Muse Wonder：扎克伯格与王慧文的千亿美元AI

主动记忆。

Active memory.

主动专注是另一个方面，实际上就是模型在执行任务时能够持续跟进的能力。

Active focus is another, which is effectively just the ability for the model to, you know, when conducting tasks be able to sort of carry them through.

特别是我们看到Muse用户和Muse代码用户给出的任务越来越复杂、越来越模糊、越来越有挑战性。

Especially as we've seen Muse users as well as Muse code users give sort of like higher and higher complexity and more and more ambiguous and more and more complex tasks.

我们希望模型能够非常有效地完成这些任务。

We want the model to be able to conduct those and do those very effectively.

实际上，在早期的内部测试和试用中，很多人注意到这个模型在这方面很特别，相比他们尝试或使用过的其他模型，表现很突出。

And this is one of the things actually in early dogfooding and testing that a lot of people have noticed about the model that is sort of special and and stands out with respect to other models that they've tried or used.

第三个是领域……

The the third one is domain...

是……

Is

在AI术语里，专注是什么意思？

focus in AI terms?

是不是就是把大量计算资源投入到任务上？

Is it like putting a lot of compute on the task?

是不是会被其他类似的东西分心……

Is this getting distracted with other like...

有点像持久性？

Kind of persistence?

基本上就是能够在很长的周期内完成完整任务的能力。

Like it's basically like the ability to conduct complete tasks over very long horizons.

好的。

Okay.

然后，当遇到各种挫折时，是的。

And then, when faced with, various setbacks along the way Yes.

能够以一致的方式找到其他解决方案的能力。

The ability to sort of identify other solutions to those problems in an aligned way.

好的。

Okay.

然后还有……

Then there's there's...

这实际上是编码能力的一大步提升，也就是前沿编码能力。

It's actually a big step up in coding capabilities, so frontier coding capabilities.

所以它在深度SWE和SWE马拉松上确实是目前最先进的。

So it is actually state of the art on deep SWE and SWE marathon.

在其他基准测试上，比如终端台四和其他很多测试，也表现很好。

And does very well on other benchmarks like terminal bench four and many others.

所以我们认为，用它来编码的用户——虽然只占我们所有用户的一小部分——但那些用户会看到很大的改进和提升。

So we think that users who use it for coding, which will be a minority of all of our users, but the ones who do will see a big improvement and a big step up.

好的。

Okay.

那我们是不是也应该在Stripe用Muse Wonder来写代码？

So we should also just be using Muse Wonder full for coding at Stripe?

有可能。

Potentially.

我是说，我们出去试试看……

I mean, let's go out there and...

在SWE基准测试上。

On sweet bench.

不。

No.

在编程方面，这是一个很大的进步。

It's a big it's a big improvement on coding.

我是说，显然，你知道，我们希望人们去尝试，看看效果如何……

I mean, obviously, you know, we want people to try it and and see how do...

他们的体验怎么样。

How they experience it.

但我们在内部测试中一直非常满意。

But we've we've been very pleased in our own internal testing.

好的。

Okay.

所以这是……

So this is...

我觉得大家会在很多方面喜欢它的。

I think people will like it for a lot of things.

好的。

Okay.

所以这既是你们最新的前沿模型，开发者可以直接来找你们，也可以在开放路由器上使用，就像直接把它当作一个模型来用；同时它也是 Muse 内部的一次重大升级，作为 Muse 的活跃用户，大家会直接感受到这种提升，对吧？

So this is both just your latest frontier model, like people can come, you know, developers can come directly to you, they can use it on an open router, like they can just consume it as a model, and it's also a big upgrade inside of Muse that people will just feel as, you know, active users of Muse?

嗯，我们在 Muse 里打算这么做，因为它规模更大，服务成本也更高。

Well, what we're gonna do in Muse, because it's a lot bigger and it's more expensive to serve.

对。

Yep.

我们打算主要先从付费订阅开始。

We're we're going to start off primarily in paid subscriptions.

对。

Yep.

所以我们想做的Muse这个产品，是希望每个人都能免费使用。

So what we want to do with Muse as a product, we want it to be free for everyone to use.

所以我们提供了大量的免费额度，大家可以用Muse Spark这个模型。

So we have this huge amount of free tokens that people can use with with Muse Spark as the model.

而且……

And...

但如果你想用Wonder来处理大部分复杂任务，那至少目前，你需要使用付费版本。

But if you want to use Wonder for most of the hard tasks, then at least for now, you'll you'll use the paying version.

