Yuandong Tian on Recursive Self-Improvement — 文稿

M10:13

嘿，大家好，欢迎回到 Information Bottleneck 播客。

Hey everyone, and welcome back to the Information Bottleneck podcast.

今天我们有位特别嘉宾，Yongjun Tian。

And today we have a special guest, Yongjun Tian.

嘿 Yongjun，谢谢你加入我们。

Hey Yongjun, thank you for joining us.

M20:24

嘿 Ravid。

Hey Ravid.

嗯，嘿 Alan。

Yeah, hey Alan.

谢谢你们的邀请。

Thanks for inviting me.

M10:26

所以 Yongjun，你在 Meta 的 FAIR 待了很长时间，最近你又是 Recursive Superintelligence 的联合创始人，这家公司专注于自我递归和自动化研究。

So Yongjun, so you spend a lot of time at Meta, at FAIR, and recently you are the co-founder of Recursive Superintelligence that focused on self-recursive and automated research.

这些我们待会儿会聊到，但让我们先从最开始说起吧。

So we will get there, but let's start from the beginning.

比如，你是怎么开始进入深度学习和机器学习的？

Like, let's start like how you actually started in deep learning and machine learning.

是怎么开始的？

How does it start?

M20:59

嗯，谢谢你的问题。

Yeah, thanks for this question.

其实我是 2008 年来到美国的，对吧？

So actually, I came to the United States in 2008, right?

那时候是机器人研究所的博士生，对吧？

So, and as a PhD student in Robotics Institute, right?

所以 2008 年的时候，我当时真的非常痴迷机器学习。

So in 2008, and that time I was like really into machine learning.

那个时候，我觉得人们其实并不认为机器学习是一条可行的路。

So at that time, I think people actually don't think machine learning is kind of a valid path.

大家真的会反驳这个概念，说，哦，机器学习没那么好用，因为它会过拟合，根本不懂到底发生了什么。

People really push back on this concept saying that, oh yeah, machine learning doesn't work that well because it's overfitting, does not understand what's going on.

但我还是在博士期间做了一些优化和机器学习的工作。

But I managed to do some of the optimization and machine learning work.

基本就是在计算机视觉领域里，研究一些复杂的非线性非凸优化问题。

Back in the PhD.

而且我成功证明了一些有趣的定理，表明即使问题是非凸的，你仍然可以通过增加数据，在某种程度上证明它的最优性。

And basically on some complicated nonlinear nonconvex optimization problems in the computer vision domains and managed to prove some of the interesting theorems that show even if the problem is nonconvex, you can still prove its optimality in some extent by adding more data.

所以这是数据驱动方法和优化的一种很有意思的结合。

So it's an interesting combination of data-driven approach plus optimization.

到现在我仍然为这类工作感到非常自豪。

So I'm still really proud of this kind of work.

然后我快要毕业的时候，2013 年，或者说 2012 年，我看到了一个大新闻，对吧？

And when I was about to graduate in 2013, 2012, I mean, I see this big news, right?

AlexNet 出现了，大概是在 2012 年的时候。

So AlexNet coming, let's say, dated back to 2012.

我当时真的给 Alex 发了封邮件，要他的代码库，然后稍微玩了一下。

And I actually sent an email to Alex asking for his codebase back at that time and played with that a little bit.

当然啦，那时候我完全不知道 CUDA 编程是什么。

I mean, of course, I have no idea what CUDA programming is at the time.

然后我就觉得，哦，这真的很棒，而且它跟数据的分层处理有关。

And then I decided that, oh yeah, that's something very nice and something that is related to the hierarchical kind of processing of data.

用神经网络来处理数据，它确实会给你一些有趣的中间表示，也会给你一些不错的结果。

Into neural networks so that it actually gives you some interesting intermediate representation and gives you some nice results.

所以当时我在想，哦，也许这就是我在追寻的东西，只是换了一种形式。

So at that time I was thinking, oh yeah, maybe that one is something that I'm chasing after in a different form.

因为在我读博的时候，我做的论文课题之一，就是用一种分层的方式来优化计算机视觉里的一些非凸优化问题。

Because when I was in PhD, one of my theses, the thesis I'm working on is a kind of hierarchical way of optimizing some non-convex optimization problems in computer vision.

基本上，我就是试图去除水导致的图像畸变。

So basically I try to remove the distortion of the image caused by the water.

所以这种大的方向。

So this kind of high directions.

所以我尝试用分层的方法来做这件事，并且实际上证明我们在理论上可以做得更好。

So I'm trying to use a hierarchical approach to do that and actually prove that this is something we can do better theoretically.

从经验上看，效果也真的很好。

Empirically, it also works really well.

所以这就建立了这些联系，而那也是我对深度学习最初信念的来源。

So this makes these connections, and that's where I have this initial belief in deep learning.

但那时候，我得说有很多有趣的事情在发生，因为人们并不相信那个。

But at that time, I would say there were a lot of interesting things happening because people didn't believe in that.

人们会说，‘哦，这不过是过拟合而已。’

People would say, 'Oh, this is just overfitting.'

所以大家来到办公室，我们在CMU的Smith Hall里进行了很多激烈的辩论，讨论AlexNet的结果到底是胡扯还是真正的突破。

So everyone just came to the office and we actually had lots of heated debates in the Smith Hall at CMU, talking about whether these AlexNet results are bullshit or a real breakthrough.

所以当时这件事刚发生，大多数人就说那只是胡扯，因为它根本说不通。

So at that time it just happened and most people just said it was just bullshit because it just didn't make sense.

但我觉得当时我其实在想，也许我应该转向深度学习。

But I think at that time actually I was thinking maybe I should switch to deep learning.

所以我面试了几个地方，包括一些真的给我发了offer的地方，比如NEC Labs。

So I interviewed at a few places, including some places that actually gave me an offer, like NEC Labs.

我想做深度学习，但最后去了谷歌自动驾驶汽车团队。

I wanted to do deep learning, but ended up with the Google self-driving car team.

因为首先，那是一个非常有名的地方。

Because first of all, it's a very prestigious place.

其次，我有好朋友在那里，他们推荐了我，我觉得在某种程度上不应该破坏这种联系。

And second, I have good friends there who referred me, and I don't think I should break that tie, to some extent.

所以我去了那里，而且我当时也确实想做深度学习。

So I went there, and also I wanted to do deep learning at that time.

但不幸的是，在那一年零三个月的时间里，我没有机会做那个，因为那更像是一个创业公司，你真的得让Zeppelin汽车跑起来，对吧？

But unfortunately, in this one year and three months period, I didn't get a chance to do that because it's more like a startup and you really need to make the Zeppelin car work, right?

所以最好是用一些非常成熟的方法，而科学探索可能不是我刚加入时该做的事。

So it's better to use some very mature approach, and scientific exploration probably is not the right thing to do when I joined there.

所以我决定跳槽，对吧？

So I came to decide to jump ship, right?

所以我搬到了FAIR。

So I moved to FAIR.

那其实是我当时最好的选择之一，我会这么说。

That was actually one of my best choices at that time, I would say.

嗯。

Yeah.

所以我到了FAIR之后就开始做研究，真的非常有趣也非常好。

So I moved to FAIR and then I started to do research, which was super interesting and nice.

我想我是从这个用深度学习做计算机围棋的项目开始的，对吧？

And I think I started with this deep learning for Computer Go project, right?

那个项目其实在某种程度上相当成功，因为我们在AlphaGo之前就取得了一些非常好的成果。

So that was actually pretty successful to some extent because we actually got some very nice results before AlphaGo.

2015年，我们做出了第一个叫Dark Forest的机器人，它能用卷积神经网络击败网上的顶尖业余棋手。

In 2015, we had this first bot called Dark Forest that could beat top amateur players on the internet with the convolutional neural network.

所以那其实挺有意思的。

So that's actually interesting.

当然，我是说，AlphaGo就要来了——它是在2016年春天出现的，对吧？

Of course, I mean, AlphaGo was coming— it came in the spring of 2016, right?

而且它击败了职业棋手。

So, and that was beating the professional players.

当然，那非常非常好。

Of course, that's very, very good.

而那时候，我想我的团队只有两个人，对吧？

And at that time, I think my team only had two people, right?

我自己加一个实习生。

Myself plus an intern.

M15:48

但是你看AlphaGo的时候，你有没有想过，怎么说呢，就是我怎么没想到这个？

But did you think, like, when you saw AlphaGo, and did you think about, like, I don't know, like, how I didn't think about this?

或者就是，哦，它们就是AI而已。

Or like, oh, they are just AI.

M25:59

首先，我觉得AlphaGo确实有很多创新，对吧？

Think first of all, I mean, AlphaGo definitely have many innovations, right?

其中一个他们有而我们没有的东西，就是value network，对吧？

So one of the things that they have but we don't have is the value network, right?

当时我们只有policy network。

So at that time, we only have a policy network.

而他们有value network。

So they have a value network.

所以这让下棋的水平强了很多。

So this actually makes the play much stronger.

这是优势之一。

That's one of the advantages.

第二个区别是，我基本上没有任何reinforcement learning的背景。

And the second difference is that I was basically not having any background on reinforcement learning.

基本就是边做边学。

I basically do while I'm learning.

那是一段非常有趣的经历。

It's a very interesting experience.

M16:28

就是自己边做边学reinforcement learning？

Like reinforcement learning yourself?

对。

Yeah.

M26:30

对。

Yeah.

所以那是一段很有意思的经历。

So it's a very interesting experience.

所以这就是为什么我现在——那之后我基本上转向了reinforcement learning这边，之后做了很多reinforcement方面的工作。

So that's why I'm now, I mean, after that, I basically switched to more like reinforcement learning side, doing many reinforcement work after that.

对。

Right.

M16:41

但那时候，我还想再花一分钟说说这个。

But at that time, like, well, so I wanted to stay one more minute on it.

所以当时，这件事让你惊讶吗？

So at that time, was it surprising?

因为reinforcement learning在很长一段时间里都是那种——对，它确实是个很酷的想法，对吧？

Because reinforcement learning for a long time was something like, yeah, it's kind of like, it's a really cool idea, right?

但它其实不太管用，我们只能解决一些玩具式的问题。

But it's not really work and we can solve some toyish problems.

但对你来说，这让你惊讶吗？

But was it surprising for you?

M27:08

我觉得我们之所以选这个课题，是因为我感觉到convolutional neural network确实能捕捉棋盘上的很多模式。

I would say the reason we chose this topic is because I sensed that the convolutional neural network can actually capture many patterns in the board.

