的东西。我们也没有在未来 6 个月内这样做的计划。但我认为那封信的框架是错误的。重要的是审计、红队测试(red teaming)、以及模型在训练前必须通过的安全标准。如果我们暂停 6 个月,那么我真的不确定之后我们做什么。我们再暂停 6 个月吗?我们到那时才想出一些规则吗?我们已经制定并用于 GPT-4 部署的标准——我们希望在此基础上继续,但我们认为那是正确的方向,而不是按日历时钟的暂停。可能有些时候——我预计会有这样的时刻,当我们发现某些我们不理解的东西、确实需要暂停时,但我们现在还没看到那种情况,更不用说所有的益处了。 So, first of all, after we finished training GPT-4, we waited more than 6 months to deploy it. We are not currently training what will be GPT-5. We don't have plans to do it in the next 6 months. But I think the frame of the letter is wrong. What matters is audits, red teaming safety standards that a model needs to pass before training. If we pause for 6 months, then I'm not really sure what we do then. Do we pause for another six? Do we kind of come up with some rules then? The standards that we have developed and that we've used for GPT-4 deployment--we want to build on those, but we think that's the right direction, not a calendar-clock pause. There may be times--I expect there will be times when we find something that we don't understand and we really do need to take a pause, but we don't see that yet, never mind all the benefits. ¶
altman 抱歉。 I'm sorry. ¶
altman 我们没有看到不训练一个新模型的理由。至于部署,正如我提到的,我认为有各种各样的风险行为,我们设置了限制。我们有时不得不收回一些东西,添加新的。我的意思是,我们没有看到某种会阻止我们训练下一个模型的东西,以至于我们会担心即使在那个过程中也会创造出危险的东西,更不用说部署了。 We don't see reasons to not train a new one. For deploying, as I mentioned, I think there's all sorts of risky behavior, and there's limits we put. We have to pull things back sometimes, add new ones. I meant we don't see something that would stop us from training the next model, where we'd be so worried that we'd create something dangerous even in that process, let alone the deployment. ¶
altman 但那可能会发生。 But that may happen. ¶
altman 我是说,请原谅我的无知。难道人们不能起诉我们吗?这不像是——【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 mean, please forgive my ignorance. Can't people sue us? It's not like---- ¶
altman 哦。嗯,我认为,有很多其他法律,如果,你知道的,技术伤害了你,有一些标准可以据此起诉我们,除非我真的误解了事情运作的方式。如果问题是:「关于这项技术具体细节和消费者保护的更清晰法律是好事吗?」我会说绝对是。 Oh. Well, I think there's, like, a lot of other laws where if, you know, technology harms you, there's standards that we could be sued under, unless I'm really misunderstanding how things work. If the question is, ``Are clearer laws about the specifics of this technology and consumer protections a good thing?'' I would say definitely yes. ¶
altman 是的。我是说,再说一次,这在很大程度上是我们创办 OpenAI 的原因。我们对此有巨大的担忧。我认为重要的是让这些系统的输入民主化,即我们将要对齐(align)的价值观。我也认为重要的是让人们广泛使用这些工具。当我们启动 API 战略时——这是我们让系统可供任何人使用的重要方式之一——对此有大量的怀疑。它确实伴随着挑战,这是肯定的。但我们认为,把它交到很多人手中、而不是少数公司手中,真的相当重要,我们正看到由此产生的创新繁荣。但绝对真实的是,能够训练真正前沿模型的公司数量将会很少,仅仅因为所需的资源。所以我认为需要对我们和我们的竞争对手进行难以置信的审视。我认为正在发生一个丰富而令人兴奋的行业,有非常优秀的研究和新初创公司,它们不只是使用我们的模型,还在创造自己的模型。我认为重要的是,确保无论发生什么监管、无论可能或可能不会成立什么新机构,我们保留那团火,因为那至关重要。 Yes. I mean, again, that's so much of why we started OpenAI. We have huge concerns about that. I think it's important to democratize the inputs to these systems, the values that we're going to align to. And I think it's also important to give people wide use of these tools. When we started the API strategy, which is a big part of how we make our systems available for anyone to use, there was a huge amount of skepticism over that. And it does come with challenges, that's for sure. But we think putting this in the hands of a lot of people and not in the hands of a few companies is really quite important, and we are seeing the resultant innovation boom from that. But it is absolutely true that the number of companies that can train the true frontier models is going to be small, just because of the resources required. And so I think there needs to be incredible scrutiny on us and our competitors. I think there is a rich and exciting industry happening of incredibly good research and new startups that are not just using our models but creating their own. And I think it's important to make sure that whatever regulatory stuff happens, whatever new agencies may or may not happen, we preserve that fire, because that's critical. ¶
altman 是的。 Yes. ¶
altman 所以,我刚才关于使用模型、并在其上构建的这一点,作为一个——这是一个真正的新平台,对吧?谈论谁将创造这些模型,确实很重要。我想那样做。我也认为,决定我们将这些模型对齐(align)到谁的价值观,这一点同样重要。但在使用模型方面,那些在 OpenAI API 之上构建的人做出了不起的事情。而且,你知道的,人们经常评论说:「我不敢相信你用这么少的钱就能得到这么多技术。」所以人们——公司,人们正在建立的、把 AI 到处应用、使用我们 API 的东西,它确实让我们能够设置安全保障——我认为那相当令人兴奋。我认为这就是它正在被——不是将会被,而是现在正在被——民主化的方式。有一场全新的寒武纪大爆发(Cambrian explosion),由许多不同的公司在这些模型之上创造新企业、新产品、新服务。 So, this point that I made about use of the model and building on top of it, as a--this is really a new platform, right? It is definitely important to talk about who's going to create the models. I want to do that. I also think it's really important to decide to whose values we're going to align these models. But in terms of using the models, the people that build on top of the OpenAI API do incredible things. And, you know, people frequently comment, like, ``I can't believe you get this much technology for this little money.'' And so what people are--the companies people are building, putting AI everywhere, using our API, which does let us put safeguards in place--I think that's quite exciting. And I think that is how it is being--not how it's going to be, but how it is being democratized right now. There is a whole new Cambrian explosion of new businesses, new products, new services happening by lots of different companies on top of these models. ¶
书面证词 萨姆·奥尔特曼(Sam Altman) OpenAI 首席执行官 呈交美国参议院司法委员会 隐私、技术与法律小组委员会 Written Testimony of Sam Altman Chief Executive Officer OpenAI Before the U.S. Senate Committee on the Judiciary Subcommittee on Privacy, Technology, & the Law ¶
Blumenthal 主席、Hawley 参议员,以及司法委员会的各位委员,感谢你们给我机会今天就大型神经网络作证。我是 Sam Altman,OpenAI 的首席执行官。OpenAI 是一家研究、构建并部署人工智能(AI)的公司,开发了 ChatGPT、Whisper 和 DALL·E 2 等 AI 工具。OpenAI 的创立基于这样一个信念:安全且有益的 AI 为人类提供了巨大的可能性。我很感激有机会讲述我们在开发前沿 AI 技术和研究 AI 安全方面的经验,以及我们与各国政府协作、确保开发并广泛提供安全且有益的 AI 工具的意愿。我们认为,制定既能激励 AI 安全、又能确保人们获得这项技术诸多益处的监管规则至关重要。 Chairman Blumenthal, Senator Hawley, and members of the Judiciary Committee, thank you for the opportunity to testify today about large neural networks. I am Sam Altman, Chief Executive Officer of OpenAI, a company that studies, builds, and deploys artificial intelligence (AI) and has created AI tools such as ChatGPT, Whisper, and DALL·E 2. OpenAI was founded on the belief that safe and beneficial AI offers tremendous possibilities for humanity. I am grateful for the opportunity to speak about our experiences developing cutting-edge AI technology and studying AI safety, and our interest in working collaboratively with governments to ensure the development and widespread availability of safe and beneficial AI tools. We believe it is essential to develop regulations that incentivize AI safety while ensuring that people are able to access the technology’s many benefits. ¶
## 关于 OpenAI About OpenAI ¶
OpenAI 是一家总部位于旧金山的公司,创立于 2015 年,宗旨是确保通用人工智能(AGI)惠及全人类。OpenAI 的工作由我们的章程驱动:我们承诺致力于让 AGI 的利益得到广泛分配,致力于最大化 AI 系统的长期安全,致力于与其他研究及政策机构合作,并致力于作为 AI 领域的技术领导者来达成这些目标。 OpenAI is a San Francisco based company created in 2015 to ensure that artificial general intelligence (AGI) benefits all of humanity. OpenAI’s work is driven by our charter, in which we commit to working toward the broad distribution of the benefits of AGI, to maximizing the long term safety of AI systems, to cooperating with other research and policy institutions, and to serving as a technical leader in AI to accomplish these objectives. ¶
