递归自我改进人工智能的竞赛

这里的核心理念是打造能够自我改进的AI，我觉得我们正在接近这个目标。

The fundamental idea here is to build AI that improves itself, and I think we're getting there.

我们第一次可以把优化器当作被优化的对象，这样认知能力的提升就能反过来促进自身的进步。

For the first time, we could swap the optimizer as the optimizee, and this leads to a reinvestment of cognitive improvement into itself.

你可以从技术栈的各个层面来理解RSI。

You can think of RSI at many different levels of the stack.

从提示词、框架、权重，一直到芯片层面。

Prompts, harness, weights, and you can go all the way to basically chips.

我们的目标就是优化技术栈的每一个层面。

Our goal here is to optimize all the levels of the stack.

我们实现RSI的方法是让一个多智能体系统协同工作，并不断进行递归改进。

Our approach to RSI is that we have a multi-agent system working together that's being recursively improved.

你对这个模型自我改进的想法如此坚信，让我印象深刻——

I'm so impressed by your belief that this idea of a model improving itself-

但这并不是信念。

But it's not a belief

……

...

这是个很惊人的说法，我觉得这是我第一次听你这么说。

It is a dramatic statement, which is, I think it's the first time I've heard you say it.

因为我们一直非常严谨，在没有完全确认正确之前，我们是不会发布的。

Because we have been so thorough, we have not released things until we're super sure it's correct.

你有没有设想一个系统，随着改进，投入的能量会相对减少？

Do you envision a system where the amount of energy that goes into it somehow decreases over time relative to the improvement?

我们不能透露太多，但我们一直努力做的一件非常重要的事情是——

We can't share too much, but one of the very important things that we've tried to do-

a16z普通合伙人Martin Casado与MakerMaker.AI联合创始人Dhaval Adjodah和Owen He坐下来，深入探讨递归自我改进（RSI），并将这个越来越流行的术语与AI系统真正自我改进的实际含义区分开来。

a16z general partner Martin Casado sits down with MakerMaker.AI co-founders Dhaval Adjodah and Owen He to unpack recursive self-improvement, or RSI, and separate the increasingly popular term from what it would actually mean for an AI system to improve itself.

Dhaval和Owen将真正的RSI与仅仅使用AI来自动化AI开发的部分环节区分开来。

Dhaval and Owen distinguish true RSI from simply using AI to automate pieces of AI development.

他们的目标是形成一个闭环，在这个闭环中，被改进的系统同时也是进行改进的系统，涉及模型、框架，以及最终更多的AI技术栈组件。

Their goal is a closed loop, where the system being improved is also doing the improving across components including models, harnesses, and eventually more of the AI stack.

他们解释了为什么即使闭环形成，也不一定意味着失控式的飞速发展，以及为什么衡量真正的进展需要同时考虑能力提升和实现这些提升所需的资源。

They explain why closing that loop still doesn't guarantee runaway takeoff, and why measuring real progress requires accounting for both capability gains and the resources required to achieve them.

他们还分享了在研究中的发现，包括一些早期迹象表明，自我改进系统可以在提升效率的同时提高性能。

They also share what they're seeing in their own research, including early signs that self-improving systems can increase performance while improving efficiency.

他们还深入探讨了证明这些进步真实存在的实际挑战，从方差和奖励黑客到区分真正改进与噪声的巨大困难。

And they dig into the practical challenges of proving those gains are real, from variance and reward hacking to the enormous difficulty of distinguishing genuine improvement from noise.

欢迎两位。

Welcome, both of you.

我们要聊的是大家最喜欢的话题，就是RSI，递归自我改进。

We're going to be talking about everybody's favorite topic, which is RSI, recursive self-improvement.

也许先请你们做个自我介绍，讲讲你们的背景吧。

Maybe to start, it'd be great to just introduce yourselves and your background.

我们先简单聊聊你们现在在做什么，然后再深入讨论内容。

So we'll talk a little bit about what you're working on now, and then we'll dig into the content.

我觉得自己算是某种物理学家。

I think of myself as a physicist of some kind.

我学了那个，然后在麻省理工学院媒体实验室读了强化学习的博士。

I studied that, then I did a PhD in reinforcement learning at the MIT Media Lab.

我特别喜欢动手做东西，而且很高兴能和Owen还有Grace一起创办MakerMaker这家很酷的初创公司。

I really like building things, and I'm very excited to be building a cool startup, MakerMaker, with Owen and Grace.

我叫Owen。

My name is Owen.

我一开始学的是数学，后来有一天我意识到，数学家做的所有事情就是证明定理。

I was first studying math, and then one day I realized that all mathematicians do is are proving theorems.

