OpenAI 的计算策略与缩放定律

Greg，感谢你再次来到这里。我想我们从未向你收过租金，所以也许以后会给你寄张账单。你参与了两家非常出色的公司：Stripe，你是第四号员工，也是第一任 CTO。我最近听说他们处理了全球 GDP 的 1.6%，你一定为此感到自豪。更令人骄傲的是，OpenAI 现在每周活跃用户接近或超过十亿。这一切都非常令人兴奋，展示了技术的力量。你不仅是联合创始人兼总裁，还是首席构建者。我听说这是你在 OpenAI 的头衔之一，虽然不确定是否正式，但就这么说吧。在座的都是优秀的构建者，所以我们从最底层开始。OpenAI 的业务有多层，其中一层是算力，你们在获取算力上非常激进。为什么？

So, Greg, thank you for coming back here. I don't think we ever charged you for rent. So, maybe we'll send you an invoice later. But Greg, you've been part of two really spectacular companies, Stripe as employee number four, and then the first CTO. I just recently heard that they process 1.6% of the global GDP. You must be proud of that. That's amazing. You must be even more proud of the fact that OpenAI has almost a billion or maybe more than a billion weekly active users at this point. I mean, it's all very exciting. It shows you what technology can do. And you're not just co-founder and president, but you're also chief builder. At OpenAI, I heard that was one of your titles. I'm not sure if it was ever an official title, but let's just say that. Well, you have an audience of great builders here, so we'll start from all the way at the bottom of the stack. OpenAI has multiple stacks of the business, one of which is compute, and you guys have been very aggressive on securing compute. Why is that?

从很多方面来说，我们的业务非常简单。我们购买、租赁、构建算力，然后加价转售。仅此而已。只要利润率是正的，你就想扩大规模，因为解决问题的需求、对智能的需求是无限的。我们现在的 AI 确实能够应对你抛出的几乎任何问题。

Well, in many ways, we have a very simple business. We buy, rent, build compute, and we resell it at a margin. That's it. As long as the margin's positive, then you want to scale it because the demand for solving problems, the demand for intelligence, that's unlimited. And the AIs that we have right now really are able to rise to the challenge of effectively any kind of problem that you want to throw at them.

你们有足够的算力吗？

Do you have enough compute?

没有。

No.

真的吗？

Really?

是的，绝对不够。

Yeah, definitely not.

我刚和 Matt Garman 聊过，他说 2026 年的 GPU 算力可用性几乎为零。你们不是已经占有了所有算力吗？

I just spoke with Matt Garman, and he says the GPU compute availability in 2026 rounds to zero. Don't you guys have all of it?

我们当然想要更多。老实说，我们一直在到处寻找更多算力。我告诉你，当我们第一次推出 ChatGPT 时，我记得和团队通电话，他们问：“好吧，我们应该买多少算力？”我说：“全部。”他们说：“不，不，我们是认真的，到底该买多少？”我说：“无论我们多快增加算力，我保证我们无法跟上需求。”从那以后一直如此。

I mean, we would love more. We're constantly out there hunting for more, honestly. And I'll tell you, when we first launched ChatGPT, I remember being on a call with my team, and they're like, 'All right, how much compute should we buy?' And I said, 'All of it.' And they're like, 'No, no, no, we're serious, how much should we buy?' I'm like, 'No matter how fast we try to ramp compute, I guarantee you we're not going to be able to keep up with demand.' And that has been true ever since.

这很有趣。从算力往上，我不知道在座有多少人能帮你获取更多算力，因为他们大多是初创公司创始人。关于架构和缩放定律，我们现在处于什么位置？它们还在每年翻倍吗？你们在改变架构吗？在研究前沿你们在推动什么？

That's fascinating. Moving up from compute, since I don't know if much of this audience can help you with securing more compute, because most of them are founders of startups. About architecture and scaling laws, where are we in the scaling laws? Are they still doubling each year? Are you changing architecture? What are you guys pushing on the frontier on the research side?

首先，缩放定律是一个深刻而非常美丽的谜。它们感觉非常基础。就像你思考物理学和牛顿定律一样，它们某种程度上是宇宙的真理。而且它们是经验性的。我们不一定有完整的理论来解释为什么有效，但对我来说最美妙的是，神经网络实际上是在 1940 年代设计的，那时还没有计算机。不知何故，我们能够采用当时开发的精确想法，并应用越来越多的计算，随着你向模型投入更多算力，它们相应地变得更强大，而且一直在持续。没有天花板。这是一件美妙的事情。

Well, I would say first of all, the scaling laws are a deep and very beautiful mystery. They feel deeply fundamental. It's like a scientific truth that, just like you think about physics and Newton's laws and things like that, they're somehow this truth of the universe. And they're empirical. We don't necessarily have all the theory to explain exactly why it works, but to me the most beautiful thing is that neural networks were really designed in the 1940s, before there were computers. And somehow we've been able to take the exact ideas that were developed back then and apply increasing amounts of computation, and as you pour more compute into the models, they get correspondingly more capable, and it just keeps going. There's no wall. And that's a beautiful thing.

是否有更多研究或算法在进行中？因为过去我们有神经网络，如你所说，在 1940 年代，但我们没有计算机来实现。现在我们有了计算机，我们只是在推动同样的东西，还是有新的架构和新想法出现？

Are there more research or more algorithms that are in the works? Because in the past we had neural networks, to your point, in the 1940s, but we couldn't have the computer for it. Now that we have the computer for it, are we just pushing the same things, or are there new architectures and new ideas coming up?