而这些模式正是人类下棋、玩这个游戏时会利用的东西。

And this pattern is something that humans leverage to play the board, to play the game.

所以从某种程度上说，这并不那么令人惊讶。

So to some extent, it's not that surprising.

但真正让人惊讶的是，它居然能击败职业棋手。

But what makes things very surprising is it can actually beat professional players.

那才是让人惊讶的部分，对吧？

That is the surprising component, right?

所以如果你真的去看这个方法、这套方法论，它其实并没有那么疯狂。

So if you actually check the method, methodology, it's actually not that crazy.

所以它算是convolutional network成功的一个自然结果。

So it's kind of a natural consequence of the success of the convolutional network.

M17:41

我很好奇，你觉得那个年代有没有什么方法，在今天被忽视了？

I'm curious, were there any methods from this era that you think are being ignored today?

M27:48

对，有很多方法，对吧？

Yeah, there's many approaches, right?

所以在那个时候，我认为人们开发了很多关于强化学习的东西，对吧？

So at that time, I think people developed many things about reinforcement learning, right?

这些方法中很多在今天的的大模型时代可能还没有被充分利用。

Many of these approaches are probably not yet being leveraged in today's large model era.

所以例如，人们正在考虑使用来自未来的更密集的奖励，对吧？

So for example, people are thinking about using more dense rewards from the future, right?

这叫做 distributed RL，对吧？

So it's called distributed RL, right?

所以不仅仅是得到一个单一的数字，而是希望得到所有奖励的分布，然后利用它来调整你的模型。

So instead of having just a single number, you want to have a distribution of all the rewards and then use that to basically tune your models.

所以我认为当前的 RL 可能也在使用类似的方法，只是形式不同，对吧？

So that's something that I think the current RL, they maybe use it, but in a different form, right?

所以例如，他们使用语言而不是传播回来的标量奖励，对吧？

So for example, they use languages rather than scalar rewards propagated back, right?

所以还有很多不同的想法。

So there's also many different ideas.

并且也有人在使用 Q-learning，对吧？

And also people are using Q-learning, right?

他们使用不同的公式。

They're using different formulations.

但现在，大多数时候我们主要使用 policy gradient，对吧？

But right now, most of the time we are mostly using policy gradient, right?

所以这是非常不同的领域。

So it's very different areas.

所以是的，回到2016年，然后在2018年，我们实际上复现了 AlphaZero。

So yeah, so back to that, 2016, and then in 2018, we actually reproduced AlphaZero.

你可能知道 AlphaZero，这意味着没有人类玩家参与。

You probably know AlphaZero, which means that there's no human plays.

基本上从一开始就是自我对弈，对吧？

You basically do this self-play from the beginning, right?

我们确实复现了它，并且在某种程度上使其更高效，创建了一个名为 OpenGo 的机器人。

And we actually reproduced that and make it a little bit more efficient to some extent to create this bot called OpenGo.

那个机器人实际上轻松击败了专业玩家。

And that bot actually beat professional players easily.

我们与韩国围棋协会进行了一场正式比赛。

And we had an official match with the Korean Go Association.

我们以5比0击败了他们的4名选手，这意味着他们连一局都没赢，即使我们只用一个GPU（V100）来运行这个游戏。

We beat 4 of their players with 5-0, which means that they don't even win a single game, even if we only use a single GPU, which is V100, to play this game.

那，我认为那可能是第一次，人们使用单个GPU击败专业玩家。

That's, I think that's probably the first time that we use, people use single GPUs to beat professional players.

嗯，就是这样。

Yeah, so that.

M39:48

你当时感到惊讶吗？

Was— it surprising for you?

比如，你是否认为，嗯，当你参加这种比赛时，你是否觉得我们肯定会击败所有人，我们就是比他们强很多，还是说你有过怀疑？

Like, did you think that, like, I don't know, when you go to this, uh, like, competition, did you think that, like, yeah, we will beat everyone, like, we're just so much better, or like you had some doubt?

M210:01

是的，在我们正式开始这场正式比赛之前，我们实际上找了一些同样是职业玩家的朋友，嗯，都是半职业玩家，来评估这个机器人的实力。

Yeah, before we actually start this, uh, uh, official match, uh, we actually asked some of our friends who are also professional players, right, so all semi-professional players, to evaluate how strong it is.

然后他们，我的意思是，这个过程其实非常有趣。

And they, I mean, it's, the game is actually really interesting.

一开始还好。

At the beginning, things are fine.

之后，就没人知道发生了什么了。

After that, no one knows what's going on.

他们会说，哦是的，这看起来非常强，非常强。

They will say, oh yeah, this looks very strong, very strong.

但我不知道它到底强不强，因为我看不太懂。

But I don't know whether it's actually strong or not because I cannot see the game very well.

所以那时我会说，好吧，我们打个赌。

So at that time I would say, okay, let's take a bet.

让我们真的和Korean Go Association的专业棋手对弈，因为我们在那边有一些人脉。

Let's actually play with real professionals from the Korean Go Association because we have some connections there.

结果发现质量真的非常好。

And then it turned out that the quality is really, really good.

M110:41

那么你——我很好奇，你对gradient-free优化方法的看法是什么？

So do you— I'm curious, what's your opinion on gradient-free-based optimization methods?

我倾向于认为那是一组在历史上有点被忽视的技术，甚至到今天也是如此，尤其是。

I tend to think that that's a set of techniques that have been kind of ignored a little bit as well, historically, and, or even up to today, especially.

是的，

Yeah,

M211:02

我觉得这其实和我之前在FAIR的一个研究方向有关，对吧？

I think this actually connects with one of my research directions back in FAIR, right?

在reinforcement learning、游戏之类的方向之后。

So after the reinforcement learning, on games kind of directions.

我们试图思考如何让reinforcement learning在更实际的问题上更具通用性。

And we are trying to think about how to make reinforcement more general on more practical problems.

这实际上意味着我们也可能会探索一些gradient-free的方法。

So this actually involves that we might also explore a little bit about gradient-free approach.

因为reinforcement learning，你可以把它看作是一种基于gradient的方法。

Because reinforcement, you can think of this as one way to do gradient-based approach.

但当然信号非常稀疏，很难高效地获取信号。

But of course the signal is very sparse and it's very hard to get the signal in a very efficient way.

所以我们想着也许这个局限，也许我们应该超越它。

So we are thinking about maybe this limitation of it, maybe we should go beyond that.

对于gradient-free方法，我觉得它肯定提供了更多的灵活性。

So for gradient-free approach, I would say it definitely offers more flexibility.

而且随着当前更大的语言模型，它可能有更多潜力，因为更大的模型现在对系统、对要优化的目标函数有更好的高层次理解。

And with the current larger language models, it might have more potentials because larger models now know better about high-level understandings about the systems, about the objective function to be optimized.

所以这种高层次的理解实际上比局部gradient信息更有信息量。

So this high-level understanding is actually much more informative than the local gradient information.

所以。

So.

M312:09

在gradient-free方法中到底发生了什么？

What's actually happening in gradient-free approach?

M212:12

是的，我觉得这取决于你所说的gradient-free方法是什么。

Yeah, I think it depends on what gradient-free approach you're talking about.

所以，好的，如果你只是说。

So, okay, if you just talk.

M312:20

关于。

About.

M212:20

零阶，比如零阶，evolutionary algorithms，零阶，那么没有任何高层次信息，实际发生的是你在gradient不太有信息量的地方进行探索。

Zeroth order, for example, zeroth order, so evolutionary algorithms, zeroth order, then without any kind of high-level information, then what really happens is that you do exploration where the gradient is not super informative.

所以当函数非常不光滑，对吧？

So when the function is very chunky, right?

所以，好的。

So, okay.

所以你可能会有有趣的局部最优突然出现。

So you may have like interesting local optimums appear out of nowhere.

那里没有gradient指向那个方向。

There's no gradient going that way.

在这种情况下，gradient信息可能非常有误导性，对吧？

So in that case, the gradient information may be very misleading, right?

所以它会给你错误的方向，让你陷入local minima。

So it'll give you the wrong place to go and it will get you stuck in local minima.

但通常gradient-free方法会提出探索技术，对吧？

But usually the gradient-free approach will come up with exploration techniques, right?

所以这是我们实际讨论的内容，比如open-ended search，对吧？

So this is something that we actually talk about, like, for example, the open-ended search, right?

我们既有探索，也有利用。

We have exploration and we also have exploitation.

利用呢，就是说我们实际上是在按照某种贪婪的信息走。

So by exploitation, we are actually following some greedy information.

探索呢，就是说，哦对，也许局部搜索真的没那么好。

By exploration, we are saying that, oh yeah, maybe the local search is really not that great.

所以让我，让我们试试不一样的东西。

So let me, let's try something different.

对。

Right.

所以这种跳跃，是贪婪方法从来不会用的。

So this jump is something that the greedy approach never ever used.

这是巨大的优势。

That's something that's a huge advantage.

M113:29

对。

Yeah.

对。

Yeah.

所以呢，我早期学习神经网络怎么运作时，很多尝试都是从零开始写代码，然后发现我根本没希望从零开始做自动微分或者通用情况的反向传播。

And so I'm, a lot of my early forays into learning how neural nets worked was trying to code them up from scratch and then realize that I had no hope of doing autodifferentiation or doing the backward pass in the general case from scratch.

所以我就想，好吧，那就看看只基于前向传播的方法。

And so I was like, well, let's look at forward pass only based methods.

然后我开始发现你提到的所有这些事情，比如如果你不假设值或奖励的形状，那就能得到很好的探索。

And then I started discovering all these things you talked about where, like, if you don't assume the shape of the values or the rewards, then you get great exploration.

我其实非常在意创造力。

I actually care a lot about creativity.

所以我想问的是，你觉得这些无梯度的方法是不是让我们得到了，或者说，带来了更多样化的结果？

And so I guess what I would ask is, do you think that these gradient-free based methods are why we get or, you know, lead to more diverse outcomes.

另外，你信不信——我知道扩散模型里有梯度，但我认为当你观察它们的优化方式时，比如去噪器，我的理解是其中有很多无梯度的方法。

And furthermore, do you believe— I'm aware that diffusion models do have, you know, gradients in them, but I claim that when you look at how they're optimised, for example, denoisers, for example, my understanding is that there are plenty of gradient-free approaches for those.

所以呢，我认为，扩散模型和扩散语言模型，我听很多人说它们更多样化。

And, and, and so I claim that, you know, diffusion models and diffusion language models, I've heard many others argue that they're more diverse.