不过，你知道的，我们会……

But, you know, I mean, we'll...

你也知道这个领域变化有多快。

You you know how quickly the field moves.

我们希望，在不久的将来，能有一个和Wonder一样好的模型，但以Spark的形式呈现，那样就能服务全球数亿用户，那将是非常棒的事情。

We hope to, you know, at some point not too far from now, have a model that is as good as Wonder, but in the Spark form factor, and that will be a great thing to be able to serve hundreds of millions of people around the world.

好的。

Okay.

所以你去Muse，订阅那个，叫什么来着，高级版还是专业版，对。

So you go to Muse, you subscribe to the, what's it called, like the premium or pro Yeah.

要么付费。

Plan or pay money.

一直操作到掏出信用卡为止。

Just don't stop until you've taken out your credit card.

然后你就能……

And then you'll be able to...

你会选择它，或者它会自动……

You'll select it or it'll just magically...

哦，它会变得更好。

Oh, it just gets better.

对你来说会变得更好。

It just gets better for you.

然后……

And...

然后它也会在……上提供。

And then it'll also be available on the...

通过API。

Via API.

在Metal模型API和OpenRouter上。

On the metal model API and OpenRouter.

而且，它也会对Muse代码用户开放。

And and then it'll also be available for Muse code users.

所以……

So...

然后，Muse Wonder完整版将对Muse代码的订阅者开放。

And then Muse Wonder full will be available to subscribers of Muse code.

所以有……

So there's...

这真的很令人兴奋。

It's it's pretty exciting.

我觉得我们想……

Think we wanna...

你知道，就像马克提到的，这是一个更大的模型，而且

You know, as Mark mentioned, is a much larger model and

对，是的。

so Yeah.

我们正在研究如何最有效地把它提供给尽可能多的人。

We're working through how to serve that most effectively to as many people as possible.

但我们很兴奋能让尽可能多的人用上它。

But we're excited to get it in the hands of as many people as possible.

因为这是一次进步。

Because it's a step forward.

我的意思是，这就像……

I mean, this is like...

而且我们真的，对于Muse Spark，我们对到目前为止能够提供的性能感到非常满意。

And we really, Muse Spark, we were very happy with the performance that we were able to deliver as we were getting up to this point.

但这只是开始，是的。

But like this is the beginning of like Yes.

MSL 开始处于前沿位置。

MSL kind of being at the frontier.

我们觉得这些模型和市面上最好的模型相比，非常有竞争力。

We feel like the models are quite competitive with the best models that are out there.

而且我觉得这……

And I I think that this...

好的。

Okay.

所以你觉得 Astra、Fable 和 Wonder 一起处于前沿？

So you think it's like Astra, Fable, and Wonder altogether at the at the frontier?

嗯，你可以……

Well, mean you can...

你是根据基准测试的实际结果来判断的吗？

Just empirically by the benchmarks you think?

按照……

By the...

按照基准测试。

By the benchmarks.

是的。

Yes.

可以得出这个结论。

One would conclude that.

对。

Yes.

我明白了。

I see.

是的。

Yes.

对。

Yep.

而且我觉得，其实我们特别自豪的一点是，你知道，我们花了很多时间，就像我们之前提到的，针对Muse的产品形态去优化模型。

And and I think actually one of the one of the things that we're especially proud of is, you know, we have taken the time kind of as we've mentioned to optimize models for the product form factor of Muse.

所以这是最先进的模型。

So this is the state of the art model.

我们认为它真的是世界上最好的模型，是的。

We think it really is the best model in the world Yes.

针对个人专业代理这种使用场景。

For the use case of personal professional agents.

公司里的人喜欢抱怨古德哈特定律，你知道，一个指标一旦成为目标，就不再是个好指标了。

People inside companies like complaining about Good Heart's Law, you know, a metric that becomes a target ceases to become a good metric.

你会听到有人抱怨说模型都是刷榜刷出来的，而且到了一定程度，过度拟合基准测试反而会让模型变得更差。

You hear complaints that models are bench maxed, and that at a certain point kind of overfitting to the benchmarks actually makes the models slightly worse.

作为真正要训练AI模型的人，这是真的吗？

As people who actually have to train AI models, is that true?

这些基准测试真的能很好地评估模型吗？

Are the benchmarks pretty good at grading models?

你们会不会开始看到一些奇怪的边缘失败案例，比如新的……

Do you start to see like weird edge failure cases from the new...

基准测试必须

Benchmarks have to

不断更新。

keep moving.

