从个人 AI 到 Codex：OpenAI 的内部演变

2025 年发生的一个重大变化是，人们开始将 ChatGPT 用于更个人化、非常私密的应用。例如，我妻子患有复杂的医疗状况，包括活动过度综合征，我们花了很多年才确诊。当我们把这些症状输入 ChatGPT 时，它几乎能立刻判断出来。但问题是，每个医生都有自己的专长，没有一个医生能通览全局，而她一直用 ChatGPT 来管理自己的健康。

One thing that's really changed over the course of 2025 was people started to use ChatGPT for much more personal, very intimate applications. For example, my wife has complex medical conditions including hypermobile syndrome which took many years for us to get diagnosed, and as we put those symptoms into ChatGPT, it would be able to figure out pretty immediately. But the thing is that every doctor has their own specialty rather than one doctor who can see across everything, and she uses ChatGPT to manage her health all the time.

很好。

Great.

对我来说，有趣的是，我通常是我们自己技术的晚期采用者，我通常会测试它们，并以各种方式对它们进行压力测试。有趣的是，对于早期版本的模型，我通常会试图破坏它们的一些过滤器，所以我会对它们骂脏话、大喊大叫，而我妻子总是说‘他只是开玩笑’，你知道，告诉我要对 AI 友好。我认为，就我而言，我实际上是一个非常固执己见的人。第一次改变是最近，通过 Codex，真正开始是在 12 月。我一直是个老顽固。我有自己的一套做事方式。我使用我的终端，我使用我的 Emacs，就像所有我从小使用的工具一样，而现在我已经完全抛弃了它们。

For me, the funny thing is I'm actually a late adopter of our own technologies usually, and I usually test them and I stress test them in all sorts of ways. And it's funny, for early versions of our models, I would usually try to break some of their filters, so I would swear at them and yell at them, and my wife is always like, 'He's just kidding,' you know, telling me to be nice to the AI. And I think that for me, I actually have been someone who's almost very set in my ways. And the first time this has changed is very recently with Codex, and really starting in December. Like, I've been a curmudgeon. I've got my way of doing things. I use my terminal, I use my Emacs, like all these tools that I grew up with, and I've just abandoned all of that now.

哇。

Wow.

我现在只用 Codex。这真是革命性的。

I'm just using Codex. That's revolutionary.

确实如此。

It really is.

听起来你以前很固执。

You sound like you were set in your ways.

我确实如此。是的。

I really was. Yes.

每个新模型有什么变化？

What changes with each new model?

一切都在变。从外部看，人们会觉得‘哦，你们只是在扩大模型规模，只是在做这种傻事’。但在内部，流程的每一个部分，我们都在不断升级。机器学习中有效且被奖励的是对细节的关注。所以你真的要确保所有的 Scaling（规模扩张）都是正确的。你要确保系统，比如 GPU 故障，这种情况会发生。那么，如果你有 10 万个 GPU 在运行，如何检测哪个 GPU 是坏的呢？这并不容易，对吧？你不能直接说‘那个’。所以你需要一个近乎物理的过程。还有软件过程。还要理解你的数据是否良好，而仅仅通过查看数据来理解其中的内容，并确保格式正确、分词恰当，就能获得巨大的回报。所以输入的每一个部分我们都在不断改进，我们也非常关注输出。我们试图看看这里的改进如何与评估指标的变化联系起来。我在 OpenAI 多年的一个观察是，如果我们有一些生命迹象，一些现在勉强能用的应用，一年后你应该期待它变得出色。所以我们正处于指数级增长中，能够随着时间的推移做出非常非常复杂的改进。

Everything. And the way to think about it is that from the outside perception, it's, 'Oh, you're just scaling up the models. You're just doing this kind of dumb thing.' On the inside, every single part of the process, we are always upleveling. The thing that works in machine learning and that machine learning rewards is attention to detail. So you really want to make sure that all of the scaling is right. You want to make sure that the systems, for example GPUs failing, that happens. So how do you detect if you have a run of 100,000 GPUs, how do you detect which GPU is the broken one? It's not easy, right? You can't just be like, 'That one.' So you need this almost physical process to it. There's the software process to it. There's understanding if your data is any good, and there's so much reward to just actually looking at the data to understand what's in there and to make sure that you format it correctly and it's tokenized properly. So there's just every single part of the input we are constantly improving, and we pay attention a lot to the output too. We're trying to see how does improvement here connect to an eval shifting. And one observation I have across many years of OpenAI is that if we have some signs of life, some application that kind of works right now, one year from now you should expect it to be excellent. So we are on this exponential and able to make these very, very sophisticated improvements over time.

跟我谈谈你的联合创始人 Sam 吧。他似乎在世界各地引起了强烈的反响。

Tell me about your co-founder Sam. He seems to generate strong reactions out in the world.

我认为 Sam 被世界严重误解了。我对 Sam 的看法是，他是一个非常好的人，而这种善良反而成了他的负担。这非常符合‘好心没好报’。而且我认为，一些批评者，我用‘每个指控都是自白’的心理过滤器来看待，对吧？我认为人们将自己在自己身上看到的、他们感到不安的、或者他们自己想要的东西投射到他身上。我认为他非常有韧性。他经历了很多，我真的很感激他，老实说，他处于这么多关注的中心，并且继续前行，因为他对于 OpenAI 成为今天的模样至关重要。而且我认为没有人能像他那样扮演好这个角色。

I think Sam is very misunderstood by the world, and I think that the way I think about Sam is that he is a very good person, and that goodness is something that gets held against him. It's very much 'no good deed goes unpunished.' And I think that some of the critics, I run through my mental filter of 'every accusation is a confession,' right? I think that people project onto him things that they see in themselves, things they're insecure about, or that they want for themselves. And I think he's very resilient. He's been through a lot, and I'm very grateful for him honestly being at the center of so much attention and continuing to go, because he's been very critical to OpenAI becoming what it is today. And I think that there's no one who could have filled that role as well as he has.

是啊。而且他能承受所有压力，这样你就能继续工作，这也很不错。

Yeah. It's also nice that he can take all the heat so you can continue doing the work.

我本来不想这么说，但这是事实，我对此深表感激。

I wasn't going to say it that way, but it is true and it's something I'm deeply grateful for.

跟我讲讲那次解雇事件吧。那是个大新闻。

Tell me the story of the firing. It was a big story.

这比我预想的要大得多。

It was a much bigger story than I would have ever expected.

是什么导致了它，又发生了什么？

What led to it and what happened?

我想说，在解雇事件发生前的一年，也许一年半的时间里，我们做错的一件事就是让冲突酝酿。我在 OpenAI 整个过程中学到的一个教训是，进行艰难的对话，真正解决冲突，人们意见不合，这很正常。这会发生。但如果你让它恶化，那以后总会更痛苦。我认为我们一次又一次地看到了这一点。而解雇事件就是其中最突出的例子之一。这是我从整个经历中真正学到的一个教训。任何事物发展得如此之快，都会伴随着很多混乱和复杂。

I would say for the year, maybe year and a half leading up to the firing, I think that one thing we did wrong was that we let conflict brew. One of the lessons I have throughout the course of OpenAI is that having the hard conversation and actually really addressing when there are conflicts, people disagree, that's normal. That happens. But if you let it fester, that is always going to be more painful down the road. And that's something that time and time again, I think that we've seen. And I think that the firing was the most salient one of those examples. And that's one lesson I really take away from the whole experience. Anytime anything grows so quickly, there's a lot of confusion and complication that comes with the territory.

我认为确实如此。是的。

I think that's true. Yes.

所以我想，在公司发生的爆炸性增长中，你们发展的速度，难免会遇到困难。

So I imagine in the explosion that was going on at the company, the rate that you were growing, there are bound to be difficulties.

是的。我认为存在多个层面的问题，对吧？只是不同人之间的冲突，不同的运作方式，决策问题。也许有一些根本性的问题，比如‘ChatGPT 是不是个好主意’，你知道，我们的两位董事会成员曾谈到过。我认为这最终导致了 11 月的那件事。

Yeah. I think that there were, I think at multiple levels, right? Just conflict between different people, different ways of operating, just decisions. Maybe there's some fundamental ones around 'was ChatGPT a good one or not,' you know, that two of our board members have spoken about. And I think that this led ultimately to that event in November.

描述一下。告诉我那件事是什么。它是怎么发生的？

Describe it. Tell me what the event was. How did it happen?

我当时正在写代码。我一直在写代码。我正在准备一个我已经研究了一段时间的改动。我非常兴奋地想要合并并发布它，然后我收到一条消息，让我加入一个视频通话。于是我进入视频通话，Google Meet 的预览屏幕显示了谁在里面，我注意到除了 Sam 之外，董事会成员都在。我非常惊讶。于是我点击加入，然后他们告诉了我这个消息。

I was coding. I was coding away. I was preparing this change that I've been working on for some time. I was very excited to get it merged and shipped, and I got a message asking me to hop on a video call. So I go into the video call, and the Google Meet preview screen shows who's in there, and I noticed it was the board except for Sam. I was very surprised. So I click join, and then they tell me the news.

他们告诉你什么？

What do they tell you?

嗯，他们告诉我的基本上和公开新闻稿里的信息一样，说 Sam 被解雇了，说 Mira 是临时 CEO，还说我已经被从董事会除名。他们告诉我，我对公司非常重要，我是一个能成事的人，他们希望我留在公司。

Well, they tell me essentially the same information that was in the public press release, saying that Sam's been fired, saying that Mira is the interim CEO, and saying that I've been removed from the board. And they tell me that I'm very important to the company, that I am someone who can get things done, and that they want me to stay at the company.

你被从董事会除名了，但他们希望你留下。

You're off the board, but they want you to stay.

没错。是的。而且他们甚至没有以‘你是否愿意留下’的方式来问。只是说这是你的新角色。对我来说，我在那一刻就知道这不对。

That's accurate. Yeah. And it wasn't even really framed in a way of 'will you stay or not stay.' It was just saying this is your new role. And for me, I knew in that moment it wasn't right.

这是不是突如其来？

Did it come out of the blue?

是的。

Yes.

哇。你震惊吗？

Wow. Were you shocked?

是的。

Yes.

我能想象。

I would imagine.

是的。

Yes.

但那是这样一个时刻：因为我了解所有人，了解所有动态，了解那些一直在酝酿的冲突。对我来说，好吧，我明白发生了什么。走到这一步很遗憾，但我理解。

But it was one of those moments where because I know all the people, I know all the dynamics, I know the conflicts that have been brewing. For me it was, okay, I see what happened. It's sad that it came to this, but I understand.

是啊，如果早点有人关注这件事，也许可以有不同的处理方式。

Yeah, it could have been handled differently earlier if somebody was focused on that.

当然可以更早地以不同方式处理。我想每个相关的人现在都会这么说。我记得我要求了解更多信息，但他们当时不愿分享。我再次理解了，好吧，事情就是这样发展的。挂断电话后，我告诉妻子：‘我们得离开。’她说：‘是的，我们必须离开。’我说：‘你要知道，我们应该假设我们至今拥有的所有股权——我们什么都没卖——都会消失，对吧？董事会会以某种方式对立。我们只需要为此做好准备。’我告诉她估值金额，然后说你就假设这一切都归零。

Could have been handled differently earlier for certain. I think I would expect everyone involved would say the same at this moment. And I remember asking for more information. They wouldn't share it at the time. And again, I understood, okay, this is the way that things are going. And after hanging up the call, I told my wife and said, 'We have to leave.' And she said, 'Yes, we do.' And I said, 'Just so you know, we should assume that all of the equity that we have to date, we haven't sold anything, will go away, right? The board will be adversarial in some way. We just need to be prepared for that.' I told her the amount that it was valued at and I said you just assume all of this goes to zero.

嗯。

Yeah.

还想这么做吗？她说是的。

Still want to do it? And she said yes.

这是你从一开始就共同创立的，那时已经投入了多少年？八年。

And this is something that you co-founded from the beginning and have spent at that time how many years? Eight years.

八年完全投入其中。是的。我们推迟了要孩子，这样我才能专注于这件事。OpenAI 一直是我们生活中如此重要的一部分。

Eight years totally devoted to it. Yes. We had put off children so I could really focus on this. OpenAI had been such a big part of our life.

嗯。

Yeah.

而且我认为我们都非常相信这个使命。人工智能帮助人类的潜力，我妻子热爱动物，真正帮助它们，每一个生命，那是我们真正关心的事情。只是意识到所发生的事情不对，你可以看到一个不同的世界，在那里你试图以某种方式利用它，但我甚至没有想过。我感受到的就是这是错的。

And I think we both just believe in the mission so much. The potential for AI to help humans, and my wife loves animals, to really help them, like every living being, that's something we really care about. Just realizing that it wasn't right what had happened and that you could see a different world where you try to take advantage of it in some way, but it didn't even occur to me. The thing that I just felt was this was wrong.

是啊。于是我打电话给 Sam，问他打算做什么。他说：‘嗯，我想我会去创办另一家公司。’我说：‘不，Sam，我们要创办一家新公司。’所以那天，我辞职了。我认为在人们记得的叙述中，人们忘记了那个时刻。一切都模糊在一起了。

Yeah. And so I called Sam, asked him what he was planning on doing, and he said, 'Well, I'm going to go start another company, I guess.' I said, 'No, Sam, we are going to start a new company.' And so that day, I quit. I think that in the narrative that people remember, people forget that moment. It all kind of blurs together.

我不知道故事的那部分。之前发生的事情，我认为看起来像是被解雇了，一定发生了什么可怕的事情。而改变局势的是我辞职的时候。我写了一条很简短的消息，只是说当我今天听到这个消息时。

I didn't know that part of the story. The thing that happened was prior to then I think that it looks like there's a firing, something terrible must have happened. And the thing that changed it was when I quit. I wrote a very short message just saying that when I heard the news today.

他们没想到你会辞职吗？

Were they not expecting you to quit?

我认为他们没有预料到。

I do not think they expected it.

决定是由董事会做出的吗？

Was the decision made by the board?

是的。董事会有多少人？

Yes. How many people are on the board?

所以当时，我们放任的一个问题是，那一年我们一直在失去董事会成员，每个人都有自己的合理理由。一个想竞选总统，其他人各有原因。所以董事会剩下六名成员，包括 Sam 和我。我认为对我来说，只是努力度过这一切，而董事会在那通电话中对我说的一部分是，在那些真正能用 OpenAI 做成事情的人中，我在这方面有独特的能力。

So at the time, one of the things that we had let fester was that we had been losing board members that year for reasonable reasons each of them. One wanted to run for president, other reasons for each. And so left on the board were six members including Sam and myself. I think that for me just really trying to operate through this and this is part of what the board said to me during that call is that of people who could really get things done with an OpenAI that I was uniquely capable there.

是的。所以我一直关注的是我们需要做什么？我们怎么做？如何组织人员？如何把他们团结起来？而这种推动有时本身就会引发冲突，对吧？但总是因为你在做需要做的事情来推动它前进。

Yeah. And so a lot of what I always focused on was what do we need to do? How do we do it? How do we organize the people? How do we bring them together? And that motion sometimes could cause conflict itself, right? But it was always because you're doing what you need to do to push it forward.

有道理。

Makes sense.

然后接下来发生了什么？

And then what happened next?

嗯，我们开始接到很多人的电话，让我非常惊讶的是，有这么多人联系我说：‘我不知道你下一步是否打算做什么，但我想跟你走。’

Well, so we started to get a lot of calls from people and it was truly surprising to me the number of people reached out saying, 'I don't know if you're planning on doing something next, but I want to go with you.'

公司内部还是外部？

Within the company or outside of the company?

公司内部。

Within the company.

真的吗？

Really?

是的。这真的让人感到谦卑。这不是我要求或期望的事情。

Yeah. It was truly humbling. It was not something I ever would have asked or expected.

嗯。但这确实发生了。我记得那天我说，我们有 10% 的机会拿回公司。

Yeah. But it was something that really happened. And I remember saying that day, there is a 10% chance that we get the company back.

真的吗？

Really?

不是零，但不超过 10%。

Not zero, but not more than 10%.

嗯。嗯。

Yeah. Yeah.

因为董事会掌握所有牌。他们需要决定他们认为对公司、对使命正确的方向。我离开的部分原因是因为对我来说，使命比公司实体更重要。

Because the board has all the cards. They need to decide that that's the direction they think is right for the company, for the mission. And part of why I left is because to me, the mission is bigger than the corporate entity.

对。

Right.

使命是我追求并关心的事情，我想以我认为最能实现它的形式去追求。所以那天，我的一些亲密合作者，Yakob、Chimone、Alexander 也辞职了。所以我们有了一支小队伍。

That the mission is something I'm pursuing that I care about and I want to pursue in the form that I believe can most accomplish it. So that day some of my close collaborators, Yakob, Chimone, Alexander quit as well. So we had a small band.

有点像叛变。

It's like a mutiny a little bit.

那些人就是觉得，好吧，我们看到发生了什么。这不对。我们需要去做些不同的事情。

It was the people who were just like, all right, we see what happened. This is not right. We need to go do something different.

嗯。

Yeah.

然后我们开始思考，好吧，那会是什么？老实说，能量很棒。因为我们有了一张白纸，对我们想要实现的目标有完整的愿景，我们在白板上规划，思考所有可能的组合方式。

And then we just started thinking about, okay, well what's it going to be? It was great energy, honestly. Because we had this blank slate where we had a whole vision of what we wanted to accomplish and we were charting out on a whiteboard and thinking about all the ways it could fit together.

还有这么多人联系过来。所以我第二天在 Sam 家安排了一个会议，我说：‘来吧，我们都见个面。我们会向大家展示愿景。我们会讨论事情。我们会集思广益，看看这一切将如何运作。’

And there were all these people reaching out. And so I set up a meeting the next day at Sam's house where I said, 'Come, let's all meet. We'll show everyone the vision. We'll talk through things. We'll brainstorm how it's all going to work.'

第二天就安排，真了不起。

It's amazing that it's the next day.

是啊，了不起。

Yeah, amazing.

没有时间了。我们都有精力。我们对此非常兴奋，因为这是一个真正重新思考任何我们想要的东西的机会。现在就是那个时刻，我们对那些我们任其酝酿的冲突有了所有这些反思。所以我带回来的一个价值观就是进行艰难的对话。

There's no time. We all had energy. We were so jazzed about this because it was an opportunity to really rethink anything we wanted. Now was the moment and we had all these reflections on the conflicts we let brew. So one of the values that I actually came back with was have the hard conversation.

嗯。

Yeah.

那是从整件事中得出的一个非常明确的教训。所以第二天，顺便说一句，没人真的在睡觉。大家都熬夜到很晚。可以说是靠咖啡因撑着。

And that was one very clear takeaway from the whole thing. So next day, and by the way no one's really sleeping. Like people are staying up super late. It's just a lot of caffeine energy so to speak.

嗯。第二天我们在下午一点左右开会。那天又是惊人的能量。我们花了很多时间规划新公司会是什么样子。那天晚上，执行团队过来了，那是一次非常愉快的重聚，因为我们都感觉我们在一起，我们将一起努力达成一个好的结果。第二天，与董事会进行更多对话，更多尝试找出如何度过难关。哇。我无法捕捉到这段时间的紧张程度，因为这是我们倾注了心血建立的公司。

Yeah. And the next day we have the meeting around 1 pm. That day was again amazing energy. We spent a lot of time on charting out what the new company would be. That night the executive team came over and so it was a very nice reunion as we all kind of felt like we're in this together that we're all going to try to get to a good outcome together. Next day more conversations with the board, more trying to figure out how to get through. Wow. I cannot capture the intensity of this time because it was this company that we had built that we had poured our heart and soul into.

嗯。

Yeah.

那是一个考验时刻：它能否生存？而且就在感恩节前夕。

That there was this crucible moment of is it going to survive? And it was right before Thanksgiving.

人们挤满了办公室。

People packed the office.

人们取消了回家的航班，又回来了，就在那里。并不是因为他们觉得自己能做什么积极的事，只是他们想在那里表示支持，试着待在一起。

People canceled their flights home, came back, were just there. It wasn't that they felt they could do anything active, but they just wanted to be there showing support and to try to just be together.

真是个疯狂的故事。

It's a wild story.

这绝对疯狂。所以我们都在顶层，时不时有人下楼去拿水什么的，每个人都会问：‘有消息吗？有消息吗？’你知道，会有掌声。就像每个人都真的想参与其中。我认为这是没人真正谈过的，就是人们的支持在这个时刻有多重要。那天晚上，那个周日晚上，董事会决定，嘿，我们要用另一个临时 CEO 替换 Mura 作为临时 CEO。公司炸了，直接拒绝了，说这绝对错误。这时一片混乱，人们开始离开办公室，涌出办公室。几乎就像新 CEO 来了，每个人都觉得我们得离开这里，消失。有各种故事，人们去了不同的房子，几百人挤在后院，有人演讲，那晚发生了很多疯狂的事。每个人都在想，这家公司是我还想待的地方吗？很多竞争对手在周围转悠，提供 offer，说我们会给 OpenAI 的任何人提供 offer，诸如此类。

It's absolutely wild. And so we were all up on the top floor and every so often someone would go down the stairs to get a water or something and everyone would be like, 'Any news? Any news?' You know, there'd be applause. It was just like everyone really wanted to be a part of it. And I think that is something no one's really talked about, just how much the support of the people mattered in this time. And that night, that Sunday night, the board decided that hey, we're going to replace Mura as interim CEO with a different interim CEO. And the company went wild and just rejected it and said this is absolutely wrong. And at this point it was chaos and the people started to just leave the office, streaming out of the office. It was almost like the new CEO coming and everyone's like we just need to get out of here and be gone. There were all these stories of people going over to different houses, there were like hundreds of people packed into backyards and people were making speeches and there were a bunch of crazy things that happened that night. Everyone was trying to figure out, is this company somewhere I want to be? There were lots of competitors swirling around making offers, saying we'll make offers to anyone at OpenAI, all these things.

嗯。

Yeah.

与此同时，我们试图为人们搭建一个救生艇，我们已经有了一家公司的计划。它就像一个小救生筏。我们迅速扩展，一天之内就变成了能容纳全部 770 人的规模。这是与微软合作的。微软说，是的，我们会接收所有人。微软最终是来支持的，如果那是我们想做的，他们就在那里帮助 Sam、我、Jakub 和其他辞职的人，我们试图弄清楚我们想要什么结构？我们想要一家独立公司吗？我们想要微软资助吗？我们想成为那个企业实体的一部分吗？诸如此类。所以这真的是我们的决定，要弄清楚如何推进使命。我记得我很晚才睡，因为我在安慰所有试图弄清楚发生了什么的人，说：‘我们有个计划。是这样的。’那晚有一份给董事会的请愿书，说：‘嘿，请撤销这个。’最终，95%的员工签了名，全是自发的，人们互相传阅。太多人在编辑那份请愿书的 Google 文档，结果它把 Google 文档搞崩溃了。

And then meanwhile we were trying to set up a lifeboat for people and we already had this plan for a company. It was like a small life raft. We rapidly extended to be a full size for all the 770 people in a day. And this was with partnership with Microsoft. Microsoft says yes, we will take everyone. Microsoft was there ultimately to support, like if that's what we wanted to do, they were there to help Sam and I and Jakub and the others who had quit, that we were trying to figure out what is the structure we want? Do we want an independent company? Do we want to be funded by Microsoft? Do we want to be actually part of that corporate entity? Those kinds of things. And so it's really our decision to figure out how to carry the mission forward. And I remember I went to sleep very late because I was comforting all the people who were trying to figure out what was going on and saying, 'We have a plan. Here's how it looks.' And that night there was a petition to the board saying, 'Hey, please undo this.' And in the end, 95% of employees signed and it was all self-organized, people sending it around. So many people were editing that Google Doc for the petition that it actually crashed Google Docs.

嗯。嗯。

Yeah. Yeah.

所以后来他们不得不指定一些人作为添加 Google 文档的人。这整件事太疯狂了。

And so then they had to have some people designated as the person to add to the Google doc. It was a whole crazy thing.

听起来像一场革命。

It sounds like a revolution.

确实是。就是那种感觉。

It really was. It had that feeling.

这就是你在描述的。

That's what you're describing.

就是那种感觉。百分之百是那种感觉。

That's how it felt. That's 100% how it felt.

那天早上我记得大约 5 点醒来，我查看 Twitter，看到 Mira 的一条帖子。对我来说，即使只是谈论这个，我都感到情绪激动。

And that morning I remember waking up 5:00 a.m. or so and I check Twitter and I see a post from Mira. And for me, I feel emotional even just talking about this.

是的。我记得解雇发生时，我只是感到这种距离，对吧？就像这种情绪化的事情，你只觉得‘你怎么能这样对我？’

Yeah. I remember when the firing happened, I just felt this distance, right? It was like this emotional thing and you just feel like 'how can you do this to me?'

完全正确。百分之百。

Exactly. 100%.

在那一刻我感到宽恕，对吧？我觉得好吧。

And in that moment I felt forgiveness, right? I felt like okay.

很美。

Beautiful.

是的。是的。

Yeah. It was.

所以整个故事就是 48 小时吗？

So was 48 hours the whole story?

稍微长一点。大概 72 小时左右。

A little bit longer. That was probably 72 or so.

仍然是很短的时间。

Still a tiny amount of time.

是的。但我们又多了两天来弄清楚如何回到公司，对吧？因为接下来的两天我们真的在试图找出正确的配置，因为我们显然需要在某些方面有所不同，因为我们不希望这种事再次发生。所以找到一套对旧董事会有效、对我们有效的董事会成员，这是一个过程。我们一起有了新旧成员，能够向前推进。

Yes. But we got two more days to figure out how to get back into the company, right? Because those next two days we were really trying to figure out what is the right configuration, because we clearly needed it to be different in some way, because we don't want this to happen again. And so finding a set of board members that worked for the old board, that worked for us, that was a process. And together we had like new and old and were able to move forward.

你觉得从那件事之后，公司有什么不同？

How would you say the company is different since that event?

非常不同。

Extremely.

在哪些方面？

In what way?

我们花了一段时间才真正践行，进行艰难的对话，真正弄清楚如何提前解决冲突。我不认为我们在这方面做得完美。仍然有我看到的时候，但我认为我们回来时带着一种死亡的感觉，这个企业实体不是注定的。它的成功不是预先注定的。它不会仅仅因为势头而生存。它生存并繁荣是因为人们努力工作，因为人们让它如此。所以我认为这种理解，即我们需要努力工作来实现这一点，而最可能阻碍成功的是我们自己绊倒自己。就像那些是人的因素。

It took us time to really live by, have the hard conversation and to really figure out how do you get ahead of conflict. I don't think we're perfect at it. There's still moments where I see it, but I think that the feeling that we came back with is this feeling of mortality that this corporate entity is not destined. It's not pre-ordained that it succeeds. It does not survive simply because of momentum. It survives and thrives because people work hard, because people make it be so. And so I think that that understanding that we need to work hard to achieve this and that the thing that is most likely to get in the way of success is tripping over ourselves. Like those are the human component.

因为你会说不满是关于人的事情。不是技术问题。也不是业务在做什么。

Because would you say the discontent was about human things. It wasn't technical. It wasn't what the business was doing.

是的。

Yes.

只是人的事。

It was just human stuff.

总是人的事，就像婚姻一样。

Always human stuff, like in a marriage.

是的。是的。是的。

Yes. Yes. Yes.

当然，它以各种不同的方式和形式表现出来，但基本上我认为对我来说，得到一个一致的董事会、领导团队、朝着一个方向前进的公司，有愿景并执行它，这是最重要的事情。我觉得在那一刻，我们的人生在眼前闪过，所以我认为从那以后你永远不能把它视为理所当然。

And of course, it plays out in all sorts of different ways and forms, but fundamentally I think that for me getting to an aligned board, leadership team, company that's moving in one direction, that has a vision and is executing on it, that's been the most important thing. And I felt like in that moment that we had our life flash before our eyes and so I think since then you can never take it for granted.

你现在对发生的事情有很好的理解吗？

And do you feel like you have a good understanding of what happened now?

我有。

I do.

你有。

You do.

回来的条件之一是让我们对所有事件进行独立审查。所以有一家律师事务所进来，查看了所有事情，说这里有什么需要改变的吗？一家律师事务所做了这件事，说我们理解董事会为什么这么做。这是他们的权利，但他们绝对没有必要这样做。所以有所有这些旁观者，他们把自己的解释投射到发生的事情上。我认为到目前为止，他们对所看到的都很失望。这很不舒服。这很不舒服，因为我认为这是一个我们的内部动态外泄到外部世界的时刻。

Part of the deal for coming back was to say let's do an independent review of all events. So there's a law firm that came in, took a look at everything and said like is there something that needs to be different here? And a law firm did that and said that we understand why the board did this. It was within their right, but they definitely didn't have to. And so there are all these onlookers who then project their own explanation onto what happened. I think they've all been pretty disappointed so far in terms of what they've seen. And it's uncomfortable. It is uncomfortable because I think that it's a moment where our internal dynamics spilled out into the outside world.

嗯。我在 OpenAI 的一个方法一直是，我有个口头禅：保持团队团结。

Yeah. One of my approaches throughout OpenAI has been I had this mantra of keep the band together.

嗯。

Yeah.

在我们经历的各种分歧中，我采取的方法是：我们有这么多人，观点不同，存在冲突，想往不同方向推进，不同的人认为自己应该负责等等。但这些都是聪明人，我希望把这些观点提炼成一个关于需要做什么的一致图景。对我来说，这就是我在 OpenAI 的核心工作——我们有一个宏伟的使命，但如何实现它？

And through various splits that we had, that was the approach I took is to say, hey, we've got all these people. We have different views. We're having conflict. We want to push in different directions. Different people think they should be in charge, whatever it is. But these are all smart people and I'd love to distill all these views into consistent picture of what needs to happen. And that to me is the core thing that I have feel like I do at OpenAI is try to figure out we have this grand mission but how do we do it?

嗯。

Yeah.

什么是正确的决定？在每一个这样的时刻——不仅仅是解雇事件，回顾我们与埃隆的时期，与 Anthropic 创始人的时期——我都认真思考过，我们如何共同努力才能取得好的结果。

And what is the right decision? And in each of those instances and it's really not just for the firing it's like looking back to our time with Elon to our time with the anthropic founders for each of these moments I have really gone through each of them trying to say like how can we all work together to get to a great outcome.

是的。最终，我们总是以分道扬镳告终，要么是因为对方决定不再继续，这通常会带来后续的痛苦。如果你看通常发生的情况，我们让别人讲述故事，而我们继续前进。部分原因我认为我们只是不想——如果我们是在推动公司的人，我们希望其他人能够拥有他们自己的故事。

Yeah. In the end, we've ended up with departures either because the other person decides they don't want to do it anymore and that there's usually downstream pain and that if you look at what what happens normally we let other people tell the narrative like we keep going and partly I think we just don't want to you know if we're the ones who have who are pushing on the company we want other people to be able to own their story. Yeah.

但有时人们会利用这一点来打击我们。

But then sometimes people use that to kind of whack us.

是的。告诉我你记得的关于我们那次大卫与歌利亚的对话。

Yeah. Tell me what you remember about our David and Goliath conversation.

嗯？因为我觉得这有关联。

Huh? Because I feel like this relates.

是的，我们确实讨论过，我在努力回忆具体是在什么背景下。但我认为核心是，OpenAI 就像——看看我们所在的行业，有 Google、Meta，这些都是成熟的公司，市值巨大，人员众多，用户广泛，算力充足。而我们是挑战者。我们是刚刚起步的那个。

Yeah, we did we did we did talk about it and I'm trying to remember exactly what context it was in. But I think that the core is that open AI like if you look at the industry that we're in, right? There's Google, there's Meta, and these are established companies with very, very huge market caps and many people and all sorts of users and lots of compute. And we are the challenger. Like, we are the ones who are just getting started.

起步。

Start.

没错。但不知何故，我认为因为 ChatGPT 太成功了，人们就认为我们是品类领导者，因此我们是歌利亚，是既得利益者。而在我看来，AGI 才是真正的目标，这项技术最终可能会极大地惠及这些现有巨头。我们的精神始终是——我们创办公司的全部原因就是思考如何为每个人引导一个更好的方向。在这方面，我认为我们非常像大卫，身处一个歌利亚林立的行业。我认为与埃隆的情况也是如此，对吧？埃隆·马斯克，世界上最富有的人之一，最有权势的人之一，他在起诉我们，说我们欺骗了他。当我回顾我们所做的一切时，我付出了巨大的努力，巨大的努力试图与他合作。而且我们非常透明。

That's right. And somehow I think that because trackbt has been so successful that people then think of us, well, because we're the category leader, therefore we're the Goliath. We're the establishment and where I sit I say the AGI that's the real prize and this technology is something that I think could play out such that it really benefits these incumbents and our ethos has always been the whole reason we started was to think about how can we steer this in a better direction for everyone and that that is something where I think we are very much David in an industry of Goliaths and I think the same is true with Elon, right? That Elon Musk, you know, the richest man in the world, one of the most powerful men in the world, and that he's suing us, right? Saying that we tricked him. And when I look back at what we did, I spent so much effort, I spent so much effort to try to make it work with him. And we were very transparent.

愿景上的分歧是什么？

What was the difference in vision?

控制权。我们无法与他达成协议的原因是，我们都同意我们需要比通过慈善方式获得的资本多得多的资金。所以，我们需要一个营利性实体。我们就条款进行了谈判，

So control. The reason that we could not get to a deal with him because we all agreed that we're going to need far more capital than we can get through philanthropic means. So, we're going to need a for-profit entity. And we negotiated over the terms and

他也同意这一点。

and he agreed with that as well.

哦，他同意。是的。他告诉我们时机已到，一旦我们取得第一个重大成果，这就是触发事件。好吧。我们都明白这是唯一的前进道路，然后细节谈判开始了。他要求多数股权，他需要绝对的初始控制权，他需要担任 CEO。我们无法在控制权上让步。我记得伊利亚和我在思考：我们是否应该把公司给他？想象一下，如果它真的成功了。想象你真的实现了使命，真的构建了 AGI，而只有一个人拥有绝对控制权。不管那个人是谁，你感觉好吗？

Oh, he agreed with that. Yeah. He told us the time is now to to go. He said this is the triggering event once we did our first big result. Okay. We all got that this was the only way forward and then the negotiation over the details began and he needed majority equity. He needed absolute initial control and he needed to be CEO and we could not give on the control. I remember Illy and I were thinking about it. Do we f do we give this to him? Do we give him the company? And then you start thinking about well imagine it actually works. Imagine that you actually achieve the mission. Imagine you actually build an AGI and there's one person who has absolute control over it. Doesn't matter who that person is. Do you feel good?

