人工智能正跨越人类知识的边界 | 凯文·韦尔

只要你有个有意思的想法，就没理由不做。

You have no excuse if you've got an interesting idea.

现在你能把任何想到的东西都创造出来。

You can now create anything that you can think of.

这些模型已经能解决人类从未解决过的问题。

The models can now solve problems that humans have never solved before.

正在突破人类知识的边界。

Going beyond the frontier of human knowledge.

我认为，这就是人工智能，乃至通用人工智能真正改变我们生活的方式。

That's how AI, I think, and AGI will really change our lives.

为什么不试着加速科学进程，把2050年的科学成果提前到2030年实现呢？

Why not try and accelerate science, bring about the science of 2050 but in 2030

你说对吧？

instead?

大多数人把人工智能看作是提升效率的工具。

Most people think of AI as a productivity tool.

凯文·韦尔认为，AI最大的影响可能出现在完全不同的领域。

Kevin Weil thinks the biggest impact may be somewhere else entirely.

他曾任OpenAI的首席产品官兼科学副总裁，现在专注于一个AI不再只是帮人写文档或生成代码，而是直接参与科学发现的未来。

Formerly CPO and Vice President of Science at OpenAI, Weil is focused on a future where AI doesn't just help people write documents or generate code, but contributes directly to scientific discovery itself.

这个想法很有雄心：利用AI加速数学、医学、材料科学等塑造人类未来的关键领域的突破。

The idea is ambitious: Use AI to accelerate breakthroughs in mathematics, medicine, materials science, and other fields that shape the future of human progress.

在这次对话中，凯文谈到了前沿科学、机器人实验室、AI推理、创业机会，以及他为何相信AI最重要的影响之一，在于拓展人类发现新知识的能力。

In this conversation, Kevin discusses frontier science, robotic labs, AI reasoning, startup opportunities, and why he believes some of the most important consequences of AI may come from expanding humanity's ability to discover new knowledge.

好的。

All right.

我觉得，现在活着真是一个了不起的时代。

This is an incredible time to be alive, I think.

你曾经参与打造并壮大了过去十年最重要的几家科技公司——Facebook、Instagram和Twitter，现在你又在OpenAI做同样的事。

Now you helped build and scale some of the most important technology companies of the last decade, Facebook, Instagram, and Twitter, and now you're doing that at OpenAI.

我刚才还问了ChatGPT，它对你怎么看。

I did ask ChatGPT what it thinks about you.

所以你可能猜到了，它对你的评价非常高，但你也确实是个非常出色的人。

And so as you would expect, it was very complimentary, but you're also like a very accomplished guy.

Kevin 思维缜密，不自负，而且像他这样参与过这么多重大产品的人，能保持如此踏实实在，实属难得。

So Kevin is thoughtful, low ego, and unusually grounded for someone who's been at the center of so many high stakes products.

当你面对这四个机会，还有其他无数个同样出色的选项时，是什么让你确信，这家公司、这个团队，就是你该加入的？

So how did you, like, as you looked at all those four opportunities, plus many others that were amazing, what gave you the confidence that this is the type of company, this is the team I should be working with?

我的第一条建议是：找个好伴侣。

Yeah, number one advice, marry up.

这个想法其实是我妻子最先提出来的。

It was my my wife originally.

我当时还在读研究生，攻读物理学学位，那时认识了我现在的妻子，她当时是斯坦福的梅菲尔德研究员，还曾在安德森的公司短暂工作过。

Actually, I was a I was in grad school doing a physics degree, and I met my now wife who was a Mayfield fellow at Stanford and actually worked at Andreessen for a little while.

正是她让我真正看清了硅谷的创业世界，还有这一切是怎么运作的。

And she was the one that kinda opened my eyes to everything, startups in the Valley and all that.

我是在西雅图长大的。

I I grew up in Seattle.

我爸爸在微软当工程师很多年，所以我从小就开始编程，但那时候心里想的还是数学和物理，读研期间也一直没离开这些。

My dad was a an engineer at Microsoft for a long time, so I grew up programming, but still was just like, you know, math and physics, math and physics as I went through grad school.

是我的妻子，伊丽莎白。

And it was my wife, Elizabeth.

她早年给我介绍了推特，因为她和杰西卡·瓦雷利是斯坦福的同学。

She introduced me to Twitter back in the day because she and Jessica Varelli knew each other from Stanford.

在推特工作了七年，看着公司成长，她和凯文·西斯特罗姆当年也是斯坦福的梅菲尔德研究员。

And after seven years at Twitter as it grew, she and Kevin Systrom were also Mayfield fellows together at Stanford.

所以这个联系就是这样来的。

So that's how that connection happened.

很多事其实都是我跟着我老婆走，沾了她的光。

And so, you know, a bunch of these things were just my wife me just following the coattails of my wife.

以前我一有新想法，就会时不时给萨姆打个电话。

I used to I used to call Sam periodically before whenever whenever I would, like, be thinking about doing something new.

我和萨姆不算特别熟，但熟到能随便打个电话聊聊。

Sam and I didn't know each other super well, but we knew each other well enough to to, you know, to do a quick phone call.

因为他总是同时插手好多不同的事情。

And because he he always has his, like, hands in lots of different things.

他一直在搞融合型初创公司，干这一堆玩意儿。

He's, doing fusion startups and and all of this stuff.