因为，你知道……

Mean, because you...

就像每次……

Like every...

我是说，这发生得很快，但它们很快就被饱和了。

I mean, it happens very quickly, but they just get saturated so quickly.

所以，我确实认为……

So I I mean, I do think that...

你觉得那些没饱和的基准测试没问题吗？

What you think of non saturated benchmarks are fine?

我是说，随着我们不断推进，我们更多地引导模型朝着我们内部基准的方向发展，这些基准是针对我们关注的使用场景的。

I mean, as we've gone on, we steer the the models more towards our internal benchmarks for the use cases that we're focused on.

所以，早期当我们还在做科学探索，试图一步步往上爬的时候，是的。

So I mean, early on when we were kind of just doing the science of trying to climb the ladder Yep.

我们更关心外部基准，只是为了确保我们走在正确的轨道上。

We cared more about what the external benchmarks were just to make sure we were on the right track.

对。

Yes.

然后我想说，每一次发布，我们关注的重点中，更大的一部分是……

And then I'd say, you know, each release that goes by the larger share of the of the the...

我们看的是我们内部的基准。

What we're looking at are our internal ones.

我们在研究博客文章中量化的一件事是……

And one of the things that we actually quantified in our research blog post

对。

Yes.

作为Muse Wonder的一部分，这个指标就是关于大模型气味的。

As part of Muse Wonder is this metric for big model smell.

这有点像，大模型气味和刷榜的感觉，就像是，对。

Which is kind of, big model smell and feeling of bench maxing are kind of like Yeah.

相反的……

Inverse...

就是那种你能直接感觉出来的，对。

Idea that you could just like tell Yeah.

当你用一个真正的大模型的时候，就是那种，对。

When you're using a really big model that just Yeah.

你能感觉出来。

You can tell.

它就有那种新的感觉。

It just has new.

它就有那种“清洁工爪子”的感觉。

It's got that janitor claw.

所以我们……

And so we...

我们量化这一点的方法，其实基本上就是你提到的那个：分布内性能和分布外性能之间的差距。

The way we quantify that is actually almost explicitly what you're mentioning which is the gap between distribution in performance and out of distribution performance.

所以这个……

So what is the...

你知道，我们大致是在衡量……

You know, we kind of measure what is...

模型在符合我们强化学习和后训练分布的任务上，表现如何？而在分布外的任务上呢？

How well does the model perform on tasks that are in distribution of our RL and post training versus tasks that are out of distribution?

这种差距具体是什么样的……

And what does that gap look like for...

Muse Spark系列和Muse Wonder系列之间的差距又呈现怎样的态势。

What does that gap look like between the Muse Spark line and the Muse Wonder line.

我们确实看到了一个巨大的提升

And we actually see this huge step up

从Muse Spark到Muse Wonder系列，因为正如Mark提到的，这个新系列背后有一个更好的基础模型。

From the Muse Spark to the Muse Wonder line because of, as Mark mentioned, like, it's just a much better base model powering this new line.

这里面有很多说不清道不明的东西，但我觉得我们确实能够量化这种效果。

And there's an entire je ne sais quoi to that, but I think we've been able to actually quantify this effect.

是的。

Yes.

这显然是我们期待继续改进的一个方面。

And it's an obviously dimension that we're excited to continue improving on.

但如果你说外部基准已经不那么重要了，那你更关心的是你自己的内部基准。

But if you're saying that external benchmarks have ceased to become as important, then you care much more about your internal benchmarks.

这是不是一种反对刷榜的观点？你觉得仅仅在大家谈论的外部基准上拿高分，可能并不能帮助你开发出好的模型？

Is that kind of an argument against bench maxing that you think that just trying to get the top score on the external benchmarks that everyone talks about is maybe not going to lead you to develop a great model?

嗯，是的。

Well, yeah.

但我们也专注于特定的事情。

But also we're focused on specific things.

我的意思是，整个策略的一部分就是个人超级智能，以及尝试构建和别人不同的产品类型。

I mean, part of the whole strategy is personal superintelligence and just trying to build different types of products than other people are trying to build.

所以即使我们和其他人拥有完全相同的评估标准，它……

So even if we did have the exact same evals as everyone else, it...

不。

No.

它只是……

It just...

是的。

Yes.

而且我想说，个人超级智能并不是一个非常小众的兴趣。

And though I would say a personal superintelligence is not like a super niche interest.

它可能是其他模型无法……

Like it's probably something that other models won't I be able

去思考它有很多非常不同的地方。

to think that there are things that are very different about it.