OpenAI 采用了一种不同寻常的结构,以确保其始终聚焦于这一长期使命。我们有几条关键的经济与治理条款: OpenAI has an unusual structure that ensures that it remains focused on this long-term mission. We have a few key economic and governance provisions: ¶
- 第一,我们结构中的主体实体是我们的非营利组织(Nonprofit),它是一家 501(c)(3) 公共慈善机构。 - 第二,我们的营利性业务设有利润上限,并由一家完全受非营利组织控制的子公司运营。 - 第三,由于董事会服务于非营利组织,每位董事都必须履行其受托责任以推进其使命——安全的、广泛受益的 AGI。虽然营利性子公司被允许创造和分配利润,但它必须服从这一使命。非营利组织的主要受益人是全人类,而不是 OpenAI 的投资者。 - 第四,董事会保持多数独立。独立董事不持有 OpenAI 的股权。 - 第五,投资者和员工的利润受具有约束力的法律承诺所限制。非营利组织为人类的利益保留全部剩余价值。 ● First, the principal entity in our structure is our Nonprofit, which is a 501(c)(3) public charity. ● Second, our for-profit operations are subject to profit caps and under a subsidiary that is fully controlled by the Nonprofit. ● Third, because the board serves the Nonprofit, each director must perform their fiduciary duties in furtherance of its mission—safe AGI that is broadly beneficial. While the for-profit subsidiary is permitted to make and distribute profit, it is subject to this mission. The Nonprofit’s principal beneficiary is humanity, not OpenAI investors. ● Fourth, the board remains majority independent. Independent directors do not hold equity in OpenAI. ● Fifth, profit for investors and employees is capped by binding legal commitments. The Nonprofit retains all residual value for the benefit of humanity. ¶
这一结构使我们能够把安全且有益的 AI 开发置于优先位置,同时帮助我们获得开发日益强大的 AI 模型所需的资本。例如,继 2019 年和 2021 年的投资之后,微软于今年 1 月宣布对 OpenAI 进行一项多年期、数十亿美元的投资。1 这项投资为 OpenAI 开发、测试和改进我们的技术提供了必要的资本和先进的超级计算基础设施。微软是 OpenAI 的重要投资者,我们珍视他们与我们在价值观和长期愿景上的独特一致,包括他们与我们共同致力于构建值得信赖且安全的 AI 系统和产品。与此同时,OpenAI 仍然是一家完全独立的公司,由 OpenAI 非营利组织治理。微软不占有董事会席位,也不控制 OpenAI。此外,AGI 技术被明确保留给非营利组织治理。 This structure enables us to prioritize safe and beneficial AI development while helping us secure the necessary capital to develop increasingly powerful AI models. For example, in January, Microsoft announced a multiyear, multibillion dollar investment in OpenAI, following previous investments in 2019 and 2021.1 This investment provides necessary capital and advanced supercomputing infrastructure for OpenAI to develop, test, and improve our technology. Microsoft is an important investor in OpenAI, and we value their unique alignment with our values and long-term vision, including their shared commitment to building AI systems and products that are trustworthy and safe. At the same time, OpenAI remains an entirely independent company governed by the OpenAI Nonprofit. Microsoft has no board seat and does not control OpenAI. Furthermore, AGI technologies are explicitly reserved for the Nonprofit to govern. ¶
## OpenAI 的技术与工具 OpenAI Technology and Tools ¶
OpenAI 是大型语言模型(LLM)及其他 AI 工具的领先开发者。从根本上说,当前这一代 AI 模型是大规模的统计预测机器——当模型收到一个人的请求时,它会尝试预测一个可能的回应。这些模型的运行方式类似于现代智能手机、电子邮件或文字处理软件上的自动补全功能,但规模和复杂程度要大得多。2 模型通过阅读或观看关于世界的数据来学习,预测能力随之不断提高,直到能够执行诸如总结文本、写诗和编写计算机代码等任务。利用这项技术的变体,AI 工具还能够学习图像与文字描述之间的统计关系,然后根据自然语言输入生成新的图像。 OpenAI is a leading developer of large language models (LLMs) and other AI tools. Fundamentally, the current generation of AI models are large-scale statistical prediction machines – when a model is given a person’s request, it tries to predict a likely response. These models operate similarly to auto-complete functions on modern smartphones, email, or word processing software, but on a much larger and more complex scale.2 The model learns from reading or seeing data about the world, which improves its predictive abilities until it can perform tasks such as summarizing text, writing poetry, and crafting computer code. Using variants of this technology, AI tools are also capable of learning statistical relationships between images and text descriptions and then generating new images based on natural language inputs. ¶
我们的模型在范围广泛的数据上训练,其中包括公开可用的内容、获得授权的内容,以及由人类审核员生成的内容。3 创造这些模型不仅需要先进的算法设计和大量训练数据,还需要庞大的计算基础设施来训练模型,然后为数百万用户运行它们。 Our models are trained on a broad range of data that includes publicly available content, licensed content, and content generated by human reviewers.3 Creating these models requires not just advanced algorithmic design and significant amounts of training data, but also substantial computing infrastructure to train models and then operate them for millions of users. ¶
我们近期的主要发布包括能够生成图像和文本的工具。2022 年初,我们推出了 DALL·E 2 的研究预览版,这是一个能够根据自然语言描述生成逼真图像和艺术作品的 AI 系统。4 如今,数百万用户正在使用 DALL·E 创作和改进图像,并与世界分享他们的作品。自最初的预览版以来,我们不断扩展 DALL·E 的能力,包括推出 DALL·E 应用程序编程接口(API),帮助开发者将 DALL·E 集成到应用和产品中。5 Our major recent releases include tools that can generate images and text. In early 2022, we launched a research preview of DALL·E 2, an AI system that can create realistic images and art from a description in natural language.4 Millions of users are now creating and improving images using DALL·E and sharing their creations with the world. Since the initial preview, we have expanded DALL·E’s capabilities, including launching a DALL·E Application Programming Interface (API) to help developers integrate DALL·E into apps and products.5 ¶
1 「Microsoft and OpenAI Extend Partnership.」 Microsoft, 23 Jan. 2023, https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/. 2 Johnson, Steven. 「A.I. Is Mastering Language. Should We Trust What It Says?」 New York Times Magazine, 15 Apr. 2022, https://www.nytimes.com/2022/04/15/magazine/ai-language.html. 3 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 4 「DALL·E 2.」 OpenAI, https://openai.com/product/dall-e-2. Accessed 14 May 2023. 5 「DALL·E API Now Available in Public Beta.」 OpenAI, 3 Nov. 2022, https://openai.com/blog/dall-e-api-now-available-in-public-beta. 1 “Microsoft and OpenAI Extend Partnership.” Microsoft, 23 Jan. 2023, https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/. 2 Johnson, Steven. “A.I. Is Mastering Language. Should We Trust What It Says?” New York Times Magazine, 15 Apr. 2022, https://www.nytimes.com/2022/04/15/magazine/ai-language.html. 3 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 4 “DALL·E 2.” OpenAI, https://openai.com/product/dall-e-2. Accessed 14 May 2023. 5 “DALL·E API Now Available in Public Beta.” OpenAI, 3 Nov. 2022, https://openai.com/blog/dall-e-api-now-available-in-public-beta. ¶
在文本方面,我们已经训练并公开发布了多个 LLM,从 2019 年的 GPT-2 系列模型6 和 2020 年的 GPT-3 系列模型7 开始。2022 年 11 月,我们发布了 ChatGPT。8 这些模型可用于整理、总结或生成新文本。它们「理解」用户的查询和指令,然后基于这些查询生成看似合理的回应。模型通过预测回应用户请求的下一个可能的词来生成回应,然后继续预测之后的每一个词。这些模型在世界上大多数地区免费提供;我们还推出了试点订阅服务 ChatGPT Plus,为用户提供额外的权益,9 同时我们也通过 API 向开发者提供这些模型,供其构建应用和服务。 On the text side, we have trained and publicly released a number of LLMs, beginning with the GPT-2 family of models in 20196 and the GPT-3 family of models in 2020.7 In November 2022, we released ChatGPT.8 These models can be used to organize, summarize, or generate new text. They “understand” user queries and instructions, then generate plausible responses based on those queries. The models generate responses by predicting the next likely word in response to the user’s request, and then continuing to predict each subsequent word after that. The models are available for free in most of the world; we also have launched a pilot subscription service, ChatGPT Plus, that provides additional benefits to users,9 and we make the models available as an API for developers to build applications and services. ¶