权力太大了。

Too much.

太大了。所以我们极力争取某种形式的否决权。但其他人都在反对，我们无法实现。所以我们说不。然后

Too much. And so we we pushed hard for an override of some form. So if everyone else was voting against and we just couldn't get it done. And so we said no. And that

在交易没有达成的那一刻，关系就恶化了吗？

Did it end on bad terms right in that moment when the deal wasn't made?

我以为会。那非常艰难，非常消耗情感，因为我们花了六周——五周半、六周——没有做任何工作，只是在谈判这些条款。我们进行了深刻的情感对话，我以为一切都毁了，但并没有，我们仍然试图一起寻找前进的道路。但他后来回来说：“嘿，要么承诺支持非营利组织，这意味着给埃隆更多董事会席位，但会是伊利亚、格雷格、埃隆、埃隆、埃隆。你们必须承诺不挖角，并且一两年内不离职之类的。”所以承诺有非常具体的条款。我们说我们会考虑。他提出的另一件事是：“你知道吗？你们为什么不直接并入特斯拉？我对特斯拉没有完全控制权。这解决了你的担忧。在这里构建 AGI，秘密进行。你们必须这样做，因为股东不会喜欢。”这是他大力推动的事情。所以你们得到营利性实体，你们得到现金——用他的话说是“现金牛”。但我们完全不想那样做。我们认真考虑过，只是觉得不对劲。

I thought it would. It was very very tough. It was a very emotionally draining moment because we'd spent six weeks, five and a half, six weeks not doing any work, just really negotiating this these terms. We had these deep emotional conversations and I thought that it was all blown up, but it wasn't and that we still tried to find a path forward together. But the thing that he came back with was to say, "Hey, either commit to the nonprofit, which means giving Elon more board seats, but it would be, you know, Ilas, Greg, Elon, Elon, Elon. You'd have to commit to a non-solicit and to not quitting for a year or two or something like that." So, there were very specific terms for the commitment. And we said, "We're we were going to think about that." The other thing that he came back with was he said, "You know what? Why don't you just merge into Tesla? I don't have full control over Tesla. It solves your concern. Build the AGI here. Do it in secret. You got to do it because the shareholders won't like it. And that was the big thing he pushed on. So you get your for-profit, you get the cash cast out the cash cow was the term that was used. And that one we had no desire to do. Like we really thought about it. We were just like just doesn't seem right.

完全是不同的业务。

Just a different business.

这是另一回事。感觉不太可能成功，而且也不是那种透明地推动 AGI、造福人类的方式。所以我们经历了多次迭代。我们讨论过是否可以重新开启营利性谈判？有点像，也许将来可以。然后我们实际上想出了一个主意。我们进行了一次非营利筹款。所以暂时这样做，但我们需要看看实际能筹到多少。在这个过程中，我们想到也许可以做 ICO，通过某种加密货币来筹集资金。我们有一些想法，关于如何将其与我们未来生产的算力或 AGI 挂钩。他对此非常兴奋。我记得有一次对话，他说：“你解决了问题。太棒了。我们将筹集 100 亿美元。”他说他支持。我们花了很多时间思考，好吧，让我们实际规划一下。这是个好主意吗？我们想做吗？我们开始有点犹豫了。然后他发了一封邮件说：“我决定不支持 ICO。”这一切发生在 2018 年 1 月左右。那时他说，我们真的需要一条筹集更多资金的途径。

It's a different thing. It's a different thing. it doesn't feel like it's likely to succeed and it's not like pushing on AGI in like this transparent way that's going to benefit people. And so we went through multiple iterations. We talked about can we open up this for-profit negotiations again? And it was kind of like yeah, maybe we could in the future. And then we actually came up with an idea. So we pursued a nonprofit fundraiser. So let's do this for now, but we need to see what we can actually raise. And along the way that we had an idea for maybe we could do an ICO and we could actually raise capital through doing some sort of cryptocoin. And we had some ideas for how that would tie to the future compute or AGI that we produce. And he got so excited about that. I remember we had a conversation where he said, "You've solved it. It's all great. We're going to raise $10 billion." He said that he was he was on board. and we spent a lot of time thinking about, okay, let's really practically chart this out. Is this a good idea? Do we want to do it? We kind of started to get iffy on it. And then he sent an email saying, you know, I've decided I will not support the ICO. This all kind of played out in January of 2018. And then at that point, he said, we really need a path to raising more money.

我们告诉他，嘿，我们有个不涉及 ICO 的新想法。而且它仍然会是，你知道，我们要讨论这种营利性实体。所有这些想法都在酝酿中，各种我们可以尝试的不同方案。他说他不相信这能行。所以，他说他退出。他的计划是在特斯拉构建 AGI，做那件事。他实际上试图挖走我们不少人。没人跟他走。但在他离开时，他给团队开了个全员会。在那个会上，他说：‘嘿，Greg 和 Sam 看到了筹集数十亿美元的路径。我不认为它能成，但如果你们看到了路径，就得去追随。所以，我支持他们。’他说他甚至会继续提供建议。他会保持参与。他支持这一切，但他只是觉得，你知道，这需要每年数十亿美元。他还接着谈到他要在特斯拉做 AGI，他不会研究安全，因为用他的话说，如果羊在立法安全，而狼没有安全，那就没有意义。指的是 Google。所以，你知道，就是那种调调。

And we told him that, hey, we have another idea that doesn't involve an ICO. And it would still be, you know, we were gonna talk about this like for-profit entity. Again, all these ideas were just swirling around and different different things we could try. And he said that he just didn't believe it was going to work. So, he said he's out. His plan was to build AGI at Tesla and do that thing. He actually tried to recruit a bunch of our people. No one went with him. But on his way out, he did this all hands with the team. And in that all hands, he said that, 'Hey, Greg and Sam see a path to raising billions of dollars. I don't think it's going to work, but if you see a path, you got to follow it. And so, I support them.' He said he'd even keep advising. He'd stay involved. He supports all of this, but he's just like, you know, this requires billions of dollars per year. He also then went to talk about that he's going to do Tesla AGI, that he wasn't going to work on safety because, in his words, if the sheep are legislating safety and the wolves aren't having safety, then there's no point. Referring to Google. So, you know, it's like that kind of tone.

当时和 Elon 一切都还好吗？

Was it all cool with Elon at that point?

当然。是的。我的意思是，我认为非常明确，每个人都理解这就是未来的发展方向。

For sure. Yes. I mean, I think it was very explicit that everyone understood this was what was happening moving forward.

好的。那下一步是什么？

Okay. And what was the next step?

嗯，我们需要想出一个能让我们筹集更多资金的结构。我们考虑过不同的想法。最终我们实施了名为 OpenAI LP 的实体，在 2019 年年中真正启动，期间我们一直让 Elon 了解进展。我们给他发了条款清单。有很多次通话，Sam 会和他讨论我们在做什么。我记得 2018 年 12 月收到他的一封邮件，那封邮件令人崩溃。他发了一封邮件，主题大概是‘需要我提醒你吗’，正文是‘OpenAI 相对于 Google 成功的机会是 0%，除非资源和执行发生巨大变化。不是 1%。我希望情况不同，但这是事实。这需要每年数十亿美元，即使我们告诉他们计划筹集的数亿美元也不够。’我记得我当时在度假，那一整天都像被打垮了一样。是的，很明显他认为我们注定失败。

Well, we needed to figure out some structure that would actually let us raise more capital. And we thought about different ideas. We ended up implementing this entity called OpenAI LP and in mid 2019 that we actually got started and along the way we kept Elon apprised of what was happening. We sent him the term sheet. There were a bunch of calls where Sam would talk to him about what we were doing and I remember getting an email from him in December of 2018. It was devastating. He sent an email saying the subject was something like 'need I remind you' and the body was something like 'OpenAI has a 0% chance of success relative to Google without a dramatic change in resources and execution. Not a 1%. I wish it were different but that's true. This is going to require billions of dollars per year even hundreds of millions of dollars which is what we told them we were planning on raising isn't enough.' And I remember I was on vacation and I just was like kind of knocked out for that whole day. Yeah, it was very clear that he thought that we were destined to fail.

是的，我认为如果他不认为你注定失败，他不会发那封邮件。所以，是的，我同意。那是他那一整年的主题。我认为随着我们开始取得成功，情况真的发生了变化，现在他试图讲述一个非常不同的故事。

Yeah, I don't think he would have sent it if he didn't think you were destined to fail. So, yeah, I agree. That had been his theme for that whole year. And I think that things really changed as we started to be successful and that I think that now there's a very different tale that he's trying to tell.

你最后一次和他说话是什么时候？

When's the last time you spoke to him?

几个月前。很愉快。

Couple months ago. Pleasant.

非常愉快。

Very pleasant.

总是很愉快。是的。

It's always pleasant. Yeah.

随着营养科学在 20 世纪中期的发展，研究人员开始认识到，现代饮食模式、品种有限、加工食品和时间限制可能导致日常微量营养素摄入出现微小但有意义的缺口。如今，大规模人群研究证实，大约 90%的成年人缺乏一种或多种必需维生素或矿物质。AG1 正是针对这些发现而配制的。每份日服提供超过 75 种维生素、矿物质和全食物来源营养素，包括维生素 A、C、E、B6、B12、镁、锌、硒和碘。这是一个经过科学开发的配方，有营养研究支持，并以身体易于利用的形式提供。因为利用率取决于吸收，AG1 还含有益生菌，以支持肠道健康和消化功能，帮助身体有效利用所摄入的营养。AG1 不含麸质或乳制品，不使用转基因成分或人工甜味剂，适用于多种饮食方式，包括植物性、原始饮食、生酮饮食和低碳水化合物饮食。提供单份包装，方便携带。在严格的质量控制下生产，每批 AG1 都经过测试，以确认其成分、营养水平和无污染物，确保每一份的可靠性。难怪 AG1 受到世界顶级运动员和专家的信赖。曾经是一个未来主义的概念——一杯中的日常营养支持——现在已成为应用科学的成果。了解更多信息，请访问 drinkag1.com/etra。今天就行动。

As nutrition science advanced through the mid-20th century, researchers began to understand that modern eating patterns, limited variety, processed foods, and time constraints could leave small but meaningful gaps in daily micronutrient intake. Today, large population studies confirm that approximately 90% of adults fall short on one or more essential vitamins or minerals. AG1 was formulated in response to those findings. Each daily serving provides more than 75 vitamins, minerals, and whole food source nutrients, including vitamins A, C, E, B6, B12, magnesium, zinc, selenium, and iodine. A scientifically developed formula backed by nutrition research and delivered in forms the body can readily access. Because utilization depends on absorption, AG1 also contains probiotics to support gut health and digestive function, helping the body make effective use of what it receives. AG1 contains no gluten or dairy, uses no genetically modified ingredients or artificial sweeteners, and is suitable for a wide range of dietary approaches, including plant-based, paleo, keto, and low carbohydrate diets. Available in single serving packets for when you're on the go. Manufactured under strict quality controls, every batch of AG1 is tested to confirm its composition, nutrient levels, and the absence of contaminants, ensuring reliability from one serving to the next. It's no wonder AG1 is trusted by the world's top athletes and experts. What was once a futuristic concept, daily nutritional support in a glass, is now the result of applied science. Learn more at drinkag1.com/etra. Do it today.

你们最终是如何与微软开展合作的？

How did you end up getting into business with Microsoft?

微软是早期合作伙伴。我们为 Dota 项目从他们那里获得了算力捐赠。所以那是合作的开端。随着时间的推移，我认为 Sam 与 Satya、Kevin Scott 保持了良好的关系。他们真的相信我们在做的事情，我认为他们真正看到了 AGI 和构建这些强大系统的愿景。

So Microsoft was an early partner. We got compute donated from them for our Dota project. So that was the very beginning of it. And over time, I think Sam had kept a good relationship with Satya, with Kevin Scott. And they really believed in what we were doing and I think really saw the vision for AGI and building these powerful systems.

他们实际上是什么时候成为合作伙伴的？他们最终资助了 LLC 吗？

And when did they actually become partners? Did they end up funding the LLC?

是的，他们向 LP 投入了 10 亿美元，然后进行了 20 亿美元的投资，又进行了 100 亿美元的投资。所以那笔资金实际上非常可观，对吧？当时没有其他人投入那么多资本，现在不同了，但当时他们在做。所以我感激他们真正帮助我们推进使命和我们正在做的事情，在此过程中，我们当然有技术许可，我们会合作，你知道目标是嘿，让我们构建这个，然后如果我们也能将这项技术部署在微软内部，为微软创造价值，那么你知道，我认为这最终会解锁我们对模型和使命的追求。

Yeah, so they funded the billion dollars into the LP and then they did a $2 billion investment and they did a $10 billion investment. So that funding was in fact very significant, right? That no one else was putting in that, you know, that quantum of capital at the time different now, but at that time they were doing it. So I was grateful for them really helping us further the mission and what we were doing and along the way that of course that we had licensing of our technology and we would partner and you know the goal was hey let's build this and then if we can also have this technology be deployed within Microsoft and create value there for Microsoft then you know again I think it is the thing that ultimately unlocks our pursuit of the models of the mission.

当资金涌入时，你们的业务发生了怎样的变化？

How did your business change when you had the influx of cash?

我会说我的理念是始终记住我们还没有赚到任何这些。我们不是一家盈利的公司，所以我们把大量资金花在算力上，那是我们的目标。所以我认为最大的变化是我们能够做梦。在某种程度上，这并不真正关乎现金。它真正关乎的是超级计算机。

I would say that my ethos is always to remember we have not earned any of this. We are not a profitable company and so we spend a ton of on compute and that that's our goal. And so I think that the big thing that changed is that we could dream. It wasn't really about the the cash in some ways. It was really about the supercomputers.

所以我们花了很多时间设计超级计算机，梦想着这些庞大的集群。我们有各种各样的设计。非常有趣的是，这笔交易的一个论点是：‘嘿，我们不知道这些模型是否有用，但至少 OpenAI 可以帮助我们更好地为深度学习设计超级计算机。’而我们非常渴望为深度学习建造出色的超级计算机。所以这不再仅仅是‘我们怎么才能得到一个数据中心？’因为 2017 年的时候，情况就是‘我们怎么才能得到一个数据中心？数字根本对不上。我们不可能达到所需的规模。’然后由于这次合作，它开始变得可能了。

And so we spent a lot of time designing supercomputers and we were dreaming of these massive clusters. We had all sorts of designs. It was very interesting because one of the theses of the deal was: 'Hey, we don't know if these models will be useful, but at the very least OpenAI can help us figure out how to build supercomputers better for deep learning.' And we were very eager to build great supercomputers for deep learning. So it became not just 'How are we going to get a data center?' because in 2017 it was very much like 'How are we going to get a data center? The numbers just don't add up. There's no way to get to the scales that we needed to.' And then it started to become possible because of this partnership.

听起来你们是一致的。

And it sounds like you were aligned.

是的。

Yes.

这很合理。

It made sense.

是的。

Yes.

ChatGPT 相信上帝吗？

Does ChatGPT believe in God?

我会说 ChatGPT 没有一个固定的人格。我们的想法是，你应该能够根据自己的偏好来塑造它。所以你可以有一个有这种信仰的版本，也可以有一个没有的。对我来说，重要的是要认识到 ChatGPT 不是一个实体，对吧？它几乎是多种智能体的集合，可以被塑造成不同的形式。

I would say ChatGPT does not have a consistent personality. The way we think about it is that you should be able to shape it to your own preferences. So you can have one that has such a belief and you can have one that does not. To me, one thing that's important is to recognize that ChatGPT is not an entity, right? But it's almost this plurality of agents that can be shaped in different forms.

它有情感吗？

Does it have emotions?

我认为这是一个开放性问题。有一个完整的领域叫做模型福利，对吧？这些模型有情感吗？它们过得开心吗？我认为随着模型变得越来越智能，我们将不得不面对这些问题。目前我们还没有一个好的理论来解释拥有情感意味着什么。你甚至能判断另一个人是否有情感吗？你相信他们有，因为你自己有，但我们没有科学的衡量标准。所以我认为随着时间的推移，我们必须认真关注我们正在创造的东西，不仅仅是功能上，还有这个模型实际上是如何运作的。

I think this is an open question. There's a whole idea of what is called model welfare, right? Are these models, do they have emotions? Are they having a good time? I think we're going to have to encounter these questions as the models get smarter and smarter. Right now we don't have a good theory of what it means to have emotions. Can you tell even that another person does? You believe that they do because you yourself have them, but we don't have a scientific measurement of it. So I think we're going to have to pay some serious attention as time goes on to figuring out what it is that we're creating, not just functionally, but how is this model actually operating?

它睡觉吗？

Does it sleep?

ChatGPT 从不睡觉，这也是它很棒的一部分，对吧？它是一个 24/7 在你口袋里的东西。它可以是一个医生，也可以是一个老师。

ChatGPT never sleeps, which is part of what makes it great, right? It's a 24/7 in your pocket. It can be a doctor. It can be a teacher.

你会说突破来得越来越快，还是变化的速度和一开始一样？

Would you say that the breakthroughs coming are coming faster and faster, or is the rate of change the same as it's been from the beginning?

指数增长在继续，所以翻倍周期是一样的，但感觉突破的速度实际上一直相当稳定。

The exponential continues, so the doubling period is the same, but it feels like the rate of breakthroughs has actually been fairly constant.

有没有你原以为 AI 会擅长但实际上并不擅长的？有没有你原以为它不擅长但实际上很擅长的？

Is there something that you thought AI would be good at that it's not? And is there something that you didn't think it would be good at that it is?

当然有。是的。我想说模型在软件方面变得非常擅长，不仅仅是编码部分，还有所有的调试、测试、端到端。对我们来说，从 5.1 到 5.2 的转变非常快，简直是天壤之别。那只是一个月、两个月的时间，从不太行到很棒。看到这个真的很令人兴奋。而我期待模型变得出色但觉得它们还没达到的领域，更多是我所说的人类判断。比如我非常兴奋的想法是，你可以有模型帮助我们成为更好的自己。你能有模型帮助调解人们的分歧吗？你知道，也许如果我们在 OpenAI 早期就有这个，结果会不同。但想想实现像世界和平这样的目标，对吧？以及如何在不同国家之间达成更好的贸易协议？比如能够在那样的层面上进行谈判。我还没有看到模型在这方面增加价值。我不认为有什么能阻止它们达到那里，但我认为我们还有一段路要走。

For sure. Yes. I'd say that the models have gotten so good at software, not just the coding part, but all of the debugging, the testing, the end-to-end of that. It really felt like a transition very quickly for us between 5.1 and 5.2. Night and day difference. And that's in a month, two months' worth of it wasn't working. It was like kind of okay to it's great. That was a real exciting thing to see. And I think where I'm excited to see the models get great, but I don't feel they are yet, is more in what I'd call human judgment. Like the kind of thing I'm very excited about is the idea that you can have models that help us all be better versions of ourselves. And can you have models that help negotiate people of differences? You know, maybe if we had this in earlier times at OpenAI, we would have different outcomes. But think about achieving outcomes like how do you have world peace, right? And how do you actually have better trade deals between different countries? Like being able to negotiate things at that level. I don't really see the models yet adding value. I don't think there's anything that stops them from getting there, but I think that we have some ways to go.

你觉得它们只是在这些领域还没有被充分训练或测试吗？这就是原因吗？因为缺少了什么。

Do you feel like they just haven't been trained up or tested enough in those realms? Is that why? Because there's something missing.

我认为我们真的需要弄清楚如何为这些任务训练它们。

I think we just really need to figure out how to train them for those tasks.

AI 最擅长处理当前信息、历史信息还是理论信息？

Is AI best with current information, historical information, or theoretical information?

我认为 AI 最擅长理论信息，因为我们真正想要的是它成为一个推理者。我们希望它能深入思考，解决新问题。所以任何历史信息、当前信息都只是附带。我们训练它们的方式确实让它们学到了很多历史事实。一个刚训练好的模型无法获取当前信息。你需要把它连接到工具上才能实时获取。但重点是，你想要一个聪明的东西。比如你希望它能应用于一个新领域，并能完成你在那个领域关心的任何事情。

I think AI is best with theoretical information in the sense that what we really want out of it is to be a reasoner. We want it to be able to think hard, solve new problems. So any of the historical information, current information that's in there is incidental. The way we train them, it does learn a lot about historical facts. A model that you've just trained has no access to current information. You need to hook it up to tools in order to be able to get that real time. But the whole point of this is you want something that's smart. Like you want it to be able to apply to a new discipline and be able to accomplish whatever it is that you care about in that discipline.

你还在写代码吗？

Are you still coding?

是的。

I am.

对于你这种职位的人来说，继续写代码很不寻常。

It's unusual for someone who is in your position to continue coding.

是的。

Yes.

你只是觉得和它有联系吗？

Do you just feel a connection to it?

我觉得我知道正确领导方向的方式是从战壕中领导，从一线领导。我发现，在这个领域，因为事情变化太快，弄清楚该做什么——比如什么是正确的决定？有时是公司的战略决策，但有时非常微观，比如我们应该如何设计这个 API 或这个特定的软件。你需要真正感受到它。你需要尝试过，并知道什么有效、什么无效。所以我很多操作方式都是这样的，比如当我们想要创建 API 时，那是我们的第一个产品。我们知道需要创建一个产品来筹集资金，但我们思考的是，我们能构建哪些不同的东西？我们如何从无到有？这实际上是我做过的最难的项目，因为它感觉完全反了，对吧？你构建产品的方式应该是你有一个想要解决的问题。没人关心背后的技术。但在我们的情况下，我们有 GPT-3，这是一种在寻找问题的技术。所以我们说，让我们通过 API 让企业可以访问它，然后他们可以决定用它做什么。那是六个月的苦干。我记得一月份我在旧金山开车，试图找到愿意尝试这项技术的公司。

I feel that the way I know the right direction to lead is by leading from the trenches, leading from the front. The thing I have found is that in this field, because things change so quickly, figuring out what to do—like what is the right decision? Sometimes it's a strategic decision for the company, but sometimes it's just very micro, like what is the way that we should architect this API or this particular piece of software. You need to really feel it. You need to have tried things and have a sense of what works and what doesn't. So a lot of how I have operated has been like, for example, when we wanted to create the API, which was our first product. We knew we needed to create a product in order to raise capital, but we thought about, well, what are the different things we could build? How do we actually go from nothing to something? It was actually the hardest project I've ever worked on because it felt totally backwards, right? The way you're supposed to build a product is that you have a problem you want to solve. No one cares about the technology behind it. But in our case, we had GPT-3, which was a technology in search of a problem. So we said, let's just make it accessible to businesses through an API, and they can decide what to do with it. It was a six-month grind. I remember in January I was driving around San Francisco trying to find companies that would be willing to try out this technology.

我的一些朋友只是出于好意参加了会议，我们在二月做了一点，这很好，因为三月全世界因新冠疫情封锁，所以没法到处跑，但就是这种从零开始构建、获取第一批测试用户的艰苦努力。我记得告诉团队我们有两个目标：一是获得第一个付费客户，也就是有人给我们第一美元；二是找到一个我们内部每天都会使用的用例。

Some of my friends took the meeting just as a favor and we did a little bit of that in February which was very good because in March the whole world shut down for COVID so there's no driving around but it was just this like intense grind of building this thing from nothing and getting the first beta users and I remember telling our team that we had two goals. One was to get the first paying customer. So the first dollar that someone would give to us. The second was get a use case that we would all use internally every single day.

是的。

Yeah.

第一个目标我们很快就实现了，实际上只花了大约两个月。但第二个目标我们直到……所以花了大约两年。这真的是一个信念的飞跃。但我做到的方式是通过构建和推动技术，我认为没有其他人能做到。你觉得如果你从构建部分休息一年，还能回来吗？还是说如果你不一直保持领先，它发展得太快就会失控？

And that first one we got very quickly. It actually took another two months or so. But that second one we didn't really get till. So that took like another two years. So it was really a leap of faith. But the way that I did it was by building and really moving the technology in a way that I don't think that anyone else could have marshaled. Do you feel like if you took a year off from the building part, could you come back or is it something that if you're not always on top of it, is it moving so fast that it would run away?

嗯，我确实这么做过，所以我有答案。2025 年是我真正全职做管理的第一年，那时我大概有 25 个直接下属，处理各种复杂问题，包括数据中心、GPU 基础设施、大型训练运行，还有非常优秀的工程师需要我支持。今年我聘请了一些非常出色的经理来帮助我，所以我实际上已经能够重新写代码了。

Well, I actually did this, so I have an answer. So 2025 was my first year where I was really full-time management and at that point I think I had like 25 direct reports and was really running a very diverse and complicated set of problems including our data centers including GPU infrastructure including large training runs and there were really really excellent engineers that I needed to support and this year I've hired some really excellent managers to help me and so I've actually been able to get back to writing code.

感觉怎么样？

How did it feel?

感觉既有生疏又有熟悉，因为这个领域的问题在某些方面其实没有变化——我们仍然在训练神经网络，神经网络仍然有前向、后向和优化步骤。所以我们所做的基本形态是相同的，但在某些方面一切都不同了。尤其是今年，工具完全变了。好在它们对其他人来说也是新的。所以大家都在共同学习，我不觉得别人都在进步，而我发现的东西对所有人来说都是旧的。

It felt there was a mix of rustiness with familiarity because the problems in this field in some ways they actually do not change because we're still training neural nets and the neural nets still are forward a backward and an optimizer step. And so that the fundamental shape of what we're doing is the same but in some ways everything is different. And this year in particular, the tools are entirely different. Now, the one nice thing is they're different for everyone else, too. And so, everyone is learning together, and I don't feel like that everyone moves and that I'm discovering something that is old to everyone.

跟我说说 ChatGPT 刚推出的时候。看到人们体验与它对话是什么感觉？那是一件全新的事。

Tell me about when chat GPT first came out. What was it like seeing people experiencing talking to it? It was a totally new thing.

当时，我们训练了 GPT-4。

At the time, we trained GPD4.

嗯。

Yeah.

我很清楚 GPT-4 会大获成功，我们应该把它做成聊天系统。我们有一个更早的模型 GPT-3.5，之前让不同的测试者试用过。我们付钱给数百名承包商使用它，但感觉并不好。我非常支持发布 ChatGPT，先把聊天的基础设施做好，然后放入真正的模型，就会很棒。真正的惊喜是，我们已经习惯了下一代模型，以至于忘记了上一代模型其实是人们从未见过的。仅仅让它变得可访问和可用，就能让它火起来，因为对你来说它已经是老生常谈了。

It was very clear to me that GPD4 was going to be a huge hit and we should turn it into a chat system and we had an earlier model GPT 3.5 that we'd actually been having different beta testers try out. We were paying like hundreds of contractors to use it. It just didn't feel like it was great. I was very supportive of releasing chatbt where it's like let's get the infrastructure out for doing a chat thing and then we'll put in the real model and it'll be great. And so the real surprise was that we'd gotten so used to this next model that we had forgotten that the previous one was actually something people had never seen before. And just making it accessible and available that that's something that would then cause it to click cuz for you it was already kind of

老生常谈。

old hat.

哇。

Wow.

是的。在 OpenAI 工作最棒的一点就是你生活在未来。

Yes. And that's one of the amazing things about working at OpenAI is you live in the future.

跟我说说 Sora。

Tell me about Sora.

啊，Sora 是技术树的一个不同分支。有趣的是，人类并不生成视频，对吧？我们不直接输出像素。我们可以创造工具来做，但我们没有像 Sora 这样工作的东西，但也许在我们的想象中有，对吧？所以可能那里有与人类智能深刻相关的东西。而且肯定有一个非常重要的东西，那就是世界模型。我认为有很多争论关于 GPT 文本模型是否拥有世界模型。如果你问它们：桌子上有个球，它滚下了桌子，现在在哪里？它们能回答。所以我认为那里面的世界知识比你预期的要多得多。但对于 Sora 来说，这一点非常突出，对吧？它理解物理，当它犯错时你确实能感觉到，比如掉下一个 iPhone，如果它没有在正确时刻碎裂，早期版本的 Sora 会完全搞错，彻底破坏体验。但我认为 Sora 代表的是创造潜力。去年我们发布图像生成功能时，人们开始把自己的家庭照片变成吉卜力风格肖像。我发现最有趣的是，人们并不想生成与现实无关、没有人情味的图像。但一旦有了人性元素——那是你和家人的照片，不是随便什么人——人们突然就与之产生联系，这真的很重要。所以我认为这些模型的能力还处于非常早期的阶段。但对于需要世界模型的应用，比如机器人技术，或者物理世界中的任何东西，拥有这样的模型会非常有帮助。

Ah, sora is a different branch of the tech tree. And the thing that's interesting is that humans do not generate videos, right? We do not directly output pixels. We can create tools to do it, but we don't have something that works quite like Sora, but maybe we do in in our imagination, right? So maybe there is something that is deeply tied to human intelligence there. And there's definitely something very important, which is a world model. And I think that there are lots of debates about do the GBD text models have a world model. And if you ask them things about, well, there's a ball on the table and it rolls off the table. Where is it now? They can answer it. So, I think there's much more world knowledge in there than you'd expect. But for Sora, it's so front and center, right? That it has an understanding of physics and you you really feel it when it makes a mistake like dropping an iPhone and that if it doesn't shatter at the right moment, like early versions of Sora would get that totally wrong and it just breaks the experience entirely. But I think that what Sora represents in my mind is creative potential. And we saw very much with last year when we released our image gen and people started to turn photos of them themselves in their family into like studio Giblly portraits. And it was this moment where the thing that I found most interesting was that people didn't really want to just generate images that were not connected to reality in some way where there's no humanity in it. But as soon as there was an element of humanity, it's a picture of you and your family. It's not just some arbitrary people. Suddenly people connect with it and it really matters. And so I think we're still at the very early days of what these models will be capable of. But I think for applications that require a world model you think about robotics, you think about yeah anything else in the physical world, having such models can be extremely helpful for that.

它是如何学习物理世界的？

How did it learn the physical world?

和我们的文本模型学习方式一样。它通过观察视频来学习，对吧？通过观察世界并尝试预测接下来会发生什么。

Same way that our text models learn. The way that it learns is through observing video, right? Through observing the world and trying to predict what will happen next.

你看到 AI 有神秘的一面吗？

Do you see a mystical side to AI?

有的。

I do.

跟我说说。

Tell me about it.

嗯，缩放定律是一个经验观察：当我们向模型投入更多算力，以特定方式扩大模型规模时，所有参数之间会形成某种关系，对吧？数据集的大小，模型性能会遵循一条确定的曲线，你可以非常精确地预测。关于缩放定律，你可以在公开论坛上看到有人说缩放定律已死，但他们错了，绝对错了。缩放定律继续势不可挡。更难的是实现它的工程，那确实困难。但对我来说，我们正在做出一些根本性的科学发现。这是对未知的探索。这是一个前沿，你可以看到，因为我们引入了深刻的科学原理和数学方法，并且看到了实证结果。所以，这几乎就像我们的首席科学家雅各布喜欢描述的那样：构建更大的模型就像建造更大的火箭，公差变得更紧。你知道，你仍然受限于同样的火箭方程和同样的引力场等等。但我认为，随着我们不断前进，我们不仅更多地了解模型，也更多地了解我们自己。我们更多地了解智能是什么。我认为这是非常深刻的。

Well, so there's this idea of scaling laws which are an empirical observation of as we pour more compute into models that are their size is increased certain ways, you get the relationship between all these parameters, right? size of the data sets that there's some deterministic curve of how good those models are and you can project it out very precisely and there's something about the scaling laws and you can go on you know public forums and see people saying oh scaling laws are dead but they're wrong they're absolutely wrong scaling laws continue unabated the thing that's harder is the engineering to actually realize that that that is difficult but to me there is some fundamental scientific discover discovery that we are making. This is an exploration into the unknown. This is a frontier and you can see it because we're bringing like deep scientific principles and mathematical approaches and you're seeing the empirical results. And so it's almost like like the way that Yakob, our chief scientist, likes to describe it is that building a bigger model is like building a bigger rocket where it's the tolerance all become more tight. And you know, you're still constrained by the same rocket equation and the same gravitational pole and all those things. But I think that what we're doing as we keep going is that we learn more not just about the models, but we learn more about ourselves. We learn more about what intelligence is. And I think that that is something that is just very profound.

OpenAI 现在有多少员工？

How many employees are there at Open AI now?

我们大约有 5000 人。

We're about 5,000 people.

他们做什么？

And what do they do?

很多不同的事情。我们大约有 1000 到 1500 人从事研究。然后是我的组织，叫做 Scaling。这部分组织共同推动深度学习的能力，试图探索其极限，工程化到芯片，建设数据中心，整个体系。我们还有大约 500 人，也许另外 1000 人专注于部署。这涉及应用，真正将 ChatGPT、Codex 等产品变为现实。所有这些之间有着深厚的合作关系，我们无法独自完成。它确实需要端到端的构建、部署，并从部署中学习。这才是我们取得进步的方式。然后还有从事销售、财务、通信、法律等职能的人，所有这些共同使这个体系运转。他是如何学会领导这么大的公司的？

A lot of different things. So we have a about maybe a thousand maybe 1500 people that are in research. And then my organization which is called scaling. And so together that part of the organization is really pushing forward what deep learning can do trying to explore its limits engineering into silicon building the data centers that whole apparatus. We have a number of people maybe it's 500 people maybe maybe another thousand who are focused on deployment. So this is on applications and actually bringing products like chachebt Codex to life and there's deep partnership across all of these like we cannot do this alone. It it really requires the end to end of building deploying and learning from what we've deployed. That is really how we make progress. And then there's people who work on sales, finance, communications, legal really a whole host of functions that all come together to make this apparatus work. How did he learn to lead such a big company?