我觉得，个性可能更重要。

I mean, think like the personality probably matters more.

我们经常讨论的一个点就是“谨慎”这个概念。

I mean, one of the things that we talk about is this idea of discretion.

所以如果你想用模型帮你完成事情，它需要了解你很多，同时尽可能少地透露你的信息。

So if you want to like use the model to get something done, it needs to be able to know a lot about you and get things done in the world while disclosing as little as possible.

所以，比如你……

So, you know, if you're...

比如你有过敏，或者你……嗯。

I mean, you have an allergy or you're Mhmm.

怀孕了之类的，你想订个餐厅位子，然后……嗯。

Pregnant or something, you're trying to like get a, you know, like a table at a restaurant and Mhmm.

你不想非得告诉餐厅你的健康状况，对吧，就为了确保那里适合你。

You don't wanna like have to tell them your medical condition you Yes, know, to make sure that it's gonna be a good place for you.

那怎么处理呢？

Like, okay, how do you handle that?

而且，人类有常识，而模型到目前为止，如果你在构建一个编码代理，在每个人都能访问相同上下文的环境中工作，嗯。

And I mean humans have common sense and the model so far, if you're building a coding agent that works in an environment where everyone has access to the same context Mhmm.

谨慎就不是特别重要。

Discretion isn't particularly important.

所以，我觉得对我们来说，这些事情更重要。

So there are all these things that I think are just kind of more important for us.

对。

Yes.

编码的某些方面，代码需要能运行，但使用它的消费者可能根本看不到代码。

Aspects of coding where the code needs to work, but you don't, the consumers who are using it may not even ever see the code.

我觉得这在个人超级智能的背景下可能更重要，而不是服务于软件开发人员。

I mean, that probably matters more in a personal superintelligence context rather than, you know, serving software developers.

所以，嗯。

So Mhmm.

我的意思是，这些事情有点不同。

I I mean, these things are just a bit different.

归根结底，我们关心的是做出有用的模型。

Ultimately, we care about making useful models.

我们希望模型能像——对，完全正确。

We want models when like Totally.

当你把它们交到人们手中。

You put them into people's hands.

他们使用起来会有很好的体验。

They have a great experience using.

我认为这是任何模型最重要的指标，是的。

And I think that is the most important metric for any model that we Yep.

或者其他人生产的模型。

Or other people produce.

我觉得，你知道，真的真的专注于外部基准测试。

I think that that, you know, really really focusing on the external benchmarks.

比如，如果你稍微深入了解一下供应链，就像某个博士生做的这些基准测试……

Like, if you really dig into the supply chain a little bit, like, it's some PhD student who made, you know, these benchmark...

这些基准测试是某个博士生在攻读博士学位期间做的。

Benchmark as part of like their, you know, as the course of them doing their PhD.

所以这相当于把你的产品思考外包给了那个碰巧做出爆款基准测试的博士生。

And so you're kind of outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark.

很明显，这是不对的。

So clearly like that's wrong.

所以这就是为什么我们非常重视……

And so that's why we care a lot...

在Meta和MSL，很多人工作的重点就是打造，对，出色的评估体系。

Like a huge part of what a lot of people at Meta and MSL are focused on is crafting Yeah.

出色的评估体系。

Amazing evals.

现在有了这些，我们就可以自己动手了。

And now that we have them, we can work on our own.

但如果你是在构建一个……

But if you're building like a...

你在搭建自己的实验室，对吧，那些公开的基准测试其实是个不错的起点。

You're ramping up of your lab, right, it's like those public ones are not a crazy place to start.

但那不是你想最终停留的地方。

It's just not where you wanna end up.

你只是……

You're just...

那时候你只是在寻找一些生命迹象。

At that point you're just looking for signs of life.

但当我们真正进入这种状态，就像现在这样，我们想非常刻意地改进我们的模型，以打造越来越好、越来越好的产品。

But then when you're like really in this mode, as we are now where it's like, you know, we wanna be very intentional about how we're improving our model over time to build better and better better products.

就像马克刚才提到的很多案例，这些其实在电子表格里都有非常详细和严格的记录，包括产品的各种小功能和细微差别，我们认为这些对最终用户非常重要。

Like a lot of the cases that Mark just mentioned, like these are actually detailed very rigorously in like the spreadsheet of all the little capabilities and, nuances of the product that we think really really matter the end user.