当时我上了一门叫《形式语言与逻辑》的课，课上展示了怎么构建一个定理证明器。

And I was taking a class called Formal Language and Logic, and they were showing how you can build a theorem prover.

我当时就想，"哦，那太酷了。

And I was like, "Oh, that would be cool.

我可以造一台机器来帮我证明定理。"

I can just build a machine that proves theorems for me."

就这样，我进入了人工智能领域。

That's how I got into AI.

太棒了。

Amazing.

然后我读博的时候，研究的是持续学习和元学习，之后在DeepMind做了几年研究科学家。

And then for my PhD, I worked on continual learning and meta-learning, then worked for a few years at DeepMind as a research scientist.

现在我在做RSI相关的工作。

Now I'm working on RSI.

我看了很多项目提案，很多都提到RSI这个词。

I see a lot of pitches come in, and a lot of them use the term RSI.

我就会问：“哦，那你们具体做什么？”

And I'll be like, "Oh, so what do you do?"

他们会说：“我们用AI来创造AI。”

And they'll say, "We use AI to create AI."

“听起来挺有意思的。

"That sounds pretty interesting.

那你们具体做什么？”

So what do you do?"

“我们用AI来生成GPU内核。”

"We use AI to create GPU kernels."

这在我看来，跟RSI的定义不太对得上。

Which that doesn't, to me, kind of map the definition of what I would consider to be RSI.

所以，也许我们可以先请你定义一下，你认为RSI是什么，或者你会把哪些内容算作RSI？

So maybe can we start by you defining what you think or what you would include in RSI, definitionally?

是的，我觉得RSI现在确实是个被过度使用的词。

Yeah, I think RSI is definitely an overloaded term today.

很多人说RSI的时候，其实指的是不同的东西。

Many people, when they say RSI, they mean different things.

所以我觉得，区分它们的关键是要问一个问题：优化者是什么，被优化者是什么？

So I think the important thing when you are differentiating them is to ask the question, what is the optimizer and what is the optimizee?

当你在自我改进的时候，是谁在改进，对吧？

When you're self-improving, who is improving, right?

所以有些人会说：“好吧，我们在做RSI。”

So there are people who are saying, "Okay, we're doing RSI.

我们是用人工智能来开发更好的人工智能。

We have human using AI to develop better AI."

但对我们来说，这只能算是自我催化。

But that to us is just autocatalysis.

也就是说，改进者仍然是人加上工具，即使这些工具越来越先进，但想法和改进都来自人的大脑。

In a sense that the improver is still human with tools, even though these tools are getting better and better, but the ideas and the improvement come from human brains.

我不是想打断你，但是——

I don't mean to interrupt, but-

……说清楚一点，大多数技术时代都有自我催化的效应，对吧？

... to be clear, most technical epochs have had autocatalytic effects, right?

没有电脑，你根本造不出电脑芯片，对吧？

There's no way you can create a computer chip without a computer, right?

你可以说，互联网的发展之所以加速，是因为地理位置不同的人能够协作，对吧？

You could argue that the internet development was sped up by the fact that people that weren't geographically located could coordinate, right?

比如蒸汽机，你能造出更好的蒸汽机，是因为你能更快地获得煤炭，等等。

I mean, the steam engine, you could build better steam engines because you could get coal faster, et cetera.

所以，用这个东西来更快地造出这个东西，这个想法并不新鲜。

So this notion of using the thing to build the thing faster is not new.

所以你说的是另一回事。

And so you're saying something different.

对。

Yes.

我觉得对我们来说，在历史上，在文明进程中，我们一直有这种自催化现象，用电脑来造更好的电脑。

I think for us, in the history, right, over civilization, we have this autocatalysis, using computers to build better computers.

但是这种变化、这种改进的想法，是来自人脑的。

But the idea for this change, for this improvement, come from human brain.

但人脑从来不是被改进的对象，也不是被优化的对象。

But that human brain was never the improvee, never the optimizee.

哦，对。

Ah, right.

现在，我们第一次可以把优化者变成被优化的对象。

Now, for the first time, we could swap the optimizer as the optimizee.

所以智能体的核心，也就是语言模型，可以成为你优化的对象。

So the brain of the agents, which is the LM, the model, can be the thing that you're optimizing.

这就形成了一个复合效应，把认知提升重新投入到自身。

And this leads to a compound, this reinvestment of cognitive improvement into itself.

但你决定挑战这个难题，而不是像其他人那样只啃问题的一小块，肯定有你的理由。

But there has to be a reason that you decided to take on this hard problem as opposed to what anybody else would do, which is like chew off pieces of the problem.