那么你觉得，如果这是真的，它和无梯度方法在多样性上更优之间有没有关联？

And so do you think that there's a connection here between that, if that's true, and the gradient-free method stuff being better at diversity?

M214:45

我明白了。

I see.

我觉得对于基于扩散的方法，首先，我不是这方面的专家，对吧？

I think for diffusion-based approach, first of all, I'm not an expert of that, right?

其次，我目前的理解是，扩散方法好是因为它用了很多不同的掩码来训练模型，对吧？

So second, I think my understanding right now is that diffusion-based approach is good because it actually uses many different masks to train the models, right?

所以你可以把自回归方法看作是一种特例，对吧？

So you can think of the autoregressive approach as kind of a special case, right?

基本上就是只以特定方式使用注意力掩码，让它变成因果的，然后你学习一些模式，对吧？

So basically you only use the attention mask in a specific way so that it becomes causal, then you learn some patterns, right?

但如果你用基于扩散的方法，通常掩码是随机生成的，对吧？

But if you use diffusion-based approach, usually the mask is randomly generated, right?

所以你生成掩码，然后试图填补空白，对吧？

So you generate mask and then try to fill in the gap, right?

所以这当然会带来一种优势，就是你在强迫自己预测中间的内容，即使这些信息在自回归的对应方法中已经提供了。

So of course it actually gives you a kind of advantage in the case that you are forcing yourself to predict something in the middle, even if that information is already being provided in the autoregressive counterpart.

所以我认为，因为这种打乱，你可能能学到数据中隐藏的更多关联。

So I think because of this kind of shuffling, then you might be able to learn more correlations that is hidden in the data.

所以这可能给你带来优势。

So this might give you advantage.

但当然，扩散模型本身还是通过梯度下降训练的。

But of course the diffusion model itself is still trained by gradient descent.

所以不一样。

So it's different.

M315:52

好。

Okay.

所以你就花时间在理解和开发更好的围棋棋手上，对吧？好像是在2018年那会儿，对吗？

So like you spend like time on like understanding and like developing like better like Go players and right, like we are like in the it was 2018, something like that, right?

M216:06

是的。

Yeah.

M316:06

那时候，基本上语言模型还挺没用的，对吧？

Like in that time, like basically like language models were like kind of like useless, right?

从某种意义上说，对吧？

In some sense, right?

嗯，我知道。

Yeah, I know.

但我觉得当时普遍的认知是，深度学习在视觉方面非常强，比如在视觉上，而我们还在等待。

But like, but like I think like the common like understanding was that like deep learning is super strong with like vision, for example, and we are still waiting.

等待它应用到文本和语言上，对吧？

Waiting for it to come to text and language, right?

M216:35

嗯哼。

Mm-hmm.

是的。

Yeah.

M316:37

好的。

Okay.

然后接下来，你说，好吧，围棋确实是个有趣的游戏，我们开发了一些强化学习的方法去尝试打败它。

So then next, like, okay, so like now you said, okay, like, yeah, Go is actually like an interesting game and like we developed like some RL methods to, to try to beat.

然后你们尝试去其他领域推广，把它泛化到不同的强化学习任务中吗？

And then what did you try like to, to go to other domains, like to generalize it to, to, yeah, to different RLs?

M216:55

是的，我觉得那之后确实有两个方向。

Yeah, I think there's definitely 2 directions after that.

一个是，在游戏上的强化学习在某种程度上已经被解决了，因为你总是可以设定奖励，让它运行，然后它会给你一些结果，这很不错。

So one is like, okay, reinforcement learning on games to some extent is being solved because you can always set a reward, let it run, and then it will give you something, which is nice.

我觉得在2018、19年左右，人们在这方面已经达成了共识。

So that's something that people have consensus, I think back in 2018, '19-ish.

我的意思是，当然有两条路。

So I mean, of course, there's 2 paths.

一条路是，我们做搜索优化怎么样？

One path is that, okay, how about we do search optimization?

强化学习是其中一部分，对吧？

Reinforcement learning is part of it, right?

然后应用于真正困难的现实世界问题，对吧？

So to real hard kind of like real-world problems, right?

而现实问题可能很难处理，对吧？

And the real problems may be tricky to do, right?

所以我们实际上花了很多时间在这上面。

So we actually spend a lot of time on that.

因为后来我们发现，对于现实世界问题，如果你把强化学习应用上去，真正重要的并不是算法。

So because then we find that for real-world problems, what happens is that if you apply reinforcement learning to these problems, what really matters is not the algorithm.

有意思。

Interesting.

重要的是动作空间和状态空间的设计。

What matters is the design of the action space and also state space.

这是我们一直在研究的一些东西，对吧？

That's something that we have been studying a little bit, right?

所以如果你看我的简历，看我的Google Scholar，你会看到一堆神经架构搜索的论文，对吧？

So that's why if you check my resume, check my Google Scholar, we see like a bunch of neural architecture search papers, right?

你仍然需要搜索这个巨大的空间，但这种搜索可能不遵循强化学习。

So you still want to search this big space, but this search may not follow reinforcement learning.

它可能遵循其他类型的方法，比如你刚才提到的无梯度方法，对吧？

It might follow other kinds of domains like gradient-free approaches you've just mentioned, right?

但真正的问题是，我们如何自动学习设计状态和动作空间的方式，使搜索变得更加高效？

But the real question is like, how can we automatically learn the way we design this state and action space so that the search becomes much more efficient?

所以在这个方向上，我们确实有很多相关工作，对吧？

So for that direction, we actually have a bunch of work on that, right?

但这里的想法是，我们能不能自动找到正确的动作空间，从而更有效地搜索整个路径、整个目标函数。

But the idea here is that can we automatically find the right action space to take in order to search the entire path, entire objective function much more effectively.

我给你举一个例子。

I will give you one example.

比如说，如果你考虑架构搜索，人类会说，好吧，我们都知道深度很重要。

So for example, if you think about architecture search, I mean, humans will say, okay, we all know that the depth really matters.

所以浅层网络相比深层神经网络，表现确实不太好。

So the shallow network is really not doing a good job compared to the deeper neural network.

我们知道这有点像是一个我们想要在搜索中强加的启发式方法。

We know this is kind of like a heuristic that we want to impose in the search.

但另一方面，AI可能不知道。

But on the other hand, the AI probably doesn't know.

如果AI从搜索这些小细节开始，那它永远给不了你有意义的东西。

If the AI starts with a search over these small details, then it will never give you something interesting.

假设AI从搜索卷积设置中神经网络第一层的核大小开始，那影响不大，因为无论你怎么搜索，除非你遇到像1这样的极端数字，否则性能可能都差不多。

So suppose the AI starts with a search over the kernel size of the first layer of the neural network in the convolutional setting, then it doesn't matter that much because whenever you do the search, I mean, unless you hit some extreme numbers like 1, it may give you a very comparable performance.

在这种情况下，你所有的搜索分支都差不多。

So in that case, all your search branches are comparable.

如果你在此基础上计算很多价值函数，它们得出的结论都是一样的。

And if you compute a lot of value functions on top of that, they're all telling you the same story.

那你就得真正深入每一个分支。

Then you really need to get deep into each of the branches.

这会导致搜索空间呈指数级爆炸。

This will lead to an exponentially explosive search space.

所以在这种情况下，你用什么算法都不重要。

So in that case, it doesn't matter what kind of algorithm you use.

你可以用强化学习，也可以用任何无梯度的方法。

You can use reinforcement learning, you can use any kind of gradient-free approach.

这不会给你带来什么有用的东西，因为你面对的是指数级的搜索空间。

It doesn't give you anything that useful because you are facing an exponential search space.

但如果你有一个想法，基本上是重新组合你的动作空间，然后你神奇地发现深度很重要。

But if you have an idea to basically shuffle your action space and you somehow magically find that the depth is important.

然后突然间，你确实会看到深度网络和浅层网络之间非常鲜明的对比，对吧？

Then all of a sudden you actually see a very sharp contrast between a deep network and a shallow network, right?

所以你不需要遍历所有路径。

So you don't need to traverse all the paths.

你只需要几次 rollout。

You just have a few rollouts.

然后你就能清楚地看到有很大的不同。

Then you clearly see there's a big difference.

然后你就会走深度路径，这样你的搜索就取得了进展。

Then you will go through the deep paths and then you have made progress on the search.

所以这个想法可以实现，而且实际上可以用一个数学公式来表达。

So this idea can be implemented and can actually be formulated by a mathematical formulation.

这是我们开发的一种方法，叫做潜在空间蒙特卡洛树搜索。

That's something that we developed called latent space Monte Carlo tree search.

对于这个想法，你要做的是找到什么样的动作空间能带来最具区分度的搜索子空间。

So for that idea, what happens is that you want to find what is the action space that will lead you to the most discriminative search subspaces.

你可以凭直觉做到这一点。

And you can do this intuitively.

一旦你找到了，突然间你会发现一个很好的动作空间，而且保证能取得良好的进展。

And then once you have that, all of a sudden you find some nice action space, and you will guarantee to have good progress on that.

M320:47

你觉得在这个时期，我们几周前和 Frank Krueger 聊过，他非常热衷于 AutoML，对吧？

Do you think like around this time, like there was, we talked like a few weeks ago with Frank Krueger that, like, was really into AutoML, right?

就是那种可以找到超参数的想法，对吧？

Like the idea that you can, like, find the hyperparameters, right?

而且是自动地。

And like automatically.

那你怎么看？

So what is your take?

你觉得，至少在那个时期，真正需要搜索的重要组件是什么？

Do you think, like, at least in that time, what are the important components that you actually need to search?

对你的结果影响最大的组件有哪些？

What are the components that had the most influence on your result?

M221:26

那时候，重要的组件仍然是在动作空间里进行手工设计。

At that time, the important component is still manual design in the action space.

比如说，如果你去看Google的AutoML工作，通常他们做的事情是，哦，我想设计这些非常漂亮的cell。

So if you check, for example, AutoML work from Google, usually what they do is that they say, oh, I want to design these very nice cells.

然后这个cell有六到七种不同的动作序列。

And the cell has 6 or 7 different sequences of actions.

然后你复制它，基本上就是把这些cell堆叠起来，就得到一个神经网络。

And then you replicate it, basically you stack all these cells together and you get a neural network.

所以堆叠的方式是固定的。

So stacking is kind of fixed.

然后关键就是设计这种能够考虑到、并且有潜在可训练组件的cell。

And then the key is to design this kind of cells that can consider, that has a potential component to train.