这对我来说是新的。是新的，也不是我最初想做的事。

It is new to me. It is new and it's not what I set out to do either.

是啊。

Yeah.

在邓巴数以下，与每个人紧密联系，认识每个人，这对我们很多人来说很有吸引力。

Being under Dunar's number and having a tight connection across everyone, knowing everyone like that was something that appealed to many of us.

是啊。

Yeah.

我认为发生的事情是，使命很宏大，对吧？要真正交付一个惠及所有人的 AGI，你不能只靠一小部分职能。你真的需要弄清楚如何建立一个大企业？如何获得大量收入，因为你需要这些来资助算力，筹集资本，支付员工工资，尤其是在我们所在的这个竞争极其激烈的市场中。所以我认为，必要性迫使我适应。在 OpenAI 的每个阶段，我的运作方式都经历了学习、改变和成长。我觉得我今天的工作方式，回顾去年的我，回顾五年前的我，我会觉得我什么都不知道，真的什么都不知道。所以我认为，只有通过不断的迭代，不断回到最初，保持谦逊，不相信自己什么都知道。

And I think that what's happened is that the mission is big, right? To really be able to deliver an AGI that's going to benefit everyone, you don't just do it with a small set of functions. you really need to figure out how do you build a large business? How do you get large revenue because you need that to fund compute and to be able to raise the capital and to be able to pay the employees especially in like the insanely competitive market that we are in. And so I think that the necessity of it has forced me to then adapt. And at every single stage of OpenAI, the way that I operate, I've learned and I've changed and I've grown. And I think that there's so much of how I work today that I look back at last year me I look back at five years ago me and I'm like I knew nothing like really nothing. And so I think that it's just through constant iteration and constant I go back to that very beginning of having the humility to not believe that I know everything. Yeah.

我认为我的很多运作方式是提出很多问题，试图推动清晰度，真正理解我们为什么要这样做？这有意义吗？这一切如何配合？谁在做这件事？然后深入这些细节。

And I think that I a lot of how I operate is I ask a lot of questions and try to drive clarity and try to really understand why are we doing this? Does this make sense? How does this fit together? Who are the people who are working on this? and get into those details.

在 AI 中所有需要关注的事情中，你如何决定你的时间花在哪里？

Of all the things to focus on in AI, how do you decide where your time goes?

这总是一个非常难的问题。

This is always a very tough question.

是啊。

Yeah.

我认为这结合了直觉。多年来我成功做到的是，当有突破或突破的迹象时，我能感觉到这是重要的方向，我们需要真正关注它。有时这意味着我个人专注于它，有时意味着我们需要让整个公司投入其中。但我致力于不同的问题。这就是我的运作方式。所以现在，我把很多精力花在数据中心、机器学习工程上。同样，我有优秀的领导者帮助我，有优秀的工程师让一切成为可能。我也在帮助 Codex，花很多时间思考什么是正确的未来形态，比如人们尚未意识到但将成为每个人都在使用的产品。六个月、十二个月后，确保我们尽可能快、尽可能好地朝着那个方向前进。这通常需要调动公司上下的人员，但很大程度上是由我对模型现状和未来走向的看法驱动的，并将其与我们试图交付的价值相匹配。

And I think there's a combination of intuition. And I think that what I have successfully done over the years is kind of have a sense of when we have a breakthrough or the signs of life on a breakthrough that this is the significant direction and that we really need to focus on this. And sometimes it means I focus on it personally and sometimes it means it's just we really need to get the company on it. But I am dedicated to a different problem. But that is how I operate. And so right now I'm spending a lot of my my effort on the data centers, the ML engineering. Again, I have great leaders helping me with this and and great engineers who make it all possible. And I'm also helping out with Codex and spending a lot of time really trying to think about what is the right future shape like what's the product that people have not yet realized is going to be the product everyone's going to be using. it's six months, 12 months and make sure we're running at that as as fast as we can and as well as we can. And that's usually marshalling people from across the company, but a lot of it is driven by where I see the models are, where they're going, and mapping it to where is the value that we're trying to deliver.

你会说每隔几个月或每年，你关注的事情都与之前完全不同吗？

Would you say every few months or every year you're focused on something completely different than what you were before?

是的。

Yes.

你喜欢这样吗？有趣吗？

Do you like that? Is that fun?

我喜欢。是的。从不无聊。你认为公司五年后和十年后会是什么样子？

I do. Yes. It's never boring. Where do you see the company in 5 years and then in 10 years?

嗯，在这个领域，这些时间线很难预测，因为我认为世界在那时会发生巨大变化。一年前我开始说，2025、2026、2027 年将是这些变革性的时刻，我认为它们将是关键时刻，你真的可以看到。今年我们将拥有用于知识工作的智能体，覆盖每一个职能。

Well, in this field, those timelines are so hard to project because I think the world is going to be massively different in that time. I started saying this a year ago that 2025, 2026, 2027 are going to be these transformative moments and I think they're going to be these critical moments and I think you can really see it. This year we're going to have agents for knowledge work for every single function

那是新的。就在过去一年才出现。

and that's new. That's just been in the past year.

那是新的。我这个预测有点早。我以为去年是智能体之年，但很明显是今年。公平地说，我认为 12 月基本上变成了智能体时刻。所以，我们就在这里。第二件事是科学发现。我认为我们将开始看到科学发现真正被革命化。我认为会有一些早期采用者的组织看到这股浪潮，并真正围绕它进行重组。而我们在 OpenAI 所做的就是围绕这个进行重组。

That's just new. I was a little bit early on this prediction. I thought last year be the year of agents, but it is very clearly this year. And to be fair, I think December basically turned into the the agent moment. And so, here we are. Second thing is scientific discovery. I think we're going to start seeing scientific discovery be revolutionized in a real way. And I think that they're going to be organizations that are early adopters that see this wave coming and really retool around it. And what we're doing with OpenAI is we're retooling around this.

我们希望成为最拥抱 AI 的组织，既因为这有助于我们的运营，也因为我们想弄清楚如何将这项技术带给每个人？比如，如何利用所有这些极具加速性的工具来构建一家公司？我认为我们将在 2027 年看到这些工具开始非常广泛地普及。思考公司的形态将如何变化是很有趣的，因为现在创办公司变得如此容易。你可以用很少的人运营一家非常高效、收入可观的公司。因此，我们将迎来一个不同公司做不同事情的寒武纪大爆发。我认为这是技术不会仅仅被现有企业垄断的一种方式。这在我看来非常重要。所以，快进五年，我认为我们将被 AI 彻底改变。

We want to be the most AI forward organization both because we think it's helpful for our operation, but because we want to figure out how do we bring this technology to everyone? Like how do you build a company with all these tools that are so accelerative and I think that what we're going to see in 2027 is these tools start to be diffused very very widely. And it's very interesting to think about that the shape of what a company is is going to change because it's so easy to create companies now. You can run a company that's very effective and makes lots of revenue with a small number of people. And so we're going to have a Cambrian explosion of different companies doing different things. And this is one way that I think that it won't be a technology just accruing to the incumbents. And that's something I think is very important. So you fast forward five years, I think that we will be a very transformed company by AI.

是的。

Yeah.

而且我们将为非常困难的问题部署大量算力。AI 已经开始彻底改变生物学。例如，我曾与 Arc Institute 合作训练 DNA 模型。我们只是把用于语言训练的同一类模型，换成输入 DNA 序列。也就是说，不是文本，而是输入 ATGC，它学习预测下一个是什么。但由于它学习了底层规则，你可以将其应用于下游任务。你可以让它判断某个蛋白质序列是否有效。实际上任何生物学任务都可以，因为它学到了关于生物学的知识。我们看到像 AlphaFold 这样的系统已经取得了巨大进步，彻底改变了蛋白质折叠，而这只是冰山一角。所以我认为我们将迎来一个复兴时代，人们能够真正应对不同的疾病，然后你还可以利用我们正在创造的推理技术，我们已经看到它在物理学和数学领域取得了重大突破。

And that we will deploy lots of compute for very hard problems. AI is already starting to revolutionize biology. For example, I spent some time working with the Arc Institute to train DNA models. So you just take the exact same type of models that we train for language and instead you put DNA sequences in. So rather than text, you put in ATGC and it just learns it to predict what's next. But because it learns the underlying rules that you can actually apply it to downstream applications. You can ask it to determine if a particular protein sequence is valid or not. Like really any biological task because it learns something about biology and that we're seeing great strides with being able to you know there's systems like AlphaFold that have revolutionized protein folding and this is just scratching the surface. So I think we are going to head into a renaissance of how people are actually able to approach different diseases and then you can also bring to bear the reasoning technology that we're creating and we're seeing it already make massive breakthroughs in physics and strides in mathematics.

推理在 AI 中指的是什么？

What is reasoning as it relates to AI?

我认为推理型 AI 是指在给出答案之前消耗大量算力的 AI。就像你问最初的 ChatGPT 一个问题，

I would think of reasoning AI as an AI that expends a lot of compute before coming back with an answer. Like if you think about the original ChatGPT that you'd ask

下一个词。

next word.

没错。它会直接给出答案，有时会犯错，明显说错了什么，然后在写的过程中又说出与之矛盾的话。

Exactly. It would just give you the answer right away and sometimes it would make a mistake where it would clearly say something was wrong and then as it was writing it be like say something inconsistent with that.

是的。

Yeah.

因此，你真正想要的是一个 AI，如果问题很难，它可以暂停、思考、调用工具、上网搜索、运行一些实验，然后回来给出答案。所以，这就是推理技术的核心。现在，我认为这一切将改变人们的使用方式。我认为你真正想要的是一个始终在线的 AI，你可以随时与之交谈，而不必坐着等待答案，它应该能说：“哦，顺便说一下，五分钟前你问了我这个问题，我给了你一个答案，但我意识到那其实是错误的。这是正确的答案，以及我犯错的原因。”就像人类助手或同事一样，你们之间应该有非常流畅的来回互动。我认为我们现在已经拥有实现这一点的技术，我们需要推进它的产品化。

And so instead you really want an AI that if it's a hard problem it can pause it can think it can call tools it can go look on the internet it can run some experiments and then come back with an answer. So at the core that is what the reasoning technology is. Now, I think all this is going to change in terms of how people use it. I think that what you really want is you want an always on AI you can just talk to and that is not something you have to sit around waiting for answers and it should be able to say, "Oh, by the way, you asked me this question five minutes ago and I gave you an answer, but I realized actually that was the wrong answer. Here's like the correct thing and here's why I made the mistake." And just like a human assistant, a human coworker that you should have a very fluid interaction back and forth. And I think we now have the technology to do this and that we need to pursue the productization of it.

普通人使用 ChatGPT 的方式是，他们问一个问题，得到一个答案。你认为他们想要正确的答案，还是仅仅想要一个答案？

In the way normal people use ChatGPT, they ask it a question, they get an answer. Do you think they want the right answer or do you think they just want an answer?

我认为这取决于情况。我认为用户比许多人想象的要更成熟。所以我认为人们确实关心答案是否准确、可信和有用，而有时并没有正确答案。

I think it depends. I think users are more sophisticated than many people think. And so I think people do really care that the answer is accurate and trustworthy and useful and so sometimes there is no right answer.

是的。通常可能

Yeah. Often probably

通常非常频繁，人们把 ChatGPT 当作头脑风暴伙伴，用来产生新想法，比如约会之夜去哪里等等。所以我认为人们真正关心的是让 AI 成为他们意图和目标的放大器，或者是一个可以交流想法、提供第二意见的对象。

often very often right and that people use ChatGPT as a brainstorming partner to come up with new ideas to come up with you know where do I want to go on date night like all these kinds of things. So I think what people actually care about is having AI be an amplifier for their intent for their goals. or someone to bounce things off of and second opinion in a way.

是的。

Yes.

告诉我 OpenAI 的财务状况。它已经盈利了还是仍处于增长模式？

Tell me about the financial side of OpenAI. Is it profitable yet or still in growth mode?

仍处于增长模式。绝对没有盈利。有盈利的愿景吗？

Still in growth mode. Definitely not profitable. Is there a vision to profitability?

有的。但我认为我们面临的挑战是，我们处于一个算力将非常稀缺的世界，

There is. But I think the challenge that we have is that we are in a world where compute is going to be so scarce and

仅仅是因为无法更快地制造。是这个问题吗？

just because it can't be made faster. Is that the issue?

因为无法更快制造。是的，没错。所以目前能创造的算力存在物理上限。如果你去市场看供应链，看看台积电（TSMC）——他们负责大部分晶圆，也就是实际的硅片——再看看内存，甚至硬盘，这些基本上都售罄了，有时是 18 个月，有时更长，但供应链现在变得非常紧张，因为每个人都在试图构建算力。所以供应链的每一个环节都是限制因素。但重要的是，我们正在转向算力驱动的经济。比如现在你用 ChatGPT 问问题。人们开始用 Codex 之类的工具写代码。但我发现，如果我想让 Codex 做困难的事情，比如构建一个大型应用、一个大型网站之类的，我需要 10 个并行的 Codex 将任务分解成 10 个子任务，或者我需要 100 个。算力消耗得非常快。我们内部有一些应用，团队想要 100 块 GPU 或 1000 块 GPU 来为他们编写高度优化的代码。我们在 OpenAI 内部看到的是，对算力的争夺非常激烈。

Because it cannot. Yes. Exactly. So there's a physical boundary for how much compute can be created right now. And if you go out into the market and you look at the supply chain, you look at TSMC who does most of the actual wafers, so the actual silicon, you look at memory, you look at even hard drives, these are all basically sold out for sometimes it's 18 months, sometimes it's longer, but the supply chain is becoming incredibly crunched right now because everyone is trying to build, everyone is trying to build compute. And so it's every single piece of that supply chain and it's a limiter. But the reason it's important is because we're moving to a compute powered economy. Like right now you use ChatGPT to ask some questions. People are starting to use Codex and tools like that to write code. But one thing I have found is that if I want Codex to go do something hard like it's going to build a big application, a big website, something like that, I want 10 parallel Codexes that are breaking up the task into 10 sub pieces or I want a 100 of them. And that compute goes so fast. And we have these applications internally where there are teams who want 100 GPUs or a thousand GPUs each to write a very optimized piece of code for them. And the thing that we've seen within OpenAI is that the battles over compute have been intense.

我认为我们在经济中会看到同样的情况。人们的生产力将直接与他们能利用的算力挂钩。所以我认为算力真正成为一项基本人权是我预期在未来会看到的，因为这在很多方面将决定你的生活质量。

And I think that we're going to see the exact same thing in the economy. That people's productivity will be directly tied to how much compute they can harness. And so I think that compute really becoming a basic human right is something I expect to see in the future because that will determine your own quality of life in many ways.

这个领域会有某种突破，让未来不再是个问题吗？

Will there be some breakthrough in that area where it won't be a problem in the future?

不会。因为这里面的机制是杰文斯悖论。你听说过杰文斯悖论吗？就是如果你提高一项技术的效率，对它的需求可能会比效率提升的幅度增长更多。我认为我们看到的正是这样：每年我们都让模型高效得多。你想达到同样的智能水平，没问题，你可以用十分之一甚至百分之一的成本做到，但你会想用一千倍那么多。而且你想要的不是这个水平的智能，而是十倍或百倍的智能，因为与其解决你的小网站，为什么不建个大网站呢？所以我认为我们的野心、我们想解决的问题是无穷无尽的。因此，AI 的应用也有同样的特性。

No. Because the way that this works, there's Jevons paradox. Are you familiar with Jevons paradox? It's if you increase the efficiency of a technology, the demand for it can increase more than the efficiency gave you. And I think that's what we've seen is every single year we make the models way more efficient. You want to hit the same level of intelligence. Yeah, no problem. You can do it for 10x cheaper, 100x cheaper sometimes, but you're going to want to use a thousand times more. And you're going to want rather than this level of intelligence, you're going to want the 10x or 100x intelligence because rather than solve your small website, why not build a big website? And so I think our ambitions, the problems we want to solve are endless. And therefore, the applications of AI have the same property.

是啊。听起来如果你当初差点离开的时候真的走了，专注于制造更多 GPU 会是件好事。

Yeah. It sounds like if you would have left at the time that you almost left, focusing on making more GPUs would have been a good thing to do.

没错，没错。是的。而且我认为，顺便说一句，Sam 这些年来投入了很多精力，试图弄清楚如何获得尽可能多的算力。这不仅仅是为了 OpenAI 的竞争性质，当然对 OpenAI 的竞争态势也很重要，但我认为我们会发现每个人都想要算力，而没人能找到它，这将是件疯狂的事。

That's true. That's true. Yes. Yes. Yeah. And I think that this by the way is something that Sam I think has put a lot of effort in over the years to try to figure out how do we get the most compute possible. It's not just for OpenAI's competitive nature and I think it's important for OpenAI's competitive posture for sure but I think that the degree to which we're going to find that everyone wants compute and no one can find it like that is going to be just a wild thing.

告诉我一件你现在相信、但年轻时不相信的事。

Tell me something you believe now that you didn't believe when you were young.

我年轻时，真的相信人们总会直接、诚实地履行他们所处的任何正式职位。比如，我姐姐是耶鲁的导游，我记得跟她一起游览时，她讲了所有那些非常有趣的故事。之后我问她：‘哇，太酷了。你们怎么记住所有这些故事的？’她说：‘哦，我们就是编的。’我从未想过这有可能。你是导游，所以你会带领游览，你会讲述所有发生过的真实故事。我真的有过这种看法，并且非常看重诚实和忠实履行职责。我学到的一件事是，你不能只看人们表面说的话。你需要真正审视他们的动机，他们这么做的原因。这是一个很难接受的现实，对吧？这不是我喜欢的方式。我喜欢在我所做的一切中直截了当。如果导游编造他们讲的故事，我会很震惊。

When I was young, I think I really believed that people would always directly carry out, honestly carry out whatever formal position they're in. For example, my sister was a tour guide at Yale and I remember going on her tour and she said all these really interesting stories. Afterwards, I asked her, 'Wow, that's so cool. How do you guys keep track of all these stories?' And she said, 'Oh, we just make them up.' And it had never occurred to me that that would be possible. You're a tour guide, therefore you will tell the tour and you will tell the true stories of everything that happened. And I think I really had this perspective and really a deep value of this like honesty and sort of faithfully carrying out duties. And I think one thing I have learned is that you can't just take what people say at face value. You need to really look at the motivations, the reason that they're doing it. And that's been a tough realization, right? That's not how I like to operate. I like to be straightforward and direct in everything that I'm doing. I'd be shocked if a tour guide made up the story they were telling.

我知道。是啊。结果很难相信。

I know. Yeah. It turns out hard to believe.

结果确实如此。是的。这让你想知道还有什么你不知道的事。

It turns out. Yes. It makes you wonder what else is out there that you don't know about.

你的公司与 Stripe 的运作方式有何不同？

How does your company function different than how Stripe functioned?

因为我们最初是一家研究公司。我们的创始 DNA 非常不同。我认为研究公司，尤其是像我们这样的研究公司，一个很大的不同是我们总是认为我们现在所做的一切一年后都会过时。这就像电影制片厂。你有一部热门电影，那很棒，恭喜你。但你已经需要思考你的下一部热门电影了。仅仅因为你拍了一部好片，并不意味着下一部就一定能成功。我认为在 Stripe，从某种意义上说，安全感更强，因为我们有这样一个想法：我们要构建一个支付处理机器。你不断构建这台机器，而拥有这台机器意味着你拥有所有 momentum、用户等等。所以我认为在那样的世界里运作是非常非常不同的。

Because we started as a research company. We had a very different founding DNA. And I think one thing that is very different about a research company and a research company like ours is that we were always thinking about everything we've done right now will be obsolete in a year. It's like being a movie studio. It's like you have your hit. That's wonderful. Good for you. But you already need to be thinking about your next hit. And just because you made this good hit doesn't mean that your next one is really going to be a winner. And I think that at Stripe there was much more in some sense security in terms of well we had this idea of we want to build a payment processor machine. And you keep building the machine and the fact you have the machine you have all this momentum and users and those kinds of things. And so I think that learning to operate in that world has been very very different.

正是。

Exactly.

以你现在的了解，如果回到 OpenAI 的起点，你会如何不同地设立它，还是会完全一样？

Knowing what you know now, if you were going back to the beginning of OpenAI, how might you have set it up differently or would you do everything exactly the same way?

我对我们现在的处境非常满意，我觉得我们在每一步都做出了有原则、理性的决定，所以我很难去想反事实。有什么我会做得不同吗？明白。在微观层面，是的，我确实希望我犯过很多错误，从中学习并成长等等。但我觉得很难对我们现在的位置感到失望。

I'm very happy with the position that we are in and I feel that we made principled reasoned decisions at every step and so it's hard for me to think about the counterfactual. Is there anything that I would do differently? Understood. And I think in the micro yes, like I do wish there are so many mistakes that I've made and learned from and grown and all those things. But I think it's pretty hard for me to be disappointed with where we are.

你认为最初以非营利形式设立是唯一的方式吗？

Do you think setting it up as a nonprofit to start with was the only way to do it?

嗯，我不太清楚我们一开始能有哪些选择。我想说的是，人们忘了我们仍然是一个非营利组织，对吧？这个非营利组织拥有超过一千亿美元的 OpenAI 股权，我认为这非常重要，并且会产生巨大影响。现在，我们还没有盈利，对吧？但我们的运营，我们实际能够同时拥有一个自给自足的业务和一个前所未有的慈善部门，这让我非常兴奋。

Well, it's not obvious to me the set of choices we could have made for the beginning. I think that what I'll say is that people forget that we are still a nonprofit, right? That the nonprofit has over a hundred billion dollars of OpenAI equity and that that is something I think is very important and something I think will have a very large impact. Now, we're not profitable yet, right? But our operation, our ability to actually bring to bear both a business that is self-sustaining and also a philanthropic arm that is unprecedented like that is something I'm extremely excited about.

你认为未来几年 AI 格局会以任何方式变化吗，比如合并或公司倒闭？

Do you see the AI landscape changing over the next few years in any way either with mergers or companies going out of business?

我认为这将是一个非常动态的格局，各种令人惊讶的事情都会发生。我认为我们参与的世界是 AI 领域前所未有的，更广泛地说，整个行业也是如此。所以我预计会有更多公司试图参与进来。我认为这些工具会广泛传播。可能会有令人惊讶的合并，但我很难准确预测。

I think it's going to be a very dynamic landscape and I think that all sorts of surprising things are going to happen. Like I think that the world that we're signed up for is one of an unprecedented nature within AI and I think more broadly in the industry and so I expect that there will be way more companies trying to get involved. I think these tools are going to diffuse widely. I think there may be surprising mergers but it's hard for me to predict exactly what.

你认为 AI 领域会有很多赢家，还是只有一个或两个大赢家？

And do you think there'll be many winners in the AI space or do you think there will be one big winner or two big winners?

嗯，我希望会有很多赢家，我认为这项技术值得如此。我的信念是，知识工作、经济、提升所有人、机器人、数字空间——有太多领域需要覆盖，我认为不同公司能够在不同方面推动这一点非常重要。我们有一个愿景，如何让这真正对人类有益，就是专注于我们所说的韧性。我们考虑了很多关于安全的问题。

Well, I hope that there will be many winners and I think that the technology deserves it. My belief is that knowledge work, the economy, lifting up everyone, robots, digital space, like there's just so much surface area to cover that I think that the distribution of different companies being able to push on different aspects of this is very important. One vision that we have for how to make this really good for humanity, for humans, is to focus on what we call resilience. We thought a lot about safety.

你如何确保 AI 与人类对齐？你谈到集中化的 AI，它是最强大的东西，并且与人类价值观对齐。但这个故事并不令人满意，因为并不存在单一的人类价值观——你实际上是在谈论某个人的价值观，这很可怕。人类通过多样性生存，而不是单一文化。我们需要的是一个有韧性的系统。就像蒸汽机一样，我们制定了安全标准、安全带、碰撞测试——我们围绕技术构建社会，因为它有益，这建立了韧性。我们如何让社会变得更强，而不是更脆弱？如果有人想做坏事怎么办？例如，如果 AI 能入侵，我们就应该加强安全，让 AI 发现漏洞并打补丁。这正在发生。多样性和韧性是我们旨在支持的重要目标。

How do you make sure that AI is aligned with humans? You talk about centralized AI that is the most powerful thing and aligned with humanity's values. But that story is dissatisfying because there is no single humanity's values—you're really talking about someone's values, and that's scary. Humanity survives through diversity, not monoculture. What we need is a system of resilience. Like the steam engine, we built safety standards, seat belts, crash testing—we configure society around technology because it's beneficial, and that builds resilience. How do we make society stronger, not more fragile? What if someone wants to do something bad? For example, if AIs can hack, we should secure things and have AIs find vulnerabilities and patch them. That's happening now. Diversity and resilience are important objectives we aim to support.

似乎在世界的不同地方，可接受的事物不同。如果它只有一种观点，就不能适用于所有人。

It seems that in different parts of the world, different things are acceptable. If it has one point of view, it can't be for everybody.

是的，没错。这是一个多样化的世界。

Yes, that's true. It's a diverse place.

什么是无监督情感神经元？

What is the unsupervised sentiment neuron?

无监督情感神经元是 2017 年的一篇论文，对我来说是一个关键时刻——意识到这项技术会成功。我们在亚马逊评论上训练了一个语言模型来预测下一个字符。它不仅学会了语法，还学会了语义。它学到了一个最先进的情感分析分类器，能判断评论是正面还是负面。这比听起来难，因为你可以说‘这个产品很棒，但它完全坏了，我讨厌它’。所以你需要理解。通过这个，我们得到了最好的情感分类器。就像图像领域的 AlexNet。那一刻让我们清楚需要扩大规模。语言建模通向智能语言系统——那一刻非常清楚它会发生。

The unsupervised sentiment neuron was a paper in 2017 that for me was a crucible moment—realizing this technology is going to work. We trained a language model on Amazon reviews to predict the next character. It learned not just syntax but semantics. It learned a state-of-the-art sentiment analysis classifier, telling if a review was positive or negative. That's harder than it sounds because you can say, 'This product was great, but it was totally broken and I hated it.' So you need understanding. Through this, we got the best sentiment classifier. It was like AlexNet for images. That moment made it clear we need to scale this up. Language modeling leading to intelligent language systems—that was the moment it was so clear it was going to happen.

考虑到 AI 的工作方式，语言学是否发生了变化？它很大程度上基于对语言学的理解。

Have linguistics changed since AI, considering the way it works? It's so much based on understanding of linguistics.

是的。语言学一直感觉像是对语言运作方式的一种近似，因为语言是活的东西。语法有例外——我们拆分不定式。这就像试图为过于多样化的东西编码规则。这些模型捕捉到了我们用来生成语言的真正底层规则。

Yes. Linguistics has always felt like an approximation of how language works because language is a living thing. Grammar has exceptions—we split infinitives. It's like trying to encode rules for something too diverse. These models capture the true underlying rules we use to generate language.

AI 能理解讽刺、双关语之类的东西吗？

Does AI understand sarcasm, double entendres, things like that?

是的，你可以在 ChatGPT 上测试。它做得很好。

Yes, you can test it on ChatGPT. It does a great job.

它能理解诗歌吗？

Can it understand poetry?

我认为可以。它做得很好。我们还不擅长的是写笑话。那是我们需要更努力的地方。

I think so. It does a great job. What we're not good at yet is writing jokes. That's something we need to work harder on.

是的。

Yeah.

思考我们现在能做什么的方式：我们构建了学习生成语言底层规则的机器。它们理解很多，拥有世界知识。然后我们通过强化学习，根据它们完成任务的好坏给予奖励和惩罚。任何可以评分的东西，我们都能教给模型。我们从简单的事情开始，比如有确定性答案的数学题。对于笑话，你怎么评价它好不好？但我们一直在扩展可以评分的东西，因为模型本身可以是好的评分者。进步就是逐渐学习更复杂的东西。我们正在扩展可以评分的范围。笑话我们可能也能做，但我们没有专注于此，因为我们专注于知识工作。

The way to think about what we can do now: we build machines that learn the underlying rules that generate language. They understand a lot, have world knowledge. Then we do reinforcement learning with rewards and punishments based on how well they do tasks. Anything you can grade, we can teach the model. We started with simple things like math problems with deterministic answers. For jokes, how do you grade how good it is? But we've been expanding what we can grade because models themselves can be good graders. Progress is about gradually learning more complicated things. We're expanding what can be graded. Jokes are something we probably could do, but we haven't focused on it because we're focused on knowledge work.

但我认为有一天我们也会做到的。我很想看到。

But I think one day we'll get there too. I'd like to see it.

我也是。

Me too.

关键是我们要生成新东西。所以当这些模型学习时，它们在学习底层规则，压缩事物存在的原因。想想笑话或科学论文，就像我写化学教科书——我的目标是教你如何独立思考，教你底层的东西来推导其他一切。在深层，这些模型深刻理解放在它们面前的任何东西。人类学习的方式也差不多——你不会记住大多数展示给你的东西。

The point is we want something to generate something new. So what's going on is that when these models learn, they're learning the underlying rules, the compression of why something existed. If you think about jokes or a scientific paper, it's like writing my chemistry textbook—my goal was to teach you how to think for yourself, the underlying things to derive everything else. At a deep level, these models deeply understand whatever is put in front of them. How you as a human learn is not that different—you don't remember most things you've been shown.

你不记得在学校学的大部分东西，但你知道过程。你知道教训。你知道事情为什么那样做。

You don't remember most of the things you learned in school, but you know the processes. You know the lessons. You know the reasons that things were done.

是的。有时即使你弄错了，对过去的错误记忆有时也会引导你走向更深的真相。

Yeah. And sometimes even when you get something wrong, a mismemory of the past sometimes leads you to a deeper truth.

我们并不那么精确。

We're not so accurate.

是的。我认为这就是关键，对吧？

Yes. And I think this is the key, right?

讲故事机器。

Storying machines.

没错。你就像一台连接机器，对吧？你试图弄清楚一切是如何连接的，并且你一直在试图弄清楚。

Exactly. And you're like a connection machine, right? You're trying to figure out how is everything connected and you're constantly trying to figure that out.

而这正是这些模型的运作方式——它们不断试图弄清楚：我有这些参数。我试图更好地处理这些数据，预测我从未见过的下一个内容。所以唯一的办法就是在所有这些不同概念之间建立越来越好的连接，并在某处形成这些连接。这是那种如果你在十年前说它会这样工作，大多数人会说不可能的事情之一。

And that is exactly how these models operate is they're constantly trying to figure out I've got these parameters. I'm trying to do a better job of this data coming through predicting what's next which I've never seen before. And so the only way to do that is to build better and better connections between all these different concepts and to form those somewhere. It's one of these things that if you had said this was how it was going to work 10 years ago, most people would have said that's impossible.

是的。真正看到它这样实现并运作，是前所未有的。

Yeah. To really see it implemented and working like that is something that is unprecedented.

那么如果你反复问同一个问题，可能会得到不同的答案吗？

So if you keep asking it the same question might you get different answers?

会的，因为这些机器试图做的不是给你一个特定的答案。它们想给你整个可能性范围，对吧？我们可以把它看作一个分布。在技术层面上，我们认为这些模型所做的是学习一个概率分布。所以它基本上在所有可能的答案上学习，这个答案有多可能是正确答案，并将更多的概率质量放在更好的答案上。

You do because what these machines are trying to do is not give you one particular answer. They want to give you the whole range of possibilities, right? And that we can think of it as a distribution. Like at a technical level, we think of what these models do is they learn a probability distribution. And so it kind of learns over all possible answers, how likely is this to be like the correct answer to it and it puts more probability mass on the answers that are better.

你会说人工智能是我们见过的最大的现状破坏者吗？

Would you say AI is the biggest destructor of the status quo that we've ever seen?

我认为可能是的。我认为人工智能在很多方面会像印刷机一样，它激发和释放创造力、生产力、表达方式，有时甚至是异端邪说，对吧？有时难以想象，有时对人们预期的方式非常有破坏性，但没有人会想要拿走印刷机。

I think it could be. I think that AI will be in many ways like the printing press that it enables and unleashes creativity, productivity, expression in ways that are sometimes heretical, right? Sometimes unthinkable, sometimes very disruptive to how people have expected things to be, but no one would want to take away the printing press.

是的。但你有没有遇到很多阻力，因为它可能被视为异端？

Yeah. Have you gotten much push back though, the fact that it can be viewed as heretical?

我认为现在有一个时刻，人们正在试图决定和理解人工智能如何融入。比如他们是否从人工智能中受益？对我来说，今年我们能做的最重要的事情之一就是真正向人们展示人工智能是帮助他们的事情，他们可以建立企业，可以从中赚钱。它帮助他们的个人生活，有时在经济上，比如帮助他们省钱，帮助他们的健康，帮助教育。所有这些，我认为，是你应该拥有和想要人工智能的原因。我认为我们的目标之一就是让人们真正感受到它。

I think that there is a moment right now where people are trying to decide and understand how AI fits in. Like are they benefiting from AI? And to me, this is one of the most important things that we can do this year is to really show people that AI is something that helps them that they can build businesses, they can make money off of it. It's something that helps them in their personal life and sometimes financially like helping them save money, helps with their health, helps with education. All of these things, I think, are the reason that you should have AI and want AI. And I think that that is one of our objectives is for people to really feel it.

跟我说说你在麻省理工的经历。

Tell me about your experience at MIT.

我本科开始时在哈佛。在选择去哪里的时候，我当时对编程并不感兴趣。我以为我会学数学。我以为我会学哲学。我可能还会学化学。对于这些兴趣，哈佛似乎是一个相当不错的地方。

So I started out at Harvard as an undergraduate and when I was choosing where to go, I wasn't into programming at the time. I thought I was going to do mathematics. I thought I was going to do philosophy. I thought I'd do maybe chemistry. And for these interests, Harvard seemed like a pretty good place.

嗯。

Mhm.