今年 3 月,我们发布了 GPT-4——我们最先进的系统,它能够产出比以往 OpenAI 产品更有用、更有创意、更善于协作、也更准确的内容。10 GPT-4 已在 ChatGPT Plus 上提供,并且(与其他 GPT 模型一样)通过 API 向开发者开放,供其构建应用和服务。 In March of this year, we released GPT-4, our most advanced system, which is capable of producing more useful, more creative, more collaborative, and more accurate outputs than previous OpenAI products.10 GPT-4 is available on ChatGPT Plus and (as with other GPT models) as an API for developers to build applications and services. ¶
## AI 持续改善人们的生活 AI Continues to Improve People’s Lives ¶
OpenAI 的使命是确保 AI 系统被安全、有益地构建、部署和使用。我们亲眼看到这些系统在改善人们生活方面的潜在和实际的积极影响。我们听到来自世界各地的用户讲述他们有多喜爱我们的工具,以及 AI 如何在许多方面帮助他们,包括帮助他们更快地编写复杂的计算机代码、提升他们的生产力和创造力、增强他们已有的能力、帮助他们学习新技能,以及拓展他们的业务。11 以下只是客户使用我们产品的部分方式: OpenAI’s mission is to ensure that AI systems are built, deployed, and used safely and beneficially. We see firsthand both the potential and the actual positive impact that these systems have in improving people’s lives. We hear from users all over the world about how much they love our tools and how AI helps them in many ways, including helping them write complex computer code more quickly, enhancing their productivity and creativity, augmenting their existing strengths, helping them learn new skills, and expanding their businesses.11 Here are just some of the ways that customers are using our products: ¶
- 教育非营利机构可汗学院(Khan Academy)正在试点一个项目,用 GPT-4 驱动面向学生的个性化虚拟导师和面向教师的课堂助手。12 - 摩根士丹利(Morgan Stanley)正在用 GPT-4 驱动一个面向内部的聊天机器人,对财富管理内容进行全面检索,「有效释放了摩根士丹利财富管理的累积知识」,帮助他们的财务顾问更好地服务客户。13 ● Educational non-profit Khan Academy is piloting a program that uses GPT-4 to power a personalized virtual tutor for students and a classroom assistant for teachers.12 ● Morgan Stanley is using GPT-4 to power an internal-facing chatbot that performs a comprehensive search of wealth management content and “effectively unlocks the cumulative knowledge of Morgan Stanley Wealth Management,” helping their financial advisors better serve their clients.13 ¶
6 「GPT-2: 1.5B Release.」 OpenAI, 5 Nov. 2019, https://openai.com/research/gpt-2-1-5b-release. 7 「OpenAI API.」 OpenAI, 11 June 2020, https://openai.com/blog/openai-api. 8 「Introducing ChatGPT.」 OpenAI, 30 Nov. 2022, https://openai.com/blog/chatgpt. 9 「Introducing ChatGPT Plus.」 OpenAI, 1 Feb. 2023, https://openai.com/blog/chatgpt-plus. 10 「GPT-4 Is OpenAI's Most Advanced System, Producing Safer and More Useful Responses.」 OpenAI, https://openai.com/product/gpt-4. Accessed 14 May 2023. Vincent, James. 「OpenAI Announces GPT-4 — The Next Generation of Its AI Language Model.」 The Verge, 14 Mar. 2023, https://www.theverge.com/2023/3/14/23638033/openai-gpt-4-chatgpt-multimodal-deep-learning. 11 Paris, Francesca and Larry Buchanan. 「35 Ways Real People Are Using A.I. Right Now.」 New York Times: The Upshot, 14 Apr. 2023, https://www.nytimes.com/interactive/2023/04/14/upshot/up-ai-uses.html. 12 「Khan Academy.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/khan-academy. 13 「Morgan Stanley.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/morgan-stanley. 6 “GPT-2: 1.5B Release.” OpenAI, 5 Nov. 2019, https://openai.com/research/gpt-2-1-5b-release. 7 “OpenAI API.” OpenAI, 11 June 2020, https://openai.com/blog/openai-api. 8 “Introducing ChatGPT.” OpenAI, 30 Nov. 2022, https://openai.com/blog/chatgpt. 9 “Introducing ChatGPT Plus.” OpenAI, 1 Feb. 2023, https://openai.com/blog/chatgpt-plus. 10 “GPT-4 Is OpenAI’s Most Advanced System, Producing Safer and More Useful Responses.” OpenAI, https://openai.com/product/gpt-4. Accessed 14 May 2023. Vincent, James. “OpenAI Announces GPT-4 — The Next Generation of Its AI Language Model.” The Verge, 14 Mar. 2023, https://www.theverge.com/2023/3/14/23638033/openai-gpt-4-chatgpt-multimodal-deep-learning. 11 Paris, Francesca and Larry Buchanan. “35 Ways Real People Are Using A.I. Right Now.” New York Times: The Upshot, 14 Apr. 2023, https://www.nytimes.com/interactive/2023/04/14/upshot/up-ai-uses.html. 12 “Khan Academy.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/khan-academy. 13 “Morgan Stanley.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/morgan-stanley. ¶
- Stripe 正在以多种方式使用 GPT-4,包括改进其客户支持运营、帮助回答有关技术文档的支持问题,以及帮助检测其社区平台上的欺诈行为。14 - 面向法律专业人士的平台 Harvey 正在使用 GPT-4 让研究和起草等任务更高效,使律师能够把更多时间用于策略,并向更多客户提供更高质量的服务。15 - 韩国增长最快的英语应用 Speak 正在使用我们的自动语音识别 AI 系统 Whisper,来驱动一个 AI 口语陪练,提供真正开放式的对话练习。16 - Weave 正在使用我们的工具为科学家构建协作平台,特别关注肿瘤学领域的突破。 - 从电影导演到独立音乐人,创意专业人士正在使用我们的图像生成工具 DALL·E 来增强他们的创作过程,从快速绘制故事板,到创作过去不可能实现的封面艺术。 ● Stripe is using GPT-4 in a variety of ways, including improving its custom support operations, helping to answer support questions about technical documentation, and helping to detect fraud on its community platforms.14 ● Harvey, a platform for legal professionals, is using GPT-4 to make tasks such as research and drafting more efficient so they can focus more time on strategy, and deliver a higher quality service to more clients.15 ● Speak, the fastest-growing English application in South Korea, is using Whisper, our automatic speech recognition AI system, to power an AI speaking companion and provide true open-ended conversational practice.16 ● Weave is using our tools to build a collaboration platform for scientists, specifically focused on breakthroughs in oncology. ● Creative professionals from movie directors to indie musicians are using our image generation tool, DALL·E, to augment their creative processes, from rapid storyboarding to creating cover art that would not have previously been possible. ¶
我们还与非营利组织及其他机构合作,探索我们工具的社会公益用途。例如,我们的技术帮助一家名为 Be My Eyes 的非营利组织,让盲人或低视力人士借助我们的模型来描述他们「看到」的事物。这些用户通常依靠志愿者帮助完成数以百计的日常生活任务,而我们看到,一个名为「虚拟志愿者」(Virtual Volunteer)的 AI 驱动系统能够达到与志愿者相当的情境理解水平。17 The Trevor Project 使用 GPT-2 大幅扩展了其在 LGBTQ 青少年中预防自杀的工作规模;Lad in a Battle 则用 DALL·E 为儿科癌症患者带来欢乐。冰岛政府正在使用 GPT-4 来保护冰岛语,18 其他国家也表示有兴趣使用同一模型来保护资源不足的语言。 We also partner with nonprofit and other organizations to explore socially beneficial uses of our tools. For example, our technology enables a nonprofit called Be My Eyes to help people who are blind or have low vision use our models to help describe what they are seeing. These users normally rely on volunteers for help with hundreds of daily life tasks, and we’re seeing that an AI powered system called “Virtual Volunteer” can help reach the same level of context and understanding as a volunteer.17 The Trevor Project has used GPT-2 to significantly scale its efforts to prevent suicide among LGBTQ teens, while Lad in a Battle has used DALL·E to bring joy to pediatric cancer patients. The government of Iceland is using GPT-4 in its preservation efforts for the Icelandic language,18 and other countries have expressed interest in using this same model to preserve under-resourced languages. ¶