在回答你那个更重要的问题之前，我先说下我为什么这么做——我觉得这是个难题，而且解决最难的问题很有趣。

I mean, before I answer your more important question here, the reason I did this is I think it's a hard problem and it's fun working on the hardest problem.

而且我觉得——

And I think-

但你有没有一种直觉，觉得这才是真正的正确答案，而不是更务实的做法？

But do you have an intuition that this is the actual right answer as opposed to something that's more pragmatic?

对，我觉得务实一点的话，从控制理论的角度看，如果系统里有缺口，就没法形成闭环。

Yeah, I think a pragmatic thing would be, just from a control theory perspective, if you have a gap in the system, you can't have a closed loop.

而自催化过程没有闭环，对吧？

And autocatalytic has no closed loop, right?

所以无论你改进什么，总会有差距。

So whatever you improve, there's still a gap.

人类必须介入，你必须投入其他资源才能让它运转。

Humans have to come in, you have to pour in other resources to make it work.

但如果你有一个闭环，你仍然需要投入资源，但边际收益理想情况下会随时间保持恒定或增加，这就是根本的区别。

But if you have a closed loop, you still have to pour in resources, but the marginal gains ideally get either constant or larger over time, and that's the fundamental difference.

我认为这里真正不同的地方很容易解释。

I think what's really different here can be explained pretty easily.

所以如果你考虑能力——

So if you think of capability-

……我们在这里试图做的是提高能力提升的速度。

... what we're trying to do here is we're trying to improve the rate of improvement of capability.

所以能力对时间的一阶导数是每个人都在努力改进的东西。

So the first derivative of capability with respect to time is the thing that everybody's trying to improve.

现在你可以把它设为等于两个因素驱动，分子和分母。

And now you can put that to be equal, driven by two things, the nominator and the denominator.

分子就是性能——

The nominator would be performance-

——也就是大家都在努力提升的东西。

... which is what everybody is trying to improve.

然后分母是某种成本，比如说资本，对吧？

And then the denominator is some kind of cost, let's say capital, right?

或者难度。

Or difficulty.

这个是比较固定的，因为能力越高，下一步的难度就越大，对吧？

And this one is pretty set because the higher the capability, the higher the difficulty of the next step, right?

现在，分子才是关键的部分。

Now, the numerator is the interesting part.

像蒸汽机这类东西，分子是固定的。

For things like steam engines and whatnot, that was fixed.

所以最后你会得到一个非常简单的微分方程，其中——

So then you end up with a very simple differential equation where-

基本上就停滞不前了。

... you basically plateau.

现在，如果有一个系统，被改进的东西同时也是进行改进的主体——

Now, if you have a system where the improvee, the things being improved, is also the thing that's doing the improvement-

那你就有了一个新的思考角度，最简单的情况就是分子也和能力相关——

... you now have a new way to think about it, and the simplest case would be that the numerator is also correlated with capability-

这说得通，对吧？

... which makes sense, right?

因为你越强，你改进的能力也越强。

Because the better you are, the better you are at also improving.

如果这样的话，数学上就抵消了，基本上就是持续不断的进步。

And if you do that, the math cancels out and you basically get constant improvement over time.

这可以算是递归自我改进的一个层次，对吧？

That could be one level of RSI, right?

然后还有一种情况——

And then there's one more case-

……也就是说，分子可能和能力的高次幂相关。

... which is now that the numerator could be correlated with a higher power of capability.

然后就会出现指数级的起飞。

And then you get exponential takeoff.

我们能不能再深入聊聊，RSI闭环和实际收敛性质之间的关系？

Can we dig in a little bit more between RSIs having closed the loop and actual convergence properties?

不知道为什么，当人们看到一张图，上面有一个圆圈指向另一个圆圈——

For some reason, when people see a graph where there's a circle with an arrow to another circle-

对

Yes

……然后又指回来，他们看到这个闭环控制系统，就会说：“起飞了。”

... and it goes back, they look at that closed control system, and they say, "Takeoff."

这，我的意思是，如果这是一个非线性控制系统——

Which, I mean, if this is a nonlinear control system-

……很难知道——

... it's very difficult to know-

对

Yes

……到底会发生什么。

... what the hell happens.

顺便说一句，大多数人似乎不明白这个区别——

By the way, most people don't seem to understand this distinction-

……因为我们一谈到这些，就很快跳到起飞和FOOM——

... because we very quickly go to takeoff and FOOM-

……以及所有那些东西。

... and all that stuff when we talk about these things.

嗯，作为网络方面的人，你知道这取决于你函数的指数——

Well, as a network guy, you know that it depends on the exponent of your function-

……那个自引用的函数。

... that's self-referencing.