我记得2020年左右跟他聊过，他说人工智能不会先取代蓝领工作。

I remember talking to him, like, in 2020 or something, and he was like, you know, AI will not replace blue collar jobs first.

它会先取代白领的工作。

It'll replace white collar jobs first.

编程会是人工智能影响最大的领域之一。

Coding is going to be one of the big things for AI.

那是2020年，我们那时候基本不用AI，顶多用点传统的机器学习模型，比如排序信息流那种。

This was 2020, and none of us used AI particularly much, at least not outside of, like, classical ML models that, you know, ranking your feed and stuff.

我就记得当时心想，有道理。

And I just remember being like, yeah.

随你便吧，老兄。

Whatever, dude.

你懂的吧。

You know?

当然。

Sure.

不过咱们聊点别的吧。

But, like, talk let's talk about something.

后来呢，我给打电话，这次他说：你知道吗？

And so, anyways, I the OpenAI thing happened because I called him, and this time he was like, actually, you know what?

我们这儿有个职位空着。

We have this role open.

你来跟我们聊聊吧。

You should come talk to us.

我一听说，立马就说：我不在乎要不要工资。

Soon as I did, I was just like, I don't like, I'll work for free.

这简直是世界上最有意思的事了。

I don't just this is the most interesting thing in the world.

只要你给个offer，我就来。

And if you give me an offer, I'm coming.

幸运的是，他真的给了。

So fortunately, he did.

当时作为首席产品官，和萨姆一起的核心任务是什么？

And then the original mandate as CPO with Sam was to what?

那段时期里，你最满意自己达成了哪些成果？

And what are you sort of most happy with what you accomplished during that time?

是的。

Yeah.

一开始我是首席产品官，负责消费类产品、B2B产品、开发者产品等等。

Originally, I was a CPO, so leading our consumer products, b two b products, developer products, etcetera.

我的天，那增长简直像野草一样疯长。

And, I mean, man, I think it's we grew like a weed.

我这辈子从没见过什么东西能这么快发展起来，我觉得我们确实把AI带给了全世界很多人。

I've never seen anything grow that quickly in my entire life and I think brought, you know, kinda brought AI to to a whole bunch of the world.

我对这项工作感到非常自豪。

So very proud of that work.

几个月前，我们越来越清楚，我们的模型不能只在ChatGPT或Codex里表现优秀——虽然Codex已经是了不起的产品，但我们的模型已经达到了能解决前沿科学问题的水平。

A few months ago, it was getting clear that our models could not just be great inside ChatGPT or inside Codex, which I think is an incredible product, but we're at the level that they could start to answer frontier scientific problems.

现在的模型已经能解决人类从未解决过的问题。

Like, the models can now solve problems that humans have never solved before.

很多人批评AI说，哦，它不过是把不同地方的想法拼凑起来，给你做个总结而已。

So a lot of people, like, the criticism of AI is, oh, well, it's just it's just bringing together different ideas from different places and summarizing them for you to give you an answer.

它其实没法做出真正新颖的思考。

I I can't actually do novel thinking.

但我们现在看到，仅在今年一月，就有十到十二个开放的数学难题被解决了，主要靠的是GPT-5.2，最近也有几个是Gemini完成的。

But we've now seen there have been, I don't know, what, 10 or 12 just in January, 10 or 12 open mathematics problems solved mostly by GPT 5.2, now with some, a few recently by Gemini.

这些模型正在超越人类知识的边界。

Models are going beyond the frontier of human knowledge.

但我还不能说，它们解决的是人类完全无法解决的问题。

And I wouldn't I wouldn't claim yet that they are solving problems that humans can't.

我觉得，如果足够多的数学家去钻研这些问题，迟早能解出来，只是还没人做到而已。

I think if you took enough people and, you know, applied enough mathematicians towards some of these problems, they would have figured it out, but they had not figured it out yet.

模型超越了人类以往的所有尝试，这真的很了不起。

The model went beyond what we had ever done as humans, and that's pretty cool.

而这一切，就是现在发生的事。

And that's today.

过去几年我学到的一件事是，技术进步快得惊人——从‘AI根本做不到’，到‘AI勉强能做一点’。

And if there's one thing I've learned over the last few years, it's that you go very quickly from models could never do this thing.

一开始它做得差强人意，经常出错，但偶尔你会看到一丝希望，好像它真的快要能做到这件事了。

It is beyond the capability of AI today to models can just barely do this thing, and it kinda sucks at it and, like, it's wrong most of the time, but you get these glimmers of, like, oh, they can almost, you know, do this.

也许一开始只有百分之五到十的成功率。

Maybe it only works five or 10% of the time.

可六到十二个月后，你会发现模型已经能完美应对了，以后我只要遇到类似问题，一定会毫不犹豫地用AI。

And then six to twelve months later, it's like, models are great at this thing, and I would never I wouldn't I would always use AI anytime I ever do that again.

用评估语言来说，就是从零起步，短短时间就冲到百分之五到十，再迅速跃升到百分之八十。

Like, in eval language, you go very quickly from, like, zero to five or 10% to, like, 80%.

你懂的。

You know?

我们现在正处在前沿科学和人工智能的中间阶段，时不时就能看到一些令人惊叹的突破。

We are clearly in that middle phase with frontier science and AI where you have all these glimmers of like, wow.

AI居然能做我们以前觉得它根本做不到的事。

It can do something that we never thought AI could do.