我还想分享 Ben Whittle 的故事。他是一名泳池安装工和园艺师,患有阅读障碍(dyslexia)。Ben 曾担心自己的阅读障碍会损害他与客户之间的邮件沟通。Ben 的一位客户基于我们的技术开发了一个 AI 工具,通过提供建议、改进语法和添加职业化措辞,帮助 Ben 写出更好的邮件。如今 Ben 在所有工作邮件中都使用这个 AI 工具,并相信它为自己的公司赢得一份 26 万美元的合同发挥了重要作用。《华盛顿邮报》援引 Ben 的话说:「这恰好给了我所需要的东西。」19 I also want to share the story of Ben Whittle, a pool installer and landscaper with dyslexia. Ben feared that his dyslexia would harm his email communications with his clients. One of Ben's clients created an AI tool built on top of our technology to help Ben write better emails by making suggestions, improving his grammar, and adding professional niceties. Ben now uses this AI tool for all his work emails and believes it played a significant role in securing a $260,000 contract for his company. Ben was quoted in the Washington Post, saying, "this has given me exactly what I need.”19 ¶
14 「Stripe.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/stripe. 15 「Introducing Our First Investments.」 OpenAI Startup Fund, 1 Dec. 2022, https://openai.fund/news/introducing-our-first-investments. 「A&O Announces Exclusive Launch Partnership with Harvey.」 Allen & Overy, 15 Feb. 2023, https://www.allenovery.com/en-gb/global/news-and-insights/news/ao-announces-exclusive-launch-partnership-with-harvey. 16 「Introducing Our First Investments.」 OpenAI Startup Fund, 1 Dec. 2022, https://openai.fund/news/introducing-our-first-investments. 17 「Be My Eyes.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/be-my-eyes. 18 「Government of Iceland.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/government-of-iceland. 14 “Stripe.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/stripe. 15 “Introducing Our First Investments.” OpenAI Startup Fund, 1 Dec. 2022, https://openai.fund/news/introducing-our-first-investments. “A&O Announces Exclusive Launch Partnership with Harvey.” Allen & Overy, 15 Feb. 2023, https://www.allenovery.com/en-gb/global/news-and-insights/news/ao-announces-exclusive-launch-partner ship-with-harvey. 16 “Introducing Our First Investments.” OpenAI Startup Fund, 1 Dec. 2022, https://openai.fund/news/introducing-our-first-investments. 17 “Be My Eyes.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/be-my-eyes. 18 “Government of Iceland.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/government-of-iceland. ¶
这只是我们的技术在人们学习采用和使用 AI 工具时使之受益的一种方式。正是这些机会,让美国前财政部长 Lawrence Summers 说,ChatGPT 这样的 AI 工具可能像印刷机、电力,甚至轮子或火一样具有深远的影响。20 This is just one way our technology can benefit people as they learn to adopt and use AI tools. These opportunities are why former U.S. Treasury Secretary Lawrence Summers has said that AI tools such as ChatGPT might be as impactful as the printing press, electricity, or even the wheel or fire.20 ¶
能够参与建设未来,我们感到无比兴奋、充满机遇,也深感责任重大。 We feel an immense amount of excitement, opportunity, and responsibility in being involved with helping build the future. ¶
## AI 安全实践 AI Safety Practices ¶
虽然我们相信,我们已经部署的工具的益处远远大于其风险,但确保它们的安全对我们的工作至关重要,我们付出重大努力,把安全内建于我们系统的所有层面。在下文中,我将讨论我们的总体安全方针,以及我们为使模型更安全而采取的一些具体措施。 While we believe the benefits of the tools we have deployed vastly outweigh the risks, ensuring their safety is vital to our work and we make significant efforts to ensure safety is built into our systems at all levels. In the sections below, I discuss our general approach to safety and some of the specific steps we take to make our models safer. ¶
在发布模型的每个新版本之前,OpenAI 都会进行广泛的测试,聘请外部专家提供反馈,通过基于人类反馈的强化学习(RLHF)等技术改进模型的行为,并部署安全与监控系统。21 Prior to releasing each new version of our models, OpenAI conducts extensive testing, engages external experts for feedback, improves the model's behavior with techniques like reinforcement learning from human feedback (RLHF), and implements safety and monitoring systems.21 ¶
我们最新模型 GPT-4 的发布就是一个有代表性的例子。在开发出 GPT-4 之后,我们花了 6 个多月的时间对系统进行评估、测试和改进,然后才将其公开发布。22 除了我们自己的评估之外,我们还聘请了外部 AI 安全专家进行一个被称为「红队测试」(red teaming)的过程,专家们帮助识别了 GPT-4 在多个领域的潜在风险,包括生成不准确信息(即「幻觉」)、仇恨内容、虚假信息,以及与常规和非常规武器扩散相关的信息。23 这个过程帮助我们更好地理解潜在的使用风险以及化解这些风险的方法。 The release of our latest model, GPT-4, provides an illustrative example. After we developed GPT-4, we spent more than 6 months evaluating, testing, and improving the system before making it publicly available.22 In addition to our own evaluations, we engaged with external AI safety experts in a process known as “red teaming,” through which they helped identify potential concerns with GPT-4 in areas including the generation of inaccurate information (known as “hallucinations”), hateful content, disinformation, and information related to the proliferation of conventional and unconventional weapons.23 This process helped us to better understand potential usage risks and ways to address those risks. ¶
在上述每一个领域,我们都制定了缓解措施,显著提升了安全性。我们的一部分工作涉及在所谓预训练阶段调整用于训练模型的数据,其他干预措施则发生在模型初始训练之后。24 举例来说,在预训练阶段,我们减少了数据集中色情文本内容的数量。25 在预训练阶段之后,我们塑造 GPT-4 行为的主要方法是让人类对模型的回应提供反馈,以帮助教导我们的模型以更安全、更有用的方式作出回应。26 我们还教导模型尽量拒绝有害的请求,并在面对敏感请求时更恰当地回应。这些努力在实证上降低了模型生成有害或不准确内容的可能性。27 当被要求生成不被允许的内容(依据我们的使用政策定义)时,GPT-4 的拒绝率超过 99%。28 尽管我们的模型仍然存在局限,在某些情况下仍可能生成不被允许或不准确的信息,但我们通过这些安全工作取得了显著进展,并将继续在此基础上前行。 In each of these areas, we developed mitigations to increase safety in significant ways. Some of our work involved making adjustments to the data used to train the model, during what is called the pre-training stage. Other interventions took place after initial training of the model.24 At the pre-training stage, for example, we reduced the quantity of erotic text content in our dataset.25 After the pre-training stage, our primary method for shaping GPT-4’s behavior involves having people provide feedback on model responses, in order to help teach our models to respond in a way that is safer and more useful.26 We also teach the model to try to refuse harmful requests and to respond more appropriately in the face of sensitive requests. These efforts empirically reduced the likelihood that the model would generate harmful or inaccurate content.27 When asked to generate disallowed content (as defined by our usage policies), GPT-4 refuses to do so more than 99% of the time.28 While our models still have limitations and can generate disallowed or inaccurate information in some cases, we’ve made significant progress through these safety efforts, and we’re continuing to build on them. ¶
19 Harwell, Drew, Nitasha Tiku, and Will Oremus. 「Stumbling with Their Words, Some People Let AI Do the Talking.」 Washington Post, 10 Dec. 2022, https://www.washingtonpost.com/technology/2022/12/10/chatgpt-ai-helps-written-communication/. 20 Summers, Lawrence. Bloomberg: Wall Street Week, 9 Dec. 2022, https://www.youtube.com/watch?v=iR31ZAacyGM. 21 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 22 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, p. 19, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 23 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, p. 4, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 19 Harwell, Drew, Nitasha Tiku, and Will Oremus. “Stumbling with Their Words, Some People Let AI Do the Talking.” Washington Post, 10 Dec. 2022, https://www.washingtonpost.com/technology/2022/12/10/chatgpt-ai-helps-written-communication/. 20 Summers, Lawrence. Bloomberg: Wall Street Week, 9 Dec. 2022, https://www.youtube.com/watch?v=iR31ZAacyGM. 21 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 22 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. “GPT-4 System Card.” OpenAI, 23 Mar. 2023, p. 19, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 23 GPT-4 System Card.” OpenAI, 23 Mar. 2023, p. 4, https://cdn.openai.com/papers/gpt-4-system-card.pdf. ¶
