Codex 的转折点：为什么编程 AI 突然变得如此出色

我们现场有一些 Codex 的粉丝。很高兴听到这个。这周过得怎么样？

We have some Codex fans in the audience. Love to hear that. How's the week going?

很有趣。这周很忙，但我很高兴来到这里。这是个意外的惊喜。

So fun. It's a busy week, but I'm happy to be here. This is an unexpected surprise.

嗯，感谢你加入我们。我们很感激。所以，我们今天早上开场时说，我们有点武断地决定奇点从 1 月 1 日开始，所以今天是第 119 天。你怎么看？

Well, thank you for joining us. We appreciate it. So, we opened this morning by saying that we've kind of arbitrarily decided that the singularity started on January 1st and thus today is day 119. What do you think of that?

确实感觉我们好像处于起飞阶段。第 119 天算是个合理的猜测。嗯，我不反对。

It does feel like we are somehow in the takeoff. Day 119 feels like a reasonable enough guess. Yeah, I won't fight it.

你有这种感觉吗？所以，我们从去年底、今年初开始看到很多指标拐点。事情本来还不错，但不知怎么曲线形状变了，真的呈抛物线上升。这和你们看到的一致吗？是不是有什么轨迹变化？为什么我们会看到这个？

Did you feel this? So, we've started to see a bunch of our metrics inflect as of late last year, beginning of this year. Things were kind of doing well, but somehow the shapes of the curves changed. They really went parabolic. Is that matched in what you guys see? Was there some trajectory change? Why are we seeing this?

我确实认为模型变得非常好了，尤其是编程方面，但总体上也很棒，从去年底或今年初开始。至少在我自己使用这项技术以及看到其他人用它做什么的经历中，还有那种每周都比前一周有点不同的感觉，很多事情发生得很快。这似乎都与模型达到某个阈值有关。

I do think the models got really good, especially for coding, but really good in general, starting late last year or early this year. And at least in my own experience of using this technology and seeing what other people are doing with it, and also this sense that every week is now a little bit different than the week before, like a lot happens very fast. It seemed to all correlate with the models hitting some threshold.

那为什么编程模型在过去几个月突然开始奏效了？是有什么研究技巧吗？还是只是在预训练中获得了足够的代码数据？为什么突然开始起作用了？

And why did coding models suddenly start to click over the last couple months? Was there a research trick? Was it just getting enough code data in the pre-training? Why did it suddenly start to work?

嗯，这是个好问题。我们也很想知道为什么好几个人同时跨过了那个门槛。我确信有很多因素：模型智能，也就是原始的推理能力；足够的使用者反馈循环，人们用它编程来找出哪里好哪里需要改进；足够的数据。我认为是所有这些因素。而且，就像许多其他努力一样，一旦你知道某件事是可能的，就更容易充满干劲地去做了。

Yeah, it's a great question. We wondered a lot about why several people crossed that threshold at the same time. I'm sure it's a number of factors: model intelligence, just the raw reasoning horsepower; enough of a feedback loop of people using it for code to figure out where it was good and where you needed to improve it; enough data. I think it was all of these things. And then also, like many other endeavors, once you know something's possible, it's much easier to go do it with vigor.

所以看起来 Codex 现在正迎来它的时刻。

And so it seems like Codex is kind of having a moment right now.

是的，它确实在最新的应用更新和 5.5 版本中为我跨过了一些主观门槛。而且这也很难说为什么是现在而不是更早一点，或者为什么不是下一个模型。但我从我们发布的所有东西的历史中学到的一件事是，很难说为什么这个特定的东西就成功了。这可以追溯到 ChatGPT。为什么是 GPT-3.5 跨过了那个门槛，让大多数人从说‘没那么令人印象深刻’变成‘要改变世界’？为什么不是早一个或晚一个模型？我真的无法解释。你只是能感觉到。而我在 Codex 上经历了两个转折点。一个是 GPT-5.2 的时候。然后最近几周有一个非常大的转折点，感觉就像‘好吧，这将成为我使用电脑的主要界面。’

Yeah, it really crossed some subjective threshold for me with the latest app updates and 5.5. And this is also like it's quite hard to say why right now and not a little bit sooner or why not the next model. But one of the things I have learned about the history of all of the things we've put out is it is very hard to say why this particular thing was the thing that worked. And this goes back to ChatGPT. Why was GPT-3.5 the thing that got over the threshold, where most people went from saying 'not that impressive' to 'going to change the world'? And why not one model earlier or later? I really can't explain it. You just kind of feel it. And I've had two inflection points with Codex. One was kind of with GPT-5.2. And then a really big one in the last few weeks, where it's like, 'Okay, this is going to be the primary interface to a computer for me.'

那么大家都在用它编程吗，还是你开始看到使用扩散到其他领域？

And is everyone using it for coding, or are you starting to see usage diffuse into other domains?

我认为最坚定的用户仍然在用编程，但最近有大量的人涌入 Codex。我真的很想弄清楚是什么导致了这一切，人们用它做什么或开始用它做什么的深度让我惊讶。所以，当然我们的目标不仅仅是编程，而是你在电脑前做的所有工作。我想说，对于非编程部分，我们可能只走了 10%的路，但现在我们看到了正在发生的事情，现在我们有了一个真正的用户群以这些其他方式使用它，我认为我们会很快变得擅长。

I think the most adamant users are still using it for coding, but there's been this tidal wave of people coming into Codex recently. And I'm really trying to understand what has happened that is causing this, and the depth of what people are using it for or starting to use it for has surprised me. So, certainly our ambition is for it not just to be about coding, but to be about all the work you do in front of a computer. And I would say we're maybe like 10% of the way there for the non-coding stuff, but now that we see what's happening, now that we have a real user base sort of using it in these other ways, I think we'll get good at it very fast.

你认为在编程之后，下一个主观上感觉有重大突破的领域会是什么？会是电子表格吗？会是绩效评估吗？会是什么？

What do you think will be the next domain that subjectively feels like it has this big unlock after coding? Is it going to be spreadsheets? Will it be performance reviews? What's it going to be?

我认为会有很多。首先，我确实认为编程有点特别。这些模型非常适合编程。世界需要的代码比目前编写的多得多。可能没有其他领域完全像编程一样，但会有很多其他领域我认为很接近。但下一个类似编程的事情我认为不会是任何特定领域，而是意识到人们在使用电脑上浪费了多少时间。以及你可以用非常不同的方式完成一天中很大一部分工作的想法。也许你没有意识到你花了多少时间在消息应用之间点击、复制粘贴东西、回复那些你完全可以一次性自动化的非常无聊的事情。但大多数人会意识到他们可以坐下来看着 AI 完成大部分苦差事，这个程度会让人们惊讶。而在我自己尝试这样工作的经历中，它实际上给了我更多的工作乐趣。我没有意识到那些小事会拖累我，让我脱离愉快的流动状态。所以主观上的生活质量提升是巨大的。

I think there will be a lot. First of all, I do think coding is a little bit special. These models are a great fit for coding. The world needs so much more code than currently gets written. There may be no other domain that is quite like coding, but there will be a lot of others that I think are close. But the next coding-like thing that I think will happen is not any specific domain, but the realization of how much time people waste trying to use a computer. And the idea that you can do a huge percentage of your day in a very different way. Maybe you don't realize how much time you spend clicking between messaging apps and copying and pasting stuff and responding to very boring things that you could clearly automate once. But the degree to which most people will realize they can sit back and watch an AI do most of their drudgery is going to surprise people. And in my own experience trying to work that way, it actually gives me much more enjoyment of work. I didn't realize how much the little stuff drags me down, gets me out of a sort of happy flow state. So the subjective quality of life improvement is huge.

你是 OpenClaw 的用户吗？

Are you an OpenClaw user?