我们发现，猜测可能会发生什么几乎没什么用，因为这真的是一个实证问题。

And what we found is that it's almost not useful to speculate on what we think could happen, because this is really an empirical problem.

真正要知道发生了什么，就得去测试。

And the real way to know what's happening is to test things.

而测试真的非常难。

And testing things is really, really hard.

比如，奖励黑客只是我们要处理的众多问题之一。

Like, reward hacking is one of many problems that we have to deal with.

但你问的问题非常实证，对于某些算法——

But the question you're asking is very empirical, and for some algorithms-

哦，有意思。

Oh, interesting.

所以我是搞网络的，我可能会那样回答。

So I'm a networker, that's how I would've answered.

对。

Yes.

我本以为数学人会这么说：“我们有这个微分方程。”

I would've thought math people would've been like, "We've got this differential equation."

然后我拿出铅笔——

And then I took out my pencil-

……然后我写下了答案。

... and I wrote down the answer.

我是说，写出微分方程是可能的，但你仍然需要实际测量变量和指数。

I mean, it is possible to write down the differential equation, but you still have to empirically measure the variable, the exponent.

所以没有解析解。

So there's no analytical solution.

说实话，我们甚至还没有深度学习的理论基础。

I mean, honestly, we don't even have a learning theory of deep learning.

我们根本不知道它为什么有效。

We don't know fundamentally why it works.

然后现在你在深度学习之上又叠加了五六层，所以我们没有好的理论。

And then now you're five, six layers on top of deep learning, and so we don't have a good theory.

我觉得我们最接近的就是缩放定律，而这些是经验定律。

I think the closest things we have are scaling laws, and these are empirical laws.

我觉得人们对我们的RSI方法有一个误解，就是他们认为整个系统必须完全自动化和数字化。

I think one misunderstanding people have about our approach to RSI is that they think the entire system has to be fully automatic and digital.

但实际上，如果你看看我们的数学基础，我们有一个多智能体系统，不仅仅是一个数字智能体。

But actually, if you look at our mathematical foundation, we have a multi-agent system, not just one digital agent.

在这个多智能体系统中，人类研究者和工程师实际上是这些智能体的一种特殊情况。

And in this multi-agent system, human researchers and engineers are actually a special case of these agents.

正是这个由多个智能体协同工作的混合系统，在被递归地改进。

And it's this hybrid system of multiple agents working together that's being recursively improved.

明白了。

I see.

我们不想遗漏任何组件而不去优化它们，因为如果有可能优化模型，我们为什么不这样做呢？

And we don't want to leave out any component to not optimizing them, because if it's possible to optimize the model, why wouldn't we do that?

因为如果你把那些也纳入优化，你会获得更多的收益。

Because you will have more gains if you do include those.

任何你不优化的部分，最终都会成为你的瓶颈。

Anything that you don't optimize will later become your bottleneck.

所以我们这里观察到的一个现象是，有人声称速度在提升，这确实是事实，对吧？

So one observation that we have made here is there are claims of speed is improving, which is actually the case, right?

比如你看模型发布的时间间隔在缩短。

So if you look at the model release time that it is reducing.

所以一方面，你可能会想，好吧，显然我们正处于某种快速进步的阶段。

And so on one hand, you could be like, okay, so clearly we're in some sort of a rapidly improving phase.

但如果把这些除以这些公司实际筹集的资金量，情况就不一样了，对吧？

But if you divide by the actual amount of capital raised by these companies, that's no longer the case, right?

然后你就会想，也许他们只是筹了一大笔钱，然后AI让他们能把这些钱用起来。

And so you're like, well, maybe they just raise a whole bunch of money, and then AI allows you to apply that money.

那么你会怎么——这不是针对你自己——但你会怎么鼓励那些在评估这些说法或者甚至在做相关研究的人，去思考不要通过往问题上砸更多钱来欺骗自己呢？

So how would you, this is not for yourself, but how would you encourage people that are either evaluating claims about this or even working on them to think about not fooling yourself by throwing more money at the problem?

你实际上是在展示你在这个分母上取得了有意义的进步。

You're actually showing that you're making an improvement here that's meaningful with the denominator.

我觉得对我们来说，这很简单。

I think for us, it's pretty straightforward.

因为如果你取得了进步，但只是靠投入越来越多的资源，那就不算真正的递归自我改进。

It's because if you get improvement and they just come at more and more resources being poured in, it's not RSI.

你必须能够证明你声称的进步有多大。

You have to be able to back your claim of how much improvement you're getting.

但你会把投入的所有人力和实验都算进去吗？

But do you include all the man-hours and experimentation that went into that?