那你可能会说，为什么不是8个或9个呢？

So you might say, why not 8 or 9?

答案当然是，我猜这样会呈指数级地昂贵，所以就做不到了。

The answer, of course, is I guess this took exponentially expensive, then you cannot do that.

所以这些允许采取的小动作，基本上需要大量的人工设计，才能真正把它做出来。

So then all these small actions that are allowed to take, it basically deserves a lot of human design before you can actually make this work.

所以通常情况就是这样。

So usually that's what's happening.

M122:19

那说到动作空间，你认为要获得好的机器人和物理AI系统，路径是什么？

So related to action spaces, what do you think is the path to getting good robotics and physical AI systems?

而且我想特别强调一下，你确实有一篇论文是关于连续空间中的推理。

And I want to particularly highlight that you actually have a paper talking about reasoning in continuous spaces.

我在试着回忆它的确切名字。

I'm trying to get the exact name of it.

这是你的一个——

It's one of your—

M222:40

对，对。

Yeah, yeah.

就是那篇Coconut论文。

It's a Coconut paper.

M322:41

对，Coconut。

Yeah, Coconut.

M122:42

对，Coconut，没错。

Yeah, Coconut, exactly.

我想指出，Yann LeCun和其他人都在谈论JEPA，以及在潜在空间而不是token空间里推理，等等这些。

And I want to point out Yann LeCun and all the others talking about JEPA and reasoning in latent space, not token space, and all these.

所以我非常好奇你对这些的看法。

And so I'm very curious your thoughts on all of this.

M222:56

嗯，我觉得这篇论文——Coconut论文其实是我实习生做的。

Yeah, I think this paper is— the Coconut paper is actually the paper from my intern.

他一直来Meta，然后我们一起头脑风暴了一些天马行空的想法和思路。

He has been coming to Meta and then we've been brainstorming some random ideas and thoughts.

然后我认为我的一个论点是，我在思考的时候，从来不用语言来思考。

And then I think one of the arguments that I have is that when I'm doing the thinking, I never use a language to do the thinking.

我能清楚地感觉到，当我在思考时，我用的是大脑的不同部分，然后我得把想法转化成语言，才能很自如地和别人交流。

I can clearly feel that when I'm thinking, I'm using a different part of my brain and then I have to translate my thoughts into language so that I can communicate with others very freely.

对吧。

Right.

所以这就是我思考的方式。

So this is how I am doing that.

所以我感觉这基本上是基于我的直觉。

So I feel this is basically based on my intuition.

对吧？

Right?

所以，然后这就带出了这个想法：哦，那我们干脆把token整个去掉，对吧？

So, and then this comes with this idea that, oh yeah, how about we get rid of the token altogether, right?

所以我们只用latent space来做思考，因为显然latent space是一个我们应该采用的很有意思的representation。

So we just use like a latent space to do the thinking because obviously latent space seems to be an interesting representation that we should take.

然后在这个火花之后，我实际上开始做一些理论推导，那些推导确实给了你一些很好的直觉。

And then after this kind of spark, then I actually started with some of the theoretical derivations and that derivation actually gives you some good intuitions.

也许latent space的representation有点像是一种interposition，它是多种想法的superposition，然后用这个superposition来思考，你可能会得到指数级的优势，对吧？

Maybe the latent space representation is kind of interposition, it's kind of superposition of multiple ideas together and then use this superposition to do thinking, you might be able to have exponential advantage, right?

所以如果你遍历某个graph，对吧？

So if you traverse some graph, right?

所以就是类似这样的想法，对吧？

So that's some idea like that, right?

当然它可能不是指数级的，但至少它给你一些优势。

So of course it may not be exponential, but at least like it gives you some of the advantage of that.

然后呢，后来我们的实习生，对，就是Shibo，他来了，对吧？

So then the reason, then we have our intern, yeah, is coming and actually like Shibo, right?

他一直在努力实现那个想法，我们发现了一些初步的迹象，表明那是可行的。

He has been trying to execute that, and we find some initial signs of life doing that.

M124:39

当你提到你的直觉，就是你知道，你不会把大部分想法用语言表达出来，我要指出的是，我也不会在说话之前把所有事情都语言化。

So when you mentioned that your intuition about how, you know, you don't verbalize most of your thought, I'll point out that I also do not verbalize everything before I talk.

也许这就是为什么我话说得太多，事前想得不够的原因之类的。

And maybe that's a source of why I talk too much and don't think enough before, et cetera.

但我觉得人类有巨大的多样性，对吧？

But I claim that humans have huge diversity, right?

比如，我认识很多人，他们确实会把大部分，甚至所有想法都语言化。

Like, I know a lot of people who absolutely do verbalize most, if not everything.

他们称之为internal monologue。

They call it internal monologue.

对。

Yeah.

有些人有时会这样。

People who do it sometimes.

类似地，图像方面也是，对吧？

And similar with, with images, right?

有些人能很好地想象图像，还能旋转形状。

Some people visualize images really well and can rotate shapes.

其他人做不到。

Others can't.

有些做不到的人却是很棒的艺术家，对吧？

Some of those who can't are great artists, right?

所以我很好奇这一点，比如这种ontological diversity，以及你觉得这和什么相关。

And so I'm curious about this, like, ontological diversity and how you think this relates.

这是不是意味着，我猜，如果人类如此多元，我们应该假设AI也会有很多不同的路径来达到同样类型的结果？

Does it just mean, I guess, that for— if humans are so plural, we should assume that AI would also have many different paths to the same kind of outcome?

M225:30

对，我觉得完全是这样。

Yeah, I think that's totally true.

所以有可能会有一些底层的learning mechanisms，但不同的人可能有不同的方式来表达他们的想法，对吧？

So it's possible that there will be some underlying mechanisms for learning, but different people may have different ways to represent their thoughts, right?

所以这些想法在内部可能会以不同的modalities呈现，对吧？

So, and they might be presented internally as different kinds of modalities, right?

对我来说，我仍然觉得有时候我在思考的时候，我其实是在用图像的方式，或者说视觉的方式在想。

So for me, I feel still like sometimes when I'm thinking, I'm still, I'm thinking maybe it's actually picture-wise or like visual-wise.

有时候那是一种很奇怪的感觉，对吧？

Sometimes it's very, it's kind of weird feelings, right?

然后你突然就跳进去，然后你有了一个想法。

So then you suddenly kind of jump in and you have an idea.

哦，你有了一个想法。

Oh, you have an idea.

然后那个想法基本上会被外化成语言。

And the idea will be basically externalized as a verbal language.

但如果真是这样，那我觉得最好的选择还是内部隐藏表征，因为它可能涵盖了所有模态和所有可能性。

But if this is the case, then I would say the best bet is still the internal hidden representation, because that's something that is probably encapsulated all the possible modalities and all the possible possibilities.

所以我会押注这个。

So that's something that I would bet.

M326:21

但最后，这个想法没流行起来，对吧？

But at the end, it didn't catch, right?

比如现在，前沿模型都不是用连续的思维链来训练的。

Like today, models like frontier models are not training with continuous chain-of-thought.

你觉得为什么会这样？

Why do you think it's the case?

M226:35

首先，数据不够，对吧？

First of all, there's a lack of data, right?

就是说，你有大量的语言，大量的思维想法、思维模式都是通过语言来体现的，对吧？

So, I mean, you have lots of languages, lots of thinking thoughts, thinking patterns that are represented by languages, right?

所以呢，这是你能交流的方式，也是能把人的想法以文字形式存储下来的方式。

So, and this is the way you can communicate, and this is the way that you can store people's thoughts in a written form.

正因为如此，你拥有的数据比潜空间那边的数据要多得多得多。

Because of that, you have way, way, way much data compared to the latent space.

所以显然你得想办法去利用这些数据。

So then obviously you have to use that amount of data in some way.

所以那是一种做法。

So that's one way of doing that.

其次，潜在表征可能是个性化的。

And secondly, that the latent representation probably is personalized.

我的意思是，每个人都在某种程度上使用自己的潜在表征。

I mean, everyone is using their own latent representations to some extent.

就像艾伦说的，人们在这方面有很多差异。

So as Alan said, I mean, people have a lot of diversities there.

所以在这种情况下，学习起来会很棘手，因为人类的学习效率比计算机和我们的大模型要高得多，这点我得说，对吧？

So in that case, it's going to be very tricky to learn because human learning efficiency is much, much higher than computers and than our large models, I have to say, right?

所以如果我们的大模型没有这么高的新概念学习效率，那我们就得用语言作为替代品来让模型学习。

So if our large model does not really have this high efficiency in learning new concepts, then we have to use language as a surrogate to let the model learn.

M327:37

但你觉得这跟JEPA的问题有关吗？比如说。

But do you think it's related to the problem with JEPA, for example, right?

就是说，连续空间最终真的是很难正则化，对吧？

That at the end, continuous space is like it's really hard to regularize it, right?

就像所有自监督学习方法的主要问题是，你需要一个好的正则化器，有不同的方法来做，但哪种是正确的做法并不明确。

Like the main problem, like with all the SSL methods, is that like you need a good regularizer and there are different ways to do it, but it's not clear like what is the right approach.

这有点像那种定义不清的问题，基本上，你知道，你没法在一些非常明确、定义好的指标上进行爬山优化。

And like it's kind of like very ill-defined problems that you basically, you know, you can't like hill climbing on some very, very specific, well-defined metric.

你觉得这里也是同样的问题吗？

Do you think this is the same problem here?

M228:16

是的，在某种程度上是这样。

Yeah, to some extent, yes.

我觉得有两个事实，对吧？

I would say there's 2 facts, right?

首先，梯度学习可能效率不高。

First of all, I think the gradient learning probably is not that efficient.

所以你需要大量数据。

So you need a lot of data.

如果没有那么多数据，那就会正则化不足，对吧？

If you don't have lots of data, then it is underregulated, right?

所以你得加正则化器。

So you have to put regularizers.

但问题在于，你需要用哪些正则化器？

But the question really is that what are the regularizers you have to use?

有太多种正则化器了。

There are so many regularizers.

每一种都会给你带来一些不同的表征。

Each of them will give you some different representation.

所以也许它们中没有一个是真正正确的做法。

So maybe none of them are actually the right thing to do.

所以这实际上给那个带来了问题。

So that's actually causing problems for that.

第二个问题是，我们实际上并不理解什么是表征，以及表征是如何随着时间形成的。

And the second one is really that we don't understand what is the representation and how the representation forms over time.