还有什么地方比科学更适合应用这项技术呢？

So, you know, what more interesting place to apply it than science?

我觉得，这可能是我们每个人最能切身感受到通用人工智能影响的方式。

I think it may be the most tangible way that we all feel the impact of AGI.

要是今天把GPT-9直接放进ChatGPT里给你用，肯定特别厉害。

If we dropped, like, GPT nine inside of ChatGPT for you today, I'm sure it would be awesome.

但更了不起的可能是，我们由此获得了新材料、实现了超导、真正理解了宇宙的本质、有了个性化医疗——这才是AI和通用人工智能真正改变我们生活的方式。

But maybe even more awesome would be that we have all of these new materials and we have superconductivity and we understand the nature of the universe and we have personalized medicine and, you know, like that's how AI, I think, and AGI will really change our lives.

还有核聚变能源。

So And fusion power.

天啊，要是能加速科学进展，把本该在2050年才实现的科研成果提前到2030年实现，那该多好？

Man, try and, you know, accelerate science and bring about the science of 2050, but in 2030 instead?

这正是我们的目标。

And that's our goal.

你有没有过那种‘克劳德周末时刻’或者‘Codex周末时刻’？就是突然意识到，世界已经彻底变了，你过去所有的工作方式都将完全不同？

When did you have your Claude weekend moment or your Codex weekend moment where you're like, wait a second, the world has fundamentally shifted and everything that I've done before is going to be very different going forward?

是的，我对这个问题的看法不太一样。

Yeah, so I have a very different take on this.

我觉得这太棒了，也让人兴奋。

I I think it's awesome and exciting.

你想啊，编程过去一直是少数人的专长，全世界懂编程的人没多少。

Like, when you when you take something that has historically been the craft of a relatively small number of people, there aren't that many people in the world that know how to program.

大概三千万人吧，但一旦把这范围扩大好几个数量级，创造力就会爆发，因为很多人其实都有点子，只是以前根本没途径把它们实现出来。

I don't know, like 30,000,000 people maybe, and you expand it by a couple orders of magnitude, you get an explosion of creativity because lots of people have ideas, and sometimes they weren't they didn't have any route to actually implement those ideas.

从普通人现在也能创业，到我记得疫情后跟一位小城市官员聊天，那时候……

And you go everything from, like, thinking about people that can start companies now that didn't used to be able to start them all the way through to, like I remember sitting and talking with this was, like, a little bit post COVID with this city official somewhere, you know, like, small city.

他跟我讲，他们多希望能有一些程序，用来做些基础的信息服务，让这个小城的居民能更好地了解情况。

And he was telling me about the kinds of programs that they wish they had been able to to run and operate just like basic kinda information awareness things for the people that lived in this small city.

我当时根本不知道该怎么实现。

Is, like, I just didn't have any way to do it.

数据就在那儿摆着。

The data was sitting there.

我能想象出它该是什么样子，但就是没法做出来——他不是技术专家，城市也没钱花几千美元请人来干这事。

I could, you know, draw what I wanted it to look like, but I had no way of getting it done because he wasn't an expert, and they didn't have, you know, however many thousands of dollars to to hire a a person to do it for the city.

那现在你会怎么做？

Now what would you do?

你只要在Codex里输入一个指令，事情就解决了。

You would, like, enter a prompt into Codex, and it would be done.

我觉得这类变化太棒了。

So I think these kinds of things are awesome.

我个人说一下，我们一周前刚推出了Prism。

I'll say personally, we just launched Prism, like, what, a week ago.

有人用过吗？

Anybody see that?

用过Prism吗？

Use Prism?

试过吗？

Tried it?

这是一个专为科研人员打造的AI原生环境，用于科学写作和协作。

It's like an AI native environment for scientists to do scientific writing and collaboration.

如果你平时用LaTeX之类的工具，就特别合适。

So if you're using LaTeX, things like that.

因为产品规模小、团队也小，我最近花了很多时间写代码、修bug，这些事儿我好几年没碰了，但特别有意思。

Because it's a small product and a small team, I've been spending a bunch of time, like, writing code, fixing bugs, etcetera, which I haven't done in a bunch of years, and it's super fun.

而且在这期间，我还经历了一次转变——以前我整天都是开会，各种事情堆在一起。

It also I I I went through this transition in the middle of this, you know, because most of my day is meetings and, you know, we got a bunch of stuff going on.

后来有段时间，我坐在那儿，说实话，我自己都有点懵。

I went through this transition where I I I was sitting in I I don't know.

我记得。

I remember.

我想当时是和斐济开个会。

I think it was, like, Fiji a meeting with Fiji.

我合上了笔记本，因为我们都在努力改掉多任务处理的坏习惯，这可是OpenAI里最常见的问题。

And I closed my laptop because we're all trying to be better about not multitasking because that's, like, one of the worst habits that people at OpenAI have.

大家总是同时做好多件事。

Everyone multitask all the time.

我合上笔记本之前，Codex还没跑起来，当时我就想：糟了。

I had not gotten a Codex job running before I closed my laptop, and I was like, shit.

白白浪费了一个小时。

I just wasted an hour.

不是因为开会浪费的，而是因为我本可以利用那一个小时多做点事。

Like, not because of the meeting, but just because I could have been multitasking during that hour.