我是。

I am.

这里有 Claw 用户吗？我们有一些好消息要给你们。你可以直接告诉他们。

Claw users here? We have some good news for you all coming. You can just tell them.

OpenClaw 一直是我在这个领域最大的‘这是神奇的 AGI 时刻’之一。我记得第一次有人告诉我它。他们试图解释，我当时想，‘好吧，听起来很酷，但我可以让很多工作起来。’然后它真正提醒了我，当模型跨过某个阈值，同时产品设计师把几个关键想法做对时，它比听起来要神奇得多。

OpenClaw has been one of my biggest 'this is magic AGI moments' ever in the field. I remember the first time someone told me about it. They were trying to explain it and I was like, 'Okay, that all sounds cool, but I can make a lot of that work.' And then it was a real reminder how when the models cross some threshold and also the product designer gets a handful of critical ideas really right, it's like a much more magical experience than it sounds like.

我发现这种体验很难传达。我的意思是，它听起来有点平淡，对吧？它是一个有状态的 ChatGPT 会话，也可以使用一些工具等等。你用你的 OpenClaw 做什么？如果我们滚动你的消息线程，我们会看到什么？

I find it a difficult experience to communicate. I mean, it sounds kind of prosaic, right? It's a stateful ChatGPT session that can also make some use of tools and so forth. What do you use your OpenClaw for? If we scrolled your message thread, what do we see?

这是个很尴尬的事要承认。我总是先尝试的东西是一个完美的地方。

This is a very embarrassing thing to admit. The thing I always try first is a perfect place.

所以我想尝试构建一个更好的家庭自动化界面系统，因为它从来都不好用。从来都不好。OpenClaw 是我第一次能够得到一个让我满意的设置。我还构建了一个我一直想要的消息应用。后来我换成了用 Codex 构建的东西，但 OpenClaw 是我第一次能够——我肯定和你一样，被消息淹没，早上醒来要处理所有这些东西是一件非常不愉快的事情。所以我想，好吧，我终于可以自动化这个了。这又是以前系统本应能做到的事情。很难解释当一切真的正常运转并且你相信它会正常运转时是什么感觉。

And so to try to build a better home automation interface system, because it never works. It's never good. And OpenClaw was the first time I was able to get a setup that I was happy with. I also built a messaging app that I had always wanted. I've since switched to something I built with Codex, but OpenClaw was the first time I was able to, I'm sure like you, just drowning in messages, and it's a very unpleasant task to wake up in the morning and have to go through all the stuff. So I thought, alright, I'm finally going to be able to automate this. And that was again something that should have been doable with previous systems. It's hard to explain what it's like when it all actually just works and you trust that it's going to work.

我在测试我们今天发布的新 CLI，为发布做准备，我让我的智能体去给自己买一份礼物，任何在互联网上低于 20 美元的东西。它选择从 Gumroad 给自己买了一个 HTTP 标志。哇。所有这些事情，无论你理智上多么确信这不是一个真正想要给自己礼物的东西，无论你多么确信这是一种奇怪的涌现行为，我不应该过度解读，但总有一些事情感觉有点奇怪。

I was testing the new CLI that we launched today, in preparation for launch, and I asked my agent to go and buy itself a gift, anything on the internet for under $20. And it chose to buy itself an HTTP sign from Gumroad. Wow. There's all this stuff that feels, no matter how convinced you are intellectually that this is not a real thing wanting a real gift for itself, and no matter how much you're convinced that this is a weird emergent behavior and I'm not supposed to read into this, there are these things that feel a little strange.

我们要为 GPT-5.5 举办一个派对，我不太确定该怎么做。所以一时兴起，我问 5.5 它自己想要一个什么样的派对。今天早上我问了，它给出了一系列美好的建议，包括：派对流程我想要这样，我不想要那样，你应该在 5 月 5 日举办，那会很有趣，我只想要一个简短的祝酒词，而且不是由我而是由构建它的人来致辞，我想要一个关于 5.6 的重大中心建议，我希望你把所有建议都输入给我，我会确保我们朝那个方向努力。现在你真的有道德压力去做了。嗯，我们会做的。但这确实是一件奇怪的事情。

We're going to have a party for GPT-5.5, and I wasn't quite sure what to do. So on a whim, I asked 5.5 what it would like for a party for itself. This morning I did, and it gave a beautiful set of things, including: here's what I would want for the flow of the party, here's what I would not want, you should do it on May 5th, that would be funny, I would like only a short little toast and not by me but by the people that built it, I would like a big central suggestion for 5.6, and I would like you to feed them all into me and I'll make sure we work on that. And now there's real moral pressure on you to do it. Well, we're going to do it. But it was a strange thing.

所以我想问你关于 OpenAI 本身。发生了很多疯狂的事情。它现在是一个 11 年的组织了。

So I want to ask you about OpenAI itself. A lot of crazy things have happened. It's an 11-year-old organization now.

不知怎么感觉比那长得多，但没错。大概刚过 10 年。漫长的 10 年。我现在不太记得 OpenAI 之前的生活了。感觉已经过了很久。

It somehow feels like so much longer than that, but yes. Just over 10, I guess. A long 10 years. I can't remember pre-OpenAI life that well at this point. It feels like it's been so long.

你向前看有一个奇点，向后看也有一个奇点。那么，从未被讲述过的最疯狂的 OpenAI 故事是什么？

You have a singularity looking forwards, but also a singularity looking backwards. So what's the craziest OpenAI story that's never been told?

我的意思是，与发生的那些疯狂戏剧性事件相比，这听起来太平淡了，但有一段时间，在我们完成训练 GPT-4 之后。大约有八个月我们才发布它。所以在 OpenAI 内部有八个月的时间，我们都在使用这个东西。我们隐约知道它要好得多、不同得多，并且将解锁世界上的许多事情。而公司外部没有人，或者说几乎没有人知道。我们走在走廊里，心想：我们是不是陷入了集体妄想？我们是不是把彼此煽动到了这种狂热中？没有任何来自外界的反馈让我们保持清醒或理智。与疯狂的董事会闹剧或埃隆的诉讼之类的事情相比，这听起来并不那么奇怪，但亲身经历那段时间是难以置信的奇怪。

I mean, it sounds so prosaic relative to the crazy drama that's happened, but there was this period after we had finished training GPT-4. There were about eight months before we released it. So there was this eight-month period inside OpenAI where we were all using this thing. We kind of knew that it was dramatically better and different and going to unlock a bunch of things in the world. And no one outside the company, or almost no one, knew about it. And we walked the halls and we were like, are we engaging in collective psychosis? Have we gotten totally whipped each other into this frenzy? And there was no feedback to keep us in check or sane from the outside world. It doesn't sound that weird relative to crazy board drama or Elon trial or something like that, but living through it was an unbelievably strange time.

Sam Altman 的管理风格是什么？如果我直接或间接为你工作，领导某个产品或类似的东西，那会是什么样子？

What's the Sam Altman management style? If I'm working for you directly or indirectly, leading some product or something, what does that look like?