这才是真正的问题。

That's actually the problem.

而这可能正是无梯度方法非常有用的原因，因为我们在这种元层面上，被梯度下降的解困在了局部最小值里。

And this actually is probably the reason why the gradient-free approach may be very useful because we are stuck with the gradient descent solutions as the local minima in this kind of meta level.

所以我们或许应该用一些不同的方式去学习表征。

So there may be some different ways we should learn representation.

M329:14

你认为存在一种最优的表征吗？

Do you think there is an optimal representation?

你觉得是不是有点像柏拉图式的表征范式？

Do you think like kind of like Platonic representation paradigm?

M229:21

是啊。

Yeah.

我想说，这目前仍然是一个未解的问题。

So I would say it's still an unsolved problem right now.

因为你可以用很多不同的方式来表征世界。

So there's so many different ways you can represent the world.

我认为可能存在某种方式比其他方式更好。

I think there could be some way that is better than the others.

比如说，有些方式占用的比特数更少，对吧？

So there's some way that occupy less bits, right?

比如压缩，对吧？

So for example, compressions, right?

有些方式可能能更好地泛化事物。

So there's some ways that may generalize things much better.

真正的问题在于，梯度下降，也就是当前神经网络的学习方式，要怎样才能学到那种表征。

And then the real question is how the gradient descent, how the current neural network learning can learn that.

还有这些学习范式的局限性是什么？

And what is the limitation for these learning paradigms?

可能存在某种非常好的表征，但当前的梯度下降方法无法学到它。

There could be some representation that is very good, but the current gradient descent approach cannot learn that.

我们现在还不知道边界在哪里。

We don't know the boundary right now.

现在我们只是说，好吧，看起来它是有效的。

Right now we just say, okay, yeah, it looks like it works.

然后我们就把它当作我们的系统来用。

And then we just use it as our system.

对吧？

Right?

但从科学的角度看，我们完全不知道边界是什么，也不知道怎样才能到达那个边界。

So, but scientifically, we have no idea what the boundaries are and how we can reach that boundary.

是的。

Yeah.

这正是我做可解释性研究、去理解神经网络的真正动力。

That's really the motivation for me to basically work on the interpretability, work on understanding neural networks.

M330:16

你觉得这会成功吗？

And do you think it will work?

就是，你认为我们会在某个时刻获得这种理解吗？

Like, do you think that we will have at some point this understanding?

M230:21

是的，我认为在某个时刻我们应该有信心，理解终将到来，对吧？

Yeah, I think at some point we should have the confidence that there will be understanding, right?

这就像是炼金术和化学之间的比较，对吧？

So this is like alchemy versus chemistry, right?

差不多就是这么个故事，对吧？

So that's kind of the story, right?

所以这个故事还在。

So this story is still there.

但是当然，因为当前的模型非常非常强，人们会去追求模型的效果，我们可能没有足够的时间，也没有足够的精力去关注那个方向。

But of course, because the current model is very, very strong and people will chase after the results of the models and we might not have enough time, enough people's mind on that direction.

但我一直说，从长远来看，这会对人类，也许对AI来说，也是一个非常困难、非常重要的问题。

But I always say in the long run, this is going to be a very hard problem and very important problem for humankind or maybe for AIs as well.

M330:51

所以我想问你一个更个人化的问题，因为你在深度学习、视觉和强化学习方面工作了很长时间。

So I want to ask you a more personal question because you worked on deep learning and vision and RL for a long time.

然后突然，对吧，大概在2022年、2023年差不多这个时候，突然我们所有人都明白了。

And then suddenly, right, like in 2022, 2023, like around this time, like suddenly, like all of us understood it.

就是，嗯，语言和文本模型已经站稳脚跟了，而且会是下一个大事件。

Like, yeah, language and text models are like here to stay and like are the next big thing.

你是怎么应对这个的？

How you manage this?

这种转变，你如何调整你的，怎么说呢，工作，以及你在工作中真正关心的那些东西？

Transition, like how you adapt your, I don't know, work and like the thing that you actually care about them in your work?

M231:29

嗯，我觉得首先，我大概从2018年、2019年左右就开始做可解释性研究了，那时候我试图对深度神经网络做一些梯度动态分析，对吧？

Yeah, I think first of all, I've been like working on like interpretability starting maybe like 2018, 2019-ish when I was trying to do some gradient dynamics analysis on deep neural networks, right?

所以从某种意义上说，这有点像我一直非常坚持推进的那些线索。

So this is in some sense, I kind of like the threads I have been pushing very consistently.

而且那条线索在某种意义上非常连贯，对吧？

And that thread is in some sense, it's quite consistent, right?

所以对于大模型，你仍然可以做同样的事情，对吧？

So you can still do the same thing for the large models as well, right?

你改变了模型，但做分析的方式是相似的，对吧？

So you are changing your models, but the way you do analysis is similar, right?

所以在这种情况下，它并不具有颠覆性。

So in that case, it's not disruptive.

主要来说，你还是在做自己真正想做的事情。

It's mostly like you're still working on something that you really want to work on.

但当然，另一方面，在实证层面，这会变得棘手，因为现在作为研究者，我们不得不彻底改变我们的思维方式。

But of course, on the other hand, on the empirical side, it's going to be tricky because now we have to completely change the way we think as a researcher.

对。

Right.

所以以前，通常对于研究者来说，我们提出自己的想法，然后用实验来检验这些想法，然后得到一些结论，对吧？

So previously, usually for researchers, we come with our own ideas and then we test the ideas with the experiments and we get some conclusions, right?

这个过程可能跟100年前有点像，那时候人们自己写论文做研究，自己做实验，对吧？

So this kind of process is similar to maybe 100 years ago when people do the research in a paper and do experiments by their own, right?

但考虑到大模型，我觉得通常因为当模型变得越来越强时，它本身就能给你一些想法。

But given the large models, I think usually because it actually can give you some of the thoughts by itself when the model becomes stronger and stronger.

所以研究者们可能被迫要在元层面上思考，而不是仅仅花时间提出一个个具体的假设。

So then the researchers are probably forced to think about something in a meta level rather than just spending their time on coming up with individual hypotheses.

所以这是人们需要面对的第一个挑战，对吧？

So this is the first challenge that people need to face for that, right?

是的。

Yeah.

第二个挑战，我觉得当时是，首先我们需要强调，卓越的工程能力也非常重要。

And the second challenge, I think at that time is that first of all, we need to emphasize that the engineering excellence is also very important.

与研究方向并行，对吧？

In parallel to the research side, right?

因为大模型时代很多好的工作，都像是半研究、半工程。

So because much of the good work from the large model era, they are kind of like half research, half engineering.

因为如果你不擅长把想法工程化，你的结果可能就不会很好。

So because you cannot, if you cannot like engineer things very well, then your results probably are not great.

那么人们可能就不会把它当作严肃的成果，对吧？

Then people probably won't take that as a serious output, right?

所以这其实是当前研究者面临的另一个挑战，对吧？

So that's actually another challenge that the current researchers are facing, right?

所以我觉得从2020年到2022年，再到大概一年前，情况就是这样。

So I think that from 2020 to 2022 to maybe a year ago, it's like that.

但是现在，因为我们都在用AI编程代理，所以情况可能会有点不一样了。

But right now, because we are all using AI coding agents, so the thing might become a little bit different.

但是因为有了AI编程代理的帮助，那些有丰富实践经验和知识的研究者，他们可能能想出很好的想法，然后让代理非常非常快地实现出来，再找到关键问题并解决它们。

But because with the help of AI coding agents, what happened is that researchers with good hands-on experience and knowledge, they might be able to come up with nice ideas, get them implemented by an agent very, very quickly, and then find the key problems and solve them.

所以突然间，这些研究者就有了超能力，可以去探索一些真正深层次的东西。

So then all of a sudden, these researchers actually have the superpower to explore something that is really deep.

所以会有一些来回的过程。

So there's some back and forth.

M334:17

好的。

Okay.

我觉得这正是讨论递归自我改进的好时机。

So I think that's a great time to go to talk about recursive self-improvement.

就像我们看到超级用户或超级工程师一样，我们会有能够自我改进的超级代理吗？

Like we see superpower users or superpower engineers, we will have superpower agents that improve themselves?

M234:38

是的，我觉得理想情况下这就是目标，对吧？

Yeah, I think that's ideally the goal, right?

M334:41

对。

Yeah.

那告诉我，这会发生吗？

So tell me, will it happen?

M234:44

会，我觉得会。

Yeah, I think so.

首先，我想告诉你为什么我觉得现在是时候开一家公司了。

I think first of all, I want to tell you why I think it's time to open a company.

关于这个，我在Meta的最后一篇论文是我和GPT-5合作完成的。

So on that, the last paper I had at Meta is a paper that I collaborated on with GPT-5.

这篇论文是关于研究神经网络中的grokking现象。

So this paper is about studying grokking behaviors in the neural networks.

基本上，网络会从记忆开始，然后训练、训练、训练、训练、训练，突然之间就变得泛化了，对吧？

So basically the network will basically start from memorization and train, train, train, train, train, all of a sudden become generalized, right?

这种grokking现象非常非常迷人，也很有趣。

So this kind of grokking behavior is very, very fascinating and interesting.

那篇论文，我和GPT-5合作，问了很多问题，还给它们布置了作业。

And for that paper, I collaborated with GPT-5 and I asked lots of questions and assigned them homework.

我就说，请证明这个定理，请确保这个定理是正确的。

So I said, please prove this theorem, please make sure that the theorem is correct.

请帮我头脑风暴一些想法和思路。

Please help me brainstorm on some ideas and thoughts.

我发现这实际上比我以前处理类似问题的方式有效得多，对吧？

I find that it's actually much more effective than how I used to work on similar problems, right?

我觉得它实际上给了我大概六到十倍的提升。

And I would say it actually gave me maybe a 6x to 10x boost.

在效率方面，对吧？

In terms of the efficiency, right?

而且这个估计其实还是保守的，因为那时候我还是靠自己写代码，但现在我觉得代码编写和实现当然也可以交给一个强大的AI代理来做，对吧？

So, and this is actually conservative because at that time I still relied on myself writing code, but now I think code writing and implementation, of course, can be delegated by a strong AI agent as well, right?

所以这会让事情变得快得多得多。

So this will make things much, much faster.

对。

Yeah.

所以我完成这个项目之后，得到了一些不错的有趣的结果。

So after I finished this project, I got some nice, interesting results.