因为我觉得在某种程度上，他们会谈论实际投入的算力——

Because I feel like at some level, they'll talk about the actual flops that go in-

——和资金，但现实是，这是第23次实验——

... and money that go in, but the reality is this is the 23rd experiment-

——而且实际上有个人在消耗大量的脑力才达到最终的配置。

... and there's actually a human burning a lot of mental capital to get to that final configuration.

你真的会把那部分也算进分母里吗？

Do you actually include that in the denominator?

我们会算进去，而且我认为这是公司良好的纪律。

We do, and I think it's good company discipline to do that.

而且我们从一开始就非常刻意地建立了一种文化，就是追踪资源去向、成本、时间这些。

And we've tried, again, to very intentionally from the beginning, have a culture of tracking where things go, cost, time, all that kind of stuff.

我跟你说个好玩的事。

I'll share a funny thing.

我们内部有个东西，大家叫它排行榜——

Internally, we have this thing, people talk about leaderboards-

就是比分那种。

... which is like score.

你肯定猜到我要说什么了。

You know what I'm about to say.

我们有个叫“流血榜”的东西——

We have something called a bleeder board-

就是记录每个人做实验花了多少钱。

... where we keep track of the cost of experiments of everybody-

太厉害了。

That's amazing

……那些在跑实验的人。

... who's running experiments.

太棒了。

That's awesome.

而且每个人每天都能看到谁在……

And everybody any day can look at who's-

谁在排行榜上名列前茅。我也可以举一个AI的例子。

Who is high up on the leaderboard I can also give an example from AI.

不知道你知不知道这个NanoGPT基准测试——对……大家都在跑。

I don't know if you know this NanoGPT benchmark- Yeah ... people are running.

对，它基本上就是让你重新训练GPT-2，这个模型刚出来的时候，训练成本非常高，要花几周时间，我想……在多块GPU上。

Yeah It's basically asking you to retrain GPT-2, which when it just came out, is a very expensive model to train and takes weeks, I think- Yeah ... on multiple GPUs.

而现在，人们可以在单块GPU上几小时内训练完成，而且这个时间还在不断缩短。

And right now, people can train it on single GPU within hours, and that number has been shrinking constantly.

但这也是完全独立的，对吧？

But that's also totally separate, right?

这就像摩尔定律一样，不，其实是同一块GPU。

That's just like whatever Moore's law like No, it's the same GPU.

硬件是一样的。

It's the same hardware.

用同样的硬件，然后——就是……你改进预训练的技术。

Using the same hardware, and then- It's what- ... you improve the techniques of pre-training.

哦，有意思。

Oh, interesting.

改进的是算法技术、数据技术、内核，所有这些。

What has improved is algorithmic techniques, data techniques, kernels, all of that kind of stuff.

你知道训练GPT-2要花多少钱吗？

Do you know how much it costs to train GPT-2?

我觉得现在也就几百美元。

I think it's a few hundred dollars now.

就100美元，我今天早上刚看到的——对，就是今天早上。

It's $100 I just saw- Yeah ... earlier this morning.

哦，哇——对……所以还是同一个GPU。

Oh, wow- Yeah ... so the same GPU.

同一个GPU，同样的数据。

Same GPU, same data.

这纯粹是算法上的进步。

And this is just algorithmic.

那这些进步是来自人类，还是来自智能体？

And do those come from humans or do those come from- Agents.

很多都是自动研究的结果。

A lot of them are auto research results.

但这绝对是分母。

But that's definitely the denominator.

这是分母。

That's the denominator.

不是分子。

That's not the numerator.

对，你想看到的是你能用GPT-5做到这一点，但我们还没看到，对吧？

Correct What you want to see is you can do that with GPT-5- Yeah ... which we're not seeing, right?

我的意思是，嗯，我觉得这些成本正在急剧上升。

I mean- Well, I think- ... the cost of these things is going up dramatically.

我觉得这个行业里最被严密保守的秘密之一，就是你能从中获得多少计算能力。

Well, I think one of the most closely guarded secrets in the industry is exactly this thing, which is how much compute you can get out of this.

因为从某种意义上说，即使你只是在改进你的分母，你也能加快训练速度，因为现在你可以进行更多的展开。

Because in some sense, even if all you were improving was your denominator, you could get faster training because now you can get more rollouts.

是的，当然。

Yeah, of course.

这意味着现在你可以更广泛地采样你刚才描述的语言空间流形。

And that means now you can sample what you were describing, the manifold of the space of language- Yeah ... more extensively.

当然。

Sure.

然后你可以获得更好的轨迹，这意味着当你用这些来训练时，可以得到更智能的模型。

And then you can get better trajectories, which means you can get smarter models when you use that to train.