是的，我认为我们绝对有新的想法不断推动我们的工作。简单地说“让我们把 1940 年代的神经网络放进一个千兆瓦的数据中心”是过于简化了。我们做了大量创新，并且不断改进。有时是微调，比如你意识到数据格式不太对，这实际上可能非常重要。有时更大，比如从 LSTM 到 Transformer 的转变，而且我不认为 Transformer 是终点。每个人都已经超越了 2018 年论文中描述的 Transformer。所以创新不断发生。我认为那些在长期研究上投入最多的机构，比如如何改进架构、改进基础算法、实现范式转变，OpenAI 一直处于领先地位，我们继续在这方面投资，我看到很多成果即将出现。

Yeah, so I would think of it as we absolutely have new ideas that are constantly powering what we do. It's very simplified to say, 'Well, let's take a neural network from the 1940s and put it in a gigawatt data center.' We have made tons of innovations and we constantly are improving things. And sometimes these are micro tweaks. Like you just realize that the way you've been formatting data was not quite right, and that can actually be a very big deal. Sometimes it's larger. You think about the shift from the LSTM to the transformer, and I don't think the transformer is the end. Everyone's moved past the transformer as described in the 2018 paper. So, there's constant innovation happening. And I think of places that have been perhaps the most invested in long-term research on how to improve the architectures, how to improve the fundamental algorithms, and how to get the paradigm shifts. I think OpenAI has been leading the pack there, and that's something we continue to invest in, and I see lots of fruit on the horizon.

明白了。关于模型，OpenAI 对 AGI 有正式定义吗？我们接近了吗？还是没有？我、Pat、Sam 和 Ilya 发表过一篇文章，说我们在功能上已经达到了 AGI。你同意吗？还是不同意？

Got it. And on the models, does OpenAI have a formal definition for AGI? Are we close? Are we not close? I, Pat, Sam, and Ilya published this thing that we are at AGI functionally. Do you agree with that? Do you not agree with that?

我们确实有一个正式定义，但在某种程度上，我学到的一件事是每个人对 AGI 都有自己的直觉。也许你可以这样看，根据我对现状的看法，我认为我们已经走了大约 80%的路，因为我们有智能模型。它们非常能干。如果给它们正确的上下文，它们能做出惊人的事情。

Well, we do have a formal definition, but to some extent, one thing I have learned is that everyone has their own intuitions about what AGI is. Maybe you can view it as, according to my view of where we are, I think we're about 80% of the way there, in that we have models that are smart. They're very capable. They're able to, if you give them the right context, they can do amazing things.

它们比你聪明吗？

Are they smarter than you?

我的意思是，它们在编写软件方面肯定比我更有能力，对吧？如果你给它们所有上下文，那么是的，我认为它们非常能干。这真的很了不起。在座有人觉得自己写软件比 GPT-5.4 更好吗？哦。好吧，写内核。即使在那里，我们也看到了巨大的进步。在我们的一些内部结果中，如果你为问题设置正确的条件，你就能从非常低级的任务中获得巨大的成果。举个例子说明趋势，我的一位系统工程师，情况类似，他说：“嘿，我从 GPT-5、5.1、5.2 的模型中没有获得价值。”对于 5.3，他一时兴起，准备了一份关于他即将进行的非常复杂的系统优化的设计文档。他把文档交给模型，然后去睡觉，醒来时打算把任务交给团队下周处理。但当他醒来时，任务已经完成了。模型实际上实现了初始规范，发现速度慢，添加了检测工具，实际运行了代码，使用分析器找出瓶颈，并多次迭代直到得到优化结果。这太不可思议了。

I mean, they're certainly more capable than I am at writing software, right? If you give it all the context, then yes, I think that they are just so capable. It's really remarkable. Is anyone here feel better at writing software than GPT-5.4? Oh. All right, writing kernels. So even there we're seeing massive gains. For some of our internal results, if you pour the right kinds of setup for your problem, then you're able to get really massive results out of very low-level tasks. And just to give you one example of how things have been trending, one of my systems engineers, also very similar, was like, 'Hey, I haven't been able to get value out of the models for GPT-5 or 5.1, for 5.2 as well.' For 5.3, he on a lark had prepared this design document for a very complicated systems optimization he was about to do. He handed it over to the model, went to sleep, waking up intending to give this to his team to work on for the next week. And when he woke up, it was done. The model had actually implemented the initial spec, had seen that it was slow, had added instrumentation, had actually run the code, used a profiler to figure out where things were slow, and iterated multiple times until it got into an optimized result. That is incredible.

我们现在就处在这样的局面。那么，你会给这里所有的初创公司什么建议？因为模型的能力越来越强。我之前 Sam 来的时候也问过这个问题：如果你今天在构建产品，两年后新模型出来时，你需要重建吗？因为所有的功能和能力都在你周围变化。你是否需要确保自己不会挡在 OpenAI 的路上，因为模型能力太强，你们会碾压初创公司？你会如何建议一群初创公司创始人在这种环境下构建？

That's where we are. And so, how what would you advise all startups here to do? Because the models keep getting more and more capable. They're kind of, uh, I've asked this, uh, when Sam was here in the past, and, you know, what if you're building today, do you need to rebuild in 2 years when a new model comes out? Because all the functionality and all the capabilities all change around you. Um, do you need to make sure that you're not in OpenAI's way because you're going to roll, you're just going to run over for startups because the models are so much more capable. Um, how how would you recommend the, uh, a set of, um, startup founders to to build in this environment?