我突然意识到，哦，好吧，也许我的工作在五年内就会被取代。

I suddenly realized that, oh, well, okay, maybe my job will be replaced in 5 years.

对吧？

Right?

所以很明显我很快就会失去我的工作了。

So it's very clear that I'm going to lose my job very soon.

所以与其被动地失去工作，我觉得不如就朝这个方向开一家公司。

So instead of losing my job involuntarily, I think it's good to basically start a company in that direction.

好的。

All right.

所以这就是为什么你会看到这种情况，对吧？

So that's why you see that happens, right?

所以通常研究员可能不太愿意去小公司当联合创始人，因为有很多事情，很多工程上的活儿。

So usually a researcher probably don't want to go to the small companies as a co-founder because there are lots of things, lots of engineering.

你需要操心研究之外的很多事情。

You need to worry about many things beyond the research.

但我觉得现在是一个阶段性的转变。

But I would say right now it's a phase shift.

如果我不采取任何行动，那以后我可能就无关紧要了。

And if I don't do any actions, then I might be irrelevant later.

是啊。

Yeah.

M136:41

你知道，你知道，我在镜头外有时候会和Raveed进行一些激烈的争论。

You know, you know, I get into bitter debates sometimes with Raveed off camera.

而且，你知道，当然在我状态不太好的时候，我也会说类似的话，主要是惊叹于编程智能体已经变得多厉害，同时也非常看衰我自己，你知道，大概到2040年代，感觉最终AI得找我谈一次话，然后说，你知道，如果所有prompt都由我来做，可能会好一点。

And also, you know, certainly when I'm not feeling as good, where I say similar things, mostly marveling at how good coding agents have gotten, and also being very bearish on me, you know, circa the 2040s, you know, feeling like eventually the AI has to give me the talk and is like, you know, it'd be a little better if I was just doing all the prompting.

而且我对此非常担忧。

And I'm very concerned about this.

最糟糕的是，我觉得你对此的回应——你知道，去成立自己的公司——是正确的回应。

And the worst part is I think your, your, you know, response to it, to go form your own company, is the right response.

我的回应则是去试着学习怎么动手干活。

My response has been to go try to learn how to work with my hands.

而且我并不擅长动手干活。

And I am not good at working with my hands.

所以当奇点来临时，我就完蛋了。

So I'm screwed when the singularity happens.

M337:24

我只想说，这个你到处都能看到，对吧？

I will just say that, like, and you can see it like everywhere, right?

比如软件工程，对软件工程师的需求一直在往上涨，对吧？

Like software engineering like the demand for software engineers are just going up, right?

而且我们到处都能看到这一点。

And we can see it like everywhere.

就像，你知道，并不是说有更好的产品就这么出现了，对吧？

Like there are no more like, you know, like better products, right?

或者说有更好的应用之类的东西。

Or like better like applications or something like that.

我们——也许在某个时候模型会变得更好，然后它们就不需要我们了，对吧？

We like— it may be that like the models will be better at some point that like they don't need us, right?

但至少现在，我们还没看到这种情况。

But at least for now, we don't see it.

我们没看到。

We don't see it.

而且，端到端地做出好产品，现在还是做不到的，对吧？

And like, it's not possible to come up like with good products end to end, right?

所以首先，你觉得，你同意这一点吗？

So do you think first, like, do you agree with that?

然后你觉得，如果同意的话，你认为这会改变吗？

And do you think like, and if so, do you think it would change?

M238:14

是的，首先我想说，现在仍然处于一个还没有自动化的阶段，对吧？

Yeah, first of all, I would say right now it's still at a stage that it's not automatic yet, right?

所以我不认为目前的模型可以在没有任何人工干预的情况下给你非常好的结果。

So I don't think that the current model can give you very nice results without any human interventions.

我认为还没到那个阶段。

I think it's still not at that stage.

但即使是我在Meta做的那篇论文，我也是给出高层次的想法、高层次的思路，让模型去跟随，而不是让模型提出自己的想法。

But even for the paper I had back in Meta, I still gave high-level thoughts, high-level ideas so that the model can follow rather than ask the model to come up with their own ideas.

因为那部分非常难，因为如果你没有任何想法，它就会给你非常机械、平庸的想法。

Because that part is very hard because you don't have any idea, then it will give you very mechanical and mediocre ideas.

基本上所有人都知道，而且这不是一个多有意思的研究方向。

Basically everyone knows, and it's not that interesting a direction to go.

对。

Right.

所以我觉得，到目前为止，模型还是在这么做。

So I think I would say till now, the models are still doing that.

没有——我们还是需要人类去发起，需要人类去思考哪些是重要的事情、该做什么。

There's no— we still need humans to initiate, still need humans to think about what are the important things to work on.

所以这是我们必须要做的。

So that's something that we need to do.

而且我们也需要人类来验证，对吧？

And also we need humans to validate, right?

就是解决方案提出来的那些问题是什么？它们长什么样？

So what are the problems that the solutions come up with? What do they look like?

所以这些其实就是我们正在面对的问题。

So these are actually the problems that we are facing.

希望在未来，如果agent有更多自动化——我是说模型——人类在重复性工作上花的时间会越来越少，更多时间花在更有创造性的方向上。

And hopefully in the future, if there's more automation from the agent, I mean, from models, humans may spend less and less time on repetitive work and spend more time on the more creative directions.

而且人类在未来也可以有更宏大的目标。

And humans can also have more ambitious goals in the future.

人类会说，哦，也许我哪天该去趟土星。

So humans will say, oh, maybe I should go to Saturn at some point.

然后我会把我很多执行类的工程工作委派给我的工程师、我的agents。

And then I will delegate lots of my execution engineering work to my engineers, my agents.

他们会想出怎么造火箭的办法。

So they will come up with the ways how to build a rocket.

M339:52

但你怎么看待所有这些——现在有很多论文和工作都在研究自我改进、递归自我改进。

But what do you think about all— so there are a lot of papers and works right now that are trying to look at self-improvement, recursive self-improvement.

你给一个agent两个小时和两块GPU，让它做post-training，或者就是试着改进某个指标——比如NeurMotron之类的。

Now you give an agent 2 hours and 2 GPUs and let him do post-training or like just try to improve some, like, I don't know, NeurMotron or like whatever, something like that on some metric.

至少据我所知，这些工作大多表明，当然，agents什么都能做，对吧？

And at least what I am aware of is that these works mostly showed that, like, of course, agents can do everything, right?

它们无所不能。

They can do anything.

任何你想要的优化之类的事情，但它们想出来的想法空间非常狭窄。

Anything that you want in terms of optimizations and things like that, but they are coming up with a very narrow space of ideas.

到头来，它们在想法上效率很低，而且它们并不真正知道怎么泛化问题，它们就是在过拟合那个指标。

At the end, they are very inefficient in terms of ideas, and they don't really know how to generalize a problem, and they are overfitting to the metric.

现在很多Kaggle比赛都在尝试用agents，它们会克隆代码来改进，看起来就像是在过拟合训练集或验证集的指标。

A lot of Kaggle competitions these days are trying to use agents, and they kind of clone code for improvement, and it looks like they are kind of like overfitting the training or validation metrics.

你怎么看？

What is your take?

你觉得这是不是——首先，还是先问，你同意吗？

Do you think this is something that like we don't— first, again, do you agree?

你觉得这是我们不知道该怎么衡量的问题，还是说这是个暂时性问题、很快就能解决？

Do you think like this is something that we don't know how to measure, or maybe it's a kind of temporary problem that is going to be solved soon?

M241:32

是的，我觉得在某种程度上它们是暂时的，因为你总有办法让agent去关注——让模型去关注更高层次的问题。

Yeah, I think to some extent they are temporary because there's always a way you can ask the agent to focus on— ask the model to focus on more high-level problems.

对。

Right.

所以你可以这么做。

So you can do that.

我觉得当前模型之所以做不好，是因为模型被要求去做那些细节的东西。

I think the reason why the current model cannot do that very well is because the model is asked to kind of do detail stuff.

没有上下文。

There's no context.

如果模型有更多上下文，它可能会提出一些更有趣的想法，而不仅仅是微调参数、调超参数、改函数配置这些。

If the model has more context, it might come with some ideas that are more interesting than just fine-tuning parameters, like tweaking all the hyperparameters, the function configurations, et cetera.

但我觉得给模型加更多上下文很可能只是一个临时的解决办法。

But I would say adding more context for models is probably like a temporary fix.

它可能会给你一些提升，但未来我们可能需要比当前基础的transformer架构要好得多的东西。

It might give you some boost, but in the future, we probably need something substantially better than the current fundamental transformer architectures.

好的。

All right.

所以我觉得现在的关键区别在于，当前的模型都是在大量数据上训练的，而这些数据都来自互联网，对吧？

So I think the key difference right now is that the current models are trained on a lot of data, and this data are all coming from the internet, right?

而且这些数据包含了很多重复性的工作，对吧？

So, and they consist of lots of repetitive work, right?

对于重复性的工作，我觉得模型做得非常好，对吧？

So for repetitive work, I think the model is doing very well, right?

所以它实际上有足够的数据来了解情况，也有足够的数据来提出新想法。

So it actually has enough data to know what's going on and enough data to come up with new ideas.

但展望未来，如果你想让模型做非常有创意的事情，一些我认为是跳出常规的事情，那模型可能就没有足够的数据来做了，这时候你可能需要思考什么样的新架构能够带来更高效的学习，并且能通过极少的数据快速捕捉新信号。

But going forward, if in the future you want a model to do very creative things, something that is, I think, out of the box, then the model probably don't have enough data to do that, then you probably need to think about what are the new architectures that lead to much more efficient learning and can pick up new signals super quickly with a very small amount of data.

所以基本上就是两个阶段。

So that's basically 2 stages.

从我刚才说的常规、重复性工作，到下一个层次，我觉得我们需要一个大的跳跃，但当前的范式仍然会存在很长时间。

From what I just said as a routine, like a regular repetitive work to the next level, I think we need some big jump there, but the current paradigm still can exist a long time.

因为总有很多补丁可以用来让模型更强。

Because there's always lots of patches you can use to make the model stronger.

这就是我的感觉。

So that's my feeling of that.

M143:22

那么让模型变得更好，最重要的是什么？预训练、后训练、甚至推理时优化，所有这些可能的改进方式，你如何看待？

What's the most important for making models better out of pre-training, post-training, even inference time optimizations, all these kind of possible ways to improve them going forward?

M343:34

或者也许是不同的架构，对吧？

Or maybe different architecture, right?

完全不同的架构、优化器。

Totally different architecture, optimizers.