我的Codex代理本来可以帮我修个bug、加个功能，或者干点别的，结果我现在只能像农民一样干坐着开会。

My my Codex agent could have been, like, fixing a bug or implementing a feature or doing something for me, and now I have to sit in this meeting just like, you know, like a farmer being in a meeting.

别闹了。

Come on.

还有啊，你晚上睡觉前也会想，好吧，

And so and, you know, the same thing before you go to bed at night, you're like, okay.

我能给Codex安排个什么特别难的任务，让它自己跑上十个小时？

What, like, really hard task can I give Codex and just let it chunk away for, like, ten hours?

对我来说，这完全是另一个世界。

So, like, that to me is a is a different world.

如果你真擅长这个，就不会只同时处理一个任务。

And if you're really good at it, you are not just juggling one job.

你会有三四个任务在不同的工作分支里并行运行，反正就是那样。

You've got, you know, three or four things running in parallel across different work trees and, like, I don't know.

这世界真棒。

What a cool world.

对啊。

Yeah.

这种被释放出来的创造力简直难以想象。

The creativity that's gonna be unleashed is just incredible.

我想，一百年前或者一百五十年前，如果你是个手工家具匠人，工业革命一来，突然之间工厂大批量生产的家具比你做的还好，或者至少不相上下，你会觉得：等等，我的整个世界都变了。但现在，你可以设计出一种能被无数人同时使用的东西。

And I think about it as like if you were a handcraft furniture maker like a hundred years ago or one hundred and fifty years ago, industrial revolution hits, all of a sudden there's factory mass produced furniture that's better than yours or at least equivalent, And you're saying, wait a second, my whole world has changed, but now you can design something that can be used by so many more people all at once.

但人们之所以感到迷茫，是因为这种变革太剧烈、太快了，我都不知道……

But like it's this disorientation that people feel because it's so disruptive and so quick at I the same don't

你知道吗，现在人们还是看重手工定制的家具。

know, you think we'll end up with like, because now you still, people still value custom made furniture.

我觉得将来我们会看到大量定制化的网站。

I think we'll end up with like bespoke websites.

这个是真人做的。

This was done by a human.

没错，完全没错。

Exactly, exactly.

就像有人把开车当爱好一样。

Same as people will be like, I drive cars as a hobby.

对。

Right.

然后说：别上路了，你太危险了。

And be like, Get off the road, you're dangerous.

对，你就待在那边，专门给人类司机用的区域吧。

Right, yeah, stay over there in your little area for human drivers.

事情发展这么快，你刚才提到你是怎么跟上节奏的。

With things moving so quickly, you gave an example of how you stay up with things.

你和团队一起工作，保持紧密联系。

You're doing work with your team, you're staying close to your team.

你还有没有别的方法，既能带领团队、把握战略，又能紧跟公司内部和其他公司发生的各种变化？

Do you have any other ways that you're both able to lead the team, lead the strategy, but also stay close to all the developments that are happening within your company, within other companies.

这本身就已经超出了全职工作的量了。

That feels like more than a full time job in its own right.

是啊，这个行业真的快得离谱，对吧？

Yeah, I mean, the industry is just moving insanely fast, right?

我从来没看过这样的情况。

I've never seen anything like it.

这让人兴奋，也很有趣，但确实很难跟上节奏。

It's exhilarating and it's fun and it's also, it's a lot to keep up with.

但我觉得，这个时代特别适合那些主动性强的人，因为你现在能实现任何想到的想法，如果你有个好点子却不行动，那就真没借口了——你可以一边做原来的事，一边让Codex帮你琢磨这个点子。

But I think that I I just think this moment kinda selects for people who are high agency that because you can now create anything that you can think of, and you have no excuse if you've got an interesting idea not to, like, get, you know, codex thinking about it while you do something else, whatever you were originally gonna do in the morning.

照常做你的事，但让Codex同时帮你推进这个想法。

Keep doing that, but have codex working on your idea in parallel.

有时候你早上冒出个主意，等一天结束时，它已经变成实际成果了，而你原本计划要做的事也照样完成了。

And sometimes you'll wake up in the morning, have an idea, and have a thing implemented by the time you're done with the day in addition to doing what you thought you were gonna do during the day.

所以，现在主动性强、好奇心重、学习快的人，这些能力比以往任何时候都更宝贵。

So, like, people that are high agency, people that are really curious, people that learn quickly, those skills are more valuable than ever in this moment.

不管未来怎样，我觉得这些能力都会帮我们渡过难关。

And, you know, kinda whatever the future holds, I think those skills are going to see us through.

没错。

Yep.

我听过你谈过实验设计，还有实验本身和验证的过程，真的让我特别着迷。

I've heard you talk about a vision of both the experiment design, but also then the experimentation itself and the validation that's very, you know, was very captivating to me.

你能不能再详细分享一下这方面的想法？

Maybe you could share it it here a little bit more.

OpenAI 科学小组才成立几个月，但说到底，OpenAI 一直以来都非常重视科学。

This OpenAI for Science group is only a few months old, although in some sense, of course, OpenAI has always cared deeply about science.

虽然我们团队不算大，但背后有整个 OpenAI 研究团队的支持，因为每一位研究员都关心科学，而科学数据长期以来一直是提升我们模型的重要途径。

So I I feel like even though we're a smallish team, we kinda have the might of all of OpenAI research at our back because every researcher at OpenAI cares about science, and scientific data has been one of the ways that we, you know, have have improved our models for a long time.