但当我到了那里，我迷上了构建，为他人构建。所以我开始待在计算机协会，做一些创业项目。我记得在计算机协会的第一年，有两个高年级学生每次会议都会进行晦涩的技术辩论，我们其他人听着，心想总有一天我们也会这样。然后他们毕业了，突然我成了俱乐部负责人，我本该主持那些晦涩的技术辩论，但我想，我才大二，我还没准备好。我还想学习。

But by the time I got there, I was addicted to building and to building for other people. And so I just started being at the computer society, was working on some startups. And I remember that my first year at the computer society, there were these two seniors who would spend every meeting have an obscure technical debate and the rest of us would listen in thinking like one day that will be us. And then they graduated and suddenly I was running the club and I was supposed to have the obscure technical debates and I was like I'm a sophomore. I'm not ready for this yet. I still want to learn.

是的。

Yeah.

我一直看着街对面的麻省理工，我知道那里有很多人在编程方面比我强得多，我知道我可以向他们学习。所以我开始把所有时间都花在那里。我意识到我可能待错了地方。所以我转学了。

And I kept looking down the street at MIT where I knew a number of people who were far better at programming than I was and that I knew I could learn from. And so I started spending all of my time there. And I realized I'm probably in the wrong place. So I transferred.

这是过了多少年之后？

After how many years was this?

一年半。

Year and a half.

一年半。是的。那是一个艰难的决定。非常非常困难。

Year and a half. Yes. And it was a tough decision. It was very, very difficult.

绝对。就像我的大学室友开始维护一个天气预报，预测 Greg 某一天转学到麻省理工的可能性。在一个月的时间里，它从 99%降到 1%，再到 50%。最后，我决定去做。

Absolutely. Like my college roommates started to maintain this weather report of how likely Greg was to transfer to MIT any given day. And so over the course of a month, it went from 99% to 1% to 50%. And in the end, I decided to do it.

是的。

Yeah.

我这样做的原因，回想起来，我意识到我的许多人生决定都是通过梦想好处来做出的：想想如果这真的有效会怎样，如果它真的像我想象的那么好，在麻省理工我能学到所有这些，然后想想坏处是什么，最坏的情况是什么。最坏的情况是我回到哈佛。实际上，当我去告诉他们我要退学时，他们说听起来像是休学，他们不让你退学。所以感觉，好吧，这实际上是一个很好的机会。我可以试试看会发生什么。结果它没有我期望的那么好，甚至可能更好。

And the reason I did and looking back I've realized that many of my life decisions are made by me dreaming of the upside of thinking of well what if this really works right if it actually is as great as I think at MIT I'd be able to learn all these things and then thinking about well what is the downside what's the worst case scenario and made the worst case I come back to Harvard Harvard actually when I went to tell them I was dropping out they said sounds like a leave of absence to me like they do not let you drop out of Harvard and So it felt like, okay, this is actually a great opportunity. I can try it and see what happens. And it wasn't as good as I hoped. It was even better perhaps.

哇。

Wow.

因为我真的把时间都花在了他们的计算机俱乐部里。我构建了所有这些不同的服务，学到了很多。我在一家由那个计算机俱乐部的人创办的初创公司实习。从他们所有人那里，我学到了所有这些深奥的技术思想。麻省理工有一种文化，很难描述，我会尽力而为，但就像想象一个地方，它比其他人早 20 年体验计算机和互联网技术。然后有很多小的文化决策和人们说话和做事的方式都源于此。甚至是一些你不会想到的非常小的事情。例如，如果你坐在别人身后，一起做事情，他们必须输入密码，你会移开视线，对吧？就像这样的小事，你只是从文化上注意到人们就是这样做的。没有人说，这就是你做的。

Because I really got to spend that time just constantly in their computing club. I built all these different services, learned a lot. I interned at this startup that was built by people from that computing club. And from all of them, I just learned all these deep technical ideas. And there's this culture at MIT that is hard to describe and I'll do my best, but it's like imagine a place that experienced computers and internet technologies 20 years before everyone else did. And then there's like a lot of little cultural decisions and ways that people talk and operate that fall from it. And even very minor things you wouldn't think about. For example, if you're sitting over someone's shoulder, you know, doing something together and they have to type in their password, you look away, right? Just like little things like that that you just notice culturally this is how people do it. No one says that it's just what you do.

还有一个叫 Zephyr 的系统，是一个全校都在用的聊天系统。

And there was a system called Zephyr, which was a chat system that the whole university was on.

这个系统存在了很久，远在 Facebook 之前，对吧？它可能已经存在了二十年，甚至更久。经常如果你走进一个房间，看到一群人拿着笔记本电脑，他们突然都开始笑。你不会问他们笑什么。你拿出你的笔记本电脑，上 Zephyr，对吧？所以人们生活在一种虚拟和物理现实混合的状态中。我记得 IT 人员技术非常深厚。

And this had existed, this is far before Facebook, right? This had existed for probably two decades, maybe more. And that often if you go into a room, you see a bunch of people out with their laptops, they all start laughing. You don't ask them what's so funny. You pull out your laptop, you go on Zephyr, right? So there was this kind of mixed virtual and physical reality that people existed in. And I remember that the IT staff were deeply technical.

有各种有趣的事情。比如，MIT 因为很早就接入互联网，实际上拥有所有 IP 地址的 1/256。世界其他地方体会不到 IP 富足意味着什么——拥有的 IP 远超你能用到的。每个灯泡都有自己的 IP，所以你可以编程控制每个灯泡。这些小事你在那里能做，在其他地方做不到。

There were all sorts of funny things. For example, MIT, because they were so early on the internet, actually owns one 256th of all IP addresses. The rest of the world doesn't experience what it means to be IP rich—to have far more IPs than you could ever use. Every single light had its own IP, so you could actually program every light bulb. These little things you could do there that you couldn't do anywhere else.

嗯。

Yeah.

我从未体验过像 MIT 这样的黑客乐园。

I've never experienced quite such a hacker's playground as MIT.

太棒了。你有机会这样做并做出了选择，这看起来确实是个大胆的决定。

That's great. It's great you had the opportunity to do it and you made the choice, because it does seem like a bold choice.

这绝非易事，而且——

It was not an easy one at all, and it was one that—

但显然是正确的选择。

But clearly the right one.

谢谢。是的。

Thank you. Yes.

我记得和父母谈过。无论你告诉父母你要做什么，只要涉及离开哈佛，对他们来说都很难接受。

I remember talking to my parents. No matter what you tell your parents you're going to do, if it involves leaving Harvard, it's got to be tough on them.

他们支持吗？

Were they supportive?

当时他们很难理解，他们也这么说了，但他们说：‘这是你的决定。你坐在那个位置上，比我们有更多信息。’离开哈佛这个想法——似乎很难接受——但我们信任你。所以最终，我认为他们是支持的，但并没有让这件事变得轻松。

I think at the time it was hard for them to understand, and they said that, but they said, 'It is your decision. You have more information than us sitting where we're sitting.' The idea of leaving Harvard—that seems like a tough pill—but we do trust you. So in the end, I'd say they were supportive, but they didn't make it easy.

嗯。

Yeah.

但回想起来，这种迫使自己真正想清楚、真正理解自己参与的是什么的过程，反而让我从这次机会中获得了更多。

But in retrospect, having that forcing function to really think it through and really understand what I was signing up for is something that helped me even get more out of the opportunity.

跟我说说你的父母。

Tell me about your parents.

他们都是医生。我妈妈是精神科医生，我爸爸是眼科医生。我在北达科他州长大。很多人因为电影和电视剧听说过法戈。往北一小时，就到了大福克斯，我小时候那里大约有 6 万人。再往外 8 英里是汤普森，我小时候大约有一千人。再往外半英里，就是我家了。

They are both doctors. My mom is a psychiatrist. My dad is an ophthalmologist. I grew up in North Dakota. Many people have heard of Fargo because of the movie and TV show. If you go an hour north, you get to a town called Grand Forks, which when I was growing up was about 60,000 people. Then eight miles outside of that is Thompson, a town of about a thousand when I was growing up. Then about half a mile outside of that, you were at my house.

所以你会说那是乡村吗？

So would you say rural?

是的，是乡村。那个一千人的小镇是最近的。

It was rural. Yes. And the thousand-person town was your closest.

没错。

That's right.

我在汤普森上到四年级。幸运的是，我父母在大福克斯工作，我们有资格开放入学进入大福克斯公立学校，因为汤普森很小——学术上的机会不多。所以从五年级开始，我就在大福克斯上学了。

I went to elementary school through fourth grade in Thompson. I was lucky that my parents worked in Grand Forks and we were eligible for open enrollment into the Grand Forks public schools, because Thompson was very small—there weren't that many different opportunities academically. So from fifth grade onwards, I was in Grand Forks.

开启我未来学习和发展的一个重要事件是在六年级。我爸爸教了我一些代数。七年级有两条轨道：高级数学预代数和普通七年级数学。我和妈妈在学年开始前去见数学老师，说：‘嘿，他能直接跳级学代数吗？’老师用最居高临下的眼神看着我们说：‘每个家长都认为自己的孩子特别。我保证你的孩子在我的课上会有足够的挑战。’几周后，我根本没听课，就在教室后面玩计算器游戏。老师想刁难我，点名问：‘答案是什么？’我看了看黑板，然后说：‘回去玩游戏。’她说：‘好吧，我没什么可教你的孩子了。你应该去上代数课。’所以七年级时，我提前了一年。

One of the big things that unlocked a lot of my future learning and development was in sixth grade. My dad taught me some algebra. In seventh grade, there were two tracks: advanced math pre-algebra and normal seventh-grade math. My mom and I went to the math teacher ahead of the school year and said, 'Hey, can he just skip into algebra?' The teacher looked at us with this most condescending look and said, 'Every parent believes their child is special. I can guarantee you that your child will be plenty challenged in my class.' A couple weeks later, I had been paying no attention, just playing calculator games in the back of the room. The teacher would try to trip me up and call on me and say, 'What's the answer?' I'd look at the board and be like, 'To go back to the games.' She was like, 'All right, there's nothing for me to teach your child. You should go into the algebra class.' So in seventh grade, I was that one year advanced.

但到了八年级，我的中学没有更高阶的数学课了。我没有车，所以不能去高中。于是我做了在线独立学习。那一年我超级投入，一年内完成了三年的数学：几何、预微积分、代数 2。所以第二年，我上了高中，学微积分。我是个瘦弱的新生，和 seniors 一起上课，但我很酷。别担心。他们其实很好。很有趣。

But then eighth grade, no more math at my middle school. I didn't have a car, so I couldn't go to the high school. So instead, I did an online independent study. In that year, I just got super into it and did three years' worth of math in one year: geometry, pre-calc, algebra 2. So the next year, I was at the high school and in calculus. I was this scrawny freshman with the seniors, but I was cool. Don't worry. They were actually really nice. It was fun.

我还问学校：‘嘿，我能跳过九年级的科学课直接学化学吗？’他们同意了。他们人很好。所以我真的可以探索自己的兴趣。

I also asked the school, 'Hey, can I just skip ninth grade science and do chemistry?' They were accommodating. That was very nice of them. So I just really got to explore my own interests.

十年级时，我在考虑接下来做什么。高中没有更多数学课了。我现在有车了，因为北达科他州——你可以很早拿到车。而且北达科他大学就在那里。于是我和高中达成协议，只要我每学期在高中只上三门课，就可以在大学上任何我想要的课。这就是我的高中教育。

In tenth grade, I was considering what to do now. I didn't have any more math in high school. I did have a car now, because it's North Dakota—you can get a car pretty early. And I had the University of North Dakota right there. So I worked out a deal with my high school where I could take any classes I wanted at the university as long as I was taking just three classes a semester at the high school. That was my high school education.

太棒了。

Awesome.

太棒了。我和一位叫 Ryan Zir 的教授一起做研究。他非常棒。我们发表了一篇论文。我学了很多化学，还学了哲学。真的可以探索，把我的学术兴趣带到任何我想去的方向。

It was amazing. I did research with a professor named Ryan Zir. He was fantastic. We got a published paper out of it. I studied a bunch of chemistry. Did philosophy. Just really got to explore and take my academic interest in whatever direction I wanted.

太神奇了。真幸运学校系统支持你做自己的事。

That's amazing. So lucky that the school system was supportive of letting you do your thing.

没错。而且这很不寻常。如果我在东边一点的明尼苏达州——我们就在边界上——我的经历会非常不同。那里有标准化的高中生上大学项目，你不能上那么多课。如果你是 junior，只能上一门课；senior 可能两门。但因为没有固定项目，又有支持的教育者，我才能真正做到这一点。

That's right. And it was very unusual. If I was just slightly to the east in Minnesota—we're right on the border—I would have had a very different experience. There they have a standardized program for how high schoolers go to college, and you wouldn't be able to take so many classes. If you're a junior, you get one class; senior maybe two. But this idea of really being embedded into university because there was no program, but having supportive educators, I was able to really do it.

是的，这是反对标准化系统的一个很好的论据。

Yeah, that's a great argument against standardized systems in general.

是的。因为那样你就扼杀了例外。对我来说，一直是这样——我记得小学三年级时，我们写了一篇小作文，关于弗罗斯特的诗《林中路》。问题是你会走哪条路，是那条被踩过的还是没被踩过的。我说：‘我两条都不走。我要在树林里开辟自己的路。那似乎好得多。’我想这就是我生活的方式。

Yes. Because then you stop the exception. For me, it's always been like—I remember in elementary school, we had to write a little essay, as you would in third grade, about Frost's poem 'Two Paths in the Woods.' The question was which path would you take, the one that's trodden or not. I was like, 'I wouldn't take either of them. I would blaze my own path through the woods. That seems way better.' I think that's very much how I've lived my life.

我读到过你小时候对表演感兴趣。

I read that you were interested in acting when you were a kid.

首先，我在初中戏剧里演过男主角。所以我演得不错，谢谢。我真正喜欢表演的地方，尤其是即兴表演，是因为它关乎发现事物之间的联系。你和另一个人或一群人互动，你只是在想象，不受现实的任何约束，只受限于你的机智和快速反应的能力。所以我非常喜欢幽默。我妻子会说我很喜欢讲冷笑话。我很感激她的包容。对我来说，表演是这种价值观的体现：在连接不同概念的同时，感受想象变为现实。所以九年级时，我面临一个选择：是走数学科学路线，还是走表演路线，因为我想变得出色，想全身心投入这门技艺。我最终选择了数学，但我一直有些好奇。我的第一份带薪工作其实是一次表演。那是九年级，我还是个矮个子高中生，曼海姆蒸汽压路机乐队要来北达科他州大福克斯开演唱会，我们听说他们在找临时演员扮演锡兵。我想，“这听起来很棒，整晚能赚 100 美元，那是我能想象的一次性赚到的最多钱了，太棒了。”我到了现场，发现全是大学生，我想，“哦，糟了，我肯定要出问题。”显然他们都更适合这个角色。试镜的方式是，大家立正站好，朝一个方向走，然后选角的人互相比较意见。与此同时，所有大学生都聚在一起聊天。但我想，我的工作就是整场演唱会立正站着。所以整个试镜过程，别人都在放松，我却一直立正站着。之后他们叫了被选中的人，我没被叫到。我正要离开，他们说，“嘿，不，我们觉得你太棒了，只是你个子太小，不能给你那个角色。所以我们为你创造一个新角色。”于是他们给了我一个姜饼人的角色。我得到了那份工作，赚了 100 美元。正是通过那次不同的试镜方式，我学到了很多。

So, first of all, I was the male lead in my middle school play. So, I was very good. Thank you. And the thing I really liked about acting, like I really loved improv because it's about seeing the connection between things. And you have this interplay with another person or people and you're just imagining and you're really just constrained not by reality in any way, but just by your wits and how quickly you can make things happen. And so I really enjoy humor. My wife would say that I enjoy making dumb jokes a lot. I appreciate that she puts up with it. And that for me, acting was an expression of this value of really feeling the imagination turning into reality while connecting across different concepts. And so in ninth grade I faced what I felt was a choice between going the more math science route or going the acting route because I wanted to be great. I wanted to really dedicate myself to the craft and I ended up going the math route but I always kind of wondered. And my first paid job was actually an acting gig. So this was ninth grade. I was a small high schooler and the Mannheim Steamroller was coming to Grand Forks, North Dakota for a concert and we were told that they're looking for extras to act as tin soldiers for the show. And I was like, "This sounds great. You get paid like $100 for the whole night. That'd be like the most money I could have imagined to make in one sitting. Like that'd be so great." And I show up and it's all these college students and I'm like, "Oh man, like I'm going to be in trouble." Like clearly they are all better that are typecast for this role. And so the way that they had the audition go was they had everyone stand at attention, walk in a direction, and then the casting people would all compare notes. And in the meanwhile, all the college kids would hang out with each other and talk. But I was like, there's one job is to stand for the whole concert at attention. And so I spent the whole audition just standing at attention when everyone else was relaxing. And afterwards they called up every, you know, the people that they were selecting. I didn't get called. I was about to leave. They said, "Hey, no, actually we thought you were just so amazing. You're just too small, so we can't give you that role. So we're going to create something new for you." And so they instead gave me the part of being a gingerbread man. And so I got the job. I got paid the $100. And it was really through that that you know different audition approach.

你认为即兴表演和 AI 之间有关系吗？

Do you think that there's a relationship between improv and AI?

我认为有。

I do.

比如“是的，而且”？

Like the "Yes, and"?

是的。我认为 AI 在很多方面是终极的即兴表演伙伴。如果你去掉后训练，只看预训练模型，那就是它的工作。它的工作是：我处于这个情境，我需要定位。这里有文本、场景、图像、DNA 序列，无论输入数据是什么。它的工作就是弄清楚接下来会发生什么。它从未见过这个，完全是个谜。它需要自我定位，找出合理的东西，然后做正确的事。你看，这就是即兴表演。

Yes. Yeah. I think that AI in many ways is the ultimate improv partner. And if you take away the post-training and look at just the pre-trained model, that is its job. Its job is to say, I'm in this situation. I need to get oriented. Here's some text, here's some scenario, here's some images, here's some DNA sequence, whatever the input data is. And its job is just to figure out what comes next. And it's never seen this before. It's a total mystery. And it needs to orient itself and figure out what's a plausible thing and then do the right thing. And come on, like that's improv.

你搭建的第一个网站是什么？

What was the first website that you built?

我搭建的第一个有用户的网站是一次奇妙的经历。我有个想法，建一个我称之为“反向图灵测试”的东西。在图灵测试中，为了判断机器是否智能，一个人与另一个人和一个 AI 对话，目标是分辨哪个是人、哪个是 AI。我建了一个网站，把它变成竞技游戏：两个人类互相交谈，同时各自与一个 AI 交谈，他们不知道哪个终端对应谁。目标是在对方之前找出哪个终端是另一个人。最优策略是问一些能分辨对方是人还是机器的问题，同时自己表现得像机器，因为如果表现得太像人，对方就认不出你，你就会输。那是 2008 年，我自学编程，上了 W3Schools 教程，学了 HTML、JavaScript、PHP、CSS。我建了这个双人游戏。我坐在游戏大厅里，以防有人来，他们能有好体验，有人对战。我让网站有一个游戏大厅。我就坐在那里，开着屏幕，可怜巴巴地等着有人出现。大约两周，没人来。但有一天，最辉煌的一天来了，我从 StumbleUpon 获得了 1500 次点击。如果你记得 StumbleUpon，它是个早期随机推荐网站的服务。那天太棒了，总是同时有三四局游戏在进行。我坐在大厅里，几分钟内就有人加入。我记得那种感觉：这个东西在我脑子里，现在变成了现实，这些人都在享受我建的东西，我想继续追求这种感觉。

The first website I ever built that got users was this amazing experience. I had this idea to build what I called a reverse Turing test. So in the Turing test to determine if a machine is intelligent, you have a human who talks to another human and talks to an AI. And the goal is to figure out which of these is the human, which is the AI. So I built a website that turned this into a competitive game where you have both humans are talking to each other and they're each talking to an AI. They don't know which terminal is which. And the objective is to figure out which of your terminals is the other human before the other human does. And so the optimal strategy is to ask questions that kind of discern, am I talking to a human or a bot, but while still acting kind of bot-like? Because if you act too humanlike, then you'll lose because the other person won't figure out who you are. This was 2008, and I just taught myself how to code. I'd gone online to W3 Schools tutorials, did HTML, JavaScript, PHP, CSS, and I remember that I built this game and it's a two-player game. So, I was sitting in the lobby so in case anyone showed up, they'd have a good experience and have someone to play against. I made it so that there was like a game lobby just on my website. And so, I just sit there just waiting waiting with this open screen just sadly waiting for someone to show up. And for like two weeks, no one showed up. But then one day, it was the most glorious day. I got 1500 hits from StumbleUpon. Yeah. If you remember StumbleUpon, it was like an early like, you know, send people to random websites. And it was an amazing moment where that day there were like always three or four games going constantly. I'd sit in the lobby and someone would join within a couple minutes. And I remember this feeling that this thing was in my head and now it's in reality. And now these people are all enjoying what I built and I want to keep chasing it.

你玩得越多，会变得更擅长那个游戏吗？

The more you play, would you get better at that game?

是的，是的。我变得相当擅长，而且很有趣。我花了很多精力改进机器人。我的机器人非常初级，它的工作方式是：我保存了所有之前游戏的数据，然后在任何特定对话中，尝试匹配最相似的对话，然后用人类当时说过的话回复。对于任何已经出现过的闲聊内容，它确实有效，因为你有一个不错的回复数据库。但任何更复杂的内容，它当然就会失败。

Yes. Yes. I got quite good at it and it was interesting actually. I focused a lot on improving the bot. My bot was very rudimentary and the way that it would work is I kept a database of all the previous games and then I tried to in any particular conversation match that conversation to the most similar one and then reply with what the human had said then. And it actually kind of worked for any sort of chitchatty thing that kind of is done already, then you have a pretty good database of replies. But anything more sophisticated and of course it would just fall on its face.

你还记得这个游戏的想法是从哪里来的吗？

Can you remember at all where the idea came from for that game?

嗯，我对 AI 产生兴趣是因为读了艾伦·图灵关于图灵测试的论文。那是他 1950 年的论文《计算机器与智能》。我在建这个游戏前不久读了它，里面有很多启发性的想法。首先他问：机器能智能吗？他说，我不知道智能是什么意思，每个人都有自己的定义。所以我们来定义一个测试。他定义了图灵测试。然后他说：“你怎么可能为这个测试编程答案？你永远无法编程。因为要写下所有规则——这个人说这个，那个人说那个——太难了。”相反，你需要建造一台能自己学习答案的机器。你必须建造一台学习机器。

Well, the way that I got excited about AI was by reading Alan Turing's paper on the Turing test. So this is his 1950 paper called "Computing Machinery and Intelligence." I was reading it shortly before building this game and it had the most inspirational ideas in it because first he asked the question, can a machine ever be intelligent? And he says, look, I don't know what intelligence means. Everyone's got to have their own definition. So let's define a test. He defined the Turing test. But then he says, "How will you ever program an answer to this test? You will never program it. It's just too hard to write down all the rules of this person says this and that." Instead, you will need to build a machine that can learn its own answer to this. You have to build a learning machine.

所以，如果你能造出一台像人类小孩一样学习的机器，然后让一个人根据它的好坏行为给予奖励和惩罚，那么你就能通过这个测试。神奇的是，这正是我们一直在做的。就像艾伦·图灵在 1950 年预言的那样，我们首先会构建这些无监督学习模型，它们观察世界，内化所有知识，然后我们进行强化学习过程，通过奖励和惩罚来引导机器实现目标。我记得和联合创始人伊利亚聊过这个，我问：‘图灵是怎么知道的？’他说：‘图灵之所以是图灵，就是因为他太聪明了。’就是这样。

So, if you can build a machine that is a child machine that learns like a human child, and you can then have a human who gives it rewards and punishments as it does good things and bad things, then that's how you will solve this test. And here's the wild thing. That is exactly what we've been doing. This is like Alan Turing in 1950 projecting how we will first build these unsupervised models that learn. They sort of observe the world, have all this knowledge in them, and then we do this reinforcement learning process. We give the machine rewards and punishments in order to achieve the objective that you have in front of it. And I remember talking to my co-founder Ilya about this and I was like, 'How did Turing know?' and he said, 'There's a reason that Turing is Turing, right? He is just so smart.' And that's why, yeah.

你觉得为什么从 1950 年到现在花了这么长时间？

Why do you think it took so long from 1950 to now?

答案很简单：算力。就是算力不够。不管图灵多聪明，他没有一台能实现他测试的计算机。

Very simple answer: it's compute. There just was not enough compute. It's like no matter how smart Turing was, he did not have a computer that he could implement his test on.

早期你有没有觉得 AI 是个失败的东西？这些想法从 50 年代就有了，但一直没真正成功。

Was there a feeling in the early days for you that AI was sort of a failed thing? The ideas have been around since the 50s and it hasn't really worked.

对我来说，这始终是个巨大的谜团。图灵测试的愿景非常清晰。作为程序员，我必须理解问题的解决方案，但图灵的观点是，你可以让机器自己提出解决方案。我想，有多少问题我根本不知道如何解决，也许机器能做到。这就是我想做的。我想帮助那个东西成为现实。我记得做完那个游戏后，我去了哈佛大学，那是 2008 年，我很兴奋地想和一位自然语言处理教授做研究。

So this for me was always a great mystery. It was so clear. The Turing test vision was so clear. There's like that's the thing you need to do because as a programmer I have to understand the solution to the problem. But the point that Turing makes is that you can have a machine that comes up with its own solution to the problem. And I was like, think of how many problems I have no idea how to solve and maybe the machine could do it. That's what I want to do. I want to help that thing come into existence. And I remember after building this game, I showed up at college at Harvard and this is still 2008 and I was very excited to do research with a natural language processing professor.

你在哈佛学数学吗？

Were you studying math at Harvard?

我本来打算学数学，结果被计算机科学‘狙击’了，但我的首选一直是数学。我原本打算主修数学、化学和哲学三学位，但后来意识到我热爱用计算机构建的实用性。于是我去找那位教授，问他能否跟他做研究。他说：‘没问题。’然后他给了我一个解析树问题。我看着那些解析树，心想：‘这永远无法规模化。这不是……’

I was going to study math and I got computer science nerd sniped, but my number one thing was going to be math. I actually was originally intending to do math, chemistry, and philosophy triple major, but I ended up realizing that I just loved the practicality of building with computers. And so I went to this professor and I asked him if I could do research with him. He said, 'Yes, no problem.' And he gave me this parse trees problem. And I remember looking at the parse trees. I was like, 'This is never going to scale. This is not...'

那是什么？我不知道。

What is that? I don't know what that is.

解析树。那是老派的自然语言处理方法。想象一下，你拿句子，找出宾语和名词的位置，有点像小学里做的那种。

Parse trees. They were like an old school NLP, natural language processing approach. So the idea is, imagine you take sentences and you figure out where the object is and where the noun is. Kind of like the things that people do in elementary school.

像语言学。

Like linguistics.

对，就像用语言学的方法来做自然语言处理。你能得到一些看起来简单的东西，因为它能生成像样的句子，但永远无法扩展到像这样的对话。对我来说这很清楚。我想：‘这不是图灵说的。我去做有用的事情吧。’于是我转而研究编程语言。我很兴奋，上了一门课，想在那里做更多研究。编程语言的力量在于，编译器是一个计算机程序，它能把另一个程序变得更好。通常它把高级形式转换成机器能理解的形式，优化使其更快，真正把人的意图翻译成机器能处理的形式，让机器处理细节。因为对我来说，精神在于：我想要一台机器，能解决我解决不了的问题，以我无法达到的方式赋予我力量，带我达到新高度，不仅为我，也为所有人。所以我投身其中。我深入参与了哈佛计算机协会，为哈佛社区构建服务，比如托管电子邮件、网页托管和其他技术服务。我们为社区构建了各种网络应用。这就是我的理念。直到我在 Stripe 做初创公司、构建产品时，我仍然关注社区。我不断看到人们谈论深度学习。感觉每天上 Hacker News（工程师们发帖的网站），都能看到关于深度学习的文章。我好奇什么是深度学习？当时几乎无法搞清楚，因为你去 deeplearning.com 或.org 之类的网站，上面只说‘深度学习是一种新的 AI 方法’。

Yeah. Like linguistics as an approach for how you're going to do natural language processing. You can get some kind of simple looking stuff out of that because it'll make reasonable looking sentences, but that's never going to scale to having a conversation like this. It was just so clear to me. I was like, 'This is not what Turing was talking about. I'll go do things that are useful.' And instead, I actually got into programming languages. So, I was very excited about that. I took a class. I wanted to do more research there. And the thing about programming languages is that you have all this power of like a compiler is a computer program that takes a different program and makes it better in some way. So it usually takes it from a high-level form and puts it into a form the machine can understand, usually optimizes it so it's faster, and really takes the intent of a human and translates it into a form that the machine can then take care of some of the details. Because for me that was the spirit: I want a machine that can solve problems that I can't, that will empower me in ways that I am unable to reach, that will bring me to new heights, and not just for me but for everyone. And so I would do that. I got very into the Harvard Computer Society where we would build services for the Harvard community, so we would host email and web hosting and other technical services. We built different web applications for the community. And so this was very much the ethos that I had. And it really wasn't until I was already at Stripe where I was doing a startup and building things, but paying attention to the community. And I just kept seeing people talking about deep learning. I felt like every day if you went on Hacker News, which was this website that many engineers would post content on, you'd see something about deep learning for X. And I remember wondering, what is deep learning? And at the time it was basically impossible to figure out because you go to deeplearning.com or .org or whatever it was and it just said 'Deep learning is a new approach to AI.'

但没说它是什么。

But didn't say what it was.

没说它是什么，毫无意义。但我记得有个朋友在这个领域，我去找他聊，他开始把我介绍给领域里的其他人，结果我发现我大学里最聪明的朋友们都在这行。

Didn't say what it was. Made no sense. But I remember that I had a friend in the field and so I went and talked to that person and he started introducing me to other people in the field and I just kept getting introduced to all my smartest friends from college because they were all in the field.

当时全世界有多少人在这个领域？

How many people were in the field at that time in the world?

那是个很小的圈子，非常小。我不知道具体数字，最多一千人。而且它在快速增长，因为我了解到 2012 年有一个时刻真正引爆了当前的深度学习革命。很多方面都在为那个时刻积累，但那个时刻让许多人确信‘这里有真东西’。那就是 AlexNet 的诞生，一篇在某个基准上竞赛的图像识别论文。

It was a small community, very small community. I don't know the overall number, thousand at most. And it was rapidly growing because the thing that I learned is that there was this moment in 2012 that really unleashed the current deep learning revolution. And in many ways everything had been building up to that moment, but this was the moment that really cemented 'there's something real here' for many people. And this was the creation of AlexNet, which was an image recognition paper that competed on this benchmark.

解释一下那是什么。

Explain what that is.

这个想法是，大概在 2006 或 2008 年，斯坦福的一个实验室创建了一个竞赛，他们从网络上收集了数百万张高分辨率图像。现在看来我们会认为那是小数据集，但当时是前所未有的大规模。他们把图像分成一千个类别，由人类标注，比如这是某种猫，这是某种鸟，这是某种飞机。一千个图像类别。目标是创建一个程序，能把新图像归入这千个类别之一。比如，你能识别图像中是否有猫或狗吗？人们激烈竞争。它吸收了 40 年来计算机视觉研究的所有想法，很像那些解析树的感觉，对吧？有各种技术做边缘检测之类的，非常具体。

The idea is that, I think in 2006 or 2008, a lab at Stanford created a competition where they gathered millions of high-resolution images from across the web. At this point we would consider it a small data set. At the time it was massive and unprecedented. And they categorized these images into a thousand different categories that humans would label them and say this is a specific type of cat, this is a specific type of bird, this is a specific type of airplane. A thousand different categories of images. And the goal was to create a machine, a program that can categorize a new image into one of these thousand buckets. So, can you recognize whether there's a cat or dog in an image? And people would compete in this, and they would compete hard. And it would take all of these 40 years' worth of computer vision research ideas, very similar to the feeling of those parse trees, right? You would have these different techniques that would do edge detection and things like that that are very specific.

如果你想想识别一只猫的规则，比如你可能要找一只眼睛和另一只眼睛，但你怎么识别出有眼睛，也许再找鼻子，但方向问题会让编程这些关系变得非常复杂。

And if you think about what are the rules for recognizing a cat, it's like well maybe you look for an eye and another eye, but how do you recognize that there's an eye and maybe look if there's a nose and then but okay the orientation could make it very complicated to actually program relationships.

没错。因为你要处理这些关系，而且这个过程非常层级化——你得先看各个部分如何组合，再看这些部分如何与其他部分关联。非常复杂。所以人们在这方面一直没取得好结果，感觉完全不可能。

Exactly. Because you have to talk about these relationships and it's very hierarchical if you think about the process of you have to see how all these different pieces fit together and then how those pieces relate to other pieces. Very complicated. And so people were not getting very good results here. It felt like a total impossibility.

这是对真实世界的研究，不是纯技术性的，而是观察性的。

It was a study of the real world, it wasn't really technical, it was observational.

是的，因为人们通常会想：我认为人类是怎么做的？我脑子里的过程是什么？或者写下一些符号化的方法，告诉机器如何执行一个过程。这在某些领域非常成功，比如 90 年代我们造出了能下棋的机器，但在计算机视觉等其他领域却非常失败。

Yes, because the thing that people would do is that they would say how do I think humans do this or what's the process I have in my head or let's write down some symbolic way to tell the machine how to pursue a process. And this was very successful in some domains, right? For example, chess is one that in the '90s we built great machines to solve it, but very unsuccessful in other domains like computer vision.

国际象棋的规则只有一页纸。没错，规则简单，搜索空间小。所以你可以让计算机说：我要穷举所有可能性，这样我就能赢。顺便说，这对围棋是不够的，对吧？围棋规则简单但搜索空间巨大。所以你需要搜索之外更像人类直觉的东西。而计算机视觉则需要完全依赖直觉。这就是神经网络登场的地方。

There's only one page of rules. Exactly. For the game of chess, simple rules and a small search space. So you could have a computer that would basically just say, you know what, I'm going to look through all the possibilities and that is how I will win. Which by the way was not enough for Go, right? Go simple rules but massive search space. And so you needed something more like human intuition on top of the search. And for computer vision, you needed something that was entirely like intuition. And so that's where the neural nets came into play.