也许每一项对用户来说只占2%的重要性，但综合起来，正是这种执着最终会让这些模型感觉像魔法一样。

And maybe each of them matters like 2% to the user, but in aggregate, that's like the obsession that will ultimately result in these models that feel magical.

你们能给我描述一下训练一个强化学习模型需要什么吗……

Can you guys describe for me just what goes into training an RL model...

或者说，是前沿的领先模型？

Or sorry, a leading edge frontier model?

我就是好奇一下。我觉得很多听众对训练流程有个大概的了解，知道有预训练和后续训练这两个阶段。

I'm just curious in that I think a lot of people listening to this have a rough idea where there's, you know, there's a pre training and post training steps.

所以预训练就是你要摄入海量的数据。

So pre training is you kind of ingest this huge amount of data.

比如你看大量的图片，这些是猫，那些是狗。看久了之后，你就会慢慢建立起一种感觉，明白什么是猫的特征、什么是狗的特征，还有毛发的质感等等。

You know, you look at lot of pictures and these are cats and these are dogs, and after a while you start to, you know, develop a sense for what is catness and what is dogness and what is fur and, you know, everything like that.

所以这算是预训练阶段，就是把数据全吸进来。你可以讲讲数据获取这块，具体是怎么操作的。

And so that's kind of the the pre training step where you just hoover up data and maybe you can talk about the data acquisition and just kind of what goes into that.

然后是强化学习阶段，这是过去几年AI领域的新进展。听起来这个环节让模型融入了很多具体的细节，你们做得更……

And then there's this RL stage, which is the new ish advance over kind of the past few years of AI, and again it sounds like that is where a lot of the specific nuance of this model has come in, you guys have better...

选了更好的领域，给模型挑了更有挑战性的任务。

Have chosen better domains, better hills to climb for the model.

不过从宏观层面来说，能不能描述一下大概需要多少人、花多长时间、有哪些步骤，以及你们是怎么规划这些环节的？

But just at a high level, describe how many people, how long it takes, what the steps are, how you think about those in

你有多少时间？

How much time do you have?

知道。

Know.

就像……

It's like...

简单来说。

At a high level.

这恰好是我们花费所有

This happens to be like what we spend all of our

时间的地方。

time on.

没错。

Exactly.

所以总结一下……

And so summarize...

YouTube课程之类的？

YouTube courses or...

从很高的层面来看，你知道，就是这几个阶段。

Think at at a very high level, you know, these are the stages.

对。

It's Yeah.

预训练和后训练。

Pre training and post training.

讽刺的是，训练其实是所有这些的总和。

Ironically, training is the super set of all of it.

但那是……

But that's...

但整个这项工作有多少人参与？

But how many people are on this whole effort?

比如当你开发音乐的时候……

Like when you develop music...

所以我觉得你开始触及到MSL的趣味所在，以及我们采取的方法。

So this I think you're getting into like what is interesting about MSL, and the the approach that we've taken.

也就是说，从一开始，我想当我第一次联系你的时候，我不知道我们是不是因为……

Which is, I mean, from the very beginning, I think when I first reached out to you, I I don't know if we're just looking at this because we're like...

就像是……

It was like...

就是，哇。

Just like, wow.

对。

Okay.

我们走了多远？

How how far we've come?

然后我就翻出了我第一次联系Alex时的截图。

And I was just like, I pulled up the screenshot of like when I first reached out to Alex.

我第一句话就是，嘿，你有时间聊聊吗？

My first thing was like, hey, do you have time to catch up and talk?

然后就像，我们聊了

And then it's like, we talked for

四十四分钟左右吧。

forty four minutes or whatever.

然后那次通话两天后，Alex给我发了一封很长的东西，里面说，嗯，如果我是你，我会这么做。

And then then after that call two days later Alex sends me this long thing which is like, well here's what I think I I would do if I were in your if I were in your position.

而且第一点就是，你不需要很多人，人才密度才是关键。

And the very first thing is, that you don't need a lot of people and talent density is the thing that matters.

所以我觉得从计划一开始，这就是你策略的核心。

So I'd say like from the beginning of the plan, this was, like, the core of your strategy.

基本上就是，你其实不需要大量的人，但你需要的是每个人都尽可能优秀，这也就促成了我们那种疯狂的招聘行动，以及非常严格的筛选过程，从Meta内部挖了一些最优秀的人，同时通过非常严格的流程从外部招人。

It's basically, like, you actually don't need a massive number of people, but what you do need is for, like, each of the people to be as good as possible, which kind of informed this kind of crazy recruiting push and, like, very, intense filtering that we had to take, you know, some of the best people from inside Meta and a very intense process to get people from outside Meta.