## 部署安全与学习 Deployment Safety and Learning ¶
我们在部署前努力理解并防范风险。29 然而,我们无法预先预见这项技术的每一种有益用途、潜在滥用或失效情形。这在很大程度上是因为这些系统仍是由人指挥的——它们会努力遵循用户的指令去完成任务。从真实世界中真实用户的实际使用里学习并作出回应,对于创造更安全的 AI 系统至关重要。30 We work hard to understand and prevent risks before deployment.29 However, we can't anticipate every beneficial use, potential abuse, or failure of the technology. This is in large part because these systems are still human-directed—they try to follow user instructions to carry out tasks. Learning from and responding to real-world use by actual people is vital for creating safer AI systems..30 ¶
我们的部署实践是谨慎而渐进地发布新的 AI 模型——在配备充分保障措施的前提下——逐步扩大用户群体,并根据汲取的教训持续改进。我们还通过自己的服务(以及通过 API)提供我们能力最强的模型,这使我们能够监测滥用行为并采取行动,并持续构建针对人们真实滥用方式的缓解措施。 Our deployment practices involve cautiously and gradually releasing new AI models—with substantial safeguards in place—to gradually larger groups of people, making continuous improvements based on the lessons learned. We also make our most capable models available through our own services (and through an API), which allows us to monitor for and take action on misuse, and continually build mitigations that respond to the real ways people misuse our systems. ¶
正如我们的《使用政策》所述,OpenAI 明确禁止将其工具用于某些活动,包括但不限于:生成暴力内容、恶意软件、欺诈活动、大规模政治竞选活动,以及许多其他不受欢迎的领域。31 As described in our Usage Policies, OpenAI expressly prohibits the use of its tools for certain activities, including but not limited to, generation of violent content, malware, fraudulent activity, high-volume political campaigning, and many other unwelcome areas.31 ¶
24 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, pp. 21-22, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 25 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, pp. 3, 21, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 26 Ouyang, Long, et al. 「Training Language Models to Follow Instructions with Human Feedback.」 arXiv, 4 Mar. 2022, https://arxiv.org/pdf/2203.02155.pdf. 27 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, p. 20, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 28 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, pp. 22-24, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 29 「Lessons Learned on Language Model Safety and Misuse.」 OpenAI, 3 Mar. 2022, https://openai.com/research/language-model-safety-and-misuse. 30 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 31 「Usage Policies.」 OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. 24 “GPT-4 System Card.” OpenAI, 23 Mar. 2023, pp. 21-22, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 25 GPT-4 System Card.” OpenAI, 23 Mar. 2023, pp. 3, 21, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 26 Ouyang, Long, et al. “Training Language Models to Follow Instructions with Human Feedback.” arXiv, 4 Mar. 2022, https://arxiv.org/pdf/2203.02155.pdf. 27 GPT-4 System Card.” OpenAI, 23 Mar. 2023, p. 20, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 28 GPT-4 System Card.” OpenAI, 23 Mar. 2023, pp. 22-24, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 29 “Lessons Learned on Language Model Safety and Misuse.” OpenAI, 3 Mar. 2022, https://openai.com/research/language-model-safety-and-misuse. 30 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 31 “Usage Policies.” OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. ¶
我们结合使用自动化检测系统和人工审核来发现潜在的违规行为,以便警告用户或采取执法行动。我们使用我们最新的模型来帮助识别不安全的内容——这减少了人类审核员接触有害或露骨内容的需要,帮助我们快速完善审核政策,并缩短了构建安全工具所需的时间。我们还向开发者免费提供一整套审核与安全工具,供其集成到自己的产品中。 We use a combination of automated detection systems and human review to detect potentially violating behavior in order to warn users or take enforcement actions. We use our newest models to help to identify unsafe content—this reduces the need for human moderators to be exposed to harmful or explicit content, helps us to quickly refine our moderation policies, and reduces the time needed to build safety tools. We also provide a free suite of moderation and safety tools to our developers to integrate into their products. ¶
我们努力做到敏捷响应客户的关切。我们基于客户、用户、公众以及其他利益相关群体(包括各国政府)的反馈,持续更新和改进我们的模型与产品。 We strive to be agile and responsive to customer concerns. We are continuously updating and improving our models and products based on feedback from our customers, users, the public, and other stakeholder groups, including governments. ¶
迭代式部署对 AI 安全还有其他好处。我们认为,人们和我们的机构需要时间来更新并适应日益强大的 AI,每一个受这项技术影响的人都应当有权对 AI 的进一步发展发表重要意见。自去年 11 月 OpenAI 发布 ChatGPT 以来,关于生成式 AI 的公共对话已取得了巨大的进展。与人们没有第一手使用这些工具的经验相比,迭代式部署帮助我们把各方利益相关者更有效地带入了关于采用 AI 技术的对话。32 Iterative deployment has other advantages for AI safety. We believe that people and our institutions need time to update and adjust to increasingly capable AI, and that everyone who is affected by this technology should have a significant say in how AI develops further. The public dialogue on generative AI has advanced dramatically since OpenAI launched ChatGPT last November. Iterative deployment has helped us bring various stakeholders into the conversation about the adoption of AI technology more effectively than if they hadn't had firsthand experience with these tools.32 ¶
## 隐私 Privacy ¶
OpenAI 严肃对待用户隐私,并已采取多项措施来促进透明和负责任的数据使用。第一,我们不使用任何用户数据来为广告投放、推广我们的服务或向第三方出售数据而建立人物画像。我们也不使用客户通过 API 提交的数据来训练或改进我们的模型,除非客户明确要求我们这样做。我们可能会使用 ChatGPT 对话来帮助改进我们的模型,但我们为用户提供了多种方式来控制其对话的使用。任何 ChatGPT 用户都可以选择退出,不让自己的对话被用于改进我们的模型。33 用户可以删除自己的账户,34 从历史侧栏中删除特定对话,并随时停用聊天记录功能。35 OpenAI takes the privacy of its users seriously and has taken a number of steps to facilitate transparent and responsible use of data. First, we don't use any user data to build profiles of people for the purposes of advertising, promoting our services, or selling data to third parties. We also do not use data submitted by customers via our API to train or improve our models, unless customers explicitly ask us to do this. We may use ChatGPT conversations to help improve our models, but we provide users with several ways to control how their conversations are used. Any ChatGPT user can opt-out of having their conversations be used to improve our models.33 Users can delete their accounts,34 delete specific conversations from the history sidebar, and disable their chat history at any time.35 ¶
虽然我们用于训练模型的一些信息可能包含公共互联网上可获得的个人信息,但我们会尽可能从训练数据集中移除个人信息,教导我们的模型拒绝索取普通个人信息的请求,并响应个人提出的从其系统中移除其个人信息的请求。这些措施降低了我们的模型生成包含普通个人信息的回应的可能性。 While some of the information we use to train our models may include personal information that is available on the public internet, we work to remove personal information from the training dataset where feasible, teach our models to reject requests for personal information of private individuals, and respond to requests from individuals to remove their personal information from our systems. These steps reduce the likelihood that our models might generate responses that include the personal information of private individuals. ¶