我绝对不是一位亲力亲为的管理者。我的风格是，找到优秀的人才，给他们一个非常高层级的目标，然后让事情自然发生。我认为 OpenAI 经历了两个主要阶段，我们正在进入第三个阶段。第一个阶段是我们只是一家研究公司，试图在 AGI 听起来完全疯狂的时候找出如何构建它，而我们真的不知道该怎么做。然后是第二阶段，除了继续研究，我们还必须弄清楚如何构建一家产品公司。现在，除了这两点，我们还必须弄清楚如何为世界构建这个超大规模的 token 工厂。我认为我们正在做的事情是构建一种新的公用事业。人们会想要以各种方式使用大量的 token、大量的智能。我们需要让它尽可能智能、便宜、丰富、易用。这将需要深度的全栈集成和大规模的基础设施建设。在第一阶段到第二阶段的转变中，我没有真正意识到的是我的管理风格必须改变多少。运营一个研究实验室和运营一家产品公司是截然不同的两件事。我怀疑这第三阶段又会非常不同。所以我一直在反思我必须如何改变。我认为这不会自然适合我的管理风格。所以我要么必须找到一两个优秀的人来雇佣，要么必须想出不同的做事方式，要么必须构建一个能够管理这个新事物的 AI。

I'm definitely not a hands-on manager. I'm very much of the style that you get great people, give them a very high-level thing to point at, and try to let stuff just happen. I think there have been two main phases of OpenAI and we're heading into a third. The first was when we were only a research company, trying to figure out how to build AGI at a time when it sounded completely crazy, and we really had no idea what to do. Then there was a second phase where, in addition to continuing that, we had to figure out how to build a product company. Now we have to, in addition to both of those, figure out how to build this mega-scale token factory for the world. I think of what we're doing as building a new utility. People are going to want to use a lot of tokens, a lot of intelligence, in all sorts of ways. We need to make that as smart, as cheap, as abundant, as easy to use as possible. That will require deep full-stack integration and a massive infrastructure build-out. The thing I didn't really appreciate between the phase one to phase two shift was how much my management style had to change. Running a research lab and running a product company are two extremely different things. I suspect this third phase is going to be very different yet again. So I've been reflecting on how I'll have to change. I think it's not going to be a natural fit for my management style. So I either have to find someone or a few people great to hire, or I have to figure out how to do things in a different way, or I have to build an AI that can manage this new thing.

我两年前在这里采访了黄仁勋，他告诉我他有 60 个直接下属。你有类似的不寻常做法吗？

I interviewed Jensen here two years ago, and he told me about his 60 direct reports. Do you have any unusual practices like that?

我想我最接近的做法是，我可能每天通过 Slack 或短信与公司里的几百人交谈。非常简短，一两条消息。不是由智能体完成的。

I think the closest thing I have is that I probably talk to, via Slack or text, a few hundred people at the company a day. Very quick, one or two messages. Not done by an agent.

我确实会这么做。有时从中获得的背景信息以这种分散的方式非常有帮助。我觉得这是前 Slack 组织和后 Slack 组织的一个有趣分水岭，它们确实非常不同。

Like I actually do it. And the context I get from that sometimes is very helpful in these diffuse ways. I find this an interesting watershed of pre-Slack organizations, post-Slack organizations, and they're truly quite different.

完全同意。我和很多人一样讨厌 Slack，但我无法想象还得通过电子邮件或我们以前用的方式沟通。

Totally. I, like many other people, hate Slack, but I can't imagine having to still communicate via email or whatever we used to do.

Stripe 大致就是这种情况。

That's roughly where Stripe is.

是的。

Yeah.

好，我要谈谈你刚才间接提到的事情。有一种观点认为，AI 实验室会向上游发展，贪婪地吞噬价值链，这些肯定发生在软件领域，甚至可能延伸到其他领域，它们会形成一种令人难以置信的正反馈循环、失控，成为一种霸权力量，我们都应该对此非常担忧。你怎么看？

Okay, I'm going to talk about something you just elliptically referenced. There is a view that the AI labs are going to progress up the stack, gobbling up the value chain voraciously, all these things that are certainly within the software sector, but perhaps even other sectors, and that they'll be this incredible positive feedback loop and runaway and kind of hegemonic force that we should all be getting very concerned about. What's your view?

我认为有些实验室确实想要那样。但我们不想。我一直钦佩 Stripe 的一点是，它非常明确地与客户保持一致。你知道，我们收入越多，向客户收费也越多。顺便说一句，ChatGPT 推出时与 Stripe 的合作非常关键，我认为没有其他人能那么快地扩展规模，但我们做到了，我们付给你更多钱。这非常一致，我们都开心，你只是为互联网提供了一层基础设施。互联网变大，你开心，你的用户也开心，这种一致性很清楚。我还不完全知道如何做到，但我希望为 OpenAI 建立一个类似的模式。我希望我们成为基础设施提供商。我很乐意我们永远保持低利润率，只要我们能做大并快速增长，我希望我们提供一种智能计量器——我不知道该怎么称呼它——公司可以购买，用来在公司内部实现自动化和加速。他们可以用它来构建产品。人们可以购买它，随身携带。我们找到真正与全球巨大分布式经济引擎的成功保持一致的方法。我相信这会奏效。我相信 AI 的转换成本……无论如何，AI 很难有高利润率。就像你最近看到的，从竞争对手的编程产品切换到我们的产品有多容易。这实际上是 AI 变得更聪明的结果。做这类事情变得更容易。更容易直接说：‘嘿，智能体，帮我去做这件事。’但如果我们能提供一种公用事业，人们在此基础上构建，我们把自己看作那样的公司，我认为这可以非常强大且高度一致。

I think some of them do want that. We don't. One of the things that I've always admired about Stripe is it is very clear that Stripe is aligned with its customers. You know, we make more revenue, we charge our customers more. Thank you, by the way, the partnership with Stripe when ChatGPT launched was extremely critical, and I don't think anyone else could have scaled that quickly, but we scaled, we pay you more money. It's very aligned, and we're all happy, and you just provide a layer of infrastructure for the internet. Internet gets bigger, you're happy, your users are happy, it's clear what the alignment is. I don't know exactly how to do this yet, but I would like to get to a model for OpenAI that is similar. I would like us to be an infrastructure provider. I'd be happy for us to be a forever low margin as long as we can be huge and grow in a fast business, and I would like us to supply kind of an intelligence meter. I don't know what quite to call it that companies can buy that they can use to automate things, accelerate things inside their company. They can use to build products. People can buy it. People can take it with them. And we find ways to really align ourselves with the success of the entire gigantic distributed economic engine of the world. I believe that will work. I believe that switching costs of AI are... It's going to be hard to have huge margins in AI anyway. It's like you've seen recently how easy it is. Many people have seen to switch from our competitors' coding product to ours. This is actually a consequence of AI getting smarter. It gets easier to do things like this. It gets easier to just say like, 'Hey agent, go do this thing for me.' But if we can provide a utility and people build on top of that utility, and we think of ourselves as that kind of a company, I think that can be quite powerful and very aligned.

嗯，我们很乐意分享低利润率业务的许多技巧和窍门。但很多人一直含蓄地批评，或者在某些情况下明确批评 OpenAI 获取了太多算力。我想不是 Codex 用户。

Well, we're happy to share lots of tips and tricks of a low-margin business. But many people have been either implicitly critical or in certain cases explicitly critical of OpenAI for procuring so much compute. And I think not the Codex users.

对，没错。

Right, exactly.

所以，我认为你非常引人注目，早在——我不确定具体时间，但大约两三年前——你就提出了当时听起来荒谬的数字，关于所需建设的规模。显然，这些数字的荒谬性现在一天天减弱。对算力资本支出和建设有什么看法？

So, I think you were quite noteworthy for, as early as I don't know exactly, but in the order of two or three years ago, stipulating what at the time sounded like preposterous figures with respect to the magnitude of the build-out that would be required. And obviously, the preposterousness of those figures now looks less tenuous by the day. Thoughts on compute capex and the build-out?