首先，我会说要积极投入。现在的工具已经变得非常有用。如果你看看整个十二月，我认为我们从智能体式编码工具只能写 20% 的代码，发展到能写 80% 的代码。这意味着它们从配角变成了你工作的主角。我认为今年我们在人们用电脑做的所有工作中都会看到这一点。你可以看看 Codex 最近的进展。它正在从软件工程师的工具转变为任何用电脑工作的人的工具。就在过去一周，我们发布了一系列功能，让它变得更强大、更有能力。比如我们今天刚宣布的一个新工具叫 Chronicle，它集成到 Codex 中，可以实际看到你在电脑上做的一切，并形成记忆。所以你问它一个问题，它立刻就知道你在说什么。比如你问“我五分钟前在做什么？”它知道。你问“这个人刚才在说什么？”它也知道。对我来说，这真是一个警醒：你花了那么多精力向电脑解释发生了什么。为什么要向电脑解释？这毫无意义。所以我认为未来几年，模型会变得更强大，会有更好的工具，能够解决越来越难的问题，产生新知识，等等。但现在有一个一次性的转变，那就是上下文。关键在于你的 AI 是否能够——你开了这么多会，却没有让 AI 参与。这对 AI 不太公平。你让它帮你做事，但它没有信息。所以我认为要真正投入，确保 AI 在理论上拥有足够的信息来解决问题，然后相信模型会不断进步。所以这将是一个持续的改进和迭代循环，积极使用工具，和朋友们交流他们如何使用，但有一项一次性的投资，现在正是时候。

Well, first of all, I would say to lean in. The tools right now have become incredibly useful. And if you look even over the course of December, I think that we went from these agentic coding tools being like, you know, they're like writing 20% of your code to writing 80% of your code. Which means they go from being kind of a side show to being the main thing that you're doing. And I think we're doing that across all of the work that people do with computers, all computer work this year. And you can look at the recent progress on Codex. It's really changing from a tool for software engineers to a tool for anyone who's doing work with a computer. And just over the past week, we've released a bunch of features that just make it so much more powerful and capable. Um and like one thing we just announced today is a new tool called Chronicle that plugs into the Codex, where it actually can see everything you're doing with your computer and can form memories of what's going on. And so you ask it a question, you just it instantly knows what you're talking about. You're like, "Huh, what was I doing 5 minutes ago?" It knows, right? You're like, "Oh, what was this person talking about?" It knows. It's To me, it was this real wake-up call to realize you spend so much of your effort right now just explaining your computer what's going on. Like, why are you explaining to your computer what's going on? That makes no sense. And so I think what's going to happen over upcoming years is the models are going to get much more capable, will have better harnesses, will be able to be able to solve harder and harder problems, come up with new knowledge, all of these things. But there is a one-time shift that's happening now, which is really about context. It's really about is your AI able to you have all these meetings, you didn't include the AI. You know, that's not very nice to the AI. Like, you're asking it to to help you with things and it has no information. So I think really leaning into how do you make sure the AI even has enough information in theory to solve the problem and then trust the models are going to really get there and improve. So I think it will be a constant cycle of improvement and iteration and leaning into the tools and kind of talking to your friends if you're how they're using it, but that there is this investment that's a one-time investment that now is the time to make.

那么，假设我们把这些都设置好了，OpenAI 使用 Codex 的方式和你认为外界其他人使用它的方式有什么不同？

And in terms of like, let's say we you you set that all up, how do you how is OpenAI using Codex differently than you think everybody else outside is using it?

我认为在 OpenAI 工作的一大好处是你能活在未来。你能真正看到正在涌现的形态，我们可以共同设计。我们可以一起改变模型、工具等一切，以更好地满足我们看到的需求。我们采取的方法很多是从软件工程开始的，我们设定了一些明确的指导原则，比如我们仍然希望人类对所有合并的代码负责。最终，合并这段代码是好事吗？它结构良好吗？它会让我们的代码库更易于维护吗？我们确保有人签字确认。我认为这种深思熟虑不是简单地说“好吧，我们盲目使用它”，或者“哦，我们根本不想用它”。我认为这两种极端都不太对。然后我们在 OpenAI 内部按垂直领域推进，在财务、销售、IT 等部门采用这些工具。我们有一个专门的小团队，深入了解领域，与领域专家合作，构建技能，修改 Codex 的 UI，或者做任何必要的事情让它变得好用。一旦我们把它打磨好，我们就会外部化，把它交付给你们所有人。我们也在和一些客户合作。所以对于那些希望非常 AI 前沿、希望参与定义这场革命的人来说，是有空间的，我之后很乐意聊聊。但确实，这种渴望就是：“嘿，我们真的想走在 AI 前沿，真正活在未来，体验其他人一两年、三年后将会经历的事情。”

Well, I think one of the amazing things about being at OpenAI is you do get to live in the future, right? You do get to really see the shape of what's emerging, and we can co-design, right? We can really change the models, the harness, everything together in order to better serve the needs that we see. And a lot of the approach we've been taking is so we started with software engineering, and we set some clear guidelines for example saying that we still want a human to be accountable for all code that gets merged, right? So, at the end of the day, is it a good thing to merge this piece of code? Is it well-structured? Is it going to make our codebase more maintainable? We want to make sure there is a human who is signing off to say yes. And that's I think that thoughtfulness of not just saying, "Okay, let's just blindly use this." Or, you know, "Oh, we don't want to use this at all." Like I think neither extreme is quite right. And then we are also going vertical by vertical within OpenAI to adopt these tools within finance, within sales, within IT. And there we have a small dedicated team who's really deeply understanding the domain, working with the people who are the experts in it in order to build skills, in order to modify the Codex UI, whatever it is that is is needed in order to get it to be good. And then that's something we can then, once we have it in good shape, we will externalize, and that we're able to to ship that to all of you. And so we are starting to work with certain customers as well. So, for people who want to be very AI forward and want to be part of defining this revolution, that there's a place for that, and I'd love to talk afterwards. But yeah, I think that just this desire to say, "Hey, we really want to be AI forward, really live in the future, and experience what it will be like for everyone else 1 year, 2 years, 3 years down the road."