当然，改进分母是好的。

For sure, improving the denominator is good.

当然，改进分子也是好的。

For sure, improving the numerator is good.

我还没见过，不过你有了一些苗头，能同时改进分子和分母，这很棒。

I've yet to see, which is great that you have glimmers, where you can improve both the numerator and the denominator.

这是个很有意义的说法，太棒了。

That's a pretty meaningful statement, which is awesome.

我们发现，很多声称做RSI的论文，在我们内部基础设施上测试时，根本行不通。

What we found is that a lot of the papers that claim to do RSI, when we test it in our internal infra, we find that they don't work.

很多时候是因为这些想法听起来很酷，人类会试着去内省，然后说：‘哦，我觉得应该是这样。’

And a lot of it is because they are cool ideas that humans try to introspect and say, "Oh, this is what I think should happen."

它们很复杂。

They're complicated.

它们听起来很酷。

They sound cool.

它们能帮你发一篇好论文。

They get you a good paper.

但当我们大规模测试时，在基准测试和真实经济问题上，它们就撑不住了。

But then when we test it at scale, on benchmarks, on real economic problems, they don't scale.

我们发现，真正能规模化的是那种简单、能闭环的方法。

And what we found is that the simpler thing that closes the loop is the thing that scales.

我想说，我们做的一件事可能比较独特，就是我们总是从一个非常扎实的数学框架出发，这个框架基本上把整个空间的地图都画出来了。

I would say one thing that we do that's probably unique is that we always start with a very grounded mathematical framework that basically lays out the map of the whole space.

原因是这个空间太大了，如果你不这么做，没有一个很有条理的方法，你很快就会在这复杂性里迷失。

And the reason is because this space is so large that if you don't do that and you don't have this very principled way of doing it, you get lost very quickly in this complexity.

另一个原因是，当你面对一个非常复杂的系统时，构建它有两种极端方式。

And the other reason is that often when you have this really complex system, there's two extremes of how you can build it.

一种是你基本上建200个不同的系统——对……然后你希望这个，怎么说呢？

You can take an approach where you basically build 200 different systems- Yeah ... and you hope that this, what is it?

这个鲁布·戈德堡机器能凑合着运转起来。

The Rube Goldberg machine will kind of work out.

对。

Yes.

或者你走另一条路，基本上花点时间思考这200个东西，然后你可能会意识到，它们其实只是12个抽象概念，可以概括这200个东西。

Or you take the other way, and you basically spend some time thinking about these 200 things, and maybe you realize, they're actually just 12 abstractions that generalize to these 200 things.

所以我们试图建立的文化是，真正从第一性原理出发去思考。

And so what we have tried to create a culture of, let's really think from first principles.

让我们看看数学，然后我们意识到为了实用我们需要做哪些假设，因为两者都必须兼顾。

Let's look at the math, and then we realize what are the assumptions we make to be practical, because you have to do both.

然后我们看看什么有效。

And then we see what works.

当你这样做的时候，你就会发现，哦，这些多篇论文或想法其实是同一个通用框架的子类。

And when you do that, you then can see that, oh, these multiple papers or ideas are actually subclasses of the same general framework.

如果你做得好，你就能扩展。

And if you do that well, you can scale.

我认为这是我所看到的一个缺失的东西，而且很难做到，需要一群非常不寻常的人才能真正实现它。

And I think that's one thing that I see that's missing and that's hard to do, and takes a very unusual group of people to really make that work.

我记得理查德·费曼说过，科学有两种。

I think Richard Feynman said once that there are two kinds of sciences.

一种是集邮，另一种是物理学。

There's stamp collections, and then there are physics.

所以如果没有统一的数学框架，所有这些研究都只是集邮而已。

So without a unified math framework, all this research are sort of just stamp collections.

我们读的很多论文，其实只是受人类行为启发的想法，然后他们试图模仿这种行为，然后说：“看，我们有RSI。”

A lot of this paper we're reading, they're just ideas inspired by some human behavior, and then they try to mimic that and say, "Oh, look, we have RSI."

但如果你不把它们放进一个统一的系统里，你就不知道整个系统是否真的在进步，还是你只是在局部应用一个越来越好的想法。

But then if you don't put into a unified system, you don't know if the overall system is actually improving, or you're just locally applying an idea that's getting better.

所以我们做的是，把所有可能递归改进的组件都放进一个统一的框架里，从整体的角度来看，这样改进一个部分不会对其他部分产生负面影响。

So what we have is that we're putting all these components that you could possibly recursively improve into a unified framework and look at it from a holistic view, so that improving one part does not negatively impact the other part.

明白了。

I see.