是的，这需要很多钱。我认为这显然将是世界有史以来最昂贵的基础设施项目。收入正在增长以满足它，所以人们对此感觉更好。此外，我们一直在发现的效率提升是不可思议的。所以每个 GPU 的产出将比我原先预期的多得多。但正如经常被提到的，当你降低每单位智能的价格时，需求会超线性增长，特别是如果你能降低价格并加快回报速度。所以现在关于‘多少算够’这个问题，我没有好的答案。我不……从某种意义上说，我认为在足够低的价格下，对智能的需求实际上是无限的。我本来想说我们不会，但也许我们会。比如，我们不会建造一个戴森球然后把它覆盖上数据中心，但也许我们会。

Yeah, it's going to take a lot of money. I think this will clearly be, at this point, the most expensive infrastructure project that the world has ever undertaken. The revenue is ramping to meet it, so people feel better about that. Also, the efficiency gains that we've all been finding are incredible. So we're going to get way more out of each GPU than I thought we were going to. But as has often been remarked, the demand goes up more than linearly as you drop the price of each kind of unit of intelligence, particularly if you can drop the price and the sort of speed with which you get it back. So this question now of what is enough, I don't have a good answer to. I don't... In some sense, I think demand for intelligence at a low enough price is effectively uncapped. Now I was going to say we're not, but maybe we are. Like, you know, we're not going to build a Dyson sphere and then just cover it with data centers, but maybe we do.

太空数据中心？

Space data centers?

祝你好运。

Good luck with that.

我甚至不认为他对此那么认真。

I don't even think he's that serious about it.

我自己不认为我们处于资本支出泡沫中。我不是这方面的专家。这不是 Stripe 的业务，但仅从我看到的数字相对于需求规模来看，我觉得是合理的。如果我们未来处于资本支出泡沫，我们怎么知道？人们喜欢宣称泡沫。我说不清为什么，但理智上我有点理解。这感觉有趣且显得聪明。记者尤其喜欢谈论泡沫。所以当任何事情看起来有点愚蠢时，就有充分的欲望去写它。显然有时是对的。确实存在泡沫，但如何区分有人喊泡沫的次数和实际处于泡沫中的次数，我从未搞清楚过。在我之前的职业生涯中，我是一名投资者，所以我很有兴趣尝试为此想出某种框架，并弄清楚何时应该部署资本。但我从未能够弄明白。我回顾了历史上不同时期聪明人说过的话，心想：‘哦，他们说得完全正确。’但再多读一点，发现他们在之前的十年里说了十次同样的话。我不知道。我没有答案。经济学家是那些预测了最近三次衰退的人。但他们正确时非常高兴。

I don't myself think that we're in a capex bubble. I'm not an expert in this. This is not Stripe's business, but just the figures I see relative to the magnitude of demand, it looks reasonable to me. If we were in a capex bubble in the future, how would we tell? People love to proclaim bubbles. I can't articulate why, but intellectually I kind of get it. It feels fun and it feels smart. Journalists in particular love to talk about bubbles. So there is ample desire to write about this when anything looks a little bit silly. And clearly sometimes it's right. There clearly are bubbles, but how you can discern between the amount of time someone calls a bubble and the amount of time you're actually in a bubble, I have never figured out how to do. In my previous career I was an investor, so I was quite interested in trying to see if I could come up with some sort of framework for this and figure out when you're supposed to deploy capital or not. And I never was able to figure it out. I went back and I read what smart people had said at different points in history and I was like, 'Oh, they called it exactly right.' But then I read a little more and they said it like 10 more times in the 10 previous years. I don't know. I don't have an answer. Economists are the people who have called each of the last three recessions. But they're so happy when they're right.

所以，你的业务，OpenAI，在很大程度上依赖于超级有才华的人。据我所知，第 20 名最有才华的人与第 5 名、与第一名之间的差距可能非常大且影响深远。而这些超级有才华或高效的人，并非在所有情况下都容易共事。你看，有些人是很好的人，有些是最棒的合作伙伴，有些则非常反传统、意志坚强，他们容易……或者你知道的。

So, a lot of your business, OpenAI, depends in a very significant way on super talented people. And the difference, as I understand it, between the 20th most talented person versus the fifth most talented person versus the most talented person might be quite large and quite consequential. And then these super talented or effective people, they're not in every case super easy to work with. And look, some of them are wonderful people and some of them are the most fantastic collaborators and some of them are very iconoclastic and strong-willed and they get easily... or you know whatever.

这就像人类状况的全谱系。但我好奇的是，在一个对效能、技能、天赋等如此敏感，同时又与人类的各种弱点交织的领域，你怎么看待这个问题？你们容忍“大牌”吗？比以前更多还是更少？你们有特殊的管理方式吗？你如何看待管理精英技能？

It's just like the full spectrum of the human condition. But I'm curious in a domain so sensitive to efficacy, skill, talent, and so forth, intersected with all the foibles of humans as they exist, how do you think about this? Do you tolerate prima donnas? More or less than you used to? Do you manage them in a special way? How do you think about managing elite skill?

有个写 OpenAI 书的人对我说：‘我想我搞明白了你在创办 OpenAI 时真正擅长且做得特别好的事。’我完全猜不到下一句是什么。他说：‘你搞定了如何让一群都觉得自己是唯一能干或最能干的人、一切都得按他们意思来的人，一起合作足够长的时间，去实现突破。’那就是 OpenAI 的魔力。

Someone working on a book about OpenAI said to me, 'I think I figured out the thing you were really great at and did uniquely well in making OpenAI happen.' I had no idea what the next sentence would be. They said, 'You figured out how to get a lot of people who all thought they were the only capable or most capable person, and everything had to go their way, to work together long enough to figure out the breakthroughs.' That was the magic of OpenAI.

好吧，那诀窍是什么？

Okay, so what's the trick?

很多痛苦。即使人们彼此不喜欢，即使他们觉得自己更聪明或有更好的方法，我们有一些深度共享的信念。我们都相信 Scaling（规模扩张）和集中资源，我们要做这一件事，而且把它做对足够重要，以至于可以放下个人冲突。OpenAI 最不寻常的一点是，训练 GPT-3 时，我们整个组织的绝大部分算力都投入到了那一个研究项目中。我们试图从 DeepMind 挖来的人会说：‘这太疯狂了，会创造一种糟糕的文化。’他们说必须平均分配算力，否则会有有毒的竞争文化。但我们采取了有信念地下注的方法。这不会感觉完全平等，但这是正确的事。我们确实认为我们知道想去的方向。拥有一种我们说要坚定信念、做这件事、忽略干扰的文化，很棒。

A lot of pain. Even when people didn't like each other, even when they thought they were much smarter or had a better approach, we had a few deeply shared convictions. We collectively believed in scale and concentrating resources, that we were going to do this one thing, and that getting it right was important enough to put aside personal conflicts. One of the most unusual things about OpenAI was that when we trained GPT-3, the vast majority of our compute went into that single research program. People we tried to recruit from DeepMind would say, 'That's insane. It's going to create a terrible culture.' They said you have to divide compute equally, otherwise you'll have a toxic competitive culture. But we took the approach of betting with conviction on this. It's not going to feel totally equal, but this is the right thing. We do think we know the direction we want to go in. Having a culture where we said we're going to have conviction and do this, ignoring distractions, was great.

所以，约翰和我在 Stripe 已经做了很久了，从 2010 年开始。你和 Greg 从 2015 年开始。你们经历了许多考验、磨难和巨大的成功。对于与联合创始人成功合作超过十年，有什么想法？为什么能成功？你们是如何让它如此成功的？

So, John and I have been doing this at Stripe for quite a while, starting in 2010. You and Greg started in 2015. You've ventured through many trials and tribulations and stratospheric successes. Thoughts on working successfully with a co-founder for over a decade? Why has it worked? How have you made it such a success?