但我其实觉得，让MSL特别、也让很多人想加入的原因，是因为某种程度上，在一个更小规模的团队里工作反而更有趣。

But I actually think a lot of what has made MSL special and why a lot of people have wanted to join is because I think it's, in some ways, it's just more fun to work on a TalentEnd smaller effort.

感觉就像创业公司，对。

Feels like the startup Yes.

这就是在运营。

That's operating.

从某些方面来说，它既有初创公司的好处，又有大公司的稳定性和资源。

In some ways, has all the benefits of a startup with the stability and resources of a larger company.

我觉得这种环境，就是你计划里的第一点，关于……

And this was like this environment that I think this was like point one in your kind of plan for what...

在亚历克斯说‘这就是我们要做的’之前。

Before Alex was like, this is what we need to go do.

所以我不知道预训练团队应该有多大。

So I don't know the answer of like how big is a pretraining team.

它不需要那么大，但我觉得如果你看看其他实验室，可能规模相当大。所以我认为这些模型表现好的部分原因，是因为我们押注的人，他们有很多资源，因为我们把算力集中起来。

It doesn't need to be that big, but I think if you look at some of the other labs, it's probably quite big in some other So I think part of the reason why these models are doing well is because you have people who we're betting on and they have a lot of resources because we're taking our compute.

我们不是像抹花生酱一样到处分散。

We're not peanut butter.

我们不会把资源分散到很多互相竞争的项目上。

We're to cross a lot of competing efforts.

我们把资源投在关键人才上，每个研究员都有大量的算力，这样他们就能专心做研究了。

We're investing in the key people, a lot of compute per researcher, and then they can do their work.

最开始是不是Wonder在做主要的训练？

Was Wonder the one big training at first?

那是为了……

That was was for

为了……

for...

大部分训练。

Most training.

接下来是预训练。

Was the next pre train.

就像爬梯子一样，我们有分级系统，然后就是Spark。

It was climbing up this ladder, so I mean we have this tiering system and then kind of Spark.

内部来说，Avocado是Muse Spark系列的代号，Watermelon是这个系列的代号，水果都用完了，真的用完了。

I mean internally like Avocado was the code name for the Muse spark line and then watermelon was the code name for this one and we're like out of fruits We're out of fruits.

所以我们得换个新的命名方式了。

So we're gonna have to switch to a to a new convention.

但是，

But,

我觉得从技术上来说，唯一比西瓜大的水果就是榴莲了。

I think the only fruit that is technically larger is the durian.

而且你不想去...

And you don't wanna go...

榴莲比西瓜还大？

Durian is bigger than a watermelon?

我觉得是更重。

I think heavier.

那太吓人了。

That's horrifying.

确实更重。

Definitely heavier.

更扎人。

Spikier.

更扎人。

Spikier.

不。

No.

我觉得，就像马克提到的，你知道，每个……

I think I think as as Mark mentioned like, you know, each of the...

我的意思是，每个阶段……

I mean, of the stage...

每个阶段，我会说在Meta有几十个人在做，而你知道，几百人……

Stages I'll describe like it's tens of people working on it at Meta versus, know, hundreds...

所以Meta是个大公司。

So Meta's a huge company.

这就像Meta内部一个非常小的项目。

This is like a really tiny effort inside of Meta.

我知道。

I know.

研究团队不到200人。

The research team is less than 200 people.

非常非常小。

It's very it's very small.

这是故意保持得非常精简。

It's intentionally very very tight.

我觉得……

I think that the...

其实更接近100人。

It's pretty closer to a 100.

然后还有

And then there's

一群信息相关的人员。

a bunch of info folks.

那工作里确实有很多信息。

Then it's it's a lot of info at work.

所以我觉得从很高的层面来看，是的。

So I think at a at a very high level Yeah.

回到那个问题，就是“它是什么”的问题。

To go back to the to the to the what is it question.

你知道，它是怎么发生的？

You know, how does it happen?

模型是怎么产生的？

How does model happen?

这基本上就是我想问的。

That's basically what I'm asking.

模型是从哪里来的？

Where do models come from?

你拿大量的数据，

You you take a huge amount of data

用于预训练。

For pre training.

然后在这里，你是在……

And here you're you're...

在这里，大致上就是所有公开的互联网内容，还有你能找到的所有书籍，总之就是尽可能多的数据。

And here, like roughly, it's like all the, you know, public Internet and like all the books you can get your hands on and just like as much data as possible

但类型不同。

But different types.