嗯。

Yeah,

M243:39

没错。

Exactly.

是的。

Yeah.

所以是完全不同的架构、优化器，对吧？

So totally different architectures, optimizers, right?

还有数据混合等等。

So data mix, et cetera.

所以这会非常有趣，并且可以成为自我改进的主题。

So this can be very interesting and can be the subject of self-improvements.

这实际上就是我们现在在做的事情。

So this is actually what we are doing.

M343:51

那你怎么看？

And do you think, like, how do you see it?

因为，到现在为止，有很多公司、个人和创业公司试图提出不同的架构，对吧？

Because, like, right, like, till now, like, at the end, like, there are a lot of companies and people and startups that try to suggest, like, different architecture, right?

比如Transformer，有些在小规模问题上确实展示了不错的结果，但从来没有真正扩展到更大的模型上。

Transformers, and some of them actually show kind of good results in small-scale problems, but it was never actually scaled up to real bigger models.

你觉得这是为什么？你觉得这是你能解决的问题吗？

Why do you think is that, and do you think it's something that you can solve?

M244:29

嗯，这是个好问题。

Yeah, I think that's a good question.

首先，为什么你会看到很多在小规模上效果不错，但在大规模上却不行？因为通常在小规模上你缺少数据，对吧？

First of all, the reason why you see lots of good improvement on a small scale but not large scale is because usually for small scale you are lacking of data, right?

所以你基本上需要在架构中额外添加一些归纳偏置。

So you have to basically additionally add some inductive bias in the architectures.

而这些偏置在小规模上效果很好，因为你没有足够的数据。

And that bias works quite well for small scale because you don't have enough data.

但在大规模上，因为数据很多，最好的做法就是基本上让模型自己去弄清楚发生了什么，对吧？

But for large scale, because there's lots of data, and the best thing to do is to basically leave model to figure out what's going on, right?

所以我觉得一个非常有趣的例子就是这些 Vision Transformers，对吧？

So I think one very interesting example is these Vision Transformers, right?

那么当这种情况发生的时候，通常如果你读论文的话，你会发现，如果你用一些 ImageNet 级别的数据来训练模型，vision transformer 可能没有卷积网络那么强。

So when it happens, usually if you read the paper, what happens is that if you use some ImageNet-level data to train models, the vision transformer probably is not as strong as the convolutional network.

但是当你拥有更多的数据，比如两亿或者五亿，那么最终你会发现 transformer 会比卷积网络好得多，因为卷积网络有这种归纳偏置，基本上就是说——

But when you have lots more data, like 200 million or 500 million, then eventually you'll see the transformer become much better compared to the convolutional network because the convolutional network has this inductive bias, which is basically saying that—.

M345:29

你确定 Yann 会同意你的看法吗？

Are you sure that Yann will agree with you about that?

M245:32

我不确定。

I'm not sure.

嗯。

Yeah.

M345:40

你不用回答。

You don't have to answer.

没关系。

That's fine.

嗯。

Yeah.

M245:43

我觉得现在的直觉大概是这样的：transformer 基本上，你可以让它关注局部，也可以让它关注非局部的数据。

I think probably the intuition right now is like a transformer, there's basically, you can have the transformer attend to the locality, or your transformer can attend to the non-local data as well.

对。

Right.

但是对于卷积网络，你必须先关注局部的东西，然后逐渐构建。

But for the convolutional network, you have to focus on the local things first and then gradually build.

所以这两者是非常不同的。

So those two are very different.

M346:04

但我的问题是，你想要怎么做，你的策略是什么，对吧？

But my question is, how you want, what is your strategy, right?

你觉得我们实际上应该从……现在，对，你刚才说了，是的，我对不同的范式有一个新想法。

Do you think that we actually should start with, now, right, you said it like, yeah, I have a new idea about different paradigms.

那么现在我会从小规模实验开始，然后试着从那里扩展。

So now I will start with small-scale experiments and try to scale it from there.

或者你觉得从一开始就……值得吗？你知道多少？

Or do you think it's worth it to start? How much do you know?

比如说，我不知道，从一开始就搞一个一万亿参数的模型，然后在这个模型上迭代。

Like, I don't know, one trillion model from the beginning and try to iterate on this model.

M246:36

我觉得，当然，你需要先从小规模开始尝试，然后逐渐走向更大规模。

I think, of course, you need to try from small scale first and then gradually go to the larger scale.

但我认为真正重要的是，这是我个人的看法，我们在某种程度上真的需要知道网络内部发生了什么，对吧？

But I think what really matters is that, this is my personal take, is that we really need to, to some extent, know what's going on inside the network, right?

所以有某种可解释性的成分。

So there's some kind of interpretability component.

是不是我们知道它能在更大规模上很好地泛化？

Is it that we know it can generalize well in larger scale?

这可能是我们需要考虑的额外因素之一。

That's something that could be one of the additional components that we need to think about.

对。

Yeah.

而且我们知道，有一些有趣的见解是可以扩展到更大规模的，对吧？

And we know that there are some interesting insights that will scale up to a larger scale, right?

这样我们就有信心这能很好地工作。

So then we are confident that this can work well.

否则，如果我们盲目遵循 scaling laws，那么发生的情况是，你永远无法证明任何在小规模上有效的东西能泛化到大规模。如果你不这样做，如果你没有这种信心，那么当然只有拥有最大计算量的公司才会赢。

Otherwise, if we blindly follow scaling laws, then what happens is that you can never prove that anything working on a small scale can generalize to a large scale. If you don't do that, if you don't have this kind of confidence, then of course only the company with the biggest amount of compute will win.

那是唯一的解决办法。

That's the only solution.

但如果你真的在小规模实验中看到了深刻的东西，那么你有可能让某些东西在大规模上奏效。

But you actually see something deep in your small-scale experiments, and then it's possible that you can get something to work on a large scale.

M347:42

也一样。

As well.

所以如果你看这个循环，从想法到代码，运行然后分析。

So if you look at this loop of idea and code, run and analyze.

哪一部分最难自动化？

What part is the hardest to automate?

M247:57

我觉得，我的意思是，想法可能非常非常重要，也非常难，对吧？

I would say, I mean, idea is probably very, very important and very hard, right?

所以如果你问一个模型要想法，它会给一些比较平庸的想法，对吧？

So if you ask a model about ideas, I mean, it will give you some like kind of mediocre ideas, right?

所以当你把这两件事结合起来，检查什么有效什么无效，把所有东西组合在一起，做消融实验，对吧？

So when you combine all these 2 things together, check what works or not, combine everything together, do ablations, right?

这些其实不是特别有趣的想法。

So these are actually ideas that are not super interesting.

而且我猜其他人也会想到同样的事情，对吧？

And I would imagine that everyone else is going to think about the same thing, right?

所以真正的问题是，什么才是产生正确想法的正确方法？

So then the question really is, what is the right thing to come up with the right idea?

而这个想法要指向正确的洞见，对吧？

And so that idea actually points to the right insights, right?

那非常重要，对吧？

That's very important, right?

当然，你也可以做迁移，对吧？

So of course you can also do a transfer, right?

如果你有一些来自不同领域的好想法，也许你可以把它迁移到你一直在研究的领域。

So if you have some nice idea coming from different domains, maybe you can transfer to the domain that you have been working on.

那样的话，可能会非常有效。

And in that case, things might be very effective.

我看到了最近的突破，比如说来自数学领域的，对吧？

I've seen the recent breakthrough from, for example, mathematics, right?

就是AI for math。

So AI for math.

很多这些想法都是从一个领域迁移到另一个领域，然后发现这个想法确实有效。

Many of these ideas are kind of transferring from one domain to another domain and find that that idea actually works.

M348:53

比如，你之前尝试了跑NanoGPT speedrun，对吧？

So for example, you had this, you tried to run the NanoGPT speedrun, right?

M249:01

对。

Yeah.

M349:01

那你觉得这类竞赛，现在你甚至把纪录缩短了两秒，对吧？

And so do you think this type of competitions, and even now you achieved like you shaved 2 seconds off the record, right?

你觉得这是真正的发现，还是只是在一些已知技巧上做了很好的搜索，而这些技巧其实更容易被发现？

Do you think it's like a real discovery or it's just very good search over some like known tricks that actually was like that it was easier to find?

M249:27

对，我觉得从某种意义上说，这是有趣的发现和之前方法的结合，对吧？

Yeah, I think it's in some sense a combination of interesting discovery with the previous approaches, right?

比如，加入了bigram和unigram，对吧？

So for example, like the addition of like bigram and unigram, right?

在架构上，这从某种意义上说很有趣。

So in the architecture, it's in some sense interesting.

当然，有些人会说，哦，这可能来自DeepSeek，来自别的地方，也可能是Gemma 3的做法，对吧？

And of course, some people will say, oh, this is coming from maybe DeepSeek, coming from other, and maybe like the Gemma 3 approach, right?

而且实际上，我们所有人——当我在Meta的时候，我们也有一篇叫STEM的论文，也遵循了非常类似的想法，对吧？

So, and actually, we all, when I was in Meta, we also had a paper called the STEM that also followed very similar ideas, right?

所以这是一种向这个实际竞赛的过渡，而且这确实会提升性能。

So it's a kind of a transition to this actual competition, and this will actually boost the performance.

所以我觉得这基本上是现有想法的迁移，再加上一些创造力，把它们粘合在一起，让它们能很好地工作。

So I would say it's a combination of basically transferring existing ideas and basically there's some creativity in gluing them together so that they can work well.

比如，每一层可能都有不同的embedding表，对吧？

So for example, each different layer, you might have different embedding tables, right?

而且基本上组合方式也不同。

And basically the combinations are different.

M150:30

那么AI的未来是什么？

So what's the future of AI?

我的意思是，我们讨论了一些可能的候选，比如 world models，另外我们还提到了 recursive self-improvement。

I mean, we've discussed a few of these possible candidates like world models, and also, we touched on recursive self-improvement.

我们还提到了 physical AI。

We touched on physical AI.

那对于未来，我们是不是还漏掉了什么？

Are we missing anything for the future?

更多的 scaling？

More scaling?

M250:46

我是说，我觉得 scaling 大概现在就是常识，对吧？

I mean, I think scaling probably is the common sense right now, right?

所以人们会说，哦对，scale, scale, scale。

So people say, oh yeah, scale, scale, scale.

但我觉得，正因为是常识，我们才想考虑一些不同于常识的东西，对吧？

But I think because of the common sense that we want to think about something different than the common sense, right?