显然，你和约翰认识的时间比我和 Greg 更长，但 Greg 和我在 OpenAI 之前就认识很久了。拥有共同的历史真的很有帮助。在 Y Combinator，成功最大的预测因素之一是联合创始人是否认识很久，至少相对于他们的人生而言。在申请 YC 前七天通过联合创始人匹配网站凑在一起的团队很少成功。不是不可能，但很少见。所以我们认识了一段时间，有共同的价值观、历史和生态，清楚想做什么。我们有深厚的相互尊重和互补的技能组合，效果非常好。我非常感激。没有你深度信任的联合创始人，经历一个紧张的创业经历真的很难。我见过有人这么做，但非常难。所以我很感激我们能一起做这件事。

Obviously, you and John knew each other longer than Greg and I did, but Greg and I knew each other for a long time before OpenAI. Having shared history really helps. At Y Combinator, one of the biggest predictors of success was whether co-founders had known each other for a long time, at least relative to their lives. Teams that came together seven days before applying on a co-founder matching site rarely worked. It's not impossible, but rare. So we had known each other for a while, had shared values, history, and ecosystem, and were clear on what we wanted to do. We had deep mutual respect and complementary skill sets that worked really well. I'm extremely grateful. Going through an intense startup experience without a co-founder you have deep trust in is really hard. I've watched people do it, but it's very hard. So I'm extremely grateful we've gotten to do this together.

在一个相关的话题上，我们在谈论 OpenAI，但还有整个生态系统，包括公司、初创企业和企业，都在这个平台上构建。这是一个有趣的创业时刻，因为能够以前所未有的速度构建产品和产生收入。我们在 Stripe 的数据中看到这一点——企业达到有意义的门槛的速度比以往快得多。你是有史以来最多产、最成功的创业投资者之一，还运营过 Y Combinator。在这个时代，让创始人成功的特质改变了吗，还是和以前一样？

On an adjacent topic, we're talking about OpenAI, but there's this entire ecosystem of companies and startups building on the platform. It's an interesting moment in startups, given the ability to build products and generate revenue at unprecedented rates. We see this in Stripe data—businesses reaching meaningful thresholds far faster than ever before. You are one of the most prolific and successful startup investors ever, and you ran Y Combinator. Have the traits that make founders successful changed in this era, or is it the same as it's always been?

曾几何时，我们经常嘲笑‘点子王’——那些想创业的人说：‘我有最好的点子，我只需要一个程序员帮我做出来。’他们并不那么成功。这让我个人很烦，因为就像说：‘我有一首好歌的点子，我只需要那个弹吉他的家伙帮我做出来。’YC 也有一个版本：没有非技术创始人的团队很难运作。突然间，点子王们复仇了，这对世界来说其实很棒。我很高兴，我支持这一点。但很长一段时间里，我看重的最重要因素，YC 看重的，是创始团队的技术人才。这仍然非常重要，但现在那些深刻理解用户但完全不会编程的人——我想投资那些人。

There was a time when we used to make fun of the 'idea guy'—people who wanted to start a company and said, 'I have the best idea, I just need a coder to build it for me.' They weren't that successful. It was personally annoying because it's like saying, 'I have a great idea for a song, I just need that guy with the guitar to make it for me.' YC had a version of this: teams without non-technical founders are difficult to get to work. All of a sudden, it's the revenge of the idea guys, which is actually awesome for the world. I'm happy, I'm here for it. But for a long time, the most important ingredient I looked for, YC looked for, was technical talent on the founding team. That's still very important, but now people who just deeply understand their users and can't code at all—I want to fund those people.

如今如何看待创业投资？一方面，可能再过几年就到 AGI、ASI 或奇点了，谁知道呢。另一方面，投资时间跨度或基金有 10 年的期限。这如何协调？能协调吗？

How does one think about startup investing these days? Because on the one hand you have a couple of years potentially to AGI or ASI or the singularity, who knows what. And then you have investing time horizons or funds with 10-year time horizons. How does that all fit together? Does it?

我认为在当下以 10 年为期做任何事情都需要真正地搁置怀疑。但这可能是正确的生活方式。我不认为说“3 年或 5 年后就是奇点”之类的话有用。我们看不到奇点之后，所以什么都不做，或者放弃，或者发疯。你必须像事情会长期以可理解的方式继续下去那样生活。

I think to do anything at this point on a 10-year time horizon requires a real suspension of disbelief. And yet that's probably the right way to live your life. Like I don't think it works to say there's the singularity in 3 years or 5 years, whatever. We can't see past it and so we're going to do nothing or we're just going to like give up or we're going to go crazy whatever. You like you have to live as if stuff's just going to keep going in an understandable way for a long time.

OpenAI 的规划有多长远？

How far ahead does OpenAI plan?

我们签了 20 年的电力和土地协议。

I mean we signed like 20-year power and land agreements.

那产品方面呢？

And for the product?

我们对两年后的情况有清晰的愿景，但之后就会模糊很多。

I think we have like a clear vision of what things can look like in 2 years and then it gets much hazier after that.

不久前有一种说法，GPT 包装器之类的公司，正如这个贬义词所暗示的，是无差异且脆弱的，会被模型改进的浪潮冲走。而现在，这种说法在某种意义上有所转变，我们不再谈论包装器，而是谈论“驾驭器”，驾驭器被视为具有重要分量和意义。我很好奇你的看法，以及你如何看待那些以 AI 为关键赋能组件的企业及其未来的持久性。

So there was in the relatively recent past a narrative that you know GPT wrappers and companies of that ilk were I guess as the pejorative suggests undifferentiated and flimsy and be swept away by a rising tide of model improvements. Whereas now it feels like that narrative in some senses flipping somewhat for now instead of talking about wrappers we talk about harnesses and harnesses are seen as having this you know significant heft and importance. And I guess I'm curious for your view on this and how you view businesses for which AI is a critical enabling component and their prospective durability.

我一直持有同样的观点：作为企业，你应该站在希望 AI 变得更聪明的一边。在早期模型时代，如果你是一个 GPT 包装器，修补当前模型的某个弱点，而这个弱点显然会在下一个模型中得到改善，那么下一个模型更好时你会难过。如果你做的事情是随着智能提升而变得更好，比如你构建的那些利用智能的美妙服务，你会更开心。在驾驭器的世界里也是如此。我认为正确的思考方式是：数据中心、模型、驾驭器，整个东西就是一个集群，从中产生非常可用的智能。有很多事情可以构建，你会希望整个集群越来越好。如果你暗中希望它不要变好，因为你正在修补其中的某个弱点，那么很可能模型堆栈中某处的下一次迭代就会解决它。

I have kind of had the same view all the way through which is you as a business want to be on the side of hoping that AI gets smarter. And whether so like in the early model days if you were the GPT wrapper and you were like patching some kind of weakness in the current model that was clearly going to get better with the next model. The next model was much better you were kind of sad. If you were doing something that got better like you know you're making any of the wonderful like services that people were building with the models that benefited from intelligence you'd be happier. I think the same thing in the world of harnesses like I kind of think the right way to think about this is like data center model harness like that whole thing is just this one cluster out of which comes this like very usable intelligence. But there are so many things to go build where you're just like happier for that whole cluster to get better and better and better. And then if you're kind of secretly hoping it doesn't because you're patching some weakness in that, probably like the next model crank turn somewhere in that stack is just going to solve it.

当你观察那些最有效利用 AI 的组织时，你经常接触 OpenAI 的客户，大大小小都有。如果让你想想最让你印象深刻的前三名或第一名，他们具体做了什么不同的事情？这里的每个人都知道，是的，AI 很重要，我们应该热情地使用它等等。但你认为那些最有效利用 AI 的组织具体有什么不同？

When you look at the organizations that are making most effective use of AI today, like if you I mean you meet OpenAI customers constantly, large and small, if you think about the top three that have impressed you the most or the one that's impressed you the most, what specifically are they doing that's different? Like everyone here knows, yes, AI, big deal, we should make enthusiastic use of it, etc. But what specifically differentiates those which in your opinion are employing it most effectively?