你们是否因为“活在未来”而用不同的方式组织公司和工程团队？我的意思是，如果追溯到很久以前，我父亲学计算机科学时，他只是一个人，然后我们有了漫长的软件发布，变成了瀑布模型。然后当网络和云出现时，有了“两个披萨团队”和 Scrum。现在我们有了这些编码智能体，你们是如何围绕这一切进行不同组织的？

Do you guys structure your company differently with the engineering teams differently because of of the living in the future? I mean, if you have to go way back, when my father learned computer science, he was just himself, and then we had these long software releases that became waterfall. And then when the web happened and the cloud happened, when these two-pizza teams, and we have scrum. Now that we have these coding agents, is it how do you structure around everything differently?

我认为我们还在摸索中，但在某些地方已经很明显了。例如，构建原型的成本现在很低。非常低，如果你想建一个仪表盘，以前可能需要一个人一周的时间，现在你马上就能做。所以很多瓶颈转移到了分享上。比如如何让企业中的任何人都能轻松构建一个仪表盘、小工具、机器人或其他东西，然后与他人分享。这开始对良好的治理产生压力。比如你希望 IT 组织能够看到所有不同的执行线程，所有被分享的小东西，对数据来源有一定的控制。确保一切正常。一个好的例子是，我认为人们现在开始将内部知识从存储到维基中转移出来。

I think we're still figuring it out, and there's certain places where you really see it. For example, the cost of building a prototype is cheap now. It's so cheap, and if you want to build a dashboard that used to be like, uh, it'd take like someone like a week to do it, and you just do it now. And so actually a lot of the bottleneck has shifted to things like sharing. Like how do you And so we we actually have some internal work on this as well that again we will be externalizing. If How do you make it really easy for anyone in your enterprise to build a dashboard, a widget, a bot, whatever the thing is, and then share it with others. And then that starts to really put pressure on having good governance. Like you want your IT organization to be able to see all these different, you know, threads of execution that are happening, all the little things that are being shared around, have some control over data provenance, right? To really make sure that, okay, like a good example of this is, um, I think people are now starting to take their internal knowledge from storing them to wikis.

我们内部有一些非常酷的案例，你马上会想到的是，如果有人内部知识库里的一个文档不小心权限设置错了，他们意识到‘哦不，我不想让这些信息被访问到’，那他们怎么修复呢？通常他们会进入文档，修改权限，但现在有了这些衍生制品。所以你需要确保有某种方式在系统中追踪，说‘这个输出文档来自那个源文档。源文档不再可访问了，所以我们也去让它失效。’所以你必须开始真正构建你的技术架构，要意识到人们将如何使用这些信息，这确实改变了团队之间的协作方式，因为你可以直接……它确实改变了瓶颈在哪里以及什么困难。

We have some really cool ones of these internally, and the thing you immediately think about is, if someone has a document in the internal knowledge base that was accidentally permissioned incorrectly, and they realize, 'Oh no, I didn't want this information to be accessible,' how do they fix that? Normally they go into the doc and change the permissions, but now there are these derived artifacts. So you need to make sure you have some way of tracking through the system to say, 'This output document came from the source one. The source one's no longer accessible, so let's go and invalidate that as well.' So you have to start really building your technical architecture with awareness of the way that people are going to use this information, and it really changes how teams relate to each other because you can just... It really changes where the bottlenecks are and what's hard.

你认为团队规模会小很多吗？十年后我们还会有人类软件工程师吗？

Do you think team sizes are going to be a lot smaller? Are we going to still have human software engineers in a decade?

嗯，十年从现在来看是很长的时间，这项技术的上限真的很难内化。我认为很明显，公司的形态会在很多方面发生变化。我认为我们将有能力让个人创业者建立非常了不起的企业。所以任何有愿景的人，我认为都能实现它。我认为你们的工作在很多方面会变得更容易、更有趣。不过，也可能竞争更激烈，对吧？因为每个人都会拥有这些神奇的工具。所以真正弄清楚你的利基是什么、你独特的角度是什么，可能会成为最重要的核心。但目前我们运营组织的方式，几乎只有一种方式组织大量人群，那就是有团队、管理结构、范围和层级等等。也许这可以改变。也许你可以更扁平化，小团队就能做出不可思议的事情。比如，我们现在在数学领域就看到，互联网上的个人正在使用 GPT-4 Pro 解决未解的数学问题。通常需要一个数学团队，而他们一个人就做到了。

Well, a decade is a long time from now, and the ceiling on this technology is really hard to internalize. I think it is clear that what a company is will change in a lot of ways. I think we're going to have this ability for solopreneurs to build very incredible businesses. So anyone who has a vision, I think will be able to realize it. I think the jobs that you all have will become way easier in a lot of ways, way more fun. Now, might be more competitive, too, right? Because everyone's going to have these amazing tools. So really figuring out what is your niche, what is your unique angle, is probably going to become the most important core. But a lot of how we run organizations right now, there's almost only one way to organize large groups of people, where you have teams, management structures, scopes, and hierarchies, and all these things. Maybe that can change. Maybe you can be much more flat, small teams that can really just do incredible things. Like, we're seeing it right now in mathematics, where these individuals on the internet are using GPT-4 Pro to solve unsolved math problems. Normally you need a math team, and they're just doing it.