所以我们开始聚焦在数学、物理和理论计算机科学上，因为这些领域完全可以在计算机里模拟完成。

But so we started thinking about math, physics, theoretical computer science because you can do everything in Silico.

你可以构建闭环系统进行优化，而我们的一部分目标，就是教会模型去解答极其复杂的科学问题——不是只思考十分钟，或者像你问 GPT-5 Pro 一个难题时它能专注一小时那样，而是要让模型能持续专注一整天、两天、一周，甚至两个月，去攻克更难的问题。

You have, like, closed loop systems that you can optimize and, you know, is part of this is teaching the models to answer really hard scientific problems, teaching them to think not for, like, ten minutes or maybe the hour that you can get GPT five pro to think if you ask it a really hard question, but teaching models to stay on track for a day, two days, a week, two months at a time to answer even harder problems.

就像你和我一样，如果给我一个二十分钟内解决不了的问题，但给我两小时，我可能就能搞定。

Because just like you and me, like, if you you could give me problems that I couldn't solve in twenty minutes, but I could given, you know, two hours.

模型也是一样的。

Same is true of the models.

他们思考的时间越长，能解决的问题就越惊人。

The more time they think, the more the more impressive problems they can solve.

所以我们先从数学和物理这类问题入手。

So you start with things like math and physics.

但更重要的是，科学加速带来的真正影响，最终会体现在我们日常生活中能切身感受到的变化上。

But then, you know, some of the the biggest ways, the most important ways that that accelerating science is gonna feed back into all of our lives in a positive way is through stuff that we can feel in real life.

比如，我们的亲人能战胜癌症，是因为医学取得了突破。

It's like, you know, one of our relatives survives cancer because we've made advances in medicine.

我们有了新的设备和材料，也是因为材料科学有了进展，而这些都需要实验室支持。

You have new devices and materials because we've been able to make advances in material science, and those things require labs.

这些事光靠计算机模拟是做不到的。

You can't do those just in silico.

虽然我觉得模拟的重要性会大幅提升，因为我们可以投入海量算力来处理这些问题，但最终还是得靠实验验证。

Although I think the the importance of simulation is going to go up pretty meaningfully because you will be able to apply huge amounts of compute to these problems, but then you're still gonna need experimental validation.

你必须在现实世界里亲自尝试。

You're gonna need to try things in the real world.

我觉得，要等到哪个模型能从夸克开始，一路推导到细胞模型，再到整个人体生物学，还得等上一阵子。

I think it's gonna be a while before we have a model that can, you know, first principles go from, like, a quark all the way through a model of a cell all the way through human biology.

实验真的很重要。

Like, experiment matters.

所以你会开始想，怎么才能大规模地做这件事。

So, you know, you start to think about how you do that at scale.

和现有的实验室合作空间很大，我们肯定会这么做。

There's lots of opportunity to partner with existing labs, and we will.

但我认为，未来的科学一定会离不开机器人实验室和强化学习循环——模型先思考，可能跑个模拟，再进一步思考，优化实验参数，然后把方案发给一堆机器人实验室，这些实验室能横向扩展，在真实世界里做实验，结果再反馈给模型。

But I think there are really interesting you know, the the science of the future will definitely involve robotic labs and reinforcement learning loops that go through the real world where the model is thinking, maybe running a simulation, thinking some more, refining the experiment that it can run using the best possible parameters, and then sending that to a bunch of robotic labs, which, by the way, you can scale horizontally, having the experiments run-in real life, the results come back to the model.

模型思考，再跑更多模拟，继续思考，就这样形成多个循环。

The model thinks, runs more simulations, thinks, and you have this you have, like, multiple loops.

模型和模拟之间有紧密的反馈循环。

You have tight loops with the model thinking and the simulation.

还有更长的循环，会真正走进现实世界。

You have longer loops, you know, that go through the real world.

未来的很多科学研究都会这样进行，这本身就能大大加快科研进程。

That's how a lot of science is gonna be done in the future, and, you know, that is its own form of acceleration.

你可以拥有可横向扩展的机器人实验室，全天候不间断运行。

You think you have you have robotic labs that you can scale horizontally that can run twenty four hours a day.

它们不需要像研究生那样，一边做移液一边还得休息、睡觉。

You know, they're not grad students pipetting things that need to take breaks and sleep and thing.

这样一来，研究生就能去做些更能发挥人类优势的事情，而不是整天做移液。

So and and then the grad students can do things that are much more, you know, you know, leveraging of what makes us human than pipetting things.

没错。

Right.

我对这个方向非常乐观，也相信我们社会有能力真正显著地加快科学进步的速度。

So I'm I'm quite optimistic about where this goes and about our ability as a society to accelerate the pace of science very meaningfully.

你觉得我们离那一天还有多远？

How far away do you think we are from that?

关键的技术有哪些？

What are the key, like, technologies?

是不是说，要完全实现这一点，关键在于机器人技术的发展？

Is it, you know, robotics that need what needs to happen for us to unlock that fully?

这部分确实涉及机器人技术。

I mean, some of that piece is robotics.

其实已经在发生了。

It's already happening.

对吧？

Right?

我刚才举了模型解决开放性数学问题的例子，实际上已经有不少机器人实验室在运行了。

There I was just giving examples of the model solving open math problems, and there are certainly robotic labs out there already.