比如说，有些人只对harness做递归改进，而有些人只对权重做改进。

So for example, there are people who are doing recursive improvement just on harness, and then there are people who are only doing on weights.

但权重和框架并不是孤立的组件。

But weights and harnesses are not isolated component.

它们会相互影响。

They have impact on each other.

你可以从技术栈的不同层面来考虑RSI，对吧？

You can think of RSI many different levels of the stack, right?

你可以想到提示词、框架、权重——对……甚至可以一直想到芯片层面，对吧？

You can think of prompts, harness, weights- Yeah ... and you can go all the way to basically chips, right?

对，当然。

Yeah, sure.

我们的目标就是优化技术栈的所有层面，对吧？

And our goal here is to optimize all the levels of the stack, right?

之所以要高效，是因为我们非常在意分母。

Efficient because we care so much about the denominator.

然后我们开始处理权重，进行权重更新。

And then we started doing weights, weight updates.

我们发现这样做收益更大，但还是得处理框架相关的事情，对吧？

And what we found is that you get more gains, but you still have to do harness stuff, right?

你可以继续往下深入，我猜你可能需要对系统的很多不同部分进行协同优化才能让它运转起来。

And you can keep going down the stack and my guess is that you probably will need to do a lot of the co-optimization of many different parts of the system to make it work.

你能谈谈你正在攻克的问题领域，来证明这个方法有效吗？

Can you talk a little bit about the problem domain that you're going after just to prove that this works?

我们非常努力地挑选AI领域中尽可能接近AI本身的方向。

We are trying very hard to pick areas of the AI space that are as close as possible to AI itself.

再说一遍——

Again-

明白了。

I see

……我们尽量做到递归。

... we try to be recursive.

你想想看，真正非常接近AI、同时能在分子和分母上都带来改进的问题并不多。

And if you think about it, there's not that many problems that are super close to AI and gives you improvement on both the numerator and denominator.

其中一个就是推理问题。

And one of those is the inference problem.

因为推理效率提升得越多，其他方面就越能解锁，对吧？

Because the more you can get improvements and efficiency out of inference, the more everything kind of unlocks, right?

这不仅仅是因为技术原因，还因为分子和分母都有望得到改善。

And it's not just for technical reasons, but also the most promise of both the numerator and denominator being improved.

这有点让人难以置信，对吧？

It kind of breaks the mind, right?

所以，给所有听众解释一下，这里有两个改进，一个是模型推理速度更快。

So just for everybody that's listening to this, there's two improvements going on here, and one of them is the ability of the model to make inference faster.

第二个是，既然推理更快了，它就能更快地完成推理。

And then the second one is, now that the inference is faster, it can do that faster.

没错。

Exactly.

是的。

Yes.

而且再说一遍——

And again-

……这并不一定意味着爆发式增长。

... that does not necessarily mean takeoff.

它仍然可能趋于平稳——

That can still asymptote-

对

Yep

……在某个时候。

... at some point.

但想到这种复合效应，真是令人惊叹。

But it's just wild to think of this kind of compounding impact.

我们是一家初创公司。

We're a startup.

我们不是研究实验室。

We're not a research lab.

我们选择它的一个原因是，现在全世界每个人都想部署自己的模型。

And one of the reasons we chose it is because everybody in the world right now wants to serve their own model.

每个人都想用最新的，比如刚出来的开源权重模型Kimi K3，而且希望尽可能高效地部署它。

And everybody wants to use the latest, for example, open weight model, Kimi K3 or whatever just coming out, and they want to be able to serve it as efficiently as possible.

所以这现在是一个经济上极其重要的问题。

So it's an economically extremely important problem right now.

为了讨论方便，如果我想让Kimi K3的推理更快，我是用Kimi K3本身来递归地做，还是用另一个模型和系统来帮Kimi K3提速？

Just for the sake of discussion, if I wanted to make the inference for Kimi K3 faster, would I use Kimi K3 in a recursive way, or would I use a separate model and system to do that for Kimi K3?

你是说改进Kimi K3的部署栈吗？

To improve the serving stack of Kimi K3, you mean?

嗯，我们的系统不只是用一个LLM。

Well, our system is not just using one LLM.

我们有多个不同的LLM在各自做不同的事情——

You have many different LLMs that are doing many different things-

哦，所以你们是——

Oh, so are you-

……针对不同的可能性。

... for different possibilities

……为所有这些模型做推理？

... doing the inference side for all of them?

再说一次，我们选择最有效的那一个，因为它被使用的时间最长。

Again, we pick whichever is the most effective, and because it's being used the most time.

但有趣的是——

But the interesting thing is-

明白了。

I see.