我儿子是个数学迷。我刚刚告诉他，也许我们应该学点数学以外的东西。但你看，这就是问题所在，对吧？如果你看看像 AlphaGo 这样的东西，第 37 手，那一手改变了人类对围棋的理解，但令人惊讶的是，它让围棋对人类来说变得更有趣、更重要。也许在其他领域也会是这样。

My son's a math nerd. I just told him that maybe we should be studying something else besides math. But well, see, this is the question, right? If you look at something like AlphaGo, move 37, that move that just changed humanity's understanding of the game, but the thing that was surprising is it made the game more interesting and important for humans. And maybe that will be true for these other domains, too.

没错。

True.

那在构建生产级智能体工作流时，常见的失败模式是什么？你认为创始人最近在构建时经常出错的地方是什么？

What about common failure modes when you're building with production agentic workflows? What do you see as the common things that founders get wrong and they're building incorrectly these days?

嗯，我认为这些模型有如此强大的能力，真正理解如何良好地操作它们需要思考。所以我们一直在投资原语、安全原语、可观测性、良好的治理等等。但给你讲一个我觉得很有启发的小故事：我让我的同事安装一个 OpenAI 某人写的包，遇到了错误，我说，‘哦，在 Slack 上 ping 那个人，请他们帮忙。’于是 ping 了那个人。两分钟后，它说，‘这太久了。我已经升级到那个人的经理。’它真的 ping 了那个人的经理。你意识到，一方面，模型这样做是合理的。它很主动，试图解决我的问题。它不是坐等指令。但另一方面，也许它应该多等一会儿，也许应该先问问我。所以，我认为真正思考这些问题，我们还在构建模型的情商。在某些方面，它已经变得非常好。例如，一直点击‘批准、批准、批准’是我们过去的方式。人类在这方面也不擅长，对吧？他们只是默认操作。所以现在我们开始有 AI 能够真正处理标记：这是高风险操作吗？嘿，这个应该升级，这个可以自动批准。这让你意识到，人类的注意力将成为一种极其稀缺的资源，对吧？做事现在很容易。‘这是好事吗？这是我想要的吗？这符合我的价值观和愿望吗？’这将成为唯一最重要的瓶颈。所以，我认为构建考虑到这一点并真正思考人为因素的系统，是现在最重要的事情。

Well, I think that these models have such power, and really understanding how to operate them well takes thought. So we've been investing a lot in primitives, security primitives, observability, having good governance, things like that. But just to give you one anecdote that I think is evocative: I asked my co-workers to install some package that someone at OpenAI had written, ran into an error, I was like, 'Oh, ping that person on Slack and ask them for help.' So, ping the person on Slack. Two minutes later, it said, 'This is taking too long. I've escalated to the person's manager.' And it actually pinged the person's manager. And you realize it's like, on the one hand, it's kind of a reasonable thing for the model to do. It's being proactive, it's trying to solve my problem. It's not just sitting around waiting to be told what to do. But, on the other hand, maybe it should have taken a little bit longer, maybe should have checked with me. So, I think that really thinking about these questions where we're still building up the EQ of the model. And that in some places, it's getting very good. For example, clicking 'approve, approve, approve' is kind of where we've been. And humans are not very good at that, either, right? They just default. And so, now we're starting to have AIs that can actually take care of flagging: is this a high-risk action? Hey, this one should be escalated, this one's okay to auto-approve. And it really makes you realize that human attention is going to be this incredibly scarce resource, right? The doing of things now is easy. The 'Is this a good thing? Is this what I wanted? Is this aligned with my values, with my desires?' That is going to become the single most important bottleneck. So, I think building systems that take that into account and really think about the human factor, that's the most important thing to do now.

另一个人为因素：安全。你会建议人们在这个 AI 时代如何看待安全？我最近听到 Vercel 等公司频频出现安全漏洞……而且这些模型在发现安全漏洞方面非常强大。那么，你会建议这里的人们如何使用这些模型来发现安全问题？

Another human factor: security. How would you advise people to think about security in this world of AI? I've heard about breaches left and right with Vercel recently and then... And these models are incredibly powerful at finding security holes. So, how would you recommend people here use the models to find those security issues?

嗯，我认为答案有几个层面。我确实认为互联网一直以来安全都是一个越来越重要的问题。想想它从哪里开始，经历了 90 年代的病毒、蠕虫和恶意软件等等，我们已经走过了那个阶段。我认为我们现在也在走向一个更终极安全的体制，但这确实需要整个互联网的努力才能实现。所以很多方面实际上就是再次利用这项技术，让这些模型能够扫描你的代码库，它们实际上可以用于端到端的红队测试。它们可以做的事情很多，我们如何思考进一步的模型和改进，很大程度上在于如何利用可信访问计划，如何利用那些真正关心成为捍卫者、让互联网更安全的人群。

Well, I think there's a couple levels to the answer. I do think that the internet has been a place where security has been a ratcheting important concern over time. You think about where it started, going through the '90s with viruses and worms and malware and those things, and we've moved past that. I think we are also moving now to a much more ultimately secure regime, but it does require kind of an internet-wide effort to get there. So a lot of this honestly is just again leaning into the technology, having these models they can scan your code base, they can actually be used for end-to-end red teaming. There's a lot that can be done with them, and a lot of how we're thinking about further models and improvements there is really leaning into how we can leverage trusted access programs, how we leverage the community of people who really care about being defenders and making the internet more secure.