32 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 33 Markovski, Yaniv. 「How Your Data Is Used to Improve Model Performance.」 OpenAI, https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance. Accessed 15 May 2023. 34 C., Johanna. 「How Can I Delete My Account?」 OpenAI, May 2023, https://help.openai.com/en/articles/6378407-how-can-i-delete-my-account. Accessed 15 May 2023. 35 「New Ways to Manage Your Data in ChatGPT.」 OpenAI, 25 Apr. 2023, https://openai.com/blog/new-ways-to-manage-your-data-in-chatgpt. 32 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 33 Markovski, Yaniv. “How Your Data Is Used to Improve Model Performance.” OpenAI, https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance. Accessed 15 May 2023. 34 C., Johanna. “How Can I Delete My Account?” OpenAI, May 2023, https://help.openai.com/en/articles/6378407-how-can-i-delete-my-account. Accessed 15 May 2023. 35 “New Ways to Manage Your Data in ChatGPT.” OpenAI, 25 Apr. 2023, https://openai.com/blog/new-ways-to-manage-your-data-in-chatgpt. ¶
## 儿童安全 Children’s Safety ¶
我们安全工作的一个关键焦点是保护儿童。我们不允许我们的技术被用于生成仇恨、骚扰、暴力或成人内容等类别,36 并已设计缓解措施来帮助执行这些政策。与 GPT-3.5 相比,GPT-4 回应不被允许内容请求的可能性降低了 82%,我们还使用人工与自动化相结合的稳健审核流程来监测滥用。虽然这些系统并不完美,但我们已经取得了显著进展,并定期探索让系统更安全、更可靠的新方法。 One critical focus of our safety efforts is to protect children. We do not permit our technology to be used to generate hateful, harassing, violent or adult content, among other categories,36 and have designed mitigations to help enforce these policies. GPT-4 is 82% less likely to respond to requests for disallowed content compared to GPT-3.5, and we use a robust combination of human and automated review processes to monitor for misuse. Although these systems are not perfect, we have made significant progress, and are regularly exploring new ways to make our systems safer and more reliable. ¶
我们还采取了其他重要措施,最大限度地降低我们的模型生成可能伤害儿童的内容的可能性。例如,当用户试图向我们的图像工具上传已知的儿童性虐待材料(CSAM)时,我们使用 Thorn 的 Safer 服务37 来检测、审核、阻断这些行为,并向美国国家失踪与受虐儿童中心报告。 We have taken other significant measures to minimize the potential for our models to generate content that may be harmful to children. For example, when users try to upload known Child Sexual Abuse Material to our image tools, we use Thorn’s Safer37 service to detect, review, block, and report the activity to the National Center for Missing and Exploited Children. ¶
除了默认的安全护栏之外,我们还与开发者合作制定量身定制的安全缓解措施——例如与非营利机构可汗学院合作,后者构建了一个 AI 驱动的助手,既能充当学生的虚拟导师,也能充当教师的课堂助手。38 我们还在开发一些功能,允许开发者对模型输出设定更严格的标准,以更好地支持希望获得此类功能的开发者和用户。 In addition to our default safety guardrails, we work with developers such as the non-profit Khan Academy—which has built an AI-powered assistant that functions as both a virtual tutor for students and a classroom assistant for teachers38—on tailored safety mitigations. We are also working on features that will allow developers to set stricter standards for model outputs to better support developers and users who want such functionality. ¶
## 准确性 Accuracy ¶
我们的模型不是通过检索或访问数据库或网络上的数据来回答问题的;39 它们是基于词语之间相互关联出现的可能性来预测答案的。在某些情况下,最可能彼此相邻出现的词并不一定是最准确的,因此 ChatGPT 或其他 AI 工具的输出也可能不准确。 Our models do not answer queries by retrieving or accessing data in a database or on the web;39 they predict answers based, in large part, on the likelihood of words appearing in connection with one another. In some circumstances, the most likely words that appear near each other may not be the most accurate ones, and the outputs of ChatGPT or other AI tools may also be inaccurate. ¶
提升事实准确性是 OpenAI 和许多其他 AI 研究者的重要工作焦点,我们正在持续取得进展。我们已经提升了 GPT-4 的事实准确性——它产出事实性内容的可能性比 GPT-3.5 高 40%。40 我们还利用用户标记为错误的 ChatGPT 输出反馈来提升 ChatGPT 的准确性;自产品发布以来,我们已经让 ChatGPT 生成关于人物的不准确信息的可能性降低了。 Improving factual accuracy is a significant focus for OpenAI and many other AI researchers, and we continue to make progress. We have improved the factual accuracy of GPT-4, which is 40% more likely to produce factual content than GPT-3.5.40 We also use user feedback on ChatGPT outputs that were flagged as incorrect to improve ChatGPT’s accuracy, and since the launch of the product, we have made ChatGPT less likely to generate inaccurate information about people. ¶
36 「Usage Policies.」 OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. 37 「Our Approach to AI Safety.」 OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. 「Customers.」 Safer, https://safer.io/customers/. Accessed 15 May 2023. 38 「Khan Academy.」 OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/khan-academy. 39 一个例外是当用户在 ChatGPT 上启用「浏览」(browsing)功能,让模型通过搜索网络来回答问题,而不是由模型自己回答。在这些情况下,答案基于模型对网络搜索结果的总结。 40 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, p. 25, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 36 “Usage Policies.” OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. 37 “Our Approach to AI Safety.” OpenAI, 5 Apr. 2023, https://openai.com/blog/our-approach-to-ai-safety. “Customers.” Safer, https://safer.io/customers/. Accessed 15 May 2023. 38 “Khan Academy.” OpenAI, 14 Mar. 2023, https://openai.com/customer-stories/khan-academy. 39 An exception is when the user enables the “browsing” feature on ChatGPT to have the models answer queries by searching the web, rather than answering by itself. In those cases, the answers are based on the model summarizing the web search results. 40 GPT-4 System Card.” OpenAI, 23 Mar. 2023, p. 25, https://cdn.openai.com/papers/gpt-4-system-card.pdf. ¶
当用户注册使用 ChatGPT 时,我们努力明确告知:它的回答可能并不总是事实准确的。然而,我们认识到,在教育用户了解 AI 工具的局限、降低不准确的可能性方面,还有更多工作要做。最大限度地减少不准确的回应是一个活跃的研究问题,我们和其他 AI 实验室都在研究,我们对有助于解决这一问题的技术持乐观态度。 When users sign up to use ChatGPT, we strive to make it clear that its answers may not always be factually accurate. However, we recognize that there is more work to do to educate users about the limitations of AI tools, and to reduce the likelihood of inaccuracy. Minimizing inaccurate responses is an active research question that we and other AI labs are working on, and we are optimistic about techniques to help address this issue. ¶
## 虚假信息(Disinformation) Disinformation ¶
OpenAI 认识到 AI 工具有可能被用于助长虚假信息活动。打击虚假信息需要全社会共同参与的方法,OpenAI 很早就与研究者和业界同行接触,以了解 AI 可能被如何用于散布虚假信息。例如,我们最近与斯坦福大学和乔治敦大学的研究者合作发表了研究成果,指出了虚假信息活动滥用 LLM 可能带来的风险,以及一套可能有助应对该问题的潜在政策工具,如内容来源(content provenance)标准。41 如上所述,我们的《使用政策》也明确禁止将其工具用于某些活动,包括生成暴力内容、恶意软件、欺诈活动、大规模政治竞选活动及其他领域。42 OpenAI recognizes the potential for AI tools to contribute to disinformation campaigns. Fighting disinformation takes a whole-of-society approach, and OpenAI has engaged with researchers and industry peers early on to understand how AI might be used to spread disinformation. For example, we recently published work with researchers from Stanford and Georgetown Universities highlighting risks that might arise from disinformation campaigns misusing LLMs, as well as a set of potential policy tools that might help address the issue, such as content provenance standards.41 As noted above, our Usage Policies also expressly prohibit the use of its tools for certain activities, including generation of violent content, malware, fraudulent activity, high-volume political campaigning, and other areas.42 ¶
生成内容只是虚假信息生命周期中的一环;虚假或误导性信息还需要传播才能造成重大伤害。我们将继续探索与业界、研究者以及各国政府的合作,覆盖虚假信息生命周期的全部环节。 Generating content is only one part of the disinformation lifecycle; false or misleading information also requires distribution to cause significant harm. We will continue to explore partnerships with industry and researchers, as well as with governments, that encompass the full disinformation lifecycle. ¶