于是，由杰弗里·辛顿、伊利亚·苏茨克弗和亚历克斯·克里热夫斯基组成的研究团队创建了一个神经网络，赢得了那场比赛。而且不只是小胜，而是以巨大飞跃碾压了其他所有方法。

And so a team of researchers who are Geoffrey Hinton, Ilya Sutskever, Alex Krizhevsky created a neural net that won this competition. And it didn't just slightly win it, it just blew everything else out of the water like a massive jump.

他们的想法和其他人有什么不同吗？

Did they have a different vision of it than everyone else?

他们的方法就是神经网络。其他人都不相信。

Their approach was neural nets. No one else believed in it.

我明白了。

I see.

而且这个结果背后的内幕其实很有趣。亚历克斯·克里热夫斯基是杰夫·辛顿实验室的研究生，他在研究用于 GPU 的快速卷积核。他基本上就是在给图形处理单元（GPU）编程——也就是现在人们用来做深度学习的东西。大家都为他感到惋惜，觉得这只是一个工程项目，他只是在写那些快速的核，谁在乎呢？我们都在做酷炫的研究，他只是一个工程师，没什么价值。他在一个小型图像识别数据集上取得了一些不错的结果，但人们并不在意。但伊利亚看到了，他立刻知道这些核该怎么用。他意识到这是一个突破的契机，因为当 ImageNet 这个大数据集出现时，他觉得这是一个巨大的挑战，几乎不可能完成。如果能解决它，那将非常了不起，但你需要投入足够的算力。他看到这些核能高效利用 GPU，就说我们需要把这两者结合起来。不要把它用在你现在用的那个数据集上，ImageNet 才是关键。然后杰夫·辛顿的贡献是一个管理技巧：亚历克斯非常讨厌写论文，他有一篇综述论文要交，杰夫告诉他，每周你在数据集上提升 1%，我就把你的综述论文截止日期推迟一周。

And it's really, it's actually very funny the inside story on how that result came to be because Alex Krizhevsky was a grad student in Jeff Hinton's lab and he was working on very fast convolutional kernels for GPUs. So he basically was programming graphics processing units (GPUs), which now are what people use for deep learning. And everyone felt bad for him. It was like that's just an engineering project. He's just writing these very fast kernels. Who cares? Like, we're off doing all this cool research. He's just an engineer. That's not valuable. And he had some cool results on like this small image recognition data set and people didn't really care. But Ilya saw that and he instantly knew what to do with these kernels, right? That he realized this is a breakthrough in the making because when ImageNet had come out, this big data set, he had felt like this was this grand challenge that was just so impossible. If you could solve it, it would be so great, but you just need to be able to put enough compute into it. And he sees these kernels that were going to use a computer GPU very efficiently and he said we need to put these two things together. Don't apply it to this other data set that you're using. ImageNet is the thing. And then Jeff Hinton's contribution was a management trick because Alex Krizhevsky really hated writing papers. He had a review paper coming up and Jeff told him each week that you get a 1% improvement on the data set, I will push back the deadline on your review paper by one week.

他这样做了十几次，连续二十多次。

And he did this like a dozen times, two dozen times in a row.

是的。所以就是这样，亚历克斯不断钻研问题，数字越来越好。你看，这个领域的进步需要正确的理论、正确的目标和正确的基本方法。还需要正确的工程实现，你需要真正去实现它，努力攻克问题。即使感觉不可能，也不能放弃。还需要正确的精神，知道这是值得的，并且有在一切困难面前继续前进的渴望。我认为正是这三者结合在一起，才在那个特定时刻解锁了这个结果。他们提交了比赛，计算机视觉领域的每个人都在想：刚才发生了什么？这个不可能的问题基本上被解决了。

Yeah. And so it was just one of these things where just like Alex just kept grinding at the problem and the numbers got better and better. And so you see the way the progress in this field happens is you need the right theory. You need the right objective, the right sort of underlying approach. You need the right engineering, right? You need to really implement it. You need to push hard on the problem. And you need to not give up even when it feels impossible. And you need the right spirit, right? You need to know that like it's worthwhile. And you need to have that desire to keep going even in the face of everything else. And so I think those three things together were what unlocked this particular result in this particular moment and they submitted to the competition. Everyone in the computer vision field was like what just happened right that this impossible problem has basically now been solved.

大家的反应是什么？

What was the reaction to that?

在计算机视觉社区里，这简直是地震性的。我认为人们很快就从说“神经网络完全是死路，做神经网络就像骗子”变成了“只有神经网络才行”。

It was one of these things where within the computer vision community it was seismic, right? I think people very quickly went from saying neural nets are a total dead end like you're kind of a fraud if you're doing neural nets to only neural nets.

但确实有一段时间，做神经网络会被当成骗子。

But there was a time when you were a fraud if you were doing neural nets.

绝对是的。

Absolutely. Yes.

直到那次突破。

Until the breakthrough.

没错。而且这段历史也很有趣。有一篇 1995 年的论文，你可以在维基百科上找到，它讲述了深度学习繁荣与萧条的历史。那是在当前所有浪潮之前。如果你读它，里面说的那些话，和人们在 OpenAI 整个历史中对我说的话一模一样。1965 年，他们说这些搞神经网络的人没有新想法，他们只想造更大的计算机。这让你意识到历史是由胜利者书写的。当时有一场精心策划的运动反对这个方向，说符号系统可以，神经网络不行。而神经网络的人知道他们想要什么：更大的计算机、更深的神经网络。符号系统的人则与资助机构关系密切，他们毒化环境，说整个东西都是扯淡。这就是为什么在 70 年代它被扼杀了——他们声称它被过度炒作，有很多团队在研究，但没有结果。最后一击是有一个结果证明单层神经网络无法解决某个特定问题，因此整个领域都死了。当然，神经网络的人说：让我们做多层，我们知道该怎么做。但这就是体制扼杀了它，中央集权说了不。

That's right. And actually the history here is also very fascinating. So there's this paper you can find on Wikipedia somewhere from 1995 that talks about the history of the deep learning booms and busts. So it's really before all the current waves. And if you read it, the things they're saying in there are the exact same things people would say to us throughout the whole history of OpenAI. These neural net people in 1965 they would say these neural net people have no new ideas. They just want to build bigger computers. And so it makes you realize that history is written by the victors. That there was a very concerted campaign waged against this whole direction saying symbolic systems yes, neural nets no. And that the neural net people knew what they wanted to do. They wanted bigger computers, deeper neural nets, and that the symbolic systems people got in very cozy with funding agencies and really just poisoned the well and said this whole thing is kind of BS. And so that's what killed it for the '70s was that they claimed that it overhyped, there were all these groups working on it, that there's no results and that to put the nail in the coffin that there was this result that showed that a single layer neural net couldn't solve a particular problem. So therefore, the whole thing's dead. And of course the neural people were like but just let us go to not single layer like we know what to do. But it was all like one of these things where the establishment killed it. The sort of centralized said no.

最疯狂的是，这篇论文指出，80 年代神经网络复兴的原因是算力的民主化，对吧？以前是教授们把持着所有算力，突然之间所有博士生都有自己的电脑了，教授没法再禁止他们研究神经网络。于是人们又开始搞了。我觉得这个主题非常普遍，很有意思。它一直如此，不是过去十年才出现的新事物，而是已经存在了六七十年。

And then the crazy thing is in the 80s what this paper says is the reason neural nets came back was because of the democratization of compute, right? It went from being that you have these professors who guard all the compute to suddenly all these PhD students have their own computer and so professors can't tell them they're not allowed to do neural nets. And so suddenly people are doing it again. To me, it's so interesting that this theme is universal. It's always been true. It's not a new thing over the past 10 years. It's something that's been there for 60, 70 years.

那么被采用之后发生了什么？

So what happened after the adoption?

嗯，在 AlexNet 成果之后，计算机视觉领域之外的人仍然嗤之以鼻。他们会说：‘哦，它对计算机视觉有效，但神经网络和机器翻译毫无关系，对吧？因为图像是固定大小的，而机器翻译有可变窗口之类的东西。’2014 年，序列到序列模型出现了。它没有像 AlexNet 那样带来巨大的阶跃变化，但你能看到，没错，你只要推动它，这就是你唯一需要的东西。实际上，部门之间的壁垒被推倒了。这是一种美妙的统一。你原以为有计算机视觉、机器翻译、语音识别这些不同的领域，但不对，你只有 AI，你只有深度学习。

Well, so after the AlexNet result, people outside the field of computer vision would still pooh-pooh it. They would say, 'Oh, it works for computer vision, but neural nets have nothing to do with machine translation, right? Because you have these fixed-size images and stuff like that. And for machine translation, you have these variable windows and things like that.' In 2014, you have sequence-to-sequence. It didn't get the same massive step function you did with AlexNet, but you could just see, yeah, you're gonna push that and this is going to be the only thing you need. And what happened is you effectively had the walls between departments being torn down. It's this beautiful unification. You thought you had all these different domains of computer vision, machine translation, speech recognition, and it's nope, you just have AI, you just have deep learning.

听起来，每当边缘群体能够聚在一起时，就会发生更有趣的事情。

It sounds like anytime that the fringe groups can come together, something much more interesting can happen.

是的。

Yes.

你熟悉亚瑟·C·克拉克第一定律吗？

Are you familiar with I think it's Arthur C. Clarke's first law?

不熟悉。

No.

定律是：‘如果一个年长但杰出的科学家告诉你某件事是可能的，那他们几乎肯定是对的。但如果他们告诉你某件事是不可能的，那他们几乎肯定是错的。’

It's: 'If an elderly but distinguished scientist tells you that something is possible, they're almost certainly right. But if an elderly but distinguished scientist tells you that something is impossible, they're almost certainly wrong.'

太棒了。我喜欢这个。

That's great. I love that.

OpenAI 是在你旧金山的客厅里起步的。给我描述一下那个客厅。

OpenAI started in your living room in San Francisco. Describe the living room to me.

那是一个很大的开放空间，我们有一张黑色木质的大椭圆形桌子。有几张沙发。我有一台大屏幕电视。第一天没有白板。我记得两个研究员在争论什么。他们转身想在白板上写东西，但没有白板。我就想，我可以弄一块白板来。所以我觉得自己从第一天起就在创造价值。

It was a big open space and we had a black wood table that was this big oval shape. Had some couches. I had a big screen TV. Day one there was no whiteboard. And I remember two researchers were debating something. They turned to write something on the whiteboard. There wasn't one. And I was like, I could get a whiteboard. And so I felt like I was adding value from day one.

嗯。那么第一天房间里都有谁？

Yeah. So who was in the room that first day?

第一天在房间里的人有山姆·奥特曼、伊利亚·苏茨克维、沃伊切赫·扎伦巴。我想当时还有张维琪、帕姆·瓦加塔、约翰·舒尔曼、安德烈·卡帕西。他们中的一些人还在其他地方完成他们的博士学业。如果我忘了谁，我道歉，但那就是创始团队。创始愿景是我们有一个伟大的目标：我们真的想帮助构建 AGI，并让它成为一股积极的力量。我们当时并没有如何实现它的理论。那就是我们的起点。

In the room that first day would have been Sam Altman, Ilya Sutskever, Wojciech Zaremba was there. I think Vicki Cheung, Pam Vagata, John Schulman, Andrej Karpathy probably would have been there at the time. Some of them were finishing up their work elsewhere, their PhDs. I apologize if I'm forgetting anyone else, but that was, you know, the founding team. The founding vision was we had this great objective of we really wanted to help build AGI and have it be something that was a positive force for humanity. And we did not have a thesis on how we would do it. And so that was where we started.

在那个房间里，我们处于 AI 革命的哪个阶段？哪些是已知的？哪些是未知的？

And at what stage of the AI revolution were we in that room? What was known? What was not known?

那是 2016 年初。不算太久以前。10 年前。时光飞逝。那时我们已经经历了四年的深度学习革命。所以有一点很清楚：早期阶段，果实就挂在地上，因为你只需拿一块 GPU、一个神经网络，指向一个新问题，它就会工作，并给你惊人的结果。新的架构，在很多方面可以说是基础研究的全盛时期。单个研究人员能在几个月内提出一个新想法，验证它，发表一篇出色的论文，这将是前所未有的。它几乎能定义一个领域。那就是我们当时所处的时刻。还不是宏大工程的时刻。还不是大规模算力的时刻，因为你想要尽可能多的算力，但你实际上无法从更多 GPU 中获得更多收益，对吧？实际情况是，你有一块 GPU，而把许多 GPU 协调起来，我们并没有好的技术来从中获得良好的回报。

So this was the very beginning of 2016. Not so long ago. 10 years ago. Time flies. We had gone through at that point four years of this deep learning revolution. And so one thing that was clear was that it was like there was this early phase where the fruit was just hanging on the ground because you could just take a GPU, take a neural net, you point it at a new problem and it's going to work and it's going to get you awesome results. And so new architectures, it was kind of the heyday of basic research in a lot of ways. So individual researchers being able to come up with a novel idea in a few months, prove it out, get an awesome paper, it would be an unprecedented thing. It would kind of define a field. So that was the moment that we were in. It wasn't yet the moment of grand engineering. It wasn't yet the moment of large-scale compute because you wanted as much compute as you could get, but you couldn't really get more out of more GPUs, right? It really was that you had one GPU and orchestrating many of them together. We didn't have good techniques for how to actually get good returns from that.

我记得在早期，我参与第一个工程项目来支持研究人员。我看到两个研究员和两个工程师一起搭建。过程是这样的：研究员说‘这是我想要的系统，这是我的需求。’然后工程师们就去构建，几天后回来，把成果投到我的电视上。然后他们一行一行地过，花整个下午争论每一行代码。我记得看着那个场景，心想这永远没完没了。太慢了。

And I remember in the very early days working on the first engineering projects to support the researchers. And I observed two researchers building with two engineers. And the way that it would go is the researchers would say, 'Here's the system I want. Here are my requirements.' And then the engineers would go off and build something and come back a few days later and they would project it up on my TV. And then they would go line by line, spend a whole afternoon just debating every single line. I remember looking at that and thinking this is never going to end. It's so slow.

太耗时了。

Takes too long.

太慢了。行不通。所以，我转而亲自参与项目，与研究员紧密合作。我会说‘这里有五个想法。’他会说‘这四个不好。’我说‘太好了，这正是我想要的。’所以真的不是试图推销自己的想法，而是真正去学习对方的视角、对方的世界观，然后说‘好的，我可以从这些不同的角度来转化’，试图梳理出什么是真相，什么是现实。

Too long. Not going to work. So instead, I ended up working on the project and I would work in a very tight loop with the researcher and I would say here are five ideas. He would say, 'These four are bad.' And I say, 'Great, that's exactly what I wanted.' And so just really this just not trying to push my own ideas but really trying to learn the other person's perspective, the other person's view on the world and try to then say okay I could translate in all these different ways and to try to just tease out what truth is, what reality is.

通常研究员也是工程师吗？还是不是？

Typically are the researchers engineers as well or no?

在这个领域，他们更接近工程师，有些人确实处于两者的交叉点。例如我们的首席科学家伊利亚·苏茨克维。他真正与众不同的一点是，他两只脚都踏在两个世界里。他有深厚的理论理解，拥有优化方向的博士学位，但他也真正知道如何构建系统，并且多次做到过。所以这是一种独特的技能组合，非常有价值。我发现，对于工程师来说，要在这个领域增加价值，门槛相当高，因为这些研究员都会编程。他们可以自己构建东西。所以你必须做得比他们自己更好。这和你为医生构建东西不同，对吧？大多数医生可能不会编程。所以为他们构建东西时，做得比他们自己更好的门槛就降低了。

In this field they are much closer to engineers and there are some people who are really at that intersection. For example Ilya Sutskever who's our chief scientist. One of the things that has really distinguished him is that he really has his foot in both worlds. That he has deep theoretical understanding. He has a PhD in optimization, but he also really knows how to build systems and has done it many times. And so that it's a unique skill set. It's very valuable. And so the thing that I found was that for engineers to add value in this field, you have a pretty high bar because these researchers, they all know how to code. They can build their own things. So you have to do better than they would on their own. And that's a contrast if you're just building for doctors, say, right? Most doctors probably don't know how to code. So the bar to do better at building something for them than they could build on their own is relaxed.

那天房间里的那几个人彼此有多熟悉？

How well did the handful of people in the room that day know each other?

有一部分人是一起读过博士的，或者一起实习过。所以有些人彼此认识，有些人则是新认识的。但那时我们已经经历了一些 formative 的事件。2015 年下半年，我一直在招募，寻找这个领域最优秀的人。

So there was a subset of people who had they'd all gone through PhD programs together. or some of them had interned together. So there was a set of people who knew each other and there was a set of people who were newer to each other. But at this point we'd actually already gone through some formative events. So we had really throughout the back half of 2015 I'd been doing all the recruiting to find just who are the best people in the field.

第一次会议就是 OpenAI 的会议，还是一个后来变成 OpenAI 的聚会？

Was the first meeting the meeting of Open AI or was it a get together that turned into Open AI?

我觉得真正启动一切的第一次聚会是 2015 年 7 月的一次晚餐，当时有 Sam、Ilia、Elon 和其他几个人。问题在于：现在开始一个真正能实现 AGI 的实验室是不是太晚了？

I'd say that the very first moment that was really the get together that set things in motion was a dinner in July of 2015 and that that was with Sam, that was with Ilia, that was with Elon, that was with a handful of others. And the question there is it is it too late to start a lab that can actually really get to AGI, right?

为什么会太晚？

Why would it be too late?

感觉 DeepMind 已经占据了优势，对吧？DeepMind 拥有所有人才，作为 Google 的一部分，拥有所有算力，感觉 AGI 可能已经很近了。你还能聚集一群优秀的人真正去追求这个目标吗？

Well, it felt like Deep Mind kind of had it, right? That deep mind had all the talent that they had as part of Google, all the compute that it felt like maybe AGI was very close and can you actually get together a group of great people and really go for this.

你会说你们是在与 Google 竞争吗？

Did you start it in competition with Google would you say?

我不认为这是竞争，而是互补。我对 AI 发展的看法——这是非常根本的——是 AI 是每个人都应该参与的事情。

I don't think of it as competition but I do think of it as complimentary right that I think that my view on how AI should go and this is very foundational is that I think that AI is something that everyone deserves to participate in.

对我来说，我们要构建这些极其强大的系统，它们如何为人类服务、提升每个人，是可能发生的最重要的事情。为此做出贡献，帮助引导它朝着确保对所有人有益的方向发展，这就是我想做的事。

And to me it felt like we're going to build these incredibly powerful systems and that how they play out for humanity to uplift everyone is something that is the single most important thing that can happen and contributing to that and helping steer that in a direction that makes sure it actually is beneficial to everyone like that's the thing that I want to do.

所以对我来说，这不是关于谁有最好的基准测试的来回较量。而是关于我们如何构建系统以及与之融合的整体社会。这些东西将共同进化。那将是一个比今天好得多的世界。这不是任何一个团体能独自完成的。

And so to me it felt like it's not about the back and forth on who has the best benchmark. It's really on how do we build systems and overall society that integrates with those systems, right? That these things are going to co-evolve together. That is a much better world than the one that we have today. And that that isn't something that any one group can do on their own.

当时你知道这是一场多大的革命吗？你能预见到我们现在的位置吗？

How big of a revolution did you know it was then? Could you see where we are now then or no?

当时感觉，如果它要成功，就必须是这种感觉。而且我认为我们还没有完成。

Then it felt like if it was going to work at all, it'd kind of have to feel like this. And I think we're not done.

描述一下房间里每个人的性格、优点和缺点。

Describe the personalities and strength and weaknesses of every person in the room.

嗯，我觉得 Sam 是一个有远见的人。他允许人们梦想宏大的想法。而且他非常关心人。他总是非常乐观，认为我们总能找到解决问题的办法。所以你会觉得，这根本行不通，但他会找到解决方案。

Ah, well, I'd say that Sam, I think, is a visionary. And I think that Sam is someone who gives permission to dream big ideas. And I think he also cares a lot. I think he cares a lot about people. And I think that he is someone who is like always very optimistic about how we can find a way to configure any solution to a problem. And so I think that he is someone where you feel like hey this is never going to work. He will find a solution. But he is always someone.

它让你意识到问题是可以解决的。

It opens your mind that it can be solved.

没错。没错。

That's right. That's right.

而且可能还不是最好的解决方案。

And that maybe it's not the best solution yet.

没错。而且它并不脱离现实。我认为他是一位出色的推动者，为研究人员和这项技术进入世界提供了引导。Ilia 同样是一位有远见的人。我记得在最初的几天里，他说他一直在思考如何解决所谓的无监督学习，如何观察世界。

That's right. And it's not detached from reality. Right. It's like kind of connected to like I think he's been he's like an excellent sort of facilitator for researchers and for this overall sort of shepherding of this technology into the world. Ilia again I think is a visionary. I think he is someone who I remember in that very first couple days he said I have this idea I've been thinking about for how we can solve what was called unsupervised learning how we can observe the world.

那是什么？

What is that?

如果你想想人类婴儿是如何通过观察世界学习的，没有人告诉他们对错。

So if you think about how a human baby learns just by observing the world right there's no one saying this is the right thing.

没有人提供输入。

No one's doing input.

没错，它就这么发生了。对我来说，这总是一个疯狂的概念：机器怎么可能在没有被告知做得好不好的情况下学习？

Exactly it just happens rights and this was always to me this always felt like a crazy concept of how can the machine ever learn without someone telling it whether it's doing a good job or not.

但我们后来解决了。我记得他有很多想法，关于如何真正将其应用到机器中。

But we figured it out. I remember that he had a lot of ideas on how to really push it into machine.

即使在那个房间里，会有人认为这太遥远了吗？

Even in that room, would there be people who think that's too far?

在那个房间里，我们立刻开始写下想法，能量是显而易见的。大家聚到那个房间的步骤之一是 2015 年 11 月的一次外出活动。我绘制了整个领域或所有最优秀的人的图谱，不断问别人认识谁很棒，然后他们把我介绍给别人。我不断被介绍给一个叫 Wojciech 的人，我想，好吧，Wojciech 可能是我应该争取的人。

So in that room, we immediately started trying to write down ideas and that the energy was just palpable. One of the steps along the way to everyone coming to that room was this offsite in November of 2015. So I mapped out the whole field or all the best people and kind of been asking people for who do you know who's great and then they would introduce me to people. And so I just kind of kept track of I kept getting introduced to this guy called Voyche. I was like all right Voych's probably someone I should go after.

我们缩小到一群人，但他们都在问，还有谁加入？我感兴趣，但还有谁？我该如何打破僵局？我问 Sam 该怎么办，他建议把所有人带到一个外出活动。那时我非常感谢 John Schulman，他说他会加入。所以我至少不是唯一一个承诺的人，可能还有一两个人也在，但这群人还没有凝聚成一个团队。

And we had narrowed down to a set of people but they were all kind of like okay well who else is joining? Like I'm interested but who else is in? You're like how do I collapse this? And I asked Sam what to do and he suggested to bring everyone to an offsite. And at this point actually I remember was very grateful to John Schulman who had said that he would be in. So I was at least not the only one who had committed and I think maybe there was one or two others who were kind of there but it really was a group of people who had not yet coalesced into a team.

我们把所有人都带了出来。我们在我的公寓里集合，上了巴士，开车到纳帕。就在那一天，每个人都合拍了。能量非常流畅，就像人类形态的心流状态。我记得我们在一个活动挂图上写下了计划，一个三步计划。第一步是解决 RL，即强化学习，从奖励和惩罚中学习。第二步是解决 UL，即无监督学习，观察世界并吸收信息。第三步是逐渐学习更复杂的东西。这实际上就是我们十年来一直在做的事情。这真的很疯狂，对吧？我们真的设定了这个愿景。

And we brought everyone out. We were in my apartment. We got into the bus. We drive up to Napa and it was just this day where everyone clicked, right? That the energy was just so smooth, right? There's that flow state in human form. And I remember that we wrote up on this flip chart. And I have a picture of this flip chart. The plan, there's a three-step plan. Step one was solve RL, which is reinforcement learning, that is learning from these rewards and punishments. Two is solve UL, that is unsupervised learning, that is observe the world and just absorb the information. And then three is gradually learn more complicated in quotes. And this actually is what we've been doing for a decade. It's actually crazy, right? It's like we really set out the vision.

这与最初的愿景非常吻合。

It's very tight to the original vision.

这只是原始愿景的成长。

It's just the growth of the original vision.

是的，确实如此。当我回顾我们所做的一切，都是为了同一个目标，采用几乎相同的技术方法和相同的核心理念。

It is. It really is. When I look at everything we have done, it has been in service of the same goal with really the same almost technical approach and the same ethos underlying it.

房间里有没有其他人有特别的专长或观点，与其他人的不同，值得描述一下？

Any of the other people in the room who had particular either expertise or views that were different than the others that are worth describing?

是的，我会说 Ilya，我仍然和他密切合作，他也是一个独特的人物。他非常擅长产生想法，对任何问题都能提出非常有创意的点子，而且他完全不会执着于自己的想法。

Yeah, I'd say Ilya and I still work very closely with him is also a unique character. He is extremely good at idea generation and he will come up with very creative ideas to any problem and then he is also someone who is not at all attached to his own ideas.

对。因为如果你容易生气，那就很难保持那种创造力。

Right. Because if you're someone who's going to take offense, then it's really hard to be that generative.

嗯，如果你想让结果尽可能好，就不能全是你的想法。

Well, if you want it to be as good as it can be, they can't all be your ideas.

没错。我通常和他合作时会思考，好吧，我们应该考虑的边界是什么？这里是我们不想去的地方，或者这里是我们需要达到的目标。然后想法生成过程就会形成一个很好的飞轮效应。

That's right. What I usually do when working with him is really think about, okay, well, what are the bounds on what we should think about, right? Here are the places that we don't want to go or here's kind of the place that we need to end up. And then the idea generation process just ends up with this great flywheel.

你是怎么最终加入 Stripe 的？

How did you end up at Stripe?

在麻省理工期间，我做了更多项目，也尝试了更多创业。每次我都觉得自己学到了一个不该做的事。最终我觉得自己知道得足够多，可以成功创业了，但缺少一个关键要素——一个想法。

So while I was at MIT, I was building more and I did more startups. And from each one, I felt like I learned another thing not to do. And eventually I kind of felt like I knew enough that I could be successful at doing a startup, but I was missing a key component, which is having an idea.

嗯。

Yeah.

我当时在模仿那些参加过计算机俱乐部的朋友，他们创办了自己的初创公司，在 Gundam Masters 项目中想出了很酷的点子。我就想，显然这就是你要做的。你需要去读研究生。你看到了这条路。

And I was pattern matching off of my friends who had been in the computer club, started their own startup, the Gundam Masters program, come up with a cool idea there. I was like, well, clearly that's what you need to do. You need to go to grad school. You saw the path.

是的。但我看到的那条路其实太常规了，对吧？常规路线就是在博士项目里，你发明点什么，然后把技术变成初创公司。我记得当时觉得，好吧，我 21 岁，太年轻了，做不了真正的大事。这条路是唯一可能的。但至少我可以开始认识一些创业的人，因为我上过创业课，但没什么用。我就想，好吧，这不会带我去我想去的地方。

Yes. But the path that I saw was actually too much of the beaten path, right? The beaten path is in your PhD program. You invent something there. And you turn that technology into a startup. And I remember just feeling like, okay, I'm 21. I'm just too young to do real things in the world. This path is the only thing that's possible. But at the very least, I can start meeting people doing startups because I had taken a startup class and it was just not useful. I was like, 'All right, like this is not going to get me to where I want to go.'

于是我决定去见这些创业的人。结果第二天我就收到了一封来自帕洛阿尔托一个支付初创公司团队的邮件。我就想，好吧，我的新目标就是见这些人，慢慢学会模式匹配。我记得见到 Patrick 时，我们一拍即合。

And so I decided I'd meet these people doing startups. And literally the next day I got an email from some people working on a payment startup in Palo Alto. I was like, well, my new thing is to meet these people and learn to pattern match over time. And I remember when I met Patrick, we just clicked.

他是什么背景？

What was his background?

嗯，他之前在麻省理工，他和他的兄弟——

Well, so he had been at MIT and so he and his brother—

你是在麻省理工认识他的吗？

Did you know him from MIT or—

不，我不认识。但我们有共同的朋友，因为他上过麻省理工，John 上过哈佛。就是那个圈子，他们到处打听。

No, I did not. Yeah. But we had mutual friends because he had gone to MIT, John had gone to Harvard. It was that community that they were asking around.

我的名字当然在两个圈子里都出现了。

And my name of course came up in both circles.

我记得我飞过去，当时下着雨，天气有点糟糕。我打开门，我们一开始交谈，就立刻产生了共鸣，对吧？我想我们就是有这种感觉，虽然背景在很多方面不同，但我们有非常相似的技术视角，甚至对一些极客的东西也一样。我们用的都是分体式键盘，就是 Kinesis 键盘。我们都用 Dvorak 布局，但实际布局不同，我们还讨论了如何为他们系统构建防火墙，聊了内核里的各种东西，所以就是有这种技术上的连接。

And I remember I flew out and it was like raining and kind of miserable and I remember opening the door and just like when we first started talking it was just an instant connection, right? I think we just like had this, you know, different backgrounds in many ways, right? But also we had a very similar technical perspective and like even for very nerdy things. I mean we had the same like split keyboard that we used this Kinesis keyboard. We both were Dvorak the actual layout that we used was different and we were talking about how to build the firewalls for the systems that they had and we're talking about you know different things in the kernel and so it was just like we had this like technical connection.

是的。而且我真的很欣赏 Patrick 和 John 的一点是，他们和我同龄，却已经在外面做事了。他们已经在创业了，我当时觉得那不可能。我们俩中有一个是错的。我很想知道是谁。

Yeah. And I think that what I also really appreciated about Patrick and John is that they were my age and they were already out there doing things, right? That they were already doing a startup and I was like, I don't think that's possible. One of the two of us is wrong. I really want to know who it is.

嗯。然后发生了什么？

Yeah. And then what happened?

那个周末很棒。他们说，你应该加入。我说，让我回去想想。我回到了学校。John 碰巧周四在城里，所以我想，我大概应该做决定了。我就想，你知道吗，我想做这个，因为又回到了那个梦想算法：万一这能成呢？我对支付一无所知，这不是我从小热衷的问题，但这些人——

So weekend was great. They said, 'You should join.' I said, 'Let me go think about it.' I went back to school. John happened to be in town on Thursday and so I was like, you know, I probably should make my decision and I was like, you know what, I want to do this because again, back to that algorithm of dream of the what if this works. I'm like, I don't know anything about payments. Like, this is not the problem that I've grown up passionate about, but these people—

但你能解决一个真实的问题。

But you get to solve a real problem.

是的。解决一个真实的问题。没错。而且这些人是我觉得可以学习或合作的人。

Yeah. Solve a real problem. Exactly. And these are the people that I feel like I can learn from or that I can work with.

我记得 John 说，你会做吗？我说，好吧，行，我做。我只记得当时觉得，好吧，有人需要我。没错，他们想要我。

And I remember John is like, 'Will you do it?' I was like, 'All right, fine. I'll do it.' And I just remember feeling like, okay, like that there's someone. Exactly. They want me.

那感觉很好。

That's a good feeling.

那确实是一种很好的感觉。

It was a really good feeling.

于是我在周四做了决定，周五花了一天时间告诉老师们我要退学。

And so I decided that on Thursday, Friday, I spent telling my teachers that I was out.

情况怎么样？

How did that go?

呃，有点难。

Uh it was a little tough.

对你来说情绪化吗？

Was it emotional for you?

确实感觉像是什么东西的结束。

It was definitely It felt like the end of something.

嗯。

Yeah.

而且你知道，哈佛当我告诉他们我要离开时，他们说，你会回来的。

And you know, Harvard when I told them I was leaving, they said, 'You're coming back.'

你会回来的。没错。我想这大概是为了他们的数据之类的。但这确实让我觉得，好吧，我明白怎么回事了。麻省理工则有点不同，和相关负责人谈的时候，他们说，你得每六个月左右报到一次，因为如果时间太长，你可能得重新申请，我们也不确定。氛围完全不同。

You're coming back. Exactly. And I think it's probably for their numbers and you know that kind of thing. But it definitely felt like okay I see how this goes. MIT was a little bit more like okay like talking to the relevant person that they said that you know if you you have to check in every six months or so because if it too long goes by then you know maybe you'll have to reapply and we're not really sure. And it just was a very different vibe.

我记得和教授们谈过。教授们很支持，我觉得他们见过这种情况，但他们也有点难过看到我学期中离开。是的，我真的很感激他们提供的指导。我当时正在上操作系统课程，那是麻省理工一门著名的课程，由非常优秀的教授授课，那门课真的很酷，我很遗憾不能继续完成后续项目。所以这确实是在放弃某种经历，也很难和那些人告别——那些我去那里学习合作的人。有趣的是，他们很多人后来都来到了硅谷。我在 Stripe 和他们一起工作，也以其他方式和他们共度了很多时光。所以，这远没有我当时想象的那么像永别。

And I remember talking to the professors. The professors were supportive and I think this is a thing that they see and but they were also, you know, I think a little sad to see me go like mid-semester. Yeah, I really appreciated the mentorship they provided. like I was in the middle of the operating systems course which is this famous course at MIT taught by these like extremely good professors and it was just such a cool class and I was very sad to not get to implement further projects and so it was definitely kind of giving up on some sort of exposure to an experience and it was also hard to say goodbye to the people right that there were all these people that I'd gone there to learn from and work with and the funny thing is many of them ended coming out to the valley. Got to work with them at Stripe. Got to spend a bunch of time with them in other ways. So, it was much less of a goodbye than I thought it was at the time.

嗯。在你成长的地方长大，又这么快地移动，你觉得自己与众不同吗？

Yeah. Growing up where you grew up and moving as quickly as you did, did you feel different?

我确实觉得自己与众不同。是的。

I definitely felt different. Yes.

孤独吗？你觉得自己像个局外人吗？

Was it lonely? Did you feel like an outsider?

我确实觉得自己不同。我确实不合群。

Like I definitely felt different. I definitely didn't fit in.