……选择用哪个模型来处理哪件事，也不总是由我们决定的。

... the choice of which LLM to use for which thing is also not always up to us.

我们让模型自己选。

We let the model choose.

说实话，今天真正决定优化什么的并不是我们。

To be honest, today it's not really us who are deciding what to optimize.

是我们的智能体。

It's our agent.

这很有意思。

That's interesting.

因为它能递归地自我改进，我们想做的不是把它限制在单一组件上。

Because it's recursively self-improving, and what we're trying to do is not to limit itself to only one component.

所以它自己决定是否要——

So it chooses whether to-

它自己决定是否要——

It chooses whether to-

我们尽量少干预。

We try to be as hands-off as possible.

让它自己发挥。

Let it do its thing.

你们有一个总提示词吗？

Do you have a single master prompt?

是不是，比如，自我改进？

Is it, like, improve yourself?

但就连那个主提示词也在自我改进。

But even that master prompt has been improving by itself.

真的吗？

Really?

它还能改进主提示词。

It improves the master prompt.

而且它不是一个单一的RSI系统，因为你可以把它想成，有一个负责改进的东西，但你还得有一个知道该改进什么的东西，这两个都是动态问题，而且它们非常——

And it's not one RSI system, because you can think of it as there's the thing that does the improvement, but then you also have to have the thing that knows what to improve, and that's both are dynamic problems, and they are very-

当然。

Sure

……

...

相互关联。

co-linked.

你觉得这最终能有多通用？

How general do you think this ends up being?

现在你专注于加速特定领域的推理。

Right now, you're focused on speeding up inference in a specific domain.

你觉得你在这里提出的方法能通用化，还是每个目标解决方案空间都会有不同的方法？

Do you think that whatever approaches you're coming up with here generalize, or do you think there'll be a different one per target solution space?

从数学上讲，我们把它定义为一个能通用于所有事物的系统。

Mathematically, we have defined it so that it's one system that generalizes to everything.

我认为系统的一部分——

I think the one part of the system-

你说的所有事物，是指任何可以指向AI的问题领域吗？

And by everything, you mean any problem domain that you can point AI to?

对。

Correct.

但比如，还存在物理限制。

But then there are physical limits, for example.

当然，是的。

Of course, yeah.

其中一个很大的限制就是人的因素。

And one big limit of this is the human element.

我觉得很多AI公司的想法是，用AI来替代人类。

I think the way that a lot of AI companies are thinking is let us use AI to replace humans.

但我们发现，如果你有优秀的研究人员和工程师，他们能带来的价值是很大的，这些很难被智能体替代。

But what we've found is that there's so much that if you have good researchers and good engineers that they can bring to the table, that's going to be very hard to replace by agents.

所以从一开始，我们就试图把整个公司看作一个多智能体系统，人类是其中一种智能体，和其他AI智能体共存。

And so that's why, from the very beginning, we've tried to think of our basically entire company as a multi-agent system, where humans are a kind of agent that's cohabitating with other AI agents.

我们能不能至少降低一些限制？

Can we at least reduce bounds?

比如说，你有没有设想一个系统，其中投入的能量相对于改进来说，会随着时间减少？

So for example, do you envision a system where the amount of energy that goes into it somehow decreases over time relative to the improvement?

我们不能透露太多，但我们在公司里尝试做的一件非常重要的事情，不仅仅是关注分子部分，也就是我之前描述的那些——

We can't share too much, but one of the very important things that we've tried to do in our company is not just work on the numerator, which is what I was describing earlier-

……也就是你的表现。

... which is your performance.

但还要在分母上下功夫。

But to work on the denominator.

也就是在同样的物理资源、资金，或者随便你怎么称呼的基础上，能获得多少产出。

Which is how much you can get out of the same amount of physics, money, whatever you want to call it.

但你觉得有没有可能做到，分子继续以同样的速度增长，而分母保持不变甚至下降？

But do you think it's possible to have a system where the numerator continues to increase at the same way with the denominator being fixed or decreasing?

不只是可能。

It's not just possible.

我觉得这正在发生。

I think it's happening.

真的吗？

Really?

所以用更少的能量就能得到更好的结果。

So it takes less energy to get better results.

随着时间推移。

Over time.

你觉得有没有可能这种情况会一直持续下去，不会趋于平稳？

Do you think there's a universe in which that continues without asymptoting?

因为如果是那样的话，那实际上就是起飞了。

Because if that was the case, then you actually have takeoff.

我是说——

I mean-

但这似乎违背了信息论和——

But it just seems to disobey information theory and-

我是说——

I mean-

……物理定律。

... the laws of physics.

我觉得到某个时候，你会达到普朗克常数。

I think at some point, you're going to reach Planck constant.