有几个不同的方向。我们共同的朋友，Shopify 的 Tobi Lütke，是我认识的第一位 CEO，他说我们将全面拥抱 AI 来运营公司。他亲自动手构建 AI 自动化一切，并让他的团队也这样做。他说，我们要弄清楚如何把所有不好的事情用 AI 变好。这不是象征性的排行榜，也不是某种游戏化的可破解的东西。就是 CEO 说，我们现在要把 AI 融入我们做的每一件事，如果你不这样做，我不会满意，或者我们不允许。这种能量现在也被其他人效仿了。当 CEO 说‘我们要自动化自己，加速自己’，无论他们内部怎么说，然后真正要求大家做到，并且自己带头，效果非常好。我们打算尝试一个新实验：派一名全职员工去与 CEO 手把手合作，无论 CEO 需要什么，帮助他们自动化工作，尽可能多地自动化。我认为如果你只为公司领导者做这件事，整个公司会产生很好的分形效应。所以这很有效，我们会帮助公司这样做。第二点是，在数据访问上要令人不安地宽松。有很多理由不这样做，我不是在推荐，只是回答你关于最有效公司的问题。这对小初创公司来说比拥有大量敏感数据和严格流程合规的公司更容易。但说‘你知道吗？我们要录制会议，让 AI 访问我们的代码库，访问每一条 Slack 消息、每一封邮件、所有东西，每个员工都可以这样使用。’看着两三个人的初创公司用 AI 做所有事情，真是令人惊叹。我不知道世界将如何权衡数据隐私与 AI 效率的取舍，我认为一些法规需要为此改变，但这非常强大。

A few different directions there. Friend of both of ours Tobi Lütke of Shopify was the first CEO I knew that just said like we are going to be all in on AI in the way we run our company. And he got himself got his hands dirty just building like AI automation of everything and made his team do it. And you know, said we're just going to figure out how we take all of these things that are bad and make them good with AI. And it was not like, you know, a token leaderboard. It was not some other kind of like gamified hackable thing. It was just like the CEO of the company said, we are now going to put AI into everything we do and I'm going to like not be happy with you, I guess, or we're not, you know, we're not going to allow it if you're not doing that. So, that energy has now been done by other people, but when the CEO of a company just says like 'We're going to automate ourselves, accelerate ourselves.' However they phrase it internally, and then really holds people to it and ideally does it themselves. That has worked very well. I think we're going to try this new experiment where we start sending like an FTE to work with like hands-on, just whatever the CEO of a company needs, work with them to automate their job. As much of it as they can and that I think will have like if you just do it for the leader of a company, there's like a nice fractal effect throughout the company. So, that works and we'll try to help companies do that. A second thing is being like uncomfortably permissive with data access. There's huge reasons not to do this and I'm not This is like I'm stopping short of recommendation. You just asked what I've seen from the most effective companies. This is easier for small startups than companies that have a lot of sensitive data and a lot of process and compliance in place. But saying like, 'You know what? We are going to record our meetings. We are going to like let this AI have access to our code base. We are going to let this have access to every Slack message, every email, every everything and every employee at the company is going to get to use it that way.' Like it is amazing watching these two or three-person startups and AI doing everything work. And I don't know how the world is going to decide the tradeoffs on data privacy versus AI efficiency. And I think there's like some regulations going to have to change for that, but it's so powerful.

Tempo 是一个新的区块链，由 Stripe 与 Paradigm 共同孵化，我们显然与你们是合作伙伴。主网刚刚启动，但项目去年夏天就启动了。所以这是一个相对较新且规模较小的团队，只有几十人。Tempo 团队在他们的 Slack 中设置了一个驾驭器工具，用于编排公司几乎所有的事情。所有事情。你可以提出任何任务，比如去阅读这些 Google 文档，把它们变成一系列线性任务，然后写一个拉取请求来实现它们，然后部署它们，并使用我们的日志分析工具来测试部署是否成功。智能体会愉快地去使用所有这些工具。看着整个组织——一个小组织，但一个由人组成的组织——在一个 Slack 频道里做所有事情，非常迷幻。

Tempo is a new blockchain that Stripe incubated with Paradigm, which obviously we're partners with you guys on. And the mainnet just launched, but the project launched back last summer. So, it's a relatively new and small team, a couple dozen people. The Tempo team set up a harness, a tool in their Slack installation for orchestrating pretty much everything at the company. Everything. You can just ask any task, go and read these Google Docs, turn those into a bunch of linear tasks, then go write a pull request to implement them, then go deploy them and use our log analysis tool to test if the deployment actually worked. And the agent will happily go and employ tools used across all of this. And it's extremely trippy watching a whole organization, a small organization, but an organization of people do everything in a single Slack channel.

我不认为这能扩展到 Stripe，但这是我第一次体验到你说的那种感觉——对他们来说确实不可思议。我不太清楚如何为我们复制，但这真的很了不起。看着真是令人惊叹。我发现很多人就是做不到——这里有一个巨大的认知落差。他们还没能理解这样一个事实：你几乎可以问我们任何问题，它很可能就会实现。我自己也仍然觉得不够信任，不太相信这真的可能。我不确定这如何应用到更大的公司。感觉我们好像还缺少一层抽象。比如人类和 AI 如何在大规模下交互。小公司的优势在于，它们只有 AI，不需要处理与所有人的接口。但我们会找到办法的。

And I don't think that would scale to Stripe, but it was the first time I had the experience you're describing, which is it's truly incredible for them. I don't quite see how to transpose it for us, but this is really something. It's really something to watch. And I find that a lot of people just can't—this is where there's a big overhang. They have not yet been able to wrap their heads around the fact that you can just kind of ask us anything and it'll probably happen. I myself still find myself like not trusting quite enough that it's going to be possible. I don't know exactly how it's going to transpose to bigger companies. It does feel like we're missing kind of like one more abstraction there. Like how humans and AIs are going to interface at massive scale. The advantage that these smaller companies have is like it's just AIs. They don't have to figure out the interface with all the people. But we'll figure it out.

开源 AI。它走向何方？有未来吗？

Open source AI. Where's it going? Does it have a future?

当然。现在，人们显然想要更智能、更快、更便宜的前沿智能，大部分需求都在那里。但对开源的需求也很大，我预计随着时间的推移，这种需求会相对增加。

For sure. Right now, people clearly want smarter, faster, cheaper frontier intelligence, and most of the demand is there. But there is also a lot of demand for open source, and I expect that to increase relatively over time.

所以，我想聊一点科学，因为我知道这是你非常兴奋的事情，但也与我花时间做的一些事有关，也就是 Arc 研究所。OpenAI 实际上是一个基金会，一个非营利组织，最近向 Arc 研究所提供了一笔资助。我们稍后可能会谈到细节，但首先，你能否谈谈 AI 在科学中的应用，你看到了什么，以及你如何看待 OpenAI 基金会背景下的捐赠？

So, I want to talk a little bit about science, because I know it's something you're very excited about, but it's also relevant to something I spend some of my time on, which is the Arc Institute. And OpenAI is in fact a foundation, a nonprofit, and recently made a grant to the Arc Institute. So, maybe we'll get to the details of that in a second, but first you just want to speak a little bit to AI as applied to science, you know, what you're seeing, and just how you think about grant making generally in the context of the OpenAI Foundation.