有很多其他孩子喜欢的事情我就是不理解。比如我记得在幼儿园的校车上，其他孩子跟着收音机唱歌，而我一个词都不知道。

And there were a lot of things that other kids were into that I just didn't understand. Like I remember being on the school bus in kindergarten and other kids were singing along to the radio and I didn't know any of the words.

嗯。

Yeah.

我只是觉得不知道如何弥合这个差距。所以我有过很多这样的时刻，就是觉得我身上有些东西不太合拍。有些是关于活动。比如很多孩子会去打猎，而我们家完全不干这个。所以就是有些很不一样。我打过一阵冰球，打得不太好，但我是守门员。所以我开始尝试做一些能合群的活动。但我记得我真正在意的一件事是给自己塑造一个身份。因为如果你与众不同，又觉得自己没有身份，那就会很孤独。但如果你与众不同，却有一个身份，有东西定义你。

And I just felt like I didn't even know how to bridge this gap. So I had a number of moments like that where I just felt like there was something about me that didn't quite match. Some of it was about activities. Like a lot of the kids would hunt and that was not something my family did at all. So there was just something very different. I did play hockey for a little bit. I was not very good, but I was goalie. So I started to try to do some activities that would match. But I remember one thing I really cared about was carving an identity for myself. Because if you're different and you don't really feel like you have an identity, then it's lonely. But if you're different and you have an identity, something that defines you.

那你就是在走自己的路。

Then you're charting your own path.

对我来说，就是当那个聪明的孩子。比如我记得小学时每周有拼写测验。流程是周一老师给你这周的 10 到 20 个单词，然后周末测试。周一有个预考，你要听写，如果写错了就得去问一个写对了的同学怎么拼。通常我全对，但有一周我错了一个词。我记得班上另一个聪明的孩子写对了。我特别羞愧，因为得去问他正确答案，那会侵蚀我的核心身份。

And for me, it was being the smart kid. Like I remember in elementary school we had a weekly spelling quiz. The way it worked was on Monday the teacher would give you the 10 or 20 words for the week, and then at the end of the week you'd be tested on them. You'd have this pre-test on Monday where you'd have to write them down as the teacher says them, and if you got it wrong you'd have to go ask one of the other kids who got it right how to spell it. Normally I'd get them all right, but I remember one week I got one of the words wrong. And I remember another one of the smart kids in the class, he got it right. And I was so ashamed that I would have to go and ask him for the right answer because that would be eroding this core identity I had.

哇。不过听起来那其实很健康，让你学会了不必什么都知道。

Wow. That sounds like it was really healthy though that you got to do that and it set you up to not have to have all the answers.

对，对。我觉得那未必是坏事。

Yes. Yes. Yeah. I think it was not necessarily a bad thing that happened.

是个好故事。

It was a good story.

没错。

That's right.

那么当你遇到像帕特里克这样的人时，是不是有种‘哦，他跟我一样’的感觉？

So then when you got to meet someone like Patrick, was it a feeling of, oh, he's like me?

是的。

Yes.

那感觉一定很棒。

Must have been a great feeling.

确实如此。确实如此。他之前做过一家创业公司，所以他和约翰以前……他 21 岁就已经做过创业公司了。

It really was. It really was. And he had done a startup before, so he and John had previously... and he was 21 and he had already done a startup.

没错。是的。

Exactly. Yes.

那是怎么发生的？

How did that happen?

他们做过一家创业公司叫……哦，我好久没想这个了，记不清所有细节了。但我想，帕特里克和约翰在爱尔兰长大，帕特里克通过 Lisp 社区——他们非常热衷的编程语言——认识了保罗·格雷厄姆，他运营着创业孵化器 Y Combinator。我想那就是他后来加入 YC、做创业公司的契机。那家公司被卖给了 Live Current Media。他们为收购方工作了一段时间，但显然那不是他们想一辈子做的事。

So they did a startup called... Oh man, I haven't thought about this for a while. I'm not going to remember all the details. But I think okay, so Patrick and John grew up in Ireland and I think Patrick had met Paul Graham, who runs Y Combinator, this startup incubator, through the Lisp community, through the programming language they were very into. And I think that was his connection to then doing YC, doing a startup. That startup got sold to a company called Live Current Media. They worked for the acquirer for some time, but clearly that was not the thing they wanted to do with their lives.

你在 Stripe 有多早？

How early were you in Stripe?

Stripe 早期，我们一共四个人：约翰、帕特里克、达拉·巴克利和我。我刚去的时候，已经有了一些基础设施和一个支付处理器。但还不清楚具体怎么推进，因为有一个想法是：我们建一些应用，然后用我们正在建的支付处理器来驱动这些应用？这样你就能真正做出东西来。最终支付处理器可能会成为核心，但这些应用也可能。所以帕特里克一直在做一个时间追踪的东西，那是潜在想法之一。所以它还处于雏形阶段，但我们能看到方向。有趣的是，我提到我在 MIT 上过创业课，其中要建一个模拟创业公司，而我建的就是一个支付处理器。所以我花了很多时间研究如何做在线支付，发现它糟透了。很痛苦。就像去 PayPal，读他们的文档，研究怎么注册。一切都那么不透明。你就会想，怎么能这么差？

So early days of Stripe, there were four of us. There was John, there was Patrick, there was Darra Buckley, and there was me. And when I first was there, there was some infrastructure and there was a payment processor. It also wasn't clear exactly how we were going to proceed because one idea was, well, what if we build some apps and then use this payment processing we're building to power those apps? And so that's how you actually build something. And eventually the payment processor will probably be the thing, but these apps could be too. So there was a time tracking thing that Patrick had been working on. That was one of the potential ideas. So it's still in this nascent form, but still something we could see the direction of travel. And the funny thing, by the way, is that I mentioned that I did this startup class at MIT and as part of that you're supposed to build a mock startup, and the one I ended up building was a payment processor. So I spent all this time trying to figure out how to do payment processing online and realizing it's horrible. It's painful. It's like trying to go to PayPal and trying to read their documentation, trying to figure out how to sign up for something. It was just so opaque. And you just realize how can it possibly be so bad?

嗯。

Yeah.

那就是 Stripe 的根本突破——意识到你可以把支付处理做得更好。我记得最初的网站写着‘支付处理不必这么烂’，对吧？那就是我们的精神：它不必这样，可以更好。我真的很喜欢那种精神。我记得有一次和一个 VC 聊天，那时还没发布，但我们已经引起了很多关注。他问：‘你们的秘诀是什么？’我说：‘我们就是做得更好。’他说：‘不，不，得了吧，我知道你对谁都这么说，但真正的秘诀是什么？’我说：‘我不知道该告诉你什么。我们做的就是关注每一个细节，把它做对。’

And that was the fundamental unlock for Stripe was this realization that you can do payments processing better. And I remember the initial website said 'payment processing doesn't need to suck,' right? And it was just like that is the ethos of just like it just doesn't need to. It can be better. And I really like that spirit. I remember talking to a VC at some point. This was pre-launch, but we had gotten a lot of buzz. And he was saying, 'Okay, what is your secret sauce?' And I was like, 'Well, we just do it better.' And he's like, 'No, no, come on. I know that's what you say to everyone, but what's the actual secret sauce?' And I'm like, 'I don't know what to tell you. That is what we do is just focus on every single detail and get it right.'

所以它其实不是创造新东西。已经有 PayPal 了，它只是一个好得多的 PayPal。没错，是这样吗？

So it really wasn't creating something new. It was already PayPal. It was just a much better PayPal. Exactly. Is that right?

是的。它真正关注的是整个端到端体验的细节，对吧？认真思考，因为如果你自己经历过，你就会意识到这些部分：你怎么注册账户？你怎么连接 API，也就是计算机之间实际通信的方式？你编程时面对哪些不同参数？你怎么决定用哪种编程语言或哪个封装库？所有这些事情，每一件都可能增加大量摩擦。所以如果你只专注于‘好，我要把这个做好，把这个做好，把这个做好，为我自己做好，为我的朋友做好。’是的。实际上，让我们做出对每个人都好的东西。

It was. And it was really focusing on the details of the whole experience end to end, right? And really thinking about just because if you've gone through it yourself and you realize like this part: how do you sign up for an account? How do you actually even connect to the APIs to the actual way the computers talk to each other? What are the different parameters you program against? How do you figure out which programming language to use or which wrapper to use? All these things. Each one of those can add a ton of friction. And so if you just focus on, okay, I'm gonna make this one good, this one good, this one good, make it good for myself, make it good for my friends. Yes. Actually, let's be something that makes it good for everyone.

那你为什么离开 Stripe？

And why did you leave Stripe?

嗯，我在 Stripe 待了差不多 5 年，四年半。大约四年的时候，我开始考虑这是不是我长期想做的事。我看事情的方式是，如果你把职业生涯分成五年一段，五年差不多是合适的长度，因为少于五年很难做出什么有意义的事。我记得当时觉得，好吧，我已经把这家公司带到了一个无论有没有我都能成功的地步。

Well, I had been at Stripe for almost 5 years, four and a half years. And about four years in, I think I started to consider whether or not this was what I wanted to do for the long term. And the way that I kind of view things is like, okay, if you think of your career in five-year chunks, and five years is about the right length because less than that, it's kind of hard to do something significant. And I remember feeling like, okay, I'd gotten this company to a place where it was going to succeed with or without me.

嗯。

Yeah.

然后问题就是，我是继续还是去做点新东西？我对创业的想法非常兴奋。

And then the question was, do I continue or do I go and do something new? And I was very excited about doing the idea of a startup.

但我记得和帕特里克聊过，他也有一些非常有说服力的理由让我留下。他提出的一个观点是，要聚集一群能在世界上做出重大成就的人非常困难。最好的情况是，你去创办一家初创公司并取得一些成功，五年后你就能组建起一个能做成事的团队，但你现在已经拥有这样的团队了。

But I remember talking to Patrick and I think he had some very convincing reasons too to stay. One point he made is that it's very hard to assemble a group of people that can do significant things in the world. Best case scenario, you go start some startup and have some success with it, and then five years later you'll have formed that group of people that can do stuff, but you already have it here.

是的。

Yeah.

没错。我们已经有了这个能做成事的紧密团队，那为什么要离开呢？所以这非常艰难，不容易。我记得我告诉帕特里克我要离开时哭了。是的，非常非常艰难。他和约翰为我办了一个果汁派对，送别很温馨。我决定的原因——我记得我仔细思考后——我想，如果组建一个能做出重大成就的团队真的那么难，那我必须现在就开始。

Right. We already had this tight group that was able to accomplish things, and so why walk away from that? So it was very tough. It was not easy. I remember I cried when I told Patrick that I was out. Yeah, it was very, very tough. He and John threw a juice party for me. It was a very nice sendoff. And the reason I decided—I remember feeling as I thought that through—I was like, well, if it really is so hard to build that group of people who can accomplish significant things, I got to get started now.

是的。听起来你离开是因为 Stripe 的使命不是你想要专注的使命。

Yeah. It sounds like you left because the mission of Stripe wasn't the mission that you wanted to focus on.

确实如此。是的，这是一个美好的使命，我非常支持。但要说这是否是我必须以某种形式追求的使命，那就不一样了。

That is true. Yeah, it's a beautiful mission and it's one I very much support. But it's different to say is this one where I will just need to pursue it in any form.

你离开时是知道接下来要做什么，还是想着边走边看？

And did you leave knowing what was going to be next, or did you leave thinking I'm going to figure it out?

更像是后者。我列了三个可能专注的领域：第一是 AI，第二是 VR/AR，第三是编程教育。对我来说，很清楚的是：如果我能以某种方式为 AI 做出贡献，那我就去做。但我不确定的是，我有没有这个技能？时机是否合适？等等。所以另外两个算是备选。

More the latter. I had a list of three different areas I might focus on. Number one was AI. Number two was VR/AR. Number three was programming education. For me, it was very clear: if I can contribute to AI in some way, okay, I'm doing that. But it wasn't clear to me, do I have the skills? Is it the time? All those things. And so the other two were kind of backup options.

我知道你曾经写过一本化学教科书。

I understand there was a point in time that you wrote a chemistry textbook.

没错。

That is true.

那是什么时候？

When was that?

那也是 2008 年。高中毕业后，我休学了一年。高中时期我花了很多时间在学术上，既上大学的课程，也参加竞赛。我非常投入数学竞赛和化学竞赛。我记得在十年级时，因为九年级学过化学，我参加了一个妈妈在网上找到的化学竞赛，抱着玩玩的心态，结果得了地区第一。我想：“好吧，挺酷的。”我所在的州没有这个比赛，所以我是在明尼苏达州参加的。我参加了全州比赛，得了州第一，然后他们邀请我参加国际化学奥林匹克训练营。那是全国前 20 名的学生。

That was also 2008. So after high school, I took a year off. I had spent much of my high school doing academic both college courses but also competitions. I got very into math competitions, got very into chemistry competitions. I remember in 10th grade, because I'd taken chemistry in 9th grade, I took some chemistry competition that my mom had found online and kind of did it on a lark, and I got best in my region. I was like, "Okay, that's cool." My state didn't really have it, so I was in Minnesota to do it. I took the statewide one and I got best in the state, and they invited me to the training camp for the International Chemistry Olympiad. So it's top 20 kids in the nation.

为此他们提前寄了一些教科书，说请阅读这本有机化学教材的第 1 到第 8 章，这是训练营要用的教材。所以我读了第 1 到第 8 章，没太当回事。我想，我就是注定会成功的，一切都会顺利，之前一直如此。我记得我到了化学训练营的比赛现场，其他孩子不仅读了前八章，他们读完了所有的书。他们不仅读完了整本书，还读了所有那些厚厚的物理化学书，以及其他各种书。

And for this they send you some textbooks ahead of time. They say please read chapters 1 through 8 of this organic chemistry textbook. These are the textbooks we're going to use for the camp. So I read through chapters 1 through 8. Didn't take it that seriously. I was like, look, I am just destined to succeed. It's just going to work. It happened so far. I remember I showed up at the chemistry competition at the chemistry camp, and the other kids had read not just those first eight chapters. They'd read all the books. They'd read not just the whole one book, but all these big fat physical chemistry books, all these other ones.

我当时就想，等等，什么？还能这样？其他人都在这么做？我记得训练营的两周里我感到非常沮丧和崩溃，因为这些人知道所有我不知道的东西。我们本该为期末考试学习，而我却在房间里玩手机游戏。我放弃了，觉得没有希望了。

And I was just like, wait, what? You can do that? These other people are doing that? I remember just feeling so demoralized and crushed for the two weeks of the training because it was just like these people knew all these things that I didn't know. We were supposed to be studying for the final exam. I was playing cell phone games in my room. I just gave up. I was like, there's just no hope.

是的。

Yeah.

然后他们宣布了去国际比赛的前四名，说的是这四个人，没有我。他们宣布了两个亚军，也没有我。我想，我显然是第 20 名。幸运的是，他们不公布其他人的排名。我觉得太糟糕了。我记得在下一学年开始时，回顾那个夏天，我觉得自己面前有一个绝佳的机会，而我却浪费了。那是一种可怕的感觉。我感觉自己一直在混日子，以为光靠天赋就能达到目标，不需要努力。我当时想，我再也不想有那种感觉了。

And so they announced the top four to go to the international. They said these four people, they weren't me. They announced two runners up. They announced those two. It wasn't me. I was like, I'm obviously 20th. Fortunately, they don't tell you the ranking of the rest. I was like, it's so bad. I remember at the beginning of the next school year looking back at how I had spent that summer and I felt like I had this amazing opportunity in front of me and I had squandered it. It was a horrible feeling. It was this feeling that I had just coasted, that I just sort of believed that talent alone would just get me to where I wanted to go and I didn't have to work hard. And I was like, I never want to feel that again.

所以那一年我认真对待了。我开始学物理化学，学有机化学，学所有这些。我花了很多时间研究过去的化学竞赛题目，学习它们，翻阅大量不同的材料，以成为一个真正的竞争者。那一年我进入了前四名，参加了国际化学奥林匹克竞赛，获得了银牌。那是一次非常棒的经历，我真的很享受。但对我来说，最大的收获是那种你必须始终努力的感觉。

And so I took it seriously that year. I started taking physical chemistry. I took organic chemistry. I took all these things. I spent a lot of time looking at old chemistry competitions and learning them and looking through a bunch of different material in order to become a real competitor. That year I made the team of the top four, went to the International Chemistry Olympiad, got a silver medal. It was a very awesome experience. I really enjoyed it. But for me, the big takeaway was this feeling of you always have to work hard.

是的。

Yeah.

你知道，我最喜欢的一句名言是自行车运动中的一句话：“它永远不会变容易，你只是骑得更快。”我认为这是我非常信奉的。

And you know, one of my favorite quotes is the cycling quote: "It never gets easier; you just go faster." I think that is something I very much live by.

那么，你最后是怎么写出那本书的？

So, how did you end up writing the book?

嗯，我觉得自己想出了一种非常独特的看待化学的方式，一种非常数学化的第一性原理方法。因为如果你读大多数化学教科书，它们基本上就是说，记住这些反应，这些是化学性质，这些是化合物。但你总是会问为什么？比如，为什么是这样？这些原子必须这样相互作用吗？或者这个化合物必须是这种颜色吗？如何从第一性原理推导出来？所以，为了在竞赛中表现出色，我采取的方法是真正去理解底层规则，而不是……

Well, I felt like I'd come up with a very unique way of looking at chemistry, a very mathematical first-principles approach to it. Because if you read most chemistry textbooks, it just says, well, basically memorize these reactions. Here are these chemical properties. Here's these compounds. But you always ask why? Like, why is it that way? Does it have to be that these atoms interact in this way, or does this compound have to be this color? How do you derive it from first principles? So the approach that I took in order to be good at the competitions was to really try to figure out the underlying rules, not the...

这在你读的教科书里是没有的。

Which wasn't in the textbook that you read.

教科书里没有。你必须自己提炼出来，对吧？也许他们试图以某种方式传达，但这并不是他们花时间去做的事情。所以我设计了这本书的结构。我觉得自己有一种不同的化学教学方法。我受到一些朋友的启发，他们在数学领域做过类似的事情，采用了一种叫做“解题的艺术”的数学形式，这是一种非常苏格拉底式的方法。所以这本书只有问题，但每个问题都有意地设定范围，层层递进。第一个问题从你应该具备的知识开始；只要你稍微思考一下，就会觉得，“哦，我能看出这是怎么回事。”然后下一个问题建立在前一个问题的基础上，再下一个又建立在前一个的基础上。在化学的情况下，你需要一些实验结果。所以你会说，“这是双缝实验。”

It's not in the textbook. You have to distill it down yourself, right? And maybe they try to communicate in some way, but it's just not the thing that they spend their time on. So I structured the book. I felt like I had a different way of teaching chemistry. I structured it in a way that was inspired by some of my friends who had done something similar in math in this math form called "The Art of Problem Solving," where it's a very Socratic method. So the book has just questions, but they're intentionally scoped so that each one builds up. The first one starts from knowledge you should have; if you just think a little bit, you're like, "Oh, I can see how this works." Then the next one builds on the previous thing. The next one builds on the previous thing. In the case of chemistry, you need some experimental results. So you say, "Here's this double-slit experiment."

然后你会想，‘好吧，粒子与波到底意味着什么？’然后，好吧，如果它意味着既是粒子又是波，那么，你知道，等等。这是我非常想传达给别人的东西。我不仅在乎这种方法存在于我的脑海中，更在乎其他人能从中受益。不过我没写完，写了大约 100 页。如果你感兴趣，可以在我的网站上找到。但我觉得这种精神是我想要继续传承的。

And then you're like, 'Okay, well, what does it mean for particle versus wave?' And then, okay, if it means it's both a particle and a wave, then, you know, and so forth. And this was something that I really wanted to communicate to others. I really cared about not just this approach living in my head, but other people being able to benefit from it. Now, I never finished. I made it through about a 100 pages. You can find it on my website if you're interested. But I felt like that ethos was something I really wanted to carry forward.

你基本上写了你自己想读的书。

You basically wrote the book that you wish you could have read.

完全正确。

That's exactly right.

太棒了。

That's great.

是的。

Yes.

编程与其他活动有何相似或不同之处？

How is coding similar or different to other activities?

我对编程的理解是，你深入理解某个过程，然后用一种非常晦涩的方式（我们称之为程序）把它写下来，然后任何人都能从中受益。对吧？人们不需要写代码，也不需要理解其中的机制。我认为还有其他非常脑力的领域也是如此，比如数学。你深入思考一个问题，然后用一种晦涩的方式（我们称之为证明）写下来。但没人读那些证明，对吧？只有大约五个关心特定领域的数学家会真正深入阅读。所以我认为编程之所以与众不同，是因为它几乎像魔法一样。你脑海中有一个愿景，只需描述它，它就以某种方式实现了。所以在某些方面它像管理，对吧？你有一台计算机，它执行你脑海中的功能、愿景，当你编写程序时，它会非常字面地执行。我认为这是我在任何其他传统领域从未见过的。感觉就像你做的许多其他事情，你无法获得同样的杠杆作用，对吧？你有一个愿景，它不知何故变成了现实，而你不需要在物理世界中移动任何东西。它就这样实现了。

So the way that I think about coding is that you deeply understand some process. You write it down in a very obscure way we call a program and then anyone can get the benefit of that thinking. Right? People don't need to write the code. They don't need to understand the sort of mechanics of what went into it. And I think that there are other very cerebral domains that are like this like mathematics, right? Where you think hard about a problem, you write it down an obscure way we call proof. But no one reads those, right? Only like the five mathematicians who care about a particular domain will really deeply read it. And so I think that what makes coding stand apart to me is it's almost like magic. You sort of have this vision in your head and you just by describing it somehow it comes to be. And so in some ways it's like management, right? that you have a computer that is there to perform the function that you have in mind, the vision that you have, and that it carries it out in a very literal fashion when you write a program. And I think that that to me is something I've never seen in really any other traditional domain. Like it feels like many other things that you might do that you just don't get that same leverage, right? Of you have this vision and somehow it comes into reality and that you don't have to physically move things in the world. It just comes to be.

你更愿意把它描述成语言还是数学？

Would you describe it more like a language or more like math?

我更愿意把它描述成数学。但关于数学，我认为有一个误解：数学不是一加一，对吧？它不是那些机械的计算。数学是关于宇宙的底层结构，对吧？它是关于理解不同对象、不同想法如何在深层概念上相互关联，以及对称性和对象之间的关系看起来非常不同。我认为编程也是如此。它实际上是关于理解一个网站应该如何工作，或者某人想要什么，以及边界情况下的各种行为方式。如果出现错误，你应该如何处理？所有这些事情感觉有些平凡，但如果你真正审视底层架构，你思考的是所有这些系统相互通信，数据以不同形式存储，像加密这样的概念被引入，你如何编排所有这些以提供有用的东西？所以对我来说，我认为这是数学之美被提炼成有用的形式。

I would describe it more like math. But the thing about math that I think there's a misconception for is that math is not one plus one, right? It's not about these like mechanical calculations. Math is about the underlying structures of the universe, right? It's about understanding how different objects, different ideas relate to each other in this deep conceptual way and the sort of symmetries and the relationships between objects look very very different. And I think that programming is like that. It's really about understanding how should a website work or what does someone want and what are all the different ways that something in a corner case should behave. If there's an error, how should you handle it? All these things they feel somewhat mundane but if you really look at the underlying architecture you're thinking about you have all these systems that are talking to each other you have data stored in these different forms you have ideas like encryption that are brought to play and how do you orchestrate all of that in a way that delivers something useful and so to me I think it's it's about the the beauty of mathematics that's raified into useful form.

数学是自然的覆盖层，还是自然是数学的覆盖层？

Is math an overlay on nature or is nature an overlay on math?

嗯，这是个好问题。我认为数学在很多方面是宇宙的织物。我喜欢数学的一点是，它是在任何现实中都成立的规则集。我记得在中学或高中开始学习生物学时，你学到了关于不同细胞类型如何工作的所有细节，以及关于叶绿素和光合作用的所有不同过程。但我记得感觉，好吧，这只是在这里对特定生物或研究方式成立的事情。但这是普遍真理吗？它必须总是成立吗？事情必须这样运作吗？我喜欢数学的是它完全与观察脱钩。它是必须成立的事物集。所以对我来说，数学是对可能性的深刻不变的理解，而自然则是特定形式的实例化。

Huh, that's a great question. I think math is the fabric of the universe in many ways. Like the thing I love about math is that it is the set of rules that are true in any reality. Like I remember in middle school or high school starting to learn biology and you learn all of these details about how different cell types work and all these different processes about chlorophyll and photosynthesis. But I remember feeling like, okay, this is something that happens to be true right here for this specific organism or way of studying things. But is this universally true? Like, does this have to be true always? Does this have to be how things work? And the thing that I love about math is that it is purely decoupled from observation. It is the set of things that must be true. And so I think for me math is this deep immutable understanding of what is even possible and then nature is an instantiation in a specific form.

你认为在理解数学之前自然可能存在，还是数学是自然存在所必需的？

You think it was possible for nature to exist before the understanding of math or was math needed for nature to exist?

我对此有两种想法，因为对我来说，数学似乎独立于其他一切而存在。它真的是第一性原理的深层真理，无论你采取什么视角。比如，如果我们遇到一个来自数百光年外的外星人，他们会有相同的数学。我们可能在那方面有共同点，而其他方面可能毫无共同之处。但与此同时，我也觉得有一个问题我有时会纠结：我们是发现数学还是发明数学？它是深层真理，还是我们的视角使它成为现实？我认为天真的观点是说，它已经存在，不需要我们。但在某些方面，如果没有人在那里欣赏它，这有点像森林中倒下的树，没有人听到，对吧？声音存在吗？所以，我认为自然以及我们在这里作为观察者、作为思考者的事实，使得数学具有意义，而不仅仅是毫无意义的抽象。

I'm of two minds at this one because to me math feels like it exists independent of everything else. It's really a first principles deep truth that doesn't matter what perspective you take. Like if we were to meet an alien from, you know, hundreds of light years away, they would have the same mathematics. We would probably have something in common in that way and maybe nothing else in common. But at the same time, I also feel like there's a question I struggle with sometimes of do we discover mathematics or do we invent it? Is it a deep truth or is it our perspective on it that makes it come into reality? And I think that a naive view is to say, well, it's already there without us. But in some ways, if there's no one there to appreciate it, it's a little bit like the tree that falls in the forest with no one to hear, right? Is the sound there? And so, I think that there is something about nature and the fact that we are here and that we are observers, that we are thinking beings that causes math to have meaning rather than just be an abstraction that with no no significance.

如果你要向一个不会编程的人解释编程过程的感觉，你会如何描述你在做什么？

If you were to explain what the coding process feels like to someone who doesn't code, how do you describe what it feels like you're doing?

编程过程中最美妙的部分是当你进入心流状态时，一切都恰到好处。你有一个目标。也许很简单，比如你想改变按钮的颜色。也许很复杂，比如你想构建一个数据库或一个大型分布式系统，许多计算机相互通信。但你的脑海中有一个你希望在现实中看到的目标，并且你有一个部分实现。也许你从零开始。也许你已经写了一些代码，它某种程度上能工作，或者实现了你所需的一个子集。然后你尝试它。你测试它，发现你脑海中的东西与你在系统中观察到的之间存在差距。有时是一个微妙的差距，通常你称之为 bug。你写了代码，它本应做某事，却做了别的事。然后在这种情况下，你开始形成一个心理模型：嗯，我知道代码是怎么写的。我看到了这个观察结果，它与预期不同，或者也许是预期的，但还不是我脑海中的样子。

So, the most beautiful part of the coding process is when you're in flow state and there everything just kind of clicks. You have some objective. Perhaps it's something simple like you want to change the color of a button. Perhaps it's something complex like you want to build a database or a big distributed system with many computers talking to each other. But you have some objective in your mind that you want to see in reality and you have a partial implementation. Maybe you're starting from a blank slate. Maybe you have written some code and it sort of works or it implements a subset of what you're looking for. And then you try it. You test it out and you see that there's a gap between what you have in your head and what you observed in the system. And sometimes it's a subtle gap which usually you think of as a bug. You wrote the code. It was supposed to do something. It did something different. And then in that case you start to form a mental model of well I know how the code is written. I see this observation that's different from what's expected or perhaps was expected but is not yet what I have in my mind.

然后你会想，差距在哪里？也许你立刻就知道，哦，是这一行代码。我可能搞砸了，这个值太高了，我得回去检查。也许你不太确定，所以你会想，什么能给我更多信息来找出问题根源。于是我们经常添加所谓的遥测，也就是增加可观测性，这样你就能从代码中间获取数字，了解不同部分的运行速度。你可以得到一些日志行，代码实际上在说：我做了这个，发生了那个，这是某个事件。然后通过查看那个追踪记录——它几乎是程序内部发生事件的历史——你可以回溯并找出问题。所以在很多方面，这就像你在非常详细地指导别人如何执行某个流程。你对此有一定理解，你想写出所有规则，精确描述它在每个极端情况下的工作方式。每当你看到不理想的结果，你就回去说：‘好吧，我在规则手册里哪里搞错了？’然后你努力真正理解导致那个结果的轨迹，以便回去做出适当的修改。

And you think about what's the gap? Maybe you know immediately, oh, there's this specific line. I probably messed it up. This value is too high. I should go back and check. Maybe you don't quite know. So you think about what would give me more information to figure out where it comes from. And so often adding what we call telemetry. So adding observability so that you can get numbers out from the middle of the code about how fast different parts are running. You can get little log lines that say the code effectively saying I did this, this happened. Here's some event. And then from looking at that trace, it's almost the history of what has happened within the program, you can go back and trace it out. And so in many ways, it's almost like you're in a very detailed fashion instructing someone else for how to perform some process. And you have some understanding of it and you want to write out all the rules for exactly how it works in every single corner case. And that whenever you see some undesirable outcome, you go back and you say, 'Well, what did I get wrong in the rule book?' and you try to really have enough understanding of the trail by which that outcome occurred so that you can go back and make the appropriate changes.

Vibe coding 是如何改变这个过程的？

How has vibe coding changed the process?

Vibe coding 是一个非常迷人的时刻，我认为现在正在发生的是软件工程正在彻底改变。我记得我第一次真正体验 vibe coding 是在我们早期一个编码模型的现场演示中。我们建了一个小网页界面，你可以和模型对话，让它写一些 JavaScript。我在直播中让它构建了一个小游戏，并让观众建议一个功能。我们仅仅通过和计算机对话就实现了那个功能。对我来说，深层的意义在于：计算机从来就是为了帮助人类而创造的，对吧？这就是全部意义。而我们在写代码时却要扭曲自己去适应机器，无论是低级编程语言，还是所有那些用来判断计算机是否做了你想要的事情的技巧和机制。而 vibe coding 的本质是让机器更接近人类。所以你指导它，你仍然需要——取决于模型的好坏和任务的难度——你需要对它将如何解决任务有一个心智模型，对吧？如果你只说‘给我建一个超棒的网站’，那什么是超棒的网站？所以，好吧，我想要这里有个按钮，或者我想要这种功能。你实际上不再是个人贡献者，不再是软件工程师，而是像一个经理，但一个仍然对结果高度负责的经理。我认为过去一年模型在显著地逐步改进。我认为 2025 年 12 月是一个真正的转折点，那是我认识的许多专家工程师第一次觉得这些模型从‘还不错、有点用’变成了‘真的能完成极其困难的工作’。所以它真的从一种演示工具——如果你不懂编程语言可以快速搞定一些东西但不太对劲——变成了驱动严肃工作、真正加速人们能力的东西。

Vibe coding is a very fascinating moment and I think that what is happening right now is software engineering is changing entirely. So, I remember the first time that I really vibe coded. This was for a live demo of one of our early coding models. We built a little web interface and you could talk to the model to ask it to write some JavaScript. And so, I on a live stream had it build a little game and ask the people watching the live stream to suggest a feature. We implemented the feature just by talking to the computer. And to me, the deep thing that's happening is that computers have always been created in order to help humans, right? That's the whole point. And that we contort ourselves to the machine when writing code, whether it's a low-level computer programming language or all this skill and machinery for how you actually tell whether computer did the thing that you wanted. And what vibe coding is is it's moving the machine closer to the human. And so you instruct it, you still need a good—depending on how good the model is, you need—and how difficult the task is, you need some mental model of how it's going to solve the task, right? If you just say, 'Build me an awesome website.' Well, what is an awesome website? So, okay, fine. I want there to be a button here or I want there to be this type of functionality. You are effectively acting as not an individual contributor, not a software engineer, but you're acting as a manager, but a manager who's still very accountable for the outcome. And I think what is happening is that the models have been incrementally getting better over the past year in a significant way. And I think there was a real turning point in December of 2025 where it was the first time that for many of the expert engineers that I know these models went from being kind of nice, kind of useful to they can actually do incredibly hard pieces of work. And so it's really shifted from being a thing for demos and just if you don't really know a programming language, you can get something done quickly, but it's not quite right to this is driving serious work and really accelerating what people can do.

还有理由用老方式编码吗？

Is there still a reason to code the old way?

我认为手工编码在某种程度上就像手写、书法、笔迹。它有一种艺术性，通过在那个层面操作，你能理解所有东西是如何组合在一起的。它有点像数学：即使你不打算做大量手算，你可能还是应该知道怎么做长乘法。编码的真正核心在于理解抽象，对吧？它真正关乎理解系统如何组合、各个部分如何协同工作。这是你想成为专家的东西。所以如果你离细节太远，只是推动系统而不真正理解内部运作，我认为那会是一个限制因素。随着模型变得更好，我认为你作为人类需要真正负责和承担责任的性质会改变。机器会在很多机械性的事情、很多设计的事情、界面如何工作等方面变得更好。这些都会随着时间推移而到来。但归根结底，这是你的愿景。如果你在乎你的愿景如何实现，那么了解机器将如何实现它的具体细节——至少有一个好的心智模型——是会有回报的。

I think that coding by hand in some ways is like handwriting, like penmanship, like calligraphy. That there is an art to it and that there is an understanding of how everything fits together by operating at that level. There's a way in which it's like mathematics where you still really want to load even if you're not going to do a bunch of hand calculations, you probably should still know how to do long multiplication. And there's something about what coding is really about is understanding abstraction, right? It's really about understanding how systems fit together, how pieces will interoperate. And that is something that you want to become an expert at. And so if you're too far from the details and that you're just pushing a system and you don't really understand how it works on the inside, I think that that is a limiting factor. As the models get better, I think that the nature of what you as a human need to really take accountability for and responsibility for will change. The machine will be much better at a bunch of the mechanical things, a bunch of the design things, how interfaces work. Each of those will come over time. But at the end of the day, it's your vision. And that if you care about how your vision is implemented, then knowing the nuts and bolts of how the machine is going to do it, at least having a good mental model of it, that is something that will pay dividends.