我认为这是每个人都扮演着角色、可以参与的事情，但首要的是要认识到这些模型非常强大，但它们不是魔法，对吧？它们只是整体韧性生态系统的一部分。我认为我们作为社会，以及每家公司，实际上都为此做出贡献，需要构建一些东西，比如如何以带来更多保障和确定性的方式整合这些模型，无论是你正在采取的特定补丁，还是思考如何确保快速滚动更新。所以我认为还有很多工作要做，但我对此非常乐观。让我们切换到速度话题。在加速变化的世界里，事情似乎越来越快。我们之前在你走过来时聊过，你如何跟上节奏。你是如何跟上所有加速变化的？你会建议这里的每个人如何跟上所有变化？

I think that's something where everyone has a role to play and can participate, but the number one thing is just sort of recognizing that these models are very powerful, but they're not magic, right? That they are just like a part of the overall resilience ecosystem and I think that we as a society and I think every company again really contributes to this have something to build in terms of how do we how do we incorporate these in a way that results in more assurance and more sort of certainty on the impacts of whether it's a particular patch that you're taking, whether it's thinking about how do you make sure that you're um yeah, just sort of rolling in updates quickly as they're being released. Um so I think that there's a lot of work to be done, but I have a lot of optimism for where this is going. Um let's switch to speed. Seems like things are moving faster and faster and faster when in the world of accelerating change. We were talking about it when we when you uh you were walking up here around how how you're trying to keep up with things. How how do you you keep up with all the accelerating change? How would you recommend everybody here keep up with everything that's changing?

嗯，我认为这是新常态，某种程度上并不完全是 AI 的原因。我认为这只是过去二十年技术的趋势。有更多人在做事情，做事比以往任何时候都容易。准入门槛降低，意味着更容易创造价值，取得巨大成功。所以，我认为真正要做的就是保持敏锐，了解正在发生的变化。某种程度上，它总是从同一件事开始：亲自使用技术。听别人描述 AI 和亲自使用它，感觉非常不同。但 AI 的美妙之处在于它非常直观。这正是关键所在：机器不再需要你扭曲自己去适应它，而是机器适应你，为你工作。它应该是一个你提问它就能做事的工具。所以，我认为真正去把握变化的脉搏、了解什么是可能的、模型在哪些方面滞后，这将是决定未来公司成功与否的核心技能。

Well, I think this is the new normal and I think to some extent it's not really because of AI. I think it's just been the trend of technology for the past two decades. There's more people doing things. It's easier to do things than ever. Barrier to entry goes down. Means it's also much more easy to build value, right? To have great successes. And so, I think that really trying to keep your ear to the ground and understand what's changing. And to some extent, it always starts with the same thing, which is play with the technology yourself. Like, it's very different to hear AI described versus to use it. But, the beautiful thing about AI is it's so intuitive. Like, that's the whole point is that rather than have the machine be something you have to contort yourself to, the machine contorts itself to you, right? It's doing work for you. And it should be something where you ask it and does something. And so, I think that just really trying to just get your finger on the pulse of what's changing, what's possible, where the models lag. That is, I think, the core skill that is going to really determine a lot of the success of of companies in the future.

另一方面，你们推迟了模型发布以便与安全机构合作。这就像是尽可能快的反面。所以，你们也在负责任地做事。那么，你如何看待这种平衡？因为你们处于竞争环境中，想要尽快发布，但同时也在努力做正确的事情。

And then, on the flip side of that, you guys have held up held back models to work with security agents. So, it's like the opposite of like going as fast as possible. So, um you're doing things responsibly, too. So, how do you kind of like think about the balance? Because you're in a competitive environment, you want to ship as quickly as possible, and yet you're trying to do the right thing, as well.

是的，我认为在价值观层面，OpenAI 的宗旨是真正将 AI 的力量交到人们手中。我们相信人们能够利用我们创造的工具来建设未来。但我们需要以深思熟虑的方式做到这一点，对吧？我们真正考虑了两方面：好处和风险。如何最大化好处？如何减轻风险？我认为在网络安全和生物安全领域，我们非常谨慎。我们已经在构建缓解措施和可信访问项目上工作了相当长的时间。我们看到未来模型将越来越强大，能力持续提升。上周我们宣布了网络安全可信访问项目的扩展。顺便问一下，这里有人申请了吗？没人？哦，我看到一只手，两只手。好吧，你们应该更多人申请。这很棒。我们真的需要帮助，因为让值得信赖、负责任并真正想推动这些模型的人参与进来非常重要，这样才能为每个人带来回报。未来几周我们还会宣布更多关于如何扩展该计划的消息，以及当我们向所有人发布模型时，我们将采取哪些缓解措施，并调整这些措施以实现真正的平衡，尽可能广泛地提供这些能力，同时确保我们考虑到风险，并能够对它们进行一定程度的观察，确保部署产生最大正面影响。所以，简短的回答是，这关乎我们的使命核心。我们非常关心我们所做事情的影响，不仅仅是孤立地构建技术，而是整个社区、整个世界的共同努力，才能达到我们需要达到的目标。