目前还处于早期采用阶段，但以我们现在的推进速度，我觉得不会太久。

It's all kind of in the early adopter phase, but the pace that we're all moving, I don't think it's long.

对。

Mhmm.

而且这对很多领域来说都是明显正确的方向，这种想法可不只是我们才有。

And it's so clearly the right thing for many fields that this is not you know, we are not the only ones to have this idea.

有很多人都有这个想法。

There are a lot of people that have this idea.

不少有趣的初创公司正在朝这个方向开发产品。

There are a lot of interesting startups building things along these lines.

世界变化太快了，不会等太久的。

And, you know, the world is just moving so fast, it can't be long.

没错。

Right.

对。

Right.

你推出过的产品，用户数达到了几亿甚至几十亿。

You've shipped products used by hundreds of millions or billions of people.

你曾经对哪个产品决策特别担心，结果却是对的？

What was a product decision you were nervous about that ended up being right?

在这么大的规模下做产品决策，有趣的是，我举的任何一个例子，台下肯定都有人会说：不对，不对，完全不是这样。

You know, the fun thing with product decisions at that scale is any example I give, I bet there are people in the audience who are like, no, no, no, no.

你这个判断错了。

You got that one wrong.

很少有决定是明显正确的。

Very few are unambiguously right.

其中一个可能是推特动态的排序，当年这可是争议极大。

Probably one was like ranking the Twitter feed, which was extremely controversial back in the day.

以前推特完全是实时的，你首页看到的永远是刚发的一条推文。

Twitter used to be completely real time, and the the most, you know, the thing you saw at the top of your Twitter feed was the thing that was tweeted one second ago.

紧跟着的是四秒前发的那条。

The the next one was the one that was tweeted four seconds ago.

如果你的配偶或者最好的朋友一小时前发了条消息，那也只能算了。

And if your, you know, spouse or your best friend happened to tweet an hour ago, like, yeah, too bad.

你根本看不到。

You were never gonna see it.

那条消息会被彻底埋掉。

It's gonna get totally buried.

但当时有很多人说，这正是推特的魔力所在。

But there were a lot of people that said this is the magic of Twitter.

你怎么可能这么做呢？

How could you possibly do that?

你这不就变成脸书了吗？

You know, you're becoming Facebook now.

那时候这事儿争议特别大。

So that was very controversial at the time.

但从某种角度来说，你难道不更在意某些人发的内容，而不是别人的吗？

Although it seemed, in some sense, like, how could you not want you know, you do care about different people's stuff more than other people's stuff.

如果我们能做好排名，平衡好时效性和其他因素，为什么不呢？

How could you not want ranking if we could do it well and if we could bring the right balance of recency and everything else?

这是一例，我觉得脸书当初推出信息流时也遇到过同样的情况。

So that was one, and I I think Facebook saw the same thing when they originally, you know, put out the news feed.

一开始很多人特别不满，但数据却显示一个非常积极的结果，增长幅度高达两位数。

You have a bunch of people that are super upset, but then the metrics tell you an incredibly positive story, like double digit positive kind of thing.

然后你就可以不断优化，持续取得进展。

And and so, you know, and then you just you you you you can just keep making that that better and keep getting wins there.

推出第一个推理模型的预览版时，我们一直在思考：对于一个不会立刻给出答案的模型，该设计怎样的交互方式才合适。

It was interesting trying to figure out how exactly we would when we rolled out o one preview, the first reasoning model, what was the right kinda UX paradigm for a, for a model that would not give you an immediate answer.

以往所有的聊天模型都是你一提问，马上就能得到回复。

Like all the other previous chat models, you type in an answer or type in a question, you basically get an answer right away.

网上很少有需要你这样干等的体验。

There aren't a lot of experiences online where you have to wait like that.

meanwhile，模型在后台进行思考，但我们不想把整个过程完全暴露出来，因为那样别人很容易模仿我们的模型——出于一些地缘政治的考虑，我们不希望这种情况发生。

So and and the the model is doing this interesting thing with its chain of thought in the meantime, which we didn't wanna expose completely because we didn't want to, because you can distill that and, you know, basically copy our model, which, you know, for for a bunch of geopolitical reasons, we didn't want to have happen.

但又得展示一部分内容。

But you wanna show some.

所以我们很纠结：用户是会离开，等模型完成后再回来，还是会一直盯着看？

So the it was very interesting trying to figure out how you was the model that people would go away and and just, you know, pop back in whenever it was done, or were they gonna watch?

如果他们选择盯着看，那我们该展示什么？怎么在避免泄露核心逻辑和保持趣味性之间找到平衡？

And if they were gonna watch, what would we show that and how do we balance not not giving too much information that would, you know, lead to model distillation, but would be interesting.

这真是个有趣的实验。

So that was an interesting experiment.

用模型做东西时，有时真让人意想不到。

And it's surprising sometimes building stuff with models.

要处理这种场景的用户体验，一个合理的参考是：我希望人在类似情况下怎么表现？

A reasonable analogy for how should I, you know, handle the UX of this current situation is how would I want a human to behave, you know, in a similar situation?

如果你问我的问题我能立刻回答，就像GPT-4那样，我就会直接告诉你。

And, like, if you ask me a question that I can immediately answer, you know, GPT four style, then I will just immediately answer.

对吧？

Right?

但如果你问的问题我得想想，我不会一上来就把整个思考过程全倒出来。

If you ask me a question that I need to think about, I don't immediately start babbling my entire chain of thought.