所以作为一种新的替代探索，对吧？

So as new alternative explorations, right?

所以我的意思是，recursive self-improvement 之所以重要，是因为模型现在已经达到了某种程度，在一定程度上理解了自身内部在发生什么，对吧？

So I mean, recursive self-improvement is important because the model now reached that point that this is already, to some extent, understand what's going on inside itself, right?

所以即使这种理解很浅，但仍然是理解，我们可以利用它来做 self-improvement。

So even if this understanding is shallow, but it's still understanding, and we can leverage that to do the self-improvement.

所以现在就是那个时机。

So that's the time.

而且它还有非常实际的使用场景，就像我们在很多场景中展示的那样，比如 kernel optimization，对吧？

And plus that it also has very practical use cases, as we've shown in many of the scenarios like kernel optimization, right?

所以这是其一。

So that's one.

当然，我常说，未来可能仍然是要找到新的架构。

Of course, I always say the future probably is still to find new architectures.

找到新的学习方式，找到更好的方法来匹配人类的数据效率。

Find a new way to learn and find a better way to match humans' data efficiency.

所以我们在哪个 modality 上工作并不重要。

So it doesn't matter which modality we are working on.

我们研究什么问题也不重要。

It doesn't matter which problem we're working on.

我认为关键仍然是这些神经网络中的学习机制和学习算法。

I think the key is still the learning mechanism and learning algorithms in these neural networks.

那大概就是关键点。

That's probably the key point.

所以一旦我们突破了这一点，对吧？

So once we crack that, right?

假设我们现在有一种新算法，它的学习效率和人类一样高。

Suppose that we now have a new algorithm that is learning as efficient as humans.

人类大概用二三十瓦的计算，而且人类只用了大约千分之一——基本上就是，对，0.1% 的数据。

A human uses like 20, 30 watts of compute and the human only uses like 1/1000 of the— basically like, yeah, so 0.1% of the data.

而人类可以在非常短的时间内搞清楚事物。

And the human can figure out things in a very short amount of time.

所以如果我们有了那种学习机制，那么一下子所有问题都能很好解决。

So if we have that, this kind of learning mechanisms, then all of a sudden all the problems can be addressed very well.

所以 robotics，现在我觉得 robotics 的问题确实是，你会说，好吧，需要大量数据。

So robotics, so right now, I think the problem with robotics is really, you say, okay, lots of data.

我觉得 Fei-Fei 确实提到过，我们在 robotics 上需要大量数据，但这只是针对当前的 paradigm。

I think Fei-Fei actually mentioned we need lots of data on robotics, but this is just for the current paradigm.

当前的 paradigm，因为你需要大量数据，所以我们真的需要提供大量数据。

The current paradigm, because you need lots of data, so we really need to provide lots of data.

但如果你想想人类是怎么做的，所有的猫、狗，所有生物，它们都在做 embodied AI，它们并不需要大量数据，对吧？

But if you think about how humans do that, how all the cats, dogs, all the creatures, they are doing the embodied AIs, they don't need a lot of data, right?

它们只有很烂的处理器，但同时它们能在这么复杂的世界里生存。

So they just have very crappy processors and at the same time they can survive in this complicated world.

这里面一定有新的算法。

There must be new algorithms there.

所以我觉得这是关键。

So I think that's the key.

而且我相信强化学习、自我改进可能是到达那里的更快方式。

And I would believe that the reinforcement learning, the self-improvement is probably the faster way to get there.

M353:06

所以最后几个问题。

So several final questions.

你觉得，像 Anthropic 和其他机构，现在对于限制最新模型有很多讨论，对吧？

Do you think, so in Anthropic and other, there is a lot of debate about it right now about restricting the latest models, right?

他们呼吁围绕递归自我改进进行协调。

And they called for coordination around recursive self-improvement.

你怎么看这件事？

What do you think about it?

M253:29

我想非常坦诚地说，因为我觉得很难阻止人们这么做，对吧？

I just want to be very honest with that because I feel like it's very hard to prevent people from doing that, right?

所以这个界限很模糊，对吧？

So the boundary is very blurry, right?

你和你的编程代理对话，带着新想法，让它们在新想法上迭代，那你就是在做自我改进。

So you talk to your coding agents, you come with new ideas and ask them to iterate on new ideas, then you are doing self-improvement.

在某种程度上，你和别人交流，他们产生想法，然后和你一起写论文。

To some extent, you talk with other people and they come up with ideas and they work together with you on some papers.

那也是自我改进。

It's also self-improvement.

所以很难划出这么一条线说，哦，你不能做这个，因为这是自我改进。

So it's very hard to draw this boundary that saying that, oh yeah, you cannot do this because it's self-improvement.

所以这很难做到。

So it's very hard to do that.

我的意思是，执行起来也很难。

I mean, it's also very hard to execute.

那界限在哪里呢？

So what is the boundary to that?

而且现在我们还有各种开源模型。

And now we also have all the open source models.

这些开源模型都非常非常好，非常强。

So in these open source models, they are all very, very good, very strong.

你也可以用它们来做一些研究、一些学习，对吧？

You can also use that to do some research, some studies, right?

那这能叫做自我改进吗？

So are they called self-improvements?

你怎么去管控它们？

How are you going to administrate them?

这些问题会很难回答。

It's going to be very hard to answer these questions.

所以并不是。

So it's not.

M354:25

那你对开源怎么看？

A— What do you think about open source?

比如，你们会开源自己的模型吗？

Like, are you going to open source your models?

我——

M254:32

我的意思是，对于我们的研究成果，我们打算开源，对吧？

I mean, I think for the outcome of our results, I mean, we are going to open source, right?

比如我们之前的博客文章，我们开源了所有的 kernels，也开源了训练这些 Nanochat 和 NanoSpeedruns 的全部配方。

For example, our previous blog post, we open source all the kernels, we open source all the recipes for training these Nanochat and NanoSpeedruns.

所以基本上，我们系统的成果，我们会开源。

So basically, the outcome of our systems, we're going to open source that.

这很重要。

So this is important.

首先，其他人可以检查，并且可以在其基础上继续构建。

First of all, everyone else can check and they can build on top of that.

那很好。

That's good.

而且我认为开源对整个社区实际上是非常好的。

And also, I think I would say open source is actually very good for the entire community.

所以最糟糕的情况可能是，一小部分精英人士控制着世界上最大最强的模型，然后说其他所有人使用它都需要支付高昂的费用。

So the worst situation that could happen is but a small elite of people and they keep— they have control of this biggest, strongest model ever in the world and say everyone else needs to pay a lot of premiums to use that.

这对整个社会来说可能不是什么好事，对吧？

That's probably not a great thing for the entire society, right?

所以我们真的想把AI的力量民主化，让每个人都有平等的机会去做好的工作，并用这些工具来提升自己。

So I think we really want to democratize the power of AI to everyone, so that everyone has the opportunity, equal opportunity, to basically do nice work and to strengthen themselves with the tools.

M355:40

你觉得未来我们会看到开源模型和闭源模型达到同样的水平吗？

Do you think in the future we will see open source models that are at the same level as the closed?

M255:47

闭源的那些？

Ones?

是的，我觉得现在已经很不错了。

Yeah, I think right now it's already pretty good.

比如Kimi K3就非常好。

For example, Kimi K3 is very good.

他们能做好多有用的工作。

They can do lots of nice work.

当然，思考过程可能比现在的闭源模型要长一些，但仍然是可用的，而且你可以在很多不同的领域利用它。

Of course, the thinking trace may be a little bit longer than the current closed source models, but still it's usable and you can leverage that in many different domains.

所以我觉得未来开源模型会变得越来越强大。

So I think I will say that in the future, the open source models will become stronger and stronger.

所以这根本不是问题。

So it's not a problem at all.

而且，作为一个编程代理，你不会真的期望AI Newton或AI Einstein来帮你写代码。

And plus that, as a coding agent, you don't really expect that AI Newton or AI Einstein will come to your place and do your coding for you.

所以其实是有一个边界的。

So there's actually a boundary.

所以有一条线，只要模型跨过了那条线，它就是可用的，你就愿意使用它。

So there's a line that as long as the model passes that line, it's usable and you're willing to use it.

所以我觉得，你知道的，很多前沿实验室已经跨过了那条线。

So I think, you know, a bunch of frontier labs, they already crossed the line.

而现在开源模型也开始跨过那条线了。

And now open-source models start to cross the line as well.

我觉得将来会有越来越多的人，越来越多的模型跨过那条线。

And I think I would say in the future, more and more people, more and more models will cross the line.

一旦达到那个程度，它就会被商品化，每个人都会使用自己最喜欢的模型，并且会非常开心。

And once you have that, then it will be commoditized and everyone's going to use their own favorite model and be very happy.

是的。

Yeah.

所以最终这会发生。

So eventually, this will happen.

所以真正的问题是，我们如何利用所有这些模型来做更好更大的事情？

So then the real question is like, how can we leverage all these models to do something better and bigger?

M356:57

好的。

Okay.

Yandong，你还有什么想补充的吗？

Yandong, do you have anything else that you want to add?

M257:01

我想基本上就是这样了。

I think that's pretty much it.

M357:04

你们在招人吗？

Are you hiring?

你想宣传一下你的公司吗？—— 是的，

Do you want to advertise your— Yeah,

M257:08

我觉得我会说，是的，如果你对我们公司感兴趣，欢迎随时给我们发送简历，对吧？

I think I would say like, yeah, if you are interested in our company, feel free to send us resumes, right?

所以我们的邮箱是 [email protected]，所有感兴趣的人，请发送给我们。

So our email address is [email protected], and everyone who is interested, please send us.

M357:23

你们在哪里招人？

Where do you recruit?

M257:29

是的，我们有一个地方——我们的主要地点是旧金山。

Yeah, so we have a place in— our main place is San Francisco.

还有伦敦。

And London.

所以我们三分之二的员工在旧金山，三分之一在伦敦。

So 2/3 of our employees in San Francisco and 1/3 in London.

M357:41

好的。

Okay.

非常感谢你加入我们。

Thank you so much for joining us.

非常荣幸。

It was a great pleasure.

M157:45

是的，真的很荣幸。

Yeah, it was a real pleasure.

谢谢。

Thank you.

M257:46

是啊。

Yeah.

非常感谢你采访我，我希望人们能从这些讨论和有趣的头脑风暴中受益。

Thanks so much for interviewing me and I hope that people benefit from the discussions and interesting brainstorming.

是的。

Yeah.

M357:54

好的。

Yeah.