Yeah, I think at a values level, like what OpenAI is about, like we really want to put the power of AI in people's hands. Like, we believe that people can We want to empower people to build the future with the tools that are being created. But, we need to do that in a thoughtful way, right? That we really think about both sides of Here are the benefits, here are the risks. How do you maximize the benefits? How do you mitigate those risks? And I think that in cybersecurity and in biosecurity, those are areas where we're very thoughtful. We've been building We've been working on these kinds of both mitigations and trusted access programs for quite a long time. And that what we see coming is models that are going to be increasingly powerful and capable in a continuous way across all dimensions of capability. And the you know, we announced last week the expansion of our trusted access for cyber program. By the way, has anyone here applied? No one. Oh, I see one hand, two hands. Okay, more of you should apply. It's great. We really need help because and it's very important that people who are trustworthy and responsible and really want to push these models are participating in this because that is how that's going to pay dividends for everyone. We're going to have more to announce over upcoming weeks on how we're expanding the program, but and also when we release models to everyone kind of the mitigations that that we have and how we're going to tune those to be both to really balance, right? To really try to bring these capabilities as broadly as possible, while also making sure that the ones that are, you know, that we're thinking about the risks and and able to have some observability over them and to ensure that this is maximally positive in terms of deployment. So, I think the short answer is like it's core to our mission. We care a lot about the impacts of what we're doing, not just building technology in isolation, but it is a whole community, a whole world effort to really get to where we need to be.

从模型层面上升到应用层，这是很多在场的人正在构建的。OpenAI 如何决定在应用层构建什么、不构建什么？

On moving up from the models to the application layer, which is what a lot of people here are building. How do you How does OpenAI decide what in the application layer you're you're going to build and what you're going to leave out?

嗯，大家可能最近经常看到“专注”这个词被用在 OpenAI 身上，可能一段时间以来第一次。这很难，因为 AI 领域充满机遇，对吧？你能想象到的任何东西都会很棒，毫无疑问。而我们作为一家公司，无论我们构建多少算力，拥有多少人，也只能做这么多。所以，我们一直在思考的是，什么是最专注的战略，能够覆盖空间中的哪些部分？也许是 80/20 法则，或者只是我们认为能产生最大影响的部分。我认为现在很清楚，我们正在经历身份转型。所以产品不仅仅是企业级与消费级之分，对吧？很明显，我们非常重视企业级市场，我们正在向大公司销售，并建立整个销售团队和销售流程。但消费级市场，消费级的定义会改变，对吧？它是一个非常宽泛的术语，包含多种东西。但消费级中有一部分不仅仅是关于生产力，而是关于目标、实现目标，甚至知道你的目标是什么，能够引出这些目标，并拥有一个可以主动做到这一点的 AI。这基本上都是一回事。最终，我们试图构建一个你可以与之交谈的 AGI，它拥有所有上下文，可以在你的个人生活、工作中使用，值得信赖，对吧？你可以向它寻求建议，它能提供有用的信息，也许是健康信息，或者财务信息，或者当你试图规划职业生涯时。

Well, people have probably seen the word focus being applied to OpenAI quite a lot recently, possibly for the first time in a while. And it's hard because the field of AI is one of opportunity, right? It's like anything you're going you can imagine is going to be great. No question it's going to be great. And we as a company, as a single company, no matter how much compute we build, no matter how many people we have, are only going to be able to do so much. And so, a lot of where we've been how we've been thinking about things is what is the sort of most focused strategy that covers the parts of the space? You know, maybe it's an 80/20 or just like the parts of the space that we think we can have most impact on. And I think there it's very clear right now we're going through this identity transition. And so products that are And it's not just about enterprise versus consumer, right? So it's like clear we are being very serious about enterprise. Like we're selling to to big companies and and building a whole muscle and sales motion there. But consumer, what consumer is is going to change, right? It's kind of a very broad term that buckets in multiple things. But that the slice of consumer that's about not just productivity, but about goals, about achieving your goal, about even knowing what is your goal, being able to elicit that and having an AI that can proactively do that. It's all kind of the same thing. Like in the end we're trying to build an AGI that you can talk to, that has all this context that you can use in your personal life, your work life, that's trustworthy, right? That you can go to it for advice, can give you useful information, maybe health information, or about finances, or you know, about if you're trying to figure out what to do with your career.

所有这些事情，它们都像梯子一样通向同一个目标。这意味着我们不得不做出一些非常痛苦的决定，关于什么不该做。但我想说的是，这就是我们看待事物的视角，那些有助于实现我们想要构建的单一愿景的事情，你可以期待我们去追求。

Like all these things, they all kind of ladder into one thing. And it's meant we had to make some very painful decisions about what not to do. But I think I would just say that that's the aperture that we look at things through and that things that accrue to that singular vision of what we want to build, you should expect us to pursue.

明白了。你认为几年后我们还会用命令行和智能体编程吗，还是说会彻底改变？

Got it. Do you think we'll be coding with command lines and agents in a few years or is it going to be completely changed?