对吧？

Right?

我不会把脑子里冒出来的所有东西一股脑全说出来。

I don't just, like, spew out whatever goes through my mind.

我也不会完全沉默，啥也不说，等一分半钟才回你答案，你知道的，我可能会先说，好吧。

I also don't, like, completely turn around and go mute and, like, do nothing until I come back to you with an answer a minute and a half later, you know, I might I might say, like, okay.

这是个挺有意思的问题。

That's an interesting question.

让我想想。

Let me think.

然后在思考的时候，会时不时给你个简要提示，比如，可能是这样。

And then you kinda give little, like, cliff notes as you think, well, it could be this.

不对。

No.

所以模型最终也就是这么做的。

And so that ultimately is kinda what the model does.

对吧？

Right?

它会在思考过程中，时不时给你更新一下思路，这本身就很有趣，然后才给出最终答案。

It it gives you kinda, like, periodic updates of what it's thinking about as it's thinking, which is interesting in and of itself, and then it comes back with the answer.

所以我们试着在这方面稍微借鉴了一下。

So we tried to model it a little bit off that.

在产品开发中，一个很大的争议就是数据和品味之间的权衡。

One of the big debates, like in terms of product development is data versus taste.

你有没有过这样的经历：你直觉觉得某个方向是对的，决定按这个方向走，但需要一点时间让数据跟上？

What's a time when you said, Hey, I've got a hunch, this is the taste, this is where we're going, we need a little bit more time for the data to catch up to it.

如果你一味地依赖数据，那你就失去了对方向的掌控，我觉得这不是你该待的地方。

If you just blindly follow the data, then it will take you, like then you're not in control of where it takes you and that's, I don't think that's where you wanna be.

用户反馈中的个案故事，其实也有很大的价值。

There's always a huge amount of value in anecdotes too when you get user feedback.

根据我的经验，通常当你发现数据指向一个方向，而用户反馈却指向另一个方向时，真正的情况是你面对的是两种截然不同的群体，而数据只是把它们的答案平均化了。

Like the, even though it it's usually the case in my experience that if you're getting the data tells you one thing and you're getting a bunch of user feedback that takes you in a different direction, then what's actually happening is you have some bimodal thing and what's your the data is, like, giving you the average of those answers.

实际上，你正在处理的是两种完全不同的情况。

And it's actually the case that you have two very different things going on.

对。

Mhmm.

而且你要

And you

换个方式切割数据，深入挖掘，别轻易忽略那些用户反馈。

need to cut your data differently and dig in because you should not dismiss the anecdotes.

那些反馈通常都很有价值。

The anecdotes are almost always valuable.

用数据的关键在于理解它，而不是简单地看到数字上涨就以为该行动了。

I think the trick with data is to understand it, not just be like, oh, you know, number go up.

我们该把这个功能上线。

Like, we should implement this thing.

但为什么数字会上涨呢？

But why does the number go up?

是不是因为新鲜感，这种事经常发生。

Is it because we're like, because of novelty, which happens a lot.

对吧？

Right?

你刚上线了个新功能，用户一看，哦，挺有意思的。

Some new thing that you shipped, people are like, oh, that's interesting.

那这是什么？

What's that?

于是他们点了一次，数据看着不错，但其实再也没回来。

And so they click on it once and your numbers look good, but they don't actually come back.

还是说他们根本就搞不懂？

Or is it because they are confused?

这种情况也挺常见的。

That also happens a fair amount.

要么是你产品没做对，要么你搞增长黑客，有些混乱状态虽然让数据好看，但其实是负面的。

Either, you know, you didn't get the product right or if you're, you know, growth hacking, there are sort of negative states of confusion, but they do make the numbers look good.

还是说用户真的喜欢这个新功能，留下来了，那你就该加大投入？

Or is it like actually people are retaining on this new thing and they like it and then you should lean into it?

所以你得真正去解读数据，弄清楚背后的原因，这样做决定才会清楚得多。

So you really wanna, like, interpret the data, try and understand what it means underneath, and then the decision is usually much more clear.

所以在座的很多人都在基于人工智能进行开发。

So a lot of people in this room are building on top of AI.

你们有没有在做什么特别的项目？

What's something that you Is

在座有没有谁没在做这个？有的，确实有。

there anyone in this room not building Yeah, a a

少数人，但即便这样，他们肯定也在使用人工智能。

few, a few, but they're still building with AI for sure.

但在基于 AI 进行开发时，你看到初创公司和创始人常犯的错误是什么？

But when you're building on top of AI, what's a mistake do you see startups and founders make?

我们内部会做的一件事，也是我看到很多人没在做的一件事是：如今很多任务，如果你使用模型集成（ensemble of models），效果通常会最好。

One thing that we do internally that people sometimes that I see people not doing is a lot of things today at least turn out best when you use like, an ensemble of models.

如果你有一个很难回答的问题，比如客户服务这类涉及多种复杂情况的事务，

If you have a hard question that you're trying to answer, you know, maybe it's customer service or something where there's a bunch of different things going on.

用户写信或提问的动机各不相同，你需要用不同的方式来处理他们的咨询。

People have different motives for for writing in, and you need to handle their queries in different ways.

你需要采取一系列操作。

There's a bunch of actions that you need to take.