## 网络安全 Cybersecurity ¶
我们理解我们的模型和工具可能对世界产生重大影响,因此我们投入大量资源,最大限度地保护 OpenAI 的技术、知识产权和数据。43 我们维持严格的内部安全控制,并不断创新以改进防御。我们定期开展内部和第三方渗透测试,并审计安全控制的适当性与有效性。我们还在构建新颖的安全控制,以保护核心模型知识产权。我们力求让 OpenAI 的安全计划尽可能透明。我们的信任门户(Trust Portal)允许客户和其他利益相关者查阅我们的安全控制和审计报告。44 We understand that our models and tools can have significant impacts on the world, so we dedicate significant resources to maximizing protection of OpenAI’s technology, intellectual property, and data.43 We maintain strict internal security controls and are constantly innovating to improve our defenses. We regularly conduct internal and third-party penetration testing, and audit the suitability and effectiveness of our security controls. We are also building novel security controls to help protect core model intellectual property. We aim to make OpenAI’s security program as transparent as possible. Our Trust Portal allows customers and other stakeholders to review our security controls and audit reports.44 ¶
我们为提升网络安全而采取的其他措施包括: Other steps we take to improve cybersecurity include the following: ¶
- OpenAI 部分出于保护知识产权的考虑,将其最强大的 AI 模型作为服务部署。我们不在 OpenAI 及我们的技术合作伙伴微软之外分发这些模型的权重,并通过 API 向第三方提供访问,使模型权重、源代码和其他敏感信息留在 OpenAI 内部。OpenAI 持续改进其防御,为 emerging 威胁做好准备。 - 我们与合作伙伴及云服务提供商合作,在数据中心层面保护我们的模型。 - OpenAI 的安全计划在设计时考虑了潜在的内部威胁,并已建立控制措施来防范和监测模型与数据的外泄。 - 我们最近宣布启动一项漏洞赏金计划,邀请独立研究者报告我们系统中的漏洞,并给予现金奖励。45 我们还设有专门渠道,用于报告模型安全问题,例如模型的回应不准确或可能有害。46 ● OpenAI deploys its most powerful AI models as services in part to protect its intellectual property. We do not distribute weights for such models outside of OpenAI and our technology partner Microsoft, and we provide third-party access via API so the model weights, source code, and other sensitive information stay within OpenAI. OpenAI continuously improves its defenses to prepare for emerging threats. ● We work with partners and cloud providers to protect our models at the data center level. ● OpenAI’s security program is built to take into account potential insider threats, and we have built controls to prevent and monitor for model and data exfiltration. ● We recently announced the launch of a bug bounty program inviting independent researchers to report vulnerabilities in our systems in exchange for cash rewards.45 We also have a dedicated channel for reporting model safety issues, such as the model’s response being inaccurate or potentially harmful.46 ¶
41 Goldstein, Josh A., et al. 「Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations.」 arXiv, Jan. 2023, https://arxiv.org/pdf/2301.04246.pdf. 42 「Usage Policies.」 OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. 43 「Security & Privacy.」 OpenAI, https://openai.com/security. Accessed 15 May 2023. 44 「Trust Portal.」 OpenAI, https://trust.openai.com/. Accessed 15 May 2023. 41 Goldstein, Josh A., et al. “Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations.” arXiv, Jan. 2023, https://arxiv.org/pdf/2301.04246.pdf. 42 “Usage Policies.” OpenAI, 23 Mar. 2023, https://openai.com/policies/usage-policies. 43 “Security & Privacy.” OpenAI, https://openai.com/security. Accessed 15 May 2023. 44 “Trust Portal.” OpenAI, https://trust.openai.com/. Accessed 15 May 2023. ¶
我们也认识到,AI 工具既可以用于防御,也可以用于实施网络攻击。我们致力于评估我们的模型本身对网络安全的影响,并努力防止其被滥用。例如,我们正在筹建一项「网络安全资助计划」(Cybersecurity Grant Program),资助研究者围绕重要的防御性课题开展安全研究,例如训练防御性网络安全 agent、缓解社会工程攻击、识别和修补代码中的安全问题、自动化事件分诊等。我们还成立了网络安全工作组(Cybersecurity Working Group),研究如何防范 AI 网络威胁;我们的产品也经过严格测试以限制危害——例如,第三方安全研究者提前获得了 GPT-4 的访问权限,以测试其安全能力。 We also recognize that AI tools can be used to both defend against, as well as carry out, cyber attacks. We are committed to evaluating the impact of our models themselves on cybersecurity and working to prevent their misuse. For example, we are in the process of establishing a Cybersecurity Grant Program which will fund researchers conducting security research on salient defensive topics such as training defensive cybersecurity agents, mitigating social engineering, identifying and patching security issues in code, automating incident triage, and other issues. We have established a Cybersecurity Working Group to research how to prevent and protect against AI cyber threats, and our products undergo rigorous testing to limit harms—for example, third-party security researchers were given early access to GPT-4 to test its security capabilities. ¶
## 安全方法的持续改进 Continuing Improvements to Our Safety Approach ¶
我们认为,我们的安全方法应当由独立专家进行外部验证,我们的决定也应当至少部分地依据独立的安全与风险评估。这很重要。例如,在筹备 GPT-4 发布时,我们促成了对齐研究中心(Alignment Research Center,ARC)对 GPT-4 执行某些自主行为能力的初步模型评估。47 我们目前正在探索更多对我们的模型进行外部验证和测试的可能性。 We think it is important that our safety approaches are externally validated by independent experts, and that our decisions are informed at least in part by independent safety and risk assessments. For example, in preparing for the GPT-4 release, we facilitated a preliminary model evaluation by the Alignment Research Center (ARC) of GPT-4’s ability to carry out certain autonomous actions.47 We are currently exploring additional possibilities for external validation and testing of our models. ¶
随着我们创建和部署能力更强的模型,我们将更加谨慎,并随着 AI 系统的演进持续加强安全预防措施。48 我们正在投资开发面向更强大模型的增强型评估,包括评估 AI 模型是否具备可能严重破坏公共安全和国家安全的能力,以便在部署前制定适当的缓解措施。解决安全问题还需要广泛的讨论、实验和参与,包括关于 AI 系统行为边界的讨论。49 我们已经并将继续推动各利益相关方之间的协作与公开对话,以创造一个安全的 AI 生态。50 We will also be increasingly cautious with the creation and deployment of more capable models, and will continue to enhance safety precautions as our AI systems evolve.48 We are investing in developing enhanced evaluations for more powerful models, including assessing AI models for capabilities that could be significantly destabilizing for public safety and national security, so we can develop appropriate mitigations prior to deployment. Addressing safety issues also requires extensive discussion, experimentation, and engagement, including on the bounds of AI system behavior.49 We have and will continue to foster collaboration and open dialogue among stakeholders to create a safe AI ecosystem.50 ¶
45 「Announcing OpenAI's Bug Bounty Program.」 OpenAI, 11 Apr. 2023, https://openai.com/blog/bug-bounty-program. 46 「Model Behavior Feedback.」 OpenAI, https://openai.com/form/model-behavior-feedback. Accessed 15 May 2023. 47 「GPT-4 System Card.」 OpenAI, 23 Mar. 2023, p. 15, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 48 「Planning for AGI and Beyond.」 OpenAI, 24 Feb. 2023, https://openai.com/blog/planning-for-agi-and-beyond. 45 “Announcing OpenAI’s Bug Bounty Program.” OpenAI, 11 Apr. 2023, https://openai.com/blog/bug-bounty-program. 46 “Model Behavior Feedback.” OpenAI, https://openai.com/form/model-behavior-feedback. Accessed 15 May 2023. 47 GPT-4 System Card.” OpenAI, 23 Mar. 2023, p. 15, https://cdn.openai.com/papers/gpt-4-system-card.pdf. 48 “Planning for AGI and Beyond.” OpenAI, 24 Feb. 2023, https://openai.com/blog/planning-for-agi-and-beyond. ¶
## 经济影响 Economic Impacts ¶
我们理解,新的 AI 工具可能对劳动力市场产生深远影响。作为我们使命的一部分,我们正在努力理解我们产品的经济影响,并采取措施尽量减少对工人和企业的任何有害影响。我们很高兴与顶尖的经济学家密切合作研究这些问题,并且最近发表了一份关于语言模型及建立在其上的软件的经济影响的初步分析。51 我们预计 AI 将在短期内产生重大经济影响,包括个人用户生产率提升、就业创造、岗位转型和岗位取代等多种因素的交织。我们正在积极设法理解这些因素的相对比例,并渴望与美国政府在这些问题上密切合作,包括共享数据和合作研究。 We understand that new AI tools can have profound impacts on the labor market. As part of our mission, we are working to understand the economic impacts of our products and take steps to minimize any harmful effects for workers and businesses. We are excited to partner closely with leading economists to study these issues, and recently published a preliminary analysis of the economic implications of language models and the software built on top of them.51 We expect significant economic impacts from AI in the near-term, including a mix of increased productivity for individual users, job creation, job transformation, and job displacement. We are actively seeking to understand the relative proportions of these factors and are eager to work closely with the U.S. government on these issues, including by sharing data and partnering on research. ¶