我的意思是，我认为我们目前的工作方式非常不自然。我们都坐在这个盒子后面，不停地打字，很明显我们的身体不是为此设计的。我们得了腕管综合征，肩膀佝偻，等等。我不认为我们想要那样。我觉得我们想要更多的自由时间。我们想要更多时间，但也不仅仅是自由时间，对吧？比如你想花更多时间与亲人在一起？是的。你想花更多时间与人交谈，想出 brilliant 的愿景，或者只是做你兴奋的事情，或者了解自己。所以，这有点像你想不想成为一个拥有 10 万个智能体的组织的 CEO？那听起来确实不错。我认为我们都能完成更多事情，但其中的机制会感觉完全不同，就像从用羽毛笔手写东西到能够发送一条短信，然后让别人代表你朝着你的目标努力。

I mean, I think that we're in a very unnatural state right now for how we work. Like we all sit behind this box and kind of type away and it's very clear our bodies were not designed for this. We got our carpal tunnel and our hunched shoulders and all these things. And I don't think we want that. I don't think any of us wanted that. Like I think that we want more free time. We want more time, but it's not even about free time necessarily, right? It's like you want to spend more time with your loved ones? Yes. You want to spend more time like talking to people and like coming up with like brilliant visions or just like what you're excited about or just understanding yourself. So, it's kind of like do you want to be a CEO of an organization of like 100,000 agents? Like that actually seems pretty good. And I think that we're all going to be able to get so much more done, but the mechanics of it are going to feel as different as like going from having to write out things with a quill to being able to just send a text message and have people go and work on your behalf on your goals.

好的。我们讨论了算力、模型、安全、智能体和应用层。我们来谈谈前沿。模型什么时候才能足够好，推动科学和物理 AI 的前沿？好像我们请过 Jen Fan。看起来 LLM 是数字智能的一个很好的缩放定律。但在机器人、物理智能、生物学和科学的某些方面，它就没那么强了，因为这些问题可能更难验证，或者需要很长时间来验证。那么，你是如何跟踪世界上的科学和物理 AI 的呢？

All right. We talked about compute. We talked about model and security and agents and app layer. Let's talk about frontier. When are the models going to be good enough to push the frontiers of science, physical AI? Seems like we had Jen Fan here. It seems like LLMs have been a great scaling law for digital intelligence. It hasn't been as strong for robotics, for physical intelligence, for aspects of biology and science where the problems are probably a lot harder to verify or takes a long time to verify. Well, how are you keeping track of science and physical AI in the world?

科学是我们真正投入的一个领域，我们看到了取得惊人进展的前景。我们开始看到一些生命的迹象，我认为在预测六个月或一年后会发生什么时，始终要立足于今天正在发生的事情。例如，我们有一个物理学成果，我们的 AI 提出了一个非常优美的公式，而研究这个问题很长时间的物理学家认为这完全不可能。他们认为这也许是一个无法解决的问题。这相当重要，对吧？真正的严肃物理学家认为这是朝着能够解决量子引力等问题迈出的一步。虽然还没到那一步，但这是一个进步。这比几个月前我们取得的进展要大得多。所以这让你不禁想知道，一年后我们会走多远？现在，像生物学这样的领域与物理和数学不同，对吧？你必须离开你美丽的模拟世界，处理混乱的现实。但我认为我们已经在其他领域学会了如何处理混乱的现实。软件工程就是一个完美的例子，我们意识到仅仅构建解决编程竞赛的东西是不够的。你需要的是见过真实世界混乱代码库的东西，人类以不同方式打断它，像这种对抗性的敲打。所以我认为在科学领域，我们将会看到一场真正的复兴。也许今年我们会看到一些重大成果。明年我认为将是一个完全疯狂的时代。我们生活在有趣的时代。

Well, science is one domain that we're really leaning into and we see line of sight to really incredible progress. And we're starting to have some signs of life and I think it's always important to ground in what is happening today when trying to predict what will happen 6 months a year from now. So, for example, we had a physics result where our AI came up with this very beautiful formula that physicists who've been working on this for quite some time thought was totally impossible. Thought it was like maybe an unsolvable problem. And like it's pretty significant, right? It's like real serious physicists who really view this as a step towards really being able to get to some sort of answer for quantum gravity and all these things. It's not there, but it's a step. That's much bigger than where we were just a couple months ago. And so it makes you really wonder a year from now, like how far will we have traveled? Now, things like biology that they are different from physics and math, right? That they are you got to leave your beautiful simulated world and deal with messy reality. But I think we've been learning how to deal with messy reality in other domains. Software engineering is a perfect example where we've really realized that just building the thing that solves programming competitions, like that's not enough. Like you need something that's seen real-world messy codebases, humans interrupting it different ways, like this adversarial banging at it. And I so I think that that on science, I expect we're going to see a real renaissance. You know, maybe we'll see some big results this year. Next year I think it's going to be a totally wild wild time. We live in interesting times.

我答应过准时放你走，因为你是个大忙人。在你离开之前，我们还有一分钟时间。既然你现在没时间，但很快你会有很多时间，你和 Anna 平时有什么娱乐活动？

I promised that I get you out on time because you're a busy man. Before we let you leave, we got 1 minute on the shot clock. What since you have no time, but soon you will have lots of time, what do you and Anna do for fun?

娱乐？和普通人一样。喜欢看电影、徒步旅行之类的。你知道，现在没那么多时间，希望 AGI 之后会有。但沿途也得享受过程。

Fun? I mean, same as anyone. Like to watch movies, go on hikes, those kinds of things. You know, not as much time for it as maybe we'll hopefully have post-AGI. But you got to kind of enjoy the ride along the way.

感谢 Greg 参加我们的节目。

Thank you, Greg, for joining us.

谢谢大家。

Thank you, everyone.