总的来说，关于 AI 和科学的问题，我希望这将是 AI、这项技术对人类生活质量最重要的贡献。如果我们能开始以更快的速度发现新科学，比如新材料、疾病疗法或其他任何东西，我相信，粗略地说，生活变得更好是因为我们更好地理解了科学，然后我们想办法用它来建造东西并分发给人们。从几个月前的模型开始，但真正到了 5.5，模型已经足够聪明，优秀的科学家们说：‘我能想出更好的点子。模型能做出一些微小但重要的发现。’科学的步伐将会加快。最终，我们将拥有自动化实验室和机器人，建造天知道什么东西。我们将能够更快地进行科学。但如果我们能开始以过去需要十年的速度在一年内完成科学，这种复合效应以及我们能做和发现的事情将极其伟大。所以，我认为这将是不可思议的。这将是 OpenAI 基金会重点关注的领域之一——基本上就是资金、专业知识和加速科学的技术，并相信这些将以美妙的方式流向世界。这将是一个大型基金会。是的，这将是——我认为它是最大的之一，也许是最大的。我认为它将成为世界上最大的基金会。所以，我们非常专注于科学，然后是 AI 韧性——帮助世界通过这个过渡期，适应这项新技术。但我们很高兴能支持 Arc。我认为它显然是最好的 AI 和生物努力，如果我们能用这项技术和基金会的资本为让人们更健康、治疗疾病做出哪怕一点贡献，随着我们更好地理解生物学，我们能做的这一整套事情，我们将非常激动。我原以为这需要更长时间，但看到 Arc 基金会所做的令人难以置信的工作，我现在认为可能不会太远了。

So, generally on the AI and science question, I hope that this will be the most important contribution of AI, of this technology, to human quality of life over time. And that if we can start to discover new science at a much faster rate, which can be like new materials or cures to diseases or any number of other things, like I believe that to a first order approximation, life gets better because we understand science better, and then we figure out how to build stuff with it and distribute it to people. Starting with the models of a few months ago, but really now with 5.5, the models have gotten smart enough that excellent scientists are saying, 'I am able to figure out better ideas. The models are able to make some small but important discoveries.' And the pace of science is going to increase. Eventually, we'll have automated labs and robots and building who knows what. And we'll be able to do science much faster. But if we can start doing like a decade of science at what it would have taken us in the old world in a year, the compounding effect there and what we'll be able to do and discover will just be extremely great. So, I think this is going to be incredible. And this will be one of the big areas of focus of the OpenAI Foundation is basically like money and expertise and technology to accelerate science and trusting that that will flow to the world in wonderful ways. And this is going to be a big foundation. Yeah, this will—I think it is one of the biggest, maybe it's the biggest. I think it will be the biggest foundation in the world. And yeah, so we're really focused on science and then AI resilience like helping the world through this transition with this new technology in it. But, you know, we were thrilled to get to support Arc. I think it's clearly the best sort of AI and bio effort and if we can make even a small contribution with this technology and with the capital in the foundation to helping make people healthier, treat diseases, this whole cluster of what we can do as we get better at understanding biology, we will be very thrilled. I thought that was going to take longer and looking at the incredible work Arc Foundation is doing, I now think it maybe won't be that far off.

你知道，在播客中，中途你可能会听到一个小插播广告。这就是你为 Arc 研究所准备的插播广告。楼下有一个 Arc 研究所的展位。你可能想知道他们为什么会在互联网经济大会上。背景是这样的：Arc 研究所是我们四年前创办的一个组织。它的目标是希望产生第一个治愈人类复杂疾病的方法。复杂疾病是指涉及一些遗传因素和一些环境因素的疾病。所以，你可以把大多数癌症、大多数自身免疫性疾病、大多数神经退行性疾病，例如，视为这种特定意义上的复杂疾病。人类从未治愈过复杂疾病，一个都没有。我们治愈过很多传染病。我们知道如何筛查单基因疾病，即由单一基因突变引起的疾病。我们从未治愈过复杂疾病。所以，我们创办 Arc 研究所时就把这作为目标。阿尔茨海默病是我们瞄准的第一个复杂疾病。我们的希望是，通过新的基因组工程技术如 CRISPR 以及 AI 的进步，我们能取得一些有意义的进展。它只有四年历史，但早期结果非常令人鼓舞。Arc 研究所目前正在寻找一位 CTO。去年我们有一位 CTO 在 Arc 做休假。他叫 Greg Brockman。他做了一些很棒的事情。他帮助我们训练了 Evo 2，这是有史以来训练过的最大的生物学基础模型。但我们正在寻找一位全职 CTO。我们想，也许那个人就在这个观众席里。如果不是，他们的朋友可能在这里。所以，如果你认识对此感兴趣的人，请去楼下 Arc 的展位看看。这就是你的插播广告。

So, you know on a podcast, midway through you might hear a little interstitial ad. This is your interstitial ad for the Arc Institute. There is an Arc Institute booth downstairs. And you might wonder why they're what they're doing at the Internet Economy Conference. And the context here is so the Arc Institute is an organization we started 4 years ago. And its goal is to produce hopefully the first cure for a complex disease in humans. So, a complex disease is one that involves some genetic factors and some environmental factors. So, you can think of most cancers, most autoimmune disease, most neurodegenerative disease, for example, as being a complex disease in this kind of specific sense. And humanity has never cured a complex disease, not one. We've cured lots of infectious diseases. We know how to screen for monogenic diseases, for this one genetic mutation that undergirds it. We've never cured a complex condition. So, we started Arc Institute with this as the goal. Alzheimer's is the first complex disease that we're targeting. And our hope is that with both new genome engineering technologies like CRISPR and then the sort of advances in AI, that we'll be able to make some hopefully meaningful progress. And it's only 4 years old, but the early results are very encouraging. The Arc Institute is currently looking for a CTO. We had one CTO do a sabbatical at Arc last year. His name was Greg Brockman. And he did some great stuff. He helped us train Evo 2, which is the largest biology foundation model ever trained. But we're looking for a full-time CTO. And we thought that, well, perhaps that person might be in this audience. And if not, their friend might be in this audience. So, if you know somebody for whom that sounds of interest, go check out the Arc stand downstairs. And that's your interstitial ad.

我觉得你把它标为插播广告而不是直接做，这样好多了。做得不错。

I think it's so much better that you labeled that as the interstitial ad rather than just doing it. That was good.

好，我们还没聊到 Stripe。你是 Stripe 的第二个投资者吧？YC 是第一个吗？

Okay, we haven't talked about Stripe. So, you were the—I think the second investor in Stripe. Was YC the first?

你和 Paul 同时投的。

You and Paul at the same time.

好的。

Okay.

实际上是在同一个厨房里。所以，也许你才是第一个。我不记得支票是按什么顺序给的。

In the same kitchen, in fact. So, maybe you were first in fact. I don't remember in which order the checks were handed out.

那是在他帕洛阿尔托的厨房里。

This was then his like Palo Alto kitchen.

是的，没错。所以，我们就像是——两个长着青春痘的青少年提议建立这个金融服务机构。

Yeah, that's right. So, you know, we were sort of—you know, we were two pimply teenagers proposing building this financial services institution.

这听起来有点荒谬。你为什么投资了？

It sounded like a bit of a ludicrous proposition. Why did you invest?

老实说，他没给我铺垫。你和约翰是我见过的最令人印象深刻的创始人，无论年龄，但尤其是满脸青春痘的青少年时期。我显然没看错。我听说过你，也听保罗谈起过你，我印象深刻的是你在为自己解决一个问题。这符合我认为会变得重要的趋势：商业大规模转向线上，涌现大量初创公司。我相信，如果你能找到非常聪明的创始人和一个会变得很大的市场，你就应该投资。就是这样。

Honestly, he didn't tee me up for this. You and John were two of the most impressive founders of any age, but certainly of pimply-faced teenager age that I had ever met. I clearly was right about that. I had known of you and heard Paul talk about you, and I was struck that you were solving a problem for yourself. It fit a trend I thought was going to be big: commerce moving online in a huge way with lots of startups. I believed that if you can find really smart founders and a market that's going to be big, you should just invest. That's kind of it.