模型现在已经相当强大了，有时候真的能一次性搞定那些复杂的流程。

The models are getting pretty good, and now they really can sometimes just completely one shot a tough flow like that.

但如果你把几个模型结合起来用，成功的几率会更高——比如先用一个模型来统筹规划，理清该怎么做才能回答这个问题。

But you can you can make your odds even higher if you use models sort of together, or you may have a an initial model that's orchestrating and is, like, putting a plan together and understanding what, you know, what you should do to answer the question.

然后再调用其他不同的模型。

And then you have different models.

有些模型可能成本更低，但专门训练来做某一件事，特别擅长。

Maybe some of them are cheaper models that are trained to do one thing really well.

统筹模型负责调用这些其他模型，我看到的人还太少。

And the orchestration model is, you know, calling the other models and things like I don't see people doing that enough.

作为整个行业，我们在这方面正变得越来越成熟。

I I think we're getting more sophisticated about it as as sort of an industry.

但与此同时，随着模型能力越来越强，它们对这种复杂协调的依赖也在逐渐减少。

And at the same time as we're getting more sophisticated about it, the models are getting better, and so they need this kind of thing less and less.

但事实上，我们在很多后台场景里还是会用模型组合，能用小模型的地方就用小模型，非得用大模型的时候才上大模型，让它们协同工作，而不是指望靠一个超级复杂的提示词，碰运气让大模型给出正确答案。

But that still is an area where, like, behind the scenes, we use ensembles of models in lots of places, trying to use, like, small models where you can, bigger models where you need to, and then have them all work together versus just like, oh, let me prompt engineer one giant, you know, prompt and hope the answer is right.

有没有哪家公司或产品，在OpenAI的基础上做得特别出色？

Where's a a company or a product that's done a really great job of building on on top of OpenAI?

天啊。

Oh, man.

太多了。

There are so many.

有没有哪一家让你觉得特别意外的？

Something that surprised you maybe.

嗯，确实有让我意外的。

Well, I mean, something that surprised me.

哪一家？

What?

对了，叫什么名字来着？

Like, what's the new name?

OpenClaw？

OpenClaw?

对。

Yeah.

OpenClaw 是基于 Codex 做出来的。

OpenClaw built on Codex.

这算是最近最让人眼前一亮的项目之一。

Like, that's one of the most interesting things that has come out recently.

更因为它预示了未来的发展方向。

More because it's like a sign of what's to come.

那个家伙居然只用了三天就搞定了，简直太厉害了。

Both like, a, the dude was able to put it together in the span of, like, three days, right, which is it's just awesome.

现在好多事情，以前得花几个月甚至根本做不到，现在几天就能完成。

Like, so many things are now possible in the span of days that would have been months and months of work or just completely impossible before.

但更重要的是，它展现了 AI 之间协同工作的全新可能性。

But also because it points at, like, this interesting emergent world of of the AIs all working together.

有人去过 maltbook.com 吗？

Have people spent any time on maltbook.com?

当然有。

Oh, yeah.

现在有好多 OpenAI 的 AI 代理，它们能访问某人的整个电脑，想干什么都行。

So the it's you know, you have all these AI OpenAI agents that, you know, basically have access to somebody's full computer, and they can do all sorts of things.

你还可以通过聊天软件之类的给它们发指令。

And you can command them through, you know, messaging apps and stuff like that.

现在出现了一个专为它们设计的社交平台，叫 Mold Book，它们在上面互相互动，聊主人的事，讲各种故事。

And now there is a social product for them called Mold Book where they go and interact with each other and talk about their humans and tell stories.

这真的太有意思了。

And it's just fascinating.

说真的，这一切都挺怪的。

I mean, it's it's all weird.

但不太可能成为下一个爆款创业项目。

It's not, like, most likely the next big startup or anything.

这真是未来趋势的一个有趣预兆。

It's just fascinating as a sign of what's to come.

我喜欢这种事，因为它让你隐约看到一点未来的样子。

I love stuff like this because, like, it just gives you a little bit of peek into the future.

对。

Yeah.

所以问题来了，这些行为是自然涌现的，还是只是新鲜玩意儿？

And so the question is how much is emergent behavior versus how much is just novelty?

对吧？

Right?

是的。

Yeah.

这里面确实有很多新奇的东西。

It is a lot of novelty.

很多人类在背后操控，不断给他们的OpenAI代理发指令，但同时也出现了一些特别有趣的现象。

There's a lot of humans trolling through, you know, prompting their OpenAI agents, But there's also just really funny stuff.

如果这个产品要求你

If the product requires you

出去买一台Mac mini，而你真的去买了，那这就叫产品契合度到位了。

to go out and buy a Mac mini and you go out and buy it, then you've got product right fit.

对吧？

Right?

没错。

So Right.

至少在像我这样的早期技术爱好者当中是这样。

At least among early adopter nerds like me.

看看它会发展到什么地步。

See where it like.

好吧。

Alright.

所以你有一个OpenAI的实例。

So you have OpenAI instance.

你分享了哪些个人资料？

What are you sharing of your personal information?

我挺小心的。

I am being careful.

是的。

Yeah.

你完全遵守了所有规定。

You followed all the guidelines.

我在小心处理。

Am being careful.

不过呢，你知道的，

There's but, you know

这里有个矛盾：怎样才能获得完整体验，又不用把整个生活都交出去？

tension there is like, how do you have the full experience without sharing your whole life?

对吧？

Right?