再加上专有的元数据。

Plus the proprietary metadata.

但还有不同的模态。

But but also different modalities.

如果你希望它在感知方面表现出色，那就要确保它有大量的图像、音频和视频之类的数据。

If you want it to be good at perception, then you wanna make sure it has a lot of images and audio and video and this kind of stuff.

你知道，如果你希望它在某些领域表现优秀，不能只在最后用强化学习来弥补。

You know, if you want it to be good at certain domains, you don't just wanna RL that in at the end.

你得从一开始就让它接触到这些。

You wanna have it have exposure to that from the beginning.

在这个阶段，你有点……

And in this stage, you're kind of...

你不太……

You're much less...

你不太在意数据具体长什么样、格式怎么样，或者关于它的一切细节。

You care a lot less exactly what the data looks like or how it's formatted or all of everything about it.

就是大量的数据。

But it's just lots and lots of data.

你说这是最好的个人代理模型，那在预训练阶段，你有没有需要获取的数据，比如大量的餐厅预订电话记录之类的？

When you talk about this being the best personal agent model, was there data you had to get for the pre train where it's like huge amounts of restaurant preservation call book...

我不知道具体是什么，但就是……

I don't know what it is, but like...

你只需要把人们关心的那些模态的大量数据灌进去。

You just kinda wanna dump in large volumes of the modalities that people are gonna care about.

我是说，其实有很多方面。

I mean, there there were a bunch of things.

我觉得我们之前聊Muse Wonder的强项时，好像没提到这个。

I mean, I don't I don't actually think we talked about this when we were talking about where Muse Wonder excels.

但有些领域……

But some of the areas that...

就是，大家关心什么呢？

It's like, alright, what do people care about?

就是……

It's...

比如购物，大家很关心，还有个人建议，大家也很关心。

Mean, there there are things like shopping that people care about a lot, personal advice that people care about a lot.

大家有个人法律问题、健康问题、税务问题。

People have personal legal questions, personal health questions, personal tax questions.

Muse Wonder在这些方面其实挺擅长的。

Muse Wonder, I mean, is like quite good at those.

我不知道我们是否想给出具体的量化数据。

I don't know if there's any quantification that we want to give.

但就是……

But it's...

比如说……

Let's say it's...

在这些个人问题上确实表现很好，是的。

It's very good on those personal Yeah.

所以这类事情，虽然可以在一定程度上进行强化学习，但你更希望直接投入大量数据。

And so those types of things, I mean, you can RL them to some degree, but you like just want to put in a large amount of data.

你可以想象，健康领域是一个非常……

Like you can imagine health is a very...

关于健康的数据非常多。

There's a lot of data out there about health.

在预训练阶段，你只需要看到所有这些不同的内容。

And in pretraining you just wanted to like see all these different things.

所以你就把这些不同的企业端数据放进去。

And so you you kinda just put these different corporate end.

所以这个……

So so the...

就像马克刚才提到的，那个……

So as as Mark was mentioning, like the...

我们……

We...

实际上，Muse Wonder 的一个亮点就是它在各个领域的专业能力。

One of the highlights of Muse Wonder actually is its ability on domain, various domain expertises.

对。

Yes.

而且我觉得我们的理念是，最好的个人助手应该让你感觉像有一支专家团队在支持你。

And I think our philosophy here is that, you know, the best personal agent should feel like a team of experts in your corner.

就像感觉它是你的法律团队一样，对。

Like it should feel like it is your legal team Yes.

还有你的财务团队、医疗团队，等等等等。

And your finance team, and your medical team, and and and and and.

所以在很多这些领域里，我们……

And so, on a bunch of these domains, we...

你知道，Muse Wonder 在很多金融评估和法律评估上，我觉得都是最先进的。

You know, Muse Wonder is actually state of the art on I think, a lot of finance evals as well as a lot of the, legal evals.

它是最先进的模型。

It is the state of the art model.

而且……

And that...

你知道，这些领域的东西，从预训练阶段就开始了。

You know, these these domain areas, that starts in pre training.

那些……

Those...

你知道，我们会找很多领域数据放进预训练里。

You know, we look for a lot of domain data to put into pre training.

对。

Yes.

但预训练的艺术，我是说预训练，有这么一个推文，说预训练充满了那些嚼着玻璃、凝视远方、凝视深渊的研究人员。

But a lot of the a lot of the art of pre training, I mean pre training is, there's this one room tweet where it's like pre training is full of the researchers who like chew glass and stare out into the distance and into the into the abyss or whatnot.