基于你的总体视角，以及 OpenAI 对 Stripe 的使用，你对 Stripe 在 AI 时代的发展有什么建议？

Based on your general perspective, and OpenAI's use of Stripe, what's your advice for Stripe on navigating the AI era?

在此之前，确认一下我的记忆：你意识到需要支付是因为你做了一个 iPhone 应用来下载整个维基百科，因为你要去一个没有网络的地方。你很难为此收款。那是个很好的记忆点。我很久没想起这件事了。我的建议更多是针对 Stripe 这家公司本身。现在世界很疯狂，很多变化。你从客户的角度看过 Stripe，所以你的抱怨和功能请求是什么？

Before that, just to check my memory: the reason you realized you needed payments was you built an iPhone app to download all of Wikipedia because you were going somewhere offline. It was hard for you to take payments for it. That was a good memory hurdle. I hadn't thought about that in a long time. My advice is more towards Stripe as a company itself. It's a crazy time in the world. A lot is changing. You've seen Stripe from the perspective of a customer, so what are your complaints and feature requests?

嗯，现在我在想我们应该更像 Stripe 的模式，我有很多要向你学习的。我记得投资者们自作聪明地说：‘哦，当大家都做大时，Stripe 会变成商品，他们会自己建支付系统。’但事实证明，如果你做出优秀的产品，人们未来仍然需要收款，如果你是一个好的生态系统合作伙伴，他们会继续用你。每家公司都需要借助 AI 提高效率，但不要假设整个社会经济体系会完全重构。世界对此有点妄想。

Well, now that I'm thinking we should think more like a Stripe-style model, I have more to learn from you. I remember when investors would outsmart themselves saying, 'Oh, Stripe's going to be a commodity when everybody gets big, they'll just build their own thing.' But it turns out that if you make a great product, people still need to take money in the future and they'll stick with you if you're a good ecosystem partner. Every company needs to get more efficient with AI, but don't assume the entire socioeconomic system completely reconfigures. The world has gotten a bit delusional about this.

除了 AI，展望未来十年，哪些技术和领域让你兴奋？

Apart from AI, as we look out over the next decade, what technologies and areas excite you?

在模型和产品层面，我对数据中心基础设施着迷。那里有很多酷事可做。物理层会有惊人的新技术：能源、机器人。那个堆栈可能是我最常思考的。世界终于在脑机接口上取得进展。我对生物技术的进步既害怕又兴奋。我希望它能快速变得更好。防御性生物技术即将变得非常重要。我对新型计算机界面感到兴奋。我们被困在旧设备和操作系统里，但我们有这种神奇的新赋能技术。Codex 很惊人，但让它通过点击为人设计的东西来使用我的电脑感觉很破碎。我们可以做得更好。还需要制定一套全新的互联网协议。

At the model and product layer, I'm obsessed with data center infrastructure. There's so much cool stuff to do there. There will be amazing new technologies at the physical layer: energy, robots. That stack is probably what I think about most. The world is finally making progress on brain-machine interfaces. I'm both afraid of and excited about progress in biotech. I'm hopeful it can get much better quickly. Defensive biotech is about to become very important. I'm excited about new kinds of computer interfaces. We're stuck with old devices and operating systems, but we have this magic new enabling technology. Codex is amazing, but it feels broken to have it use my computer by clicking around stuff made for a person. We can do much better. There's a whole new internet protocol to make too.

你认为我们什么时候会有世界上第一个盈利的核聚变反应堆？

When do you think we'll have the world's first profitable nuclear fusion reactor?

取决于数据中心需求将电价推高到什么程度。也许比我们想的更快。我猜未来 5 年内。

Depends how far electricity prices get pushed by data center demand. Maybe sooner than we thought. I'll guess in the next 5 years.

这是个大胆的预测。

It's a bold prediction.

希望如此。

I hope so.

高超音速商业航空旅行呢？

Hypersonic commercial air travel?

没有密切关注。高超音速是指 4 马赫吗？还需要相当长的时间。超过 10 年。也许少一点，但差不多。

Don't follow it as closely. Hypersonic meaning Mach 4? A good chunk of time. More than 10 years. Maybe a little less, but something like that.

有没有哪些科技领域没有被充分讨论，但会被 AI 加速并产生广泛社会影响？

Are there any domains of science and technology not in the discourse significantly that will be accelerated by AI and have broad social impact?

一个没有得到足够关注的领域是材料科学。它不酷。人们低估了世界有多少是由材料构成的，我们有多依赖它们，以及 AI 能取得多大进展。这是一个非常适合 AI 的问题。比如获得新催化剂。我预计那里会有非常快速的进展，并将积极影响我们所有人的生活。它得到的关注很少。

The one that is not getting enough attention is material science. It's not a cool thing. People underestimate how much of the world is materials. How much we depend on them and how much progress AI can make. It's such a beautifully AI-shaped problem. Just getting new catalysts. I expect very rapid progress there, and it will impact all our lives positively. It gets very little attention.

感觉某种形式的 AI 是不可避免的。你希望你的具体参与如何改变世界的轨迹？

It feels that some version of AI is inevitable. How do you hope your specific involvement changes the trajectory for the world?

我相信民主化、个人能动性和可及性，并且每个人都值得拥有真正美好的生活。

I believe in democratization and personal agency and access, and that everybody deserves a really great life.

我们在 OpenAI 历史上做出的最具争议的决定，就是我们如今所说的迭代部署。在发布 ChatGPT 时，很多人的想法是，这样做极其危险，你绝不能这么做。只有那少数一直在思考 AI 安全的人才能知道即将发生什么。告诉世界是一种信息危害，而且它太危险了，永远不能发布。我们必须把它锁起来，在象牙塔里发现这些美妙的东西，然后把成果分享给世界，但 AI 由我们掌控。这让我非常不认同。我当时认为，现在也仍然认为，避免这种权力集中至关重要，我们应该为世界构建这项技术。世界会以各种方式使用它，有些是好的，有些则不然。通过让人们探索这个非常广阔的机会空间——尽管有时会混乱，并设置合理的护栏以确保安全——我们给世界一份礼物，而世界会在其基础上为我们所有人建造一份更大的礼物。如果你不让人们使用这项技术，而是试图把它锁起来——现在这听起来可能显而易见，但在我们出现之前，这却是从事该工作的人们的粗略共识计划——我认为那将非常糟糕。我相信创业精神和创新，相信人们大多是善良的，会用工具做出惊人的事情。所以，我最大的贡献就是推动这项技术民主化，让人们能够使用并在此基础上发展。

The most controversial decision we made in the history of OpenAI is what we now call iterative deployment. At the time we released ChatGPT, a lot of the thinking was that this was insanely dangerous to do. You can't do this. Only this small set of people who've been thinking about AI safety can know what's coming. It's an info hazard to tell the world, and it's all too dangerous to ever release. We have to keep this locked up, and in our ivory tower, we will discover these wonderful things and share the fruits with the world, but we'll have the AI and we'll control it. That sat very poorly with me. I thought then and I believe now that it is extremely important to avoid that kind of power concentration and to build this for the world. The world gets to use it in many ways, some good, some not. By enabling people to explore this very wide opportunity space, messy at times, and with guardrails for reasonable safety, we give the world a gift, and the world will build a much bigger gift on top for all of us. If you don't enable people with this technology and try to keep it locked up—which now may sound obvious, but that was the rough consensus plan of people working on it before we came along—I think that would have been really bad. I am a believer in entrepreneurship and innovation, and that people are mostly good and do amazing things with tools. So, my single biggest contribution has been pushing for this to be a democratized technology that people get to use and build on.