你能看到有一天这不再必要吗？你告诉机器你想要什么，它按自己的方式编码，你仍然可以迭代和改进。但会不会有一天，编码行为就像今天的拉丁语一样？

Can you see a time when that won't be necessary? You tell the machine what you want, it codes it the way it wants and you can still do iterations and improve it. But is there a time when the act of coding will be maybe like Latin today?

我把它想象成一片正在上升的海洋，对吧？海平面是模型的能力，你有一些岛屿，时不时一个岛屿被完全淹没，但还有更高的岛屿。我认为编码就是这样，岛屿是问题的难度。所以我认为我们看到一些岛屿已经被淹没了。例如，我有一个测试提示，多年来我用来测试每个模型，构建一个特定的网站——那是我最早建的网站之一。我第一次手工建它花了几个月。然后当我用我们早期的 Codex 模型时，大概花了五六个小时。用我们最新的模型，它一分钟就写好了，我甚至不需要碰细节，而且它做得比我之前任何迭代都好得多。

I think of it almost like an ocean that is rising, right? Where the ocean level is the capability of the models and you have these islands and that every so often an island gets totally covered by the water, but there's other islands that are even higher. And I think the coding is like that where the islands are difficulty of problems. And so I think what we're seeing is some islands have already been covered. That for example, I have a test prompt that I've used for every model for a number of years to build a particular website that was one of the first websites I ever built. When I first built it by hand, it took me months. Then when I used an early one of our codex models, probably took me five hours, six hours, something like that. With our latest model, it just codes it up in a minute and I just don't even have to touch the details and it actually does a much better job than I ever did even with the previous iterations.

这真有意思。

That's really interesting.

你真的能感受到指尖的力量，以及你现在有能力做更多事情的事实。

You really feel the power of what is at your fingertips and the fact that you are now empowered to do even more.

是啊。

Yeah.

但还有一些我们尚未覆盖的高山，问题是这些高山是无限的还是有限的？我确实认为，对我们未来最好的类比是：人们将不再是个人贡献者，而是成为智能体的管理者，然后成为中层管理者，对吧？沿着食物链上升，最终成为这个智能体组织的 CEO。而我实际上没有答案的一件事是：是的，所有机械技能、所有深度调试、所有架构，所有这些，你可以看到机器会变得非常非常擅长。

But there are still mountains that we have not covered and the question is are those mountains infinite or finite? And I do think that maybe the best analogy for where we're going is that people are going to become rather than individual contributors, they'll become managers of agents and then they'll become middle managers, right? Moving up the food pyramid, eventually become CEOs of this organization of agents. And the one thing that I actually don't have an answer for is yes, all the mechanical skill, all of the deep debugging, all of the architecture, all of these things, those you can see how the machine will get very, very good at that.

但拥有结果，对吧，问责制，以及‘这是否在做你想要的’这个图景。我看不到一条清晰的路径。不是说这永远不可能发生，但我不明白这如何能脱离人类。我认为这是非常人性化的东西，是作为一个有意图、有生活、有关系的人所独有的。所以我认为你会一直关心过程中的参与。就像如果你在盖房子，你可能不太关心每颗钉子钉在哪里，但你非常关心结果。你非常关心建造者是否以你满意的方式建造，因为最终如果房子建得不好，承担责任的是你。

But owning the outcome, right, the accountability, the picture of is this doing what you want? I don't see a line of sight. Not to say that it can never happen, but I don't know how that is something that you would ever transition out of the human. And I think that that is something that is deeply human in something that is unique to being a person with an intention, with a life, with relationships. And so I think that there is an involvement in the process that you will always care about. And just like if you're someone who is having a house built, you may not care that much about where every nail is going, but you care a lot about the outcome. And you care a lot about are the people building it doing it in a way that you'd be happy with because you're the one who at the end of the day, if it's built poorly, is going to be on the hook.

一个构建糟糕的模型和一个构建良好的模型在功能上有多大区别？换句话说，如果你能描述你想要它按某种方式工作，并且它确实那样工作了，如果代码是你自己写的很优雅，而机器写的可能不那么优雅，但它仍然做了你想让它做的事，这重要吗？

How different is a poorly built model versus a well-built model in terms of how it functions? In other words, if you can describe something that you want to work a certain way and if it works that way and if the code is done elegantly if you were to do it yourself and maybe less elegantly if the machine were to do it but it still did the thing you wanted it to do would it matter?

所以这是我们在 OpenAI 内部非常具体地看到的事情，我们正在采取一种非常智能体优先的软件开发方法。

So this is something we see very concretely within OpenAI and we are taking a very agent first approach to software development.

这意味着什么？

What does that mean?

我们设定了一个目标，到 3 月 31 日（很快），我们希望实现两件事。第一，我们所有人默认使用的工具是智能体，而不是文本编辑器或终端。所以这真的成为你觉得最可靠的工具，也是你遇到每个问题时首先使用的工具。第二，人们使用这些工具的默认方式必须经过明确的安全评估。对我来说，这两个因素就是智能体优先的含义。所以这并不意味着你永远不深入细节，而是大多数时候你不需要，并且你既建立了对系统的信任，也通过架构设计使得整个组织的结果值得信任。现在，如果深入探讨我们如何指导人们的具体点，一个艰难的决定是如何确保代码库不会变成 AI 垃圾——能工作但质量不高。我们多年来一直在确保它不会变成人类垃圾，对吧？我们有写代码的人，有审查代码的人，还有一套激励体系：如果你写了好代码，随着时间的推移，代码可维护且他人可在此基础上构建，你就会得到晋升和好的绩效评估等等。我们需要将这些应用到智能体世界，所以我们有一个口号：对垃圾说不。我们告诉所有人类审查者，提交代码的人仍然需要对其负责。作为审查者，你应该对质量设定比人类更高的标准。确保那个人真正理解代码内容。不意味着他们知道每一行，但能够签字说：‘是的，这改进了我们的代码库，而不是退步了。’我们最优秀的工程师之一在使用这些模型时发现，他想要的平衡方式是控制接口。所以他仍然手动编写：这是组件，它们如何组合，也许这是文件结构，但实现细节（通常非常复杂）他外包给机器。所以我认为区别在于，如果你不注意模型所做的某个方面，当你查看底层时，它是否是你引以为豪的东西？现在我们正在评估，在什么情况下答案是‘是’或‘否’。但我们的要求是，如果有人去查看，答案应该永远是‘是，我为此自豪’。或者，你可以有一些代码段，只要满足正确性规范，你就不关心它们好不好。所以如果你有好的方法来验证那段代码是正确的，那么你可以有高度优化的代码，但很难让任何人在此基础上构建。你不把它看作是我们通常认为的代码——一个不断演化的产物。你把它看作是一次性的。好的，生成这个。我会生成一个完全不同的版本，但永远不会在它之上构建。

We have set a goal that by March 31st though very soon that we want two things. One is that the default tool that all of our people reach for is an agent rather than a text editor or a terminal. So really that this becomes the tool that you find the most reliable and that the first thing you apply to every problem. The second is that the default way that people use these tools is something that has been explicitly evaluated as safe and secure and that those two factors to me are what it means to be agent first. So it doesn't mean that you never go into the details, but it means that most of the time you don't have to and that you've built both trust with the system, but also that we've architected in a way so that we can have trust in the whole organization as an outcome. Now, if you drill into some of the individual points of how we direct people, one of the hard decisions was really about how do you make sure that the codebase doesn't turn into AI slop stuff that works but somehow is just not very good. And we have many years of making sure that it doesn't turn into human slop, right? that we have individuals who write code and then you have people who review code and that you have a large system of incentives for if you're writing good code that over time is something that is maintainable and other people can build on then you get promotions and that you have good performance reviews and all of these things and we need to bring this to bear in the agent world and so that we have a mantra of say no to slop so that we tell all the human reviewers that there's still whoever is submitting the code needs to be accountable for it. You as a reviewer should hold an even higher bar than you would for a human on quality. Make sure that the person actually understands what's there. Doesn't mean that every single line that they really know, but that they can really sign off and say, 'Yes, this improves our codebase rather than regresses it.' And one of our best engineers as he's been playing with these models has found that the way he wants to balance is that he wants to control the interfaces. So he goes and still writes by hand here's the components here's how they fit together maybe here's the file structure but the details of how it gets implemented which often are quite intricate and complicated those he outsources to the machine and so I think that the difference is really if you don't pay attention to a particular aspect of what the model's doing will it be something when you look under the hood that you're proud of and that right now we're evaluating that we're seeing in what circumstances the answer is yes versus No, but we're saying that our requirement is that if someone does go look that the answer should always be yes, I am proud of this. Or alternatively, you can have some sections of code that you don't care if they're good or not as long as they meet the correctness specification. And so if you have good ways of verifying that that section is correct, then you can have very highly optimized code that's extremely hard for anyone to build on. But you don't think of it as a as the way that we think about code is this evolving artifact. You view it as a one-off. Okay, produce this. I'll produce a totally different version, but never build on top of it.

是不是大部分情况就像在制造积木，然后它们连接的方式可能更随意？

Is most of what's happening like building blocks being made and then maybe the way they link together is more casual.

是的。

Yes.

这算是一种描述方式吗？

Would that be a way to describe it?

我认为这是一个很好的描述方式。我认为接下来会发生的是，积木的尺寸会随着时间的推移而增大。

I think that's a pretty good way to describe it. And I think what's going to happen is that the size of the building blocks is going to increase over time.

人类监督的方式也会随着时间的推移而提升抽象层次。因为现在你查看每一个单独的智能体，即使在今年内也会感觉完全原始和缓慢，因为你会想要一个监督智能体来查看所有这些不同智能体的工作，并向你标记：这个细节看起来不太对，这个智能体似乎偏离了轨道，这个看起来不像你想要的。所以我认为真正要弄清楚的是，作为人类，你如何管理越来越庞大的智能体舰队？在很多方面，根本的衡量标准是单个人类能调动多少算力，这将是未来生产力最重要的指标。

And the way in which the human oversees is also going to move up that level of abstraction over time. Because right now you look at every individual agent, that's going to feel totally barbaric and slow in even just this year because you'll want an overseer agent that's looking at the work of all these different ones and flags to you. This particular detail doesn't look quite right over here. This agent seems to have gone off the rails. This, you know, doesn't look like what you wanted. And so I think that really figuring out how do you as a human manage a larger and larger fleet of agents? And in many ways it's really about the fundamental measure is how much compute does an individual human marshall that is going to be the most important metric for future productivity.

你一生中见证的最大的技术革命是什么？

What would you say the biggest technological revolutions you've witnessed over the course of your life each one?

我记得在北达科他州长大，读到一篇《时代》杂志关于硅谷的文章，感觉自己生得太晚了。所有激动人心的事情

Well I remember growing up in North Dakota and reading like a Time magazine article about Silicon Valley and feeling like I was born too late. all the exciting things

都已经在发生了。

that was already happening.

它正在发生，而我却不在那里。是的。我太年轻了。感觉世界上只有这么多好主意，只有这么多创新可能，而我看到它正在发生，却无法参与其中。所以那是我感到错失恐惧的时刻。我觉得哇，我想我再也不会看到这样的景象了。

It was happening and I wasn't there. Yeah. I was too young. It just felt like there's only so many good ideas in the world, only so much innovation that's possible and I can see it happening right now and I am not part of it. So that was a moment that I felt the FOMO. I felt like wow I don't think I'll ever see anything like this again.

是的。

Yeah.

后来我记得像手机和那里的转变。我从未对此非常热衷。很多人想开发应用。对我来说，它从未真正让我深深着迷。我认为现代生活是由许多不同的技术片段组成的。想想看，比如 Uber。

And I remember later things like mobile phones and the shift there. It was never something I was very passionate about. Like there a lot of people wanted to build apps. For me, it never really felt like the thing that I was deeply attracted to. And I think that there's so many different pieces of technology that add up to modern life. Think about, for example, Uber.

想象一下在 1950 年向某人描述 Uber，对吧？向正在撰写 AI 论文的艾伦·图灵描述。你得解释计算机，解释互联网，解释 GPS，解释很多很多东西。然后所有这些疯狂技术创新的最终效果，就是让你能在几分钟内叫到一辆车。从某些方面看，这感觉像是一个微不足道的结果。但从另一些方面看，它又如此深刻，你会意识到所有这些神奇的技术——它们确实是魔法，对吧？如果你在它们出现之前向人们展示，他们会觉得几乎不可能。但现在我们视之为理所当然，我们可以用它们来让生活更便利，有时甚至真正增强我们、加速我们。对我来说，这就是我们创造的每一项技术的潜在主题。

Imagine describing Uber to someone in like 1950, right? To Alan Turing when he's writing his paper on AI, right? So you got to explain computers, you have to explain the internet, you have to explain GPS, you have to explain so many things. And then that the net effect of all this crazy technological innovation is so that you can get a car to you in a few minutes. And in some ways it feels like a trivial outcome. But in some ways it's so deep and you realize that all these magical pieces of technology and they truly are magic, right? If you aligned it before them that they would feel almost impossible. But now we take them for granted and now we can use them to just make our lives sometimes more convenient but sometimes it's to really enhance us and to accelerate us. And to me that was the underlying theme of each piece of technology that we've created.

在你参与的每一个开创性情境中，你事先对结果了解多少？

In each of the pioneering situations that you've been involved in, how much do you know in advance of what it's going to be?

我会说我总是带着一个愿景开始，对吧？我总是带着一种思考，先抛开所有可能失败的理由，只梦想它能成为什么。它能成为什么，在物理定律范围内什么是可能的，对吧？梦想不可能的事没有意义，但想象它如何发展。我认为 Stripe 就是这样：思考我们可以构建一个惊人的支付巨头，让支付更便捷，让更多事情发生。是的，听起来很棒。而且那里的人很优秀，你可以一起学习和成长，在世界做更多事情。所以对我来说，那感觉像是：好的，我看到了它可能的样子。很多细节可以后来解决。对于 OpenAI，你想到 AGI，如果你能真正构建出 AGI，并以一种提升所有人的方式实现，那没有比这更好的事情了，对吧？那是你在技术意义上最希望做到的最惊人的事。对于 OpenAI 内部的项目，我认为通常也有同样的感觉。我记得有些时刻，我看到一个演示或一个结果，你看到一条初始曲线，就意识到这行得通。我记得有一次，我的联合创始人 Wojciech 说，这个领域如此非凡的一点是，每个想法都行得通，只要它在理论上有动机，数学上成立，它就会发生，它会成功，你会得到好结果。但真正的挑战是找出哪些是通往目标的最快路径。所以感觉就像果实掉在地上，因为都掉在地上，有时挑战在于机会如此巨大，但找出路径需要这些证明点。我认为我们现在正处在一个这样的点上，非常明显，这些智能体不仅擅长软件，还会擅长所有知识工作。我认为今年我们将在工作方式上迎来一个非常大的转变。

I would say I always go in with a vision, right? I always go in with a thinking about the just set aside all the reasons this might fail. Just dream of what it can be. What it can be like what is even possible right within the laws of physics right you there's no point in dreaming of the impossible but how it could go and I think that stripe was like that of thinking about well we could build something that will be this amazing payment behemoth right that actually makes payments more accessible. They make payments more accessible more things happen. Yeah, that sounds great. And that the people there, great people, you can learn and grow together and you could do more things in the world. So that to me was something that felt like okay, I see the what this could be in in that way. Many details could be figured out. Open AI, you think about AGI, it's like if you can actually build an AGI into existence in a way that lifts up everyone, there's no better thing to work on, right? that is the most amazing thing you could hope to do in a technological sense. And then for projects within Open AI, I think usually it has the same kind of flavor to it where I remember there are moments where I see a demo or I see a result. You see a little initial curve and you just realize this is going to work. I remember at some point actually Voych one of my co-founders saying that the thing about this field that is so remarkable is every idea works right as long as it's theoretically motivated the math works it's going to happen right it will actually succeed you'll actually get good results but the challenge is really figuring out which ones are going to the fastest path to your objective and so it really feels like there's fruit lying on the ground and because it's all lying on the ground sometimes that's the challenge there's such a massive opportunity case, but figuring out the path through, you need these these proof points. And I think we're sitting in the middle of one right now, very very clear with these agents getting so good at not just software, but they're going to get very good at all knowledge work. I think this year we're going to have a very big transition in how work is done.

我认为 AI 的一个最大问题是，因为它能做的事情太开放了，人们很难想象它会做什么。它是一个能做任何事的工具。

I think it's one of the biggest issues with AI in general is that because it's so open-ended what it can do, it's hard for people to imagine what it's going to do. It's the tool that can do anything.

是的。你知道，这很难。

Yes. You know, it's hard.

没错。是的。我们在 ChatGPT 上就有这个问题，对吧？你打开 ChatGPT，有一个文本框，可以做任何事。

That's right. Yeah. We we have this problem with chatbt, right? You show up at CHACHBT, there's a text box that can do anything.

是的。

Yeah.

但你想让它做什么？很多人就卡在空白页上。我认为这恰恰展示了 AI 的机会。

But what do you want it to do? Many people just you're stuck on the blank page. And I think that that actually the duel of it shows the opportunity with AI.

我认为最终，AI 在我看来就是机会。这就是我们构建它的原因。机会在于，如果你有愿景，如果你有想要的东西，如果你有认为事情应该以特定方式完成的看法，如果你有希望在世界上实现的主观能动性，我们有适合你的工具。随着这些工具变得更好，会有很多问题，比如人类的位置在哪里？做人类意味着什么？所有这些。我认为主观能动性、驱动力、愿景，这些是我们作为人类必须贡献的东西。

And I think that ultimately AI in my mind AI is opportunity. Like that is why we build it. The opportunity is that if you do have a vision, if you do have something you want, if you do have a particular way you think things should be done, if you have agency that you wish to to see enacted in the world, we have the tool for you. There's lots of questions as these tools get better like where do the humans fit in? What does it mean to be human? All these things. And I think that agency, that drive, that vision, those things, those are something that we as humans have to contribute.

是的。什么是 AGI？它有没有非常清晰的界定？我认为 AGI 不仅仅是一个能完成人类任何智力任务的系统，而是能真正成为个体的力量倍增器，让他们能像有远见的 CEO 一样运作，达到那种赋能个体的水平。对我来说，它真的不仅仅是这个系统在理论上能做某事的技术抽象，而是它如何实际部署到世界中。这是一个新的定义。

Yeah. What is AGI and is it very clearly delineated? I think of AGI as not just a system that can do any intellectual task that humans can do, but that can really be this force multiplier for an individual human to the extent that they can operate as the visionary as the CEO and so that it's that level of empowerment for the individual. And to me, it's really not just the technical abstraction of this system that theoretically could do something. It's how it's actually deployed into the world. That's a new definition.

是的。我逐渐也是这么想的。

Yes. It's how I've grown to think about it.

嗯。

Yeah.

因为如果你看我们的使命，它不仅仅是一个技术问题，对吧？你可以定义技术上的 AGI，但我认为从 OpenAI 一开始，我们就对只写论文的想法不满意，对吧？你可以写所有你想要的论文，描述如何构建 AGI，但没有影响，对吧？影响在于将其实例化到人们的生活中。我认为超级智能将超越那个标志和能力。我认为我们想要的结果是帮助解决完全超出我们能力范围的问题。

Because if you look at our mission, it's not just a technical problem, right? You can define the technical AGI, but I think for the very very beginning of OpenAI, we were unsatisfied with the idea that we would just write papers, right? You could write all the papers you want that describe how an AGI would be built, but there's no impact, right? The impact is the instantiation into people's lives. And I think that what super intelligence will be is something beyond that signpost and capability. And I think that the outcomes we want from it are to help us solve problems that are totally out of our reach.

对。

Right.

对。我认为解决疾病、太空旅行，我们看到很多很多问题。

Right. And I think that solving diseases, I think that space travel, I think that there are many, many problems that we see.

ChatGPT 是 OpenAI 的主要产品吗？不，我们最终销售的是按需智能，为你的问题提供智能。ChatGPT 是其中的一个实例，非常受欢迎，每周有近十亿活跃用户。但这还不是终点。我们有一个 API，许多商业客户在此基础上构建，增长也非常惊人。我们的最新产品叫 Codex，它正在改变 OpenAI 内部乃至整个行业构建软件的方式，我们正看到它的起飞。关于 Codex，它实际上是两样东西：它是一个通用智能体框架，能够使用任何类型的工具、任何类型的应用；同时它也是一个知道如何编写代码的系统。但如果你想想第一点，它非常有价值，并且可以重新用于任何你可能想要的知识工作任务，任何你想用电脑做的事情都可以表达为一个智能体执行某些操作、编排某些工具。

Is chat GPT the primary product of open AI? No, what is the thing that we ultimately are selling is intelligence on demand for your problem. Chat GBT is one instantiation of that and it's massively popular. Almost a billion active users every single week. But that's not the end of it. We have an API that many business customers build on top of and that is also growing absolutely phenomenally. Our newest product is called Codeex and it's transforming how software is built within OpenAI and really within the industry as a whole and that again is something that we're just seeing this takeoff. And the thing about what codeex is, it's really two things. It is a general purpose agent harness so that's able to use tools for any kind of tools, any kind of application and it's a system that knows how to write code. But if you think about that first thing that is extremely valuable and repurposable for any knowledge work task you might want, anything you might want to do with your computer can be expressed in terms of an agent doing some things, orchestrating some tools.

所以我们开始把它应用到 Excel、PowerPoint 这类工具上，能够创建商务人士在各种业务职能中产出的这些制品。我们花了很多时间让我们的模型在这方面变得非常非常强大。因此，我认为到年底我们会看到一个非常不同的产品面貌，每个知识工作者都将拥有这个真正赋予他们超能力的工具。

And so we're starting to apply it to things like Excel, PowerPoint, being able to create these artifacts that people in business produce in various business functions. And we're spending a lot of time to actually make our model very very capable at this. And so I think what we're going to see by end of year is a very different product surface where every knowledge worker will have this tool that really gives them superpowers.

什么是智能体式 AI？

What is agentic AI?

我认为智能体式 AI 是一个模型，你不只是像在聊天中那样与它对话，而是它连接了工具，在实现层面上，它几乎可以像模型一样与你（人类）聊天，但也可以与系统聊天，它可以要求运行这个命令、创建这个电子表格、在某个人的电子邮件中查找这个东西，所以它需要访问外部世界，它实际上能够采取行动。它不像 ChatGPT 那样只是脑力上的影响，对吧？当你与它交谈时，它会回应你。对于智能体式 AI，它实际上嵌入在现实世界中，可以采取行动。但关于智能体需要注意的一点是，我们一直在显著增加智能体可以运行的时间。所以现在你可以让智能体在一天内完成有用的工作。而且应用智能体的空间非常大，因为你可以并行使用大量算力。你可以让许多智能体同时处理一个任务，它们能够产出人类需要非常非常长时间才能完成的东西。

I would think of agentic AI as a model that you don't just talk to like in chat but that it's hooked up to tools and that at an implementation level it's almost like the model can chat with you the human but it can also chat with the system and it can say please run this command please create this spreadsheet please look this thing up in someone's email so it needs access to the outside world it's actually able to take action. It's not just like ChatGPT its impact is more cerebral, right? As you talk to it, it talks back to you. For agentic AI it's actually embedded in the real world and can take action. But one thing to note about agents is that we've been increasing the time that an agent can run quite significantly. So you can have agents now that do useful work over the course of a day. And that the space of how you apply agents is very large because you can apply lots of compute in parallel. You can have many agents that are working on one task and that are able to produce things that would take humans very very long time to do.

跟我讲讲 ChatGPT 的机制吧。它是如何工作的？

Tell me a bit about the mechanics of ChatGPT. How does it work?

核心上，ChatGPT 由一个语言模型驱动。你应该把语言模型看作一个系统，它接收一些文本，然后输出另一些文本。这里的文本不一定非得是英语。它可以是图像，可以是视频，实际上可以是任何模态。输出也同样不一定非得是文本。

At the core, ChatGPT is powered by a language model. And you should think of a language model as a system that takes in some text and then outputs some other text. Now the text doesn't need to be literally English language. It can be images. It can be videos. It can be really any sort of modality. And the output similarly doesn't need to be literally.

所以它不只是语言。

So it's not just language.

它不只是语言。

It's not just language.

它从什么时候开始不只是语言了？

How long has it not just been language?

嗯，公平地说，即使从一开始，它也不只是语言，因为还有代码。

Well, to be fair, even from the very beginning, it wasn't just language because there was code.

你可以处理代码。我明白了。

You could do code. I see.

对。但一直都是文本。那是我们起步的方式。所以 GPT-4 有一个下游项目叫 GPT-4V，那是我们第一次拥有视觉能力，你可以让模型真正识别图像，你输入文本和图像，它会输出文本。从那以后，我们训练了能接收声音、图像、文本，输出声音等等的模型。所以实际上你可以和我们生产的模型进行完整的语音对话。这也是 ChatGPT 的一部分。所以我认为这些模型是通用智能处理器。当你创建它们时，训练它们的方式是它们观察数据，但并不是真正学习数据本身。它们学习的是生成数据的底层规则。这就是它们聪明的原因——它们是一种通用理解机器。你可以将它们指向任何你有代表性数据信息训练的任务。

Right. But it was always text. That was what we started with. So GPT-4, we had a downstream project of that called GPT-4V, which was the first time that we had vision and that you could have a model that would actually recognize images and you'd put text in and images in, it would output text. And since then, we've trained models that take in sound, take in images, take in text, output sound, etc., etc. And so you actually can have a full voice conversation with models that we have produced. And that's part of ChatGPT as well. And so I think of it as these models are general purpose intelligence processors. And when you create them, the way that you train them is that they look at data, but they don't really learn the data. They learn the underlying rules that created the data. And that's what makes them smart is that they're sort of general purpose understanding machines. And you can point them towards whatever task you have representative data information training for.

你们给这个大脑起名字了吗？

Do you have a name for the big brain?

嗯，我们给不同的模型起了名字。所以我们叫……

Well, we have names for our different models. So, we call...

但模型是使用部分，而那个学习一切的东西呢？

But the models are the use part of it, but the thing that learns everything.

是的，我们没有给整个训练系统起名字。所以各个组件有名字。例如，我们有一个训练系统负责系统的一个组件。我们有不同的系统来处理如何实际部署一个训练好的模型。但现在训练和推理也开始融合，因为你做强化学习，模型自我学习，类似于图灵所讨论的。

Yeah, we don't have a name for the overall training system. So, there are names for various components. For example, we have a training system that does a component of the system. We have different systems for how we actually take a trained model and serve it. But now also training and inference are starting to come together because you do reinforcement learning where the model teaches itself similar to what Turing was talking about.

嗯。

Yeah.

所以有一个整体的编排系统，每个组件都有自己的名字和概念，实际上整个行业都围绕着如何训练模型以特定方式帮助人们而建立起来。

And so there's kind of an overall orchestration system and each of these components they have their own names they have their own concepts and there's really a whole industry that's built up around how to train models for helping people in particular ways.

所有的知识都在一个大的基础中，然后被划分来做这些不同的事情吗？

Is all of the knowledge in one big base and then is it divided up to do these different things?

所以我们曾认为 AGI 会是一个巨大的模型。然后你提出了这些架构，比如混合专家模型，你可以把它想象成不是每次都将信息通过网络的所有部分，而是只通过较小的部分，因此在训练过程中有机会进行专门化，训练过程可以选择说，我想把这个专门用于语言，这个用于文本，或者这个用于编程，实际上这要更复杂一些。但我要说的是，最大的突破在于认识到，虽然人类为语言、视觉等创造的最有用的工具是模型，而且这些大模型运行成本很高，但它们可以使用工具。那么为什么不同时使用较小的模型呢？所以我认为我们正走向一个各种模型并存的世界。你现在看到这个领域几乎像寒武纪大爆发一样，人们训练各种开源模型，我们也训练各种不同大小的模型。所以不再只有一个系统，而是这些模型可以相互对话，并且专门用于不同目的。这确实引入了方法的多样性，意味着你可以以各种不同的方式进行专门化。

So we thought that an AGI would be literally one giant model. And you come up with these architectures that there's this thing called mixture of experts for example where you think of it as it has rather than every time you're running information through the network that it runs through just smaller parts of the network and so there's an opportunity during the training process to specialize so that the training process could choose to say well I want to specialize this for language and this for text and or this for programming in reality it's a little bit more complicated than that. But I would say that the big unlock has been to realize that while the most useful tools that humans are creating for language and for vision and things like that are models and these big models, they're expensive to run, but they can use tools. So why not use smaller models, too? And so I think that we're heading towards a world of this menagerie of different models. And you're seeing this almost Cambrian explosion right now in the field of all these open source models people are training and that we train models of all sorts of different sizes. And so there's not really just one system anymore that there's these models that can talk to other models and that are specialized for different purposes. And that really introduces a diversity of approach and means that you can specialize in all sorts of different ways.

什么是预训练，什么是后训练？

What is pre-training and what's post-training?

我认为预训练是一个阶段，模型观察世界并从中学习。在技术层面，我们实现它的方式叫做下一个词预测。所以你给模型展示一个序列，然后问接下来应该是什么，这个序列可以是任何东西。例如，它可以是互联网上某个网站的公开帖子。那么人类接下来会写什么词呢？它可以是爱因斯坦的思想，可以是某种非常深刻和重要的东西。如果你想一想，如果你能预测爱因斯坦说出的每一个词，你至少和爱因斯坦一样聪明。这里重要的是，模型被激励去学习的不仅仅是名词在哪里、逗号在哪里这些表面统计。它被激励去学习这个分布的底层规则。比如这些数据来自哪里？为什么在这里？因为如果你只是鹦鹉学舌般地重复别人说过的话，当你面对新事物时，这不会有帮助，而这正是这些模型被训练去做的事情。它真正关乎的是底层规则和生成，以及深入理解模型所处的任何新情境。

I would think of pre-training as a phase where the model observes the world and learns from it. At a technical level, the way that we implement it is with what we call next-token prediction. So you show a model a sequence and you ask what should come next and that sequence could be anything. For example, it could be a public post for some site on the internet. And so what is the word that human would write next here? It could be Einstein's thoughts. It could be something very deep and significant. And if you think about it, if you can predict every word out of Einstein's mouth, you are at least as smart as Einstein. And the important thing here is that what the model is incentivized to do is to learn not just where the nouns are, where the commas go, those kinds of surface statistics. It is incentivized to learn the underlying rules of this distribution. Like where does this data come from? Why is it here? Because if you're just sort of parroting back what someone else said, it's not going to be helpful as soon as you're looking at something new, which is what these models are trained to do. It's really about the underlying rules and generation and deeply understanding any new situation that the model's placed in.

这就是预训练。从技术层面讲，这是一个极其有趣的问题。你要把这些东西扩展到海量的计算设备上，实际要解决的问题是：把数据输入网络，然后反向传播。你大致能看到如何调整所有连接，以便在前向传播中得到稍微好一点的答案。思路是，当数据通过时，你会得到某种预测、某种输出，然后你发现——这就像你有很多电线连接起来产生某个结果，你该如何调整所有这些电线的松紧度，才能得到稍微好一点的结果。机器最终就是这样被编程的：你做前向传播，再做反向传播，然后有一个叫做优化器的步骤，你调整所有参数，然后一遍又一遍地重复。你在非常大规模上做这件事，所以任何单个数据点，甚至任何单个大规模数据源，其实都不重要，因为你在进行人类规模的学习。这个东西在观察整个世界。这有点像你童年中遗忘的部分，你不会忘记所有知识。它真正关乎的是对现实的底层背景理解。这就是预训练。在技术层面，你把计算切分到多个设备上，你总是试图提高效率，试图得到——不仅仅是更大的模型。我们在架构上也有各种创新，所以我们有形状不同的模型，你映射这些——很多工程挑战在于理解硬件如何工作、它的弱点在哪里、它如何失败，然后你如何设计最适应这些的架构。预训练的输出——通常这些运行可能持续一个月，也可能好几个月，我们最长的一次大概九个月——需要一大队人来维持运行。比如 GPT-4，我深度参与了预训练，构建了很多训练栈。凌晨两点我会因为运行中断而醒来，去修复它。这就是你必须做的。每一小时、每一分钟任务中断，你看着那些闲置的 GPU，想着浪费了多少美元。但更重要的是，这是一个错失的机会。

So that's what pre-training is. At a technical level, it is an extremely interesting problem. You scale these up to massive numbers of compute devices. The actual problem you have to solve is you put data through the network, then you pass it backwards. So you kind of see how you have to adjust all the connections in order to have gotten a slightly better answer from the forward pass. The idea is as you pass data through, you get some prediction, some output, and then you see, okay, it's almost like if you have all these wires that are connected to produce some result, and how do you adjust the tautness of all these wires in order to have gotten a slightly better result. And that's how the machine is ultimately programmed: you do this forward pass, you do this backward pass, and then you have a step called an optimizer step where you adjust all those parameters, and you do it again and again and again. You do this at very large scale, and so any individual data point, even any individual large source of data, doesn't really matter, because you're talking humanity-scale learning. This thing is observing the whole world. It's a little bit like if there was some part of your childhood that you forgot, you're not going to forget all this knowledge. It's really about the underlying background understanding of reality. So that's pre-training. At a technical level, you slice up this computation across many devices, and you're just always trying to pump more efficiency and trying to get—it's not just larger models. We also have all sorts of innovations on architecture, so that we have models that are shaped in different ways, and you map those—a lot of the engineering challenge is trying to understand how the hardware works, where its weaknesses are, how it fails, and then how do you design the architectures in a way that are most amenable to that. So the output of pre-training—and usually these runs could be a month, they could be multiple months, maybe our longest one was somewhere around nine months—big team of people in order to keep that thing running. GPT-4, for example, I was very involved in the pre-training and built a lot of that training stack. At 2 am I would wake up because the run was down, go and fix it. That's what you got to do. And it's like every hour that the job is down, every minute that it's down, you just look at the number of GPUs that are sitting idle and you think about how many dollars are being wasted. But more importantly, it's a missed opportunity.