我们认为,现在就着手为一系列潜在情景做好准备也很重要。我们正在资助对潜在政策工具和支持措施的研究,这些工具和措施可能有助于缓解未来技术颠覆带来的经济影响,例如实现失业保险福利的现代化,以及为受 AI 进步影响的工人设立调整援助计划。我们的目标当然不是直接左右这些政策,而是为政策制定者提供支持和洞见,帮助他们理解这项新技术对经济影响的潜在时间线和程度。我们还支持最大限度地提升公众对 AI 技术的广泛认知和理解,特别是通过面向那些所在岗位和职业可能受 AI 影响的工人的培训和教育项目。 We believe it is also important to begin work now to prepare for a range of potential scenarios. We are funding research into potential policy tools and support efforts that might help mitigate future economic impacts from technological disruption, such as modernizing unemployment insurance benefits and creating adjustment assistance programs for workers impacted by AI advancements. Our goal is of course not to shape these policies directly, but rather to provide the support and insights policymakers need to understand the potential timeline and extent of impacts of this new technology on the economy. We also support maximizing broad public awareness and understanding of AI technology, particularly through training and education programs for workers in roles and occupations likely to be impacted by AI. ¶
## 与各国政府合作 Working with Governments ¶
OpenAI 相信,对 AI 进行监管是必不可少的,我们渴望在决策者研究如何推动监管时提供帮助——监管应当在激励安全与确保人们能够获得这项技术的益处之间取得平衡。同样必不可少的是,像 AI 这样强大的技术必须以民主价值观为念来开发。OpenAI 致力于与美国决策者合作,在 AI 的关键领域保持美国的领先地位,并确保尽可能多美国人能够获得 AI 的益处。 OpenAI believes that regulation of AI is essential, and we’re eager to help policymakers as they determine how to facilitate regulation that balances incentivizing safety while ensuring that people are able to access the technology’s benefits. It is also essential that a technology as powerful as AI is developed with democratic values in mind. OpenAI is committed to working with U.S. policymakers to maintain U.S. leadership in key areas of AI and to ensuring that the benefits of AI are available to as many Americans as possible. ¶
我们正在积极与世界各地的决策者接触,帮助他们理解我们的工具并讨论监管选项。例如,我们赞赏美国国家标准与技术研究院(NIST)在其风险管理框架上所做的工作,目前正在研究如何将其具体应用于我们开发的这类模型。本月早些时候,我们与总统、副总统和白宫高级官员讨论了 AI,并期待与政府合作,宣布有意义的举措,在防范风险的同时,确保美国继续从 AI 中受益并保持在 AI 领域的领先地位。 We are actively engaging with policymakers around the world to help them understand our tools and discuss regulatory options. For example, we appreciate the work National Institute of Standards and Technology has done on its risk management framework, and are currently researching how to specifically apply it to the type of models we develop. Earlier this month, we discussed AI with the President, Vice President, and senior White House officials, and we look forward to working with the Administration to announce meaningful steps to help protect against risks while ensuring that the United States continues to benefit from AI and stays in the lead on AI. ¶
49 「How Should AI Systems Behave, and Who Should Decide?」 OpenAI, 16 Feb. 2023, https://openai.com/blog/how-should-ai-systems-behave. 50 「Best Practices for Deploying Language Models.」 OpenAI, 2 June 2022, https://openai.com/blog/best-practices-for-deploying-language-models. 51 Eloundou, Tyna, et al. 「GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.」 arVix, 27 Mar. 2023, https://arxiv.org/pdf/2303.10130.pdf. 49 “How Should AI Systems Behave, and Who Should Decide?” OpenAI, 16 Feb. 2023, https://openai.com/blog/how-should-ai-systems-behave. 50 “Best Practices for Deploying Language Models.” OpenAI, 2 June 2022, https://openai.com/blog/best-practices-for-deploying-language-models. 51 Eloundou, Tyna, et al. “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” arVix, 27 Mar. 2023, https://arxiv.org/pdf/2303.10130.pdf. ¶
为此,我想指出几个我认为 AI 公司与政府可以富有成效地合作的领域。 To that end, there are several areas I would like to flag where I believe that AI companies and governments can partner productively. ¶
第一,至关重要的是,AI 公司——尤其是那些研究最强大模型的公司——必须遵守一套适当的安全要求,包括发布前的内部和外部测试,以及评估结果的公布。为确保这一点,美国政府应考虑:对超过关键能力阈值的 AI 模型的开发和发布,实行许可(licensing)或注册(registration)要求的某种组合,并配套对完全合规的激励措施。 First, it is vital that AI companies–especially those working on the most powerful models–adhere to an appropriate set of safety requirements, including internal and external testing prior to release and publication of evaluation results. To ensure this, the U.S. government should consider a combination of licensing or registration requirements for development and release of AI models above a crucial threshold of capabilities, alongside incentives for full compliance with these requirements. ¶
第二,AI 是一个复杂且快速演进的领域。AI 公司必须满足的安全要求,需要一个足够灵活、能够适应新技术发展的治理机制,这一点至关重要。美国政府应考虑推动多利益相关方参与的进程,吸收广泛专家与组织的意见,为受许可或注册约束的 AI 系统制定并定期更新适当的安全标准、评估要求、披露实践和外部验证机制。 Second, AI is a complex and rapidly evolving field. It is essential that the safety requirements that AI companies must meet have a governance regime flexible enough to adapt to new technical developments. The U.S. government should consider facilitating multi-stakeholder processes, incorporating input from a broad range of experts and organizations, that can develop and regularly update the appropriate safety standards, evaluation requirements, disclosure practices, and external validation mechanisms for AI systems subject to license or registration. ¶
第三,开发这项技术的并不只有我们。决策者需要考虑如何在全球范围内实施许可监管,并确保 AI 安全方面的国际合作,包括研究潜在的政府间监督机制与标准制定,这将会很重要。 Third, we are not alone in developing this technology. It will be important for policymakers to consider how to implement licensing regulations on a global scale and ensure international cooperation on AI safety, including examining potential intergovernmental oversight mechanisms and standard-setting. ¶
## 结语 Conclusion ¶
这是从事 AI 技术工作的一个非凡时刻。六个月前,还没有人听说过 ChatGPT。如今,ChatGPT 已家喻户晓,人们正在以重要的方式从中受益。 This is a remarkable time to be working on AI technology. Six months ago, no one had heard of ChatGPT. Now, ChatGPT is a household name, and people are benefiting from it in important ways. ¶
我们也理解,人们对 AI 技术感到焦虑是理所当然的。我们非常严肃地对待这项技术的风险,未来也将继续如此。我们相信,政府和产业界携手,就能够管理好这些风险,让我们所有人都能享受这项技术的巨大潜力。 We also understand that people are rightly anxious about AI technology. We take the risks of this technology very seriously and will continue to do so in the future. We believe that government and industry together can manage the risks so that we can all enjoy the tremendous potential. ¶
## 作者简介 Biography ¶
Sam Altman:Sam Altman 是 OpenAI 的联合创始人兼首席执行官,OpenAI 是 ChatGPT 和 DALL·E 背后的 AI 研究与部署公司。2014 至 2019 年,Sam 担任早期创业加速器 Y Combinator 的总裁。2015 年,Sam 联合创立了 OpenAI——一家非营利研究实验室,使命是构建惠及全人类的通用人工智能。该公司至今仍由该非营利组织及其最初的章程治理。 Sam Altman: Sam Altman is the co-founder and CEO of OpenAI, the AI research and deployment company behind ChatGPT and DALL·E. Sam was president of the early-stage startup accelerator Y Combinator from 2014 to 2019. In 2015, Sam co-founded OpenAI as a nonprofit research lab with the mission to build general-purpose artificial intelligence that benefits all humanity. The company remains governed by the nonprofit and its original charter today. ¶