这是一个错失的机会，而且你确实能非常具体地感受到。

It's a missed opportunity and you really feel that in a very concrete way.

然后是后训练。后训练中，你拿预训练的输出。你拿这个知道很多东西的模型，它已经见识过世界，你试图告诉它该如何运用这些知识。在不同情境下什么是正确的行为？这几乎就像你有一个孩子，他观察世界并学到了很多，20 岁了，现在该上大学了。现在你专门化，为特定领域进行教学。但这些模型有一点不同：你不一定是在注入新知识。你实际上几乎是在修剪它已经知道的东西，因为预训练过程和后训练过程所用的算力通常非常非常不平衡。后训练过程通常只需要几天时间。我们教模型的方式——我们进化了这些技术，但经典做法是通过反馈。我们会用某种方式说，好吧，我们训练一个奖励模型，也就是另一个 AI，它可以评判这个 AI 在做什么，你通常通过说这些是好的、这些是坏的，或者这里有两个 AI 生成的回答，哪个更好来训练它。然后这个奖励模型评判预训练模型，给它反馈，从而能够塑造它的行为。所以我们确实有能力拿这个几乎无所不知的模型。我记得和我们的研究员 Alec Bradford 聊过，他会这样描述：这些预训练模型，与其说像一个人，不如说像整个人类。一切都在里面，然后我们尽力去引导它。我们并不总能做对，但我想这是我们一直在努力和改进的很大一部分。

And then post-training. Post-training, you take the output of pre-training. So you take this model that knows a lot of things. It's seen the world and you try to tell it how it should use that knowledge. What's the right behavior in different circumstances? So it's almost like when you have a child that's observed the world and learned a lot, 20 years old, now it's time to go to college. Now you specialize and teach for a specific domain. But one thing that's different about these models is you're not putting new knowledge in necessarily. You're really almost pruning down what it already knows because the amount of compute in the pre-training process versus in the post-training process is typically very, very out of balance. The post-training process usually takes a couple days, that kind of thing. And a lot of how we teach the model—we've evolved these techniques, but classically we do it through feedback. We would have some way of saying that, well, let's train a reward model, so another AI that can judge what this AI is doing, and you train that one usually through saying these are good, these are bad, or here's two possible generations from an AI, which one's better than the other. And then this reward model judges the pre-trained model and then gives it feedback, and from there it's able to shape its behavior. So we really have the ability to take this model that knows kind of everything. I remember talking to Alec Bradford, who is one of our researchers, who would describe it as these pre-trained models, they're less like a human and more like a humanity. Like everything's in there, and then we do our best to steer it. We don't always get it right, but I think that's a lot of what we've been working on and improving.

那么预训练就像美国国会图书馆，打个比方。

So the pre-training is like the Library of Congress, let's say.

我会说预训练就像国会图书馆，而后训练几乎就是品味，让机器知道它喜欢哪些书，或者一旦检索到信息后该怎么处理。

I'd say pre-training is like the Library of Congress, and then post-training is almost taste, giving the machine a sense of which of those books it would like or what to do with that information once it's retrieved.

对于预训练，所有这些信息来自哪里？是互联网吗？

For pre-training, where's all the information coming from? Is it the internet?

我认为经典方法一直是互联网上的公开数据。但变化在于，随着这些 AI 变得聪明得多，你实际上希望以某种形式用它们自己的数据来训练它们。例如，强化学习，机器出去尝试解决一个任务，你学习——它在尝试解决任务过程中学到的一切现在都成为其知识库的一部分。

I would think of the classic approach has been publicly available data on the internet. What's been changing is that as these AIs have gotten much smarter, that actually you really want to train them on their own data in some form. For example, reinforcement learning where the machine goes out and tries to solve a task and you learn—everything it learns in trying to solve the task is now part of its knowledge base.

就是这个意思。而且再说一次，这不仅仅是知识本身。还有技能，对吧？这才是我们真正追求的。我们的梦想一直是拥有一个推理模型，一个纯粹的推理者，能够在任何新情况下找出正确的做法并出色完成。有时背景知识有帮助，但真正重要的是那些智慧以及快速适应的能力。这才是真正带来价值的东西。

That's the idea. And again, it's not just the knowledge itself. It's also the skills, right? And that's what we're really getting at. Like our dream has always been to have a reasoning model that is just a pure reasoner and that is able to in any new situation be able to figure out the right thing to do and to do a great job there. And sometimes having background knowledge is helpful, but it's really about those smarts and the ability to adapt very quickly. That is like the real thing that delivers the value.

后训练中人的参与程度有多大？

How much of the human hand is involved in the post-training?

这也在变化。过去我们确实会有大规模的数据标注活动，有时现在也还有，但我们训练的大部分数据是——我们让人类辛苦地标注这些不同的例子。但问题是，随着模型能完成的任务越来越复杂，从大多数例子中能学到的东西越来越少。所以你确实需要深耕自己领域的专家。因此我们产生的一些任务极其复杂，比如查找某一年份的财务报告，判断这个与那个相比如何，对业务实力意味着什么，你知道，几段提示词加上一个非常具体的答案，这需要一定的领域专业知识，并且需要该领域专家花费十几个小时才能完成。所以我们确实提升了任务的复杂程度以及我们教机器的方式。这很合理，对吧？随着机器变得更聪明，我们确实需要找出它们在哪里出问题。

It's been changing as well. It really used to be that we would have these large campaigns and sometimes we still do, but that most of the data that we train on would be—we'd have humans who would painstakingly label these different examples. And the thing is that as the tasks the models are capable of has gone up, there's way less to learn from most examples. So you really need domain experts who are deep deep in their field. And so some of the tasks that we produce are these incredibly complicated like look up this finance report from this specific year and judge how this one compares to that one, was it mean for the strength of the business, you know, paragraphs of prompt with a very specific answer, and it requires some domain expertise to do and would require a dozen hours from that domain expert to accomplish. So we've really moved up the sophistication of the task and the way that we teach the machine. And it makes sense, right? As the machines get smarter, we really need to figure out where they are breaking down.

而且它变得不再那么依赖海量数据，而是更注重极高的品味、非常精准的目标，比如我们想让这个 AI 解决哪些最重要的问题，然后据此教导机器。我从 AlphaGo 故事中得到的启示是，电脑下出了一步人类绝不会下的棋，并因此获胜。

And it becomes much less about massive volumes of data and much more about this very high taste, very targeted like what are the most important problems that we want this AI to solve and trying to then teach the machine accordingly. My takeaway from the AlphaGo story was that the computer made a move that no human would have made and that's why it won.

是的。

Yes.

如果你教 AI 像人类一样更负责任地行动，那会不会削弱它下出人类不会下的那步棋的能力？这难道不会破坏整个 AI 的前提吗？

And if you're teaching the AI how to act more responsibly as a human, wouldn't that undermine its ability to make the move that the human wouldn't make? And doesn't that undermine the whole AI premise?

所以，我们研究的目标在很多方面，是的，就是要在科学、编程和其他领域实现那种 AlphaGo 时刻，因为那正是你想要的——发现新知识。你说得完全正确。如果你只是从已有的东西中学习，那怎么可能更进一步呢？但那种观点忽略的是，我们现在正在将训练扩展到强化学习。所以不仅仅是利用公开数据、让人类提供是或否，而是真正拥有让你在现实世界中测试事物的工具。就像人类如何发现新事物？有时我们深入思考，但我认为通常是通过实验，对吧？当我们尝试某件事，它没有成功，我们意识到，哦，我认为这其中的一部分原因是宇宙几乎就像一台巨大的计算机。它拥有的算力远超我们的大脑，也远超我们的 GPU。所以这就是为什么实验能带来学习，因为存在一个计算过程，其复杂程度超乎想象。而且你可以在某种程度上按自己的意愿塑造它，对吧？你可以让一个球从某个高度的斜坡上滚下来，如果 AI 能够提出这个实验并观察实验结果——也许有机器人来设置，也许由人类执行——无论哪种方式，这就是你真正发现新知识的方法。我认为，进行实验，然后将所有结果压缩到模型中，让它理解生成这些数据的底层规则，这正是我们刚刚开始看到成果的地方。

So the objective of our research in many ways, yeah, is to achieve that type of AlphaGo moment but in science, in coding, in these other domains because that is exactly what you want is you want new knowledge discovery. And you're exactly right. If all you're doing is just learning from what's been done, it feels like how are you going to go further? But the thing that that perspective misses is that the way that we are now extending the training to reinforcement learning. So it's not just take the public data, take the humans providing yes and no, but actually have tools that let you test things out in the world. Like how do humans discover new things? Sometimes we think deep thoughts, but I think that usually it's through experiment, right? As we tried something, it didn't really work. We realized, oh, and I think that part of what's going on is the universe is almost this massive computer. And that it has far more compute in it than our brains do. It has far more computing in it than our GPUs do. And so that's why there's something to be learned from experiment because there's this computational process that is just unimaginable in terms of how sophisticated it is. And that you can kind of shape it to your will, right? You can kind of have a ball that rolls down an incline of this height and that height and then if the AI is able to propose that experiment and see the experimental results maybe has a robot that sets that up for it maybe it's a human who performs it either way that is how you can actually discover new knowledge and I think that the idea of actually have experiment and then be able to compress all this into this model and for it to understand the underlying rules how all that data was generated. That is something that we are just starting to see the fruits of.

如果 AI 在物理学中证明了某些与当前教科书相悖的东西，那是一个安全问题还是一个突破？这已经发生了。

If the AI proved something in physics that negates what's in the current textbooks, is that a safety problem or is that a breakthrough? It's already happened.

给我讲讲这个故事。

Tell me the story.

有一位物理学教授，他一直公开对 AI 持怀疑态度。是的。我们说服他使用我们最新的未发布 AI 系统，他给了系统一个量子物理学中的特定假设，这个假设他原本计划与合作伙伴花一整年来研究。这是一个非常困难的问题，人们相当肯定有一个特定的答案。是的。而 RII 证明事实恰恰相反。他的反应是，这是第一次感觉系统在正确思考。里面有新知识，有非常非常新颖的东西，那篇论文正在投稿发表。但我认为这是一个非常重要的时刻，非常具有代表性。

So, there's a physics professor who has been a vocal skeptic of AI. Yes. And we convinced him to use our latest unreleased AI system and he gave it a particular hypothesis in quantum physics that he's been planning on working on this whole year with his collaborators. It's a very hard problem that people are pretty sure that there's a particular answer to it. Yeah. And RII proved that actually the opposite was true. And his reaction was that this is the first time that it's felt like the system is thinking right. There's new knowledge in there. There's something very very novel and that paper submitting it to be published. But I think it's a very significant moment and very representative of things to come.

是的。我认为 AI 最大的潜力在于它做出人类不知道正确答案的事情。

Yeah. I think the most potential in AI is when it does things that humans don't know is the right answer.

是的。这正是令人兴奋的地方。

Yes. That's what's exciting.

确实如此。

It is.

如果企业的视角是阻止这种情况发生——我能理解有人主张我们不能打破现状，这是被接受的，科学是被接受的——那正是我对 AI 最担忧的地方。最大的潜在缺点就是它无法发挥其能力。我认为这是一个非常重要的观点，而围绕它的更大框架是，我们有能力引导这项技术，这实际上是我们创建 OpenAI 的深层动机之一：我认为 AGI 无论有没有我们都会出现，但我们认为我们可以影响它朝着我们认为对世界更积极的方向发展，这就是我们的雄心和愿望。我认为通过实际创造这项技术可以带来很多价值。核心在于，我们认为它应该对所有人开放，就像人类可以质疑智慧一样——人类质疑智慧时总是很艰难，对吧？有很多抗体试图阻止这一点。但这也是社会进步的方式。我认为我们将不得不作为社会做出选择。这不只是任何一家公司或任何个人能决定的。作为社会，我们应该决定什么是规则。我们思考了很多，认为应该有社会决定的广泛界限，AI 永远不能跨越；在这个界限内，人们需要被赋予权力去选择，人们需要有一个能代表他们的 AI，无论是关于他们的价值观，还是关于质疑的能力。可能有些情况下人们不希望那样，那应该是他们的权利，他们的选择；也应该有人们试图发现新科学或其他东西的情况，那也应该是他们的选择。我认为我们确实拥有这种哲学，这与该领域的其他人非常不同，是自我赋权的，这项技术真正属于每个人。

And if the corporate perspective is to prevent that from happening, which I could see an argument for we can't rock the boat. This is accepted. The science is accepted. It's where I get most concerned about AI. The biggest potential downside is that it can't do what it's able to do. I think this is a really important point and I think that the bigger framework around it really is that this technology we have the ability to steer it and that's actually one of the deep motivations for why we created OpenAI is that AGI I think it will happen with or without us but we think that we can help be an influence on it playing out in a direction that we think is more positive for the world and that That is our ambition and our aspiration. I think there's a lot of value to be delivered by the actual creation of the technology. The core really being that we think it's something that should be available to everyone and it should be something that just like humans can question the wisdom and it's always tough when humans question the wisdom, right? There's a lot of antibodies that that try to prevent that. But that is also how society moves forward. And I think that we're going to have to make choices as a society. It's not just for any one company, anyone individual to decide. It's something that as a society we should decide what are the rules of the road and a lot of how we've thought about it is there should be broad bounds that society decides an AI can never cross within that that you really need people to have the empowerment to pick that people need to have an AI that can represent them whether it's about their values or whether it's about being able to question and there may be contexts where people don't want that that should be their right that that should be their choice and there should context where people are trying to discover new science or whatever it is and that should also be their choice and I think that we really have this philosophy it's very different from from others in the field of self-empowerment and that this technology really is for everyone

存在 AI 泡沫吗？

Is there an AI bubble?

我认为我们会发现我们在算力方面过于保守了。

I think that we will find that we were under ambitious on compute.

我认为我们正在走向一个知识工作被算力放大的世界，而且我们已经看到了证明。我看到有人发推说，如果你要找工作，问问你的预算中有多少 token。现在这还是个玩笑，但今年晚些时候就不会是玩笑了。我认为这些工具正在改变软件工程，那些还没尝试过的软件工程师还感受不到，但尝试过的人能感受到，他们从骨子里感受到。我认为我们会在金融领域看到这一点，在销售领域也会看到，人们将能够做更多的事情。我们已经看到 OpenAI 内部的个人想要 100 块 GPU，甚至一千块 GPU 专门给他们用。是的。如果你想到一个人需要一千块 GPU，那么一千个这样的人就需要一百万块 GPU，而目前全世界还没有一千万块 GPU。所以只能扩展到一定程度。所以我认为我们正处在一个看到未来、看到这项技术能力的时代。

I think that where we are going and we're seeing the proof points of it is a world where knowledge work is amplified by compute power. I saw someone tweet saying if you're taking a job ask how many tokens are in your budget. It's kind of a joke right now. It's not going to be a joke later this year. And I think that the degree to which these tools are changing software engineering like the software engineers who have not tried this don't feel it yet. But those who have, they feel it. They feel it in their bones. And I think that we're going to see that across finance. We'll see this in sales and people will be able to do so much more. And what we're already seeing is individuals within OpenAI who want 100 GPUs, a thousand GPUs dedicated just to them. Yeah. And if you think about a thousand GPUs for an individual, you have a thousand such individuals, you're at a million GPUs already and there's not 10 million GPUs in existence. So you can only scale so far. So I think that we are in a world where we are seeing what's coming, what this technology is capable of.

当我们看到模型变得多好，看到我们自己的收入曲线，以及那些因为算力不足而无法推出的东西，所有这些加在一起，就解释了为什么所有超大规模云服务商、为什么我们都在努力建设算力。我认为未来的方向是，算力将成为一项基本人权。我认为，人们要具有经济生产力，甚至为了自己的生活，拥有的算力越多，生活质量就越高。所以我认为我们将不得不进入一个每个人都拥有算力的世界。

When we look at how good the models are getting, when we see our own revenue curves and we look at the things we cannot launch because we do not have the compute for it, all of these things together mean that it is actually quite rational why all of the hyperscalers, why we are all trying to build compute. Like I think where we're going is compute will be a basic human right. Like I think that people to be economically productive and even for their own lives, the more compute they have, the higher quality of life they can have. And so I think we're going to have to be in a world where everyone has access to compute.

它会一直是 GPU 吗？还是可能有东西取代 GPU？

Will it always be GPUs or might something replace the GPU?

我认为算力是根本，而计算机一直在变化，我们已经看到很多不同的方法，比如英伟达收购了 Groq，这是一种不同的计算设备方法，人们也在采用很多有趣的方法。所以我认为 GPU 是不同类型加速器的一个很好的代表。例如，我们自己正在开发一个定制加速器，叫做智能处理器。我认为这项技术在效率和可扩展性方面还有很大的提升空间。

I think that compute is fundamental and that the computer is always changing and we're already seeing lots of different approaches and even Nvidia for example has acquired Groq which is a different approach to a compute device and there's lots of interesting approaches people are taking and so I think that GPU is a good stand-in for the different types of accelerators. We ourselves for example have a custom accelerator that we're working on called an intelligence processor. And I think that there is lots of room to improve the efficiency and scalability of this technology.

整个运营中最昂贵的部分是什么？

What's the most expensive part of the whole operation?

算力。

Compute.

电力在其中扮演什么角色？

Where does electricity fit in?

可以说，你几乎可以把人工智能看作一个从电力到智能的制造过程。我们用电作为实际进行计算的一个输入。所以电力几乎就像驱动整个系统的水。

I'd say you can almost think of AI as a manufacturing process from electricity to intelligence. And that we use electricity as one input to how we actually do the computation. So electricity is almost like the water that drives the whole system.

它会随着时间的推移变得越来越高效吗？

Does it get more and more efficient over time?

是的。但有一点要注意，随着我们提高效率，而且提高很多，如果你看同比数据，我们往往将同等智能水平的价格降低 100 倍，有时甚至更多，你可以从我们的降价中直接看到。

Yes. But one note is that we also as we increase the efficiency and we increase it a lot like if you look year-over-year we tend to cut our prices for the same level of intelligence by 100x sometimes like literally you can just see it from our price drops.

嗯。

Yeah.

比如 2020 年我们推出 GPT-3 时，那是一个 1750 亿参数的模型。现在你可以获得同等水平的智能。我甚至没查过，但类似规模的十亿参数模型，你可以在手机上免费运行，完全没问题。

So like where we were for GPT-3 in 2020 and that was a 175 billion parameter model. You can get that level of intelligence. I haven't even looked but billion parameter model something like that like it's something that you could run on your phone for free no problem whatsoever.

OpenAI 对比 Anthropic、Gemini、Grok、DeepSeek。告诉我这些公司有什么不同。

OpenAI versus Anthropic versus Gemini versus Grok versus DeepSeek. Tell me about what's different about those companies.

对我来说，我真正关注的是我们自己。我认为竞争对手的作用几乎就像一张配速卡，让你了解自己的表现。有时他们会指出，哦，这里有一个我们没想到能实现的功能，他们做到了，好吧，我们可能也能做到。所以我认为这很有帮助。但我们一贯的做法是，我们在基础研究、在真正的范式转变上投入最多。你可以从语言模型、强化学习范式中看到这一点，而且我认为还有即将到来的范式，我们正在拥抱和捕捉，并真正进行长期投资。我认为这是非常突出的一点。实际上很有趣。今天有一位候选人，我向他介绍一个长期研究项目，他说：‘哦，我很惊讶。我从外面看，以为你们总是做得很快，没想到你们会做这些长期投资，而看起来快的原因是因为我们有一个项目管道，它们一个接一个地实现。’我认为我们现在拥有最智能的模型。如果你看强化学习栈，我认为我们有非常独特和新的东西。我认为每家公司都有自己的 niche。比如在消费级聊天机器人方面，我们是迄今为止使用最广泛的。谷歌去年在构建比以往更好的模型方面做得很好，而且他们有天然的分发渠道。对我来说，谷歌拥有大量算力、大量人才、大量用户，这些天然优势始终是核心。我认为 Anthropic 正在崛起，他们非常专注于编码。我认为他们做得很好的一点是，我们在某些方面专注于基准测试。我们专注于学术编程竞赛。我们在那里取得了惊人的成绩，但我们没有太多关注这些模型将如何在现实世界中使用。

For me I really focus on us. I think that what competitors are helpful for is almost as a pace card just to get a sense of how you're doing. Sometimes they can point out that oh here's a particular feature that we didn't think we could implement, they did it, okay we could probably do it. So I think it's helpful for that. But the way that we've always proceeded is that we invest the most in basic research, in the actual paradigm shifts. And you can see this where with language models, with the reinforcement learning paradigm, and I think there's upcoming paradigms too that we are embracing and capturing and really put long-term investments in. And I think that is one thing that really stands out. It's actually interesting. I had a candidate today who was pitching him on a long-term research project and he was saying, 'Oh, I was surprised. I thought from the outside it looks like you guys are always doing things fast and I didn't realize that you make these long-term investments and the reason that it looks like that is because we have a pipeline of them and they come to fruition one by one.' And I think we have the smartest models right now. If you look at the reinforcement learning stack, I think we have something that is very unique and new. And I think each company has its own niche. Like I think when it came to consumer chatbots, we are by far the most widely used. Google did a very good job last year of building better models than they had previously and that they have a natural distribution. It's always front and center to me the fact that Google has a lot of compute, a lot of talent, a lot of users, a lot of these natural advantages. And I think that Anthropic is coming on the scene and they focused very hard on coding. And one thing that I think that they did well was that we were focused on the benchmarks in some ways. We were focused on academic programming competitions. We had amazing numbers there, but we didn't focus as much on how these models will be used in the world.

我明白了。

I see.

所以，在那些杂乱的代码库以及人们实际如何使用这些模型上进行训练。这是我们被延误的一个教训，但我们组建了一个团队，非常专注于这一点。我认为我们已经赶上了，而且人们确实看到了我们正在起飞的事实。

And so training on these like messy repos and how people are actually using it. And that was a lesson that we were delayed on, but we gathered a team focused very hard on that. I think we are caught up and I think that people are really seeing the fact that we are definitely on takeoff.

所以从某种意义上说，有其他公司确实让你知道还有其他事情需要关注。这是一个大局。

So in some ways having the other companies does give you an idea of other things to be focusing on. It's a big picture.

没错。这个世界上有太多东西，整个知识工作领域如此之大，以至于弄清楚到底要关注什么有时是最难的问题。

That's right. And there's so much in this world like the whole space of knowledge work is so large that figuring out exactly what to focus on is sometimes the hardest problem.

既然每个人的模型似乎都在针对相同的基准进行优化，这是否最终限制了所做的事情，因为每个人都专注于这一小部分测试？

Since everybody's models are being optimized it seems for the same benchmarks. Does that end up being a limitation on what's being done because everyone's focusing on this small group of tests?

有可能，而且我认为过去确实如此。如果你看事情的发展方向，我认为我们有点处于后基准世界。真正重要的是最终的基准：人们是否真的在使用它？

It can be and I think it has been in the past. If you look at where things are going, I think that we are in a bit of a post-benchmark world. What really matters is ultimately the benchmark of are people actually using it?

嗯。

Yeah.

你的收入在增长吗？就像那些无法作弊的基准。因为学术基准的问题在于，它们很容易达到 100%，对吧？你只要在测试集上训练就行了，对吧？数据污染了，数字本身不一定有多大意义。所以有时你会看到某个公司或某个模型出来，数字好得令人难以置信，而且总是如此。

Is your revenue growing? Like those kinds of benchmarks that are impossible to game? Because the problem with the academic benchmarks is they're very easy to get 100% on them, right? You just train on the test set, right? You contaminate the numbers on their own don't necessarily mean much. And so sometimes you'll see a model from a particular company or particular model comes out and it's got really good numbers that are too good to be true and it always is.

有没有理由不参与基准测试？

Would there be an argument to not participate in the benchmark tests?

实际上有两个答案。一个是，我们开发这些模型的方式是通过非常好的评估。所以我们确实想要那些告诉我们它们从不完美的基准。

So actually there's two answers. One is that the way that we develop these models is through very good evals. So that we actually do want benchmarks that tell us they're never perfect.

它们始终是代理指标，但确实能告诉我们是否走在正确的轨道上。例如，我们创建了一个名为 GDP-val 的系统，这是一个评估工具，用于衡量我们的模型在多项知识工作任务上的实用性。在这个指标上爬山是很好的，但到某个点你会饱和。通常一旦你在这些基准测试中达到 80% 或 90%，就意味着你完成了。没有必要追求 100%，因为那通常意味着你在为这个基准测试做一些非常具体且没有意义的事情。只要你在一个能提供有意义信号的范围内，那就很棒。你不应该关注太多基准测试；你需要精心构建的、能提供良好信号的基准测试。在过去的几年里，有时正确的做法是取一堆本身并不好的基准测试的平均值，这也能给出相当可靠的信号。但重要的是，你每天醒来不是想着如何提高基准测试的分数。把它当作一个代理指标、一个辅助指标。如果你过于关注它，你可能会变得‘好心办坏事’。

They're always proxy metrics, but they really tell us whether we're on track. For example, we've made a system called GDP-val, an eval that shows how useful our models are on a number of knowledge work tasks. Hill climbing on that is great, but at some point you'll saturate. Usually once you start getting into 80% or 90% on these benchmarks, it means you're done. There's no point getting to 100% because that usually means you're doing something very specific for the benchmark that doesn't make sense. As long as you're in a range where it gives meaningful signal, it's fantastic. You don't want to focus on too many benchmarks; you want well-constructed ones that give good signal. In previous years, sometimes the right answer was to take an average across a bunch of benchmarks that aren't that good individually, and that gave a pretty reliable signal. But the important thing is that you don't wake up every day thinking about how to move the number on a benchmark. Use it as a proxy, a side metric. If you focus on it too much, you can get what we call 'good-hearted'.

开发者做什么？他们是分配任务还是向你提出想法？

What do developers do? Are they assigned jobs or do they pitch you ideas?

我们做的事情有两大方面：研究和部署。在研究方面，我们创造新模型，这需要研究和工程紧密结合。在部署方面，通常是把研究成果带给世界。工程占比很高，但你需要高语境，真正理解研究，否则无法很好地将其落地。工程师在研究方面实际做的是：我们有一个想法，然后组建团队。有时是非常小的团队，两三个人，研究一个高风险的新想法，可能成功也可能失败。你只需要看到一些生命迹象，并且要有足够的耐心，因为第一次永远不会成功。有时我们知道它有效，已经有了生命迹象，然后我们扩大规模。所以一个想法的生命周期是：一旦有了想法，我们尝试，看到生命迹象，然后投入更多人。很多工程师专注于系统频繁崩溃，找出为什么太慢，或者数据导入下载率太低。他们经常构建分布式系统来处理事情，获取模型输出，观察它们，查看运行情况，构建仪表板。有很多工作要做，实际的机器学习工程是一门非常复杂的艺术，需要深厚的领域专业知识。做这个的人并不多。还有编写实际的内核——在 GPU 上运行的代码，将数据的低级表示转化为可计算的对象。这几乎就像你的眼睛将可见光转化为大脑中的信号。我们有从事低级神经元工程的人。所有这些都非常复杂，需要很长时间积累专业知识，并且有很多处于机器学习和工程交叉点的技术。例如，随着模型规模扩大，经典的做法是错误的；结果会变差，但你并不总是知道。随着模型规模扩大，曲线应该非常平滑吗？先验地，没有深层理由认为应该如此。但我们在某个时候意识到，我们需要不同参数之间的不同比例，并正确设置。我们开发了一种称为 MUP 的技术来设置初始化，这能让你在扩展时得到更平滑的曲线。所以有很多事情，即使知道哪里出错了也不容易，而深厚的专业知识和不同专长的人之间的合作才能产生伟大的成果。

There are two big aspects to what we do: research and deployment. In research, we are creating new models, which requires both research and engineering to be joined at the hip. In deployment, it's usually about taking the fruits of research and bringing it to the world. There's a very high bias of engineering, but you need to be high-context and really understand the research, because if you don't, you won't do a good job of bringing it to reality. What the engineers actually do on the research side is that there's some idea we're pursuing, and we usually form teams. Sometimes these are very small teams of two or three people working on a novel idea that's very high risk and may or may not work. There you just want to get some signs of life, and you need to have enough patience because it never works the first time. Sometimes we know it's working; we already have signs of life, and we scale it up. So the life cycle of an idea is: once we have an idea, we try it out, get signs of life, then put more people on it. A lot of the engineers are focused on the system breaking too frequently, figuring out why it's too slow, or why the data import has a low download rate. They often build distributed systems to process things, take outputs from the model, observe them, and see how things are doing, building dashboards. There's a lot of work to be done, and actual ML engineering is a very sophisticated art requiring deep domain expertise. There aren't that many people who do it. There's writing the actual kernels—the code that runs on the GPU, turning low-level representations of data into objects that can be computed upon. It's almost like how your eyes turn visible light into signals in your brain. We have people doing low-level neuron engineering. All of that is very sophisticated, takes a long time to build expertise, and there are many techniques at the intersection of machine learning and engineering. For example, as you scale up the model size, the classical way was wrong; it turned out you'd get bad results, but you wouldn't always know. Should it be the case that as you scale up the model, you get very smooth curves? There's no deep reason a priori to think so. But we realized at some point that we need to have different ratios between different parameters and get those right. We developed a technique called MUP for how we set the initialization, which gives a much straighter line as you scale. So there are many things where even knowing if something is wrong is not easy, and deep expertise and partnership between people with different expertise yields great results.

你关注过 Clawbot / Open Claw 的故事吗？

Have you followed the Clawbot / Open Claw story at all?

当然。

Of course.

说说你的想法。

Tell me your thoughts.

我很喜欢。我和 Peter 交流过；我觉得他很棒。他是 Open Claw 的开发者。对我来说，Open Claw 包含了两件事。首先，它是一个 AI 系统，你可以连接工具，它始终在线、运行并能够采取行动。我认为有一种我非常喜欢的黑客精神：我们有这些工具、这些模型，存在巨大的潜力——它们比我们目前使用的要强大得多。让我们试试，看看它能做什么。其次，有一个我们必须填补的空白：我称之为安全架构、信任架构。你有一个 AI 连接了许多东西，但你怎么知道它会做正确的事？你怎么知道它不会发送错误的信息？人们在 Twitter 上发布有趣的事情，比如一个人的妻子给他发短信，他的 Clawbot 在凌晨 2 点回复，婴儿在哭，妻子很不高兴。读起来很有趣，但想想这会导致什么。我们需要更好的护栏，为信任和安全设计的系统。这是我们关注的重点。我们不仅考虑构建技术和提升能力，还考虑如何以可扩展的方式将其带给世界。我们正在基于这项技术改造自己的企业，并希望帮助改造每个企业。所以我们对此的思考是，这是一个好兆头。要真正将其扩展到每个人，需要让人们设置起来极其简单，同时确保默认方式是安全的。这是我们大力投资的方向，我们做出了深刻的选择。例如，回顾 10 年，构建 AI 的一种愿景是保密一切，然后在无人知晓的情况下完成最后的部分，这样就没有部署压力。

I love it. I've spent time with Peter; I think he's great. He was the developer of Open Claw. To me, Open Claw encompasses two things. First, it's an AI system that you can hook up tools to, always on, running, and able to take action. I think there's a hacker spirit I really like: saying we have these tools, these models, there's this massive overhang—they're so much more capable than what we're using them for. Let's try, let's see what it can do. Second, there's a gap we have to fill: what I call the security architecture, the trust architecture. You have an AI hooked up to many things, but how do you know it will do the right thing? How do you know it won't send the wrong messages? People post fun things on Twitter, like someone whose wife was texting him, and his Clawbot replied at 2 a.m. with the baby crying, and the wife wasn't having it. It's funny to read, but think about where this goes. We need better guardrails, systems engineered for trust and safety. That's a lot of what we focus on. We think not just about building the technology and making the capability, but how to bring it to the world in a scalable way. We're transforming our own enterprise on the basis of this technology, and we want to help transform every enterprise. So how we're thinking about this is that it's a great sign of things to come. To really scale it to everyone requires making it incredibly easy for people to set up, but also making sure the default way things are done is safe. That's something we are investing in pretty heavily, and there were deep choices we made. For example, one vision of how you could build AI, rewinding 10 years, is to keep it all secret and then put together the finishing pieces with no one knowing you're doing it, so you have no pressure to deploy.

所以你真的有时间把所有事情都做对，然后按下按钮，造福世界。我一直觉得这个计划，我无法认同。感觉不对。首先，从技术角度来看，如果你从未遇到过现实情况，你怎么能确定你设置的安全系统是正确的？想想 ChatGPT，它从一个有趣的项目开始，这很好。很明显，要让它与非常复杂、重要的系统连接，并被赋予大量责任，我们需要开发新技术，但我们要在一个循环中学习和迭代。所以对我来说，要把这项技术做好，我们必须每一步都面对现实。其次是合法性：如果你要构建一项改变所有人生活的技术，我认为人们需要知道它，对吧？人们需要参与其中。所以看看 ChatGPT，我们做了一个非常深思熟虑的决定：我们认为这是应该让世界参与的技术，符合我们的理念，符合我们创办这家公司的初衷。它是为所有人服务的。

So you really have time to get it all right and then you push the button and benefit the world. And I always looked at that plan and I was like, I don't think I can sign up for this. It just feels wrong. First of all, from a technical perspective, if you've never encountered a reality, how will you be certain that you have put in place the right safety systems? You think about ChatGPT, it's nice that it's starting as a fun project. It's clear that for it to get hooked up to very sophisticated, important systems and be trusted with a lot of responsibility, we need to develop new technology, but we're going to learn and iterate in a loop. So to me, it always felt important to get this technology right, we have to encounter reality at each step along the way. But the second is legitimacy: if you're going to build technology that's going to change everyone's lives, I think people need to know about it, right? People need to be included in that. So if you look at ChatGPT, we made a very deliberate decision to say we think this is technology for the world to be included in, for our ethos, for what we started this company for. It's for everybody.

是的。

Yes.