但我觉得这里面有道理，

But but I I think there's a truth to

因为预训练确实是风险最高的研究领域之一。对。

that because it is like, it is both the highest stakes Yes.

因为我们把很大一部分算力投入到训练Muse Wonder，投入到预训练Muse Wonder上。

Of the research areas because you, you know, we invested a huge percentage of our compute into training Muse Wonder, into pre training Muse Wonder.

而且风险非常高。

And, it was very high stakes.

我觉得对。

I think Yes.

可能我们俩在那个过程中的不同阶段都...

Probably both of us at various points in that process...

你经历了悲伤的低谷。

You went through troughs of sadness.

不确定性的低谷。

Troughs of uncertainties.

你知道，悲伤的七个阶段都经历了一遍。

You know, all the seven stages of grief Yeah.

在预训练模型的时候，这些都会发生。

You know, happen, while you're while you're pre training a model.

而且，这在某种程度上是最根本的科学，因为你没有很多样本。

And, it is the most sort of like fundamentally scientific in an interesting way because you know, you just don't have very many samples.

对。

Yes.

所以，你科学地推断到底会发生什么的能力，真的取决于你能不能像大聪明一样，对过程中发生的一切进行推理。

So like your ability to infer exactly what will happen scientifically really depends on your ability to sort of like, you know, big brain reason about everything that's happening in this process.

好的。

Okay.

所以在预训练阶段，所有训练前沿模型的人基本上都在尽可能多地获取数据用于预训练，可能还会有所侧重，比如如果你想让Muse Wonder在法律、金融、税务等问题上表现出色，那你就希望在预训练中包含尽可能多的这类特定语料。

So in the pre training stage, everyone who's training frontier models basically is trying to get as much data as possible for the pre train with maybe a bent towards, if you want Muse Wonder to excel at legal and finance and tax questions and things like this, then you want to have as much of that specific corpus as possible in the pre train.

你仍然需要对模型最终要达到的目标有清晰的意图，但同样，更多的数据只会带来更多好处。

You still want as much intentionality about where ultimately you want your model to be, but again, it just benefits from a lot more data.

这就是那个海量数据、大规模计算的阶段。

And this is sort of the massive data, massive stage.

在预训练中，数据会不会达到某种饱和点，还是说最终数据越多越好？

Do you hit some saturation in the data you put in the tree train or ultimately is it just more data is better?

嗯，如果你想让模型变得更大，就需要更多的数据。

Well, relationship between if you want the model to be bigger, you need more data.

否则，就会过拟合。

Otherwise, over fits.

对。

Yes.

但我的意思是，有没有那种数据，你会觉得，哦，我们不能用那个来训练模型，因为它会分散模型的注意力？

But my point is like, is there data where it's like, oh, we can't distract the model with that.

我们不能用那些东西把它的脑子塞满，还是说数据越多越好？

We can't clutter up its brain attic with knowing that stuff or is it just more data is better?

我觉得确实存在一些这样的效应。

There there is some there are some of these effects, I think.

但这也是科学的一部分……

But this is part of the science....

秘密关闭。

Secrets off.

抱歉……

Sorry....

就像……

Just like...

我的意思是，我觉得有很多这样的问题，比如，好吧，我们放了很多这些进去。

I mean, like I think that there are all these questions about like, okay, well we put a lot of these in.

在很多地方是否存在零和效应，还是说实际上有正向迁移？

Is there any zero sumness in a bunch of places there isn't, or it's actually positive transfer?

但然后，根据它的设置方式，可能会有一些负迁移，而且

But then, like, depending on how it's set up, like, you can have some negative transfer, and

对。

Yes.

所以我不确定。

So I don't know.

这里有所有这些有趣的问题。

There there are all these interesting questions.

但总的来说，扩展更大模型的一个主要挑战不仅仅是关于有多少浮点运算和多少计算量。

But in general, one of the one of the bigger challenges of scaling bigger models is it's not just about how many flops and how much compute.

你还希望数据能够随之扩展。

You wanna be able to scale data with with that.

我的意思是，这是部分原因……

I mean, is like part of what...

当最初的预训练扩展定律被写出来的时候。

When the original kind of scaling laws for pretraining were written.

就是说，从那个研究得出的结论是……

Mean, this is sort of what came out of that is is that.

所以，如果互联网内容增长没那么快，而你又想训练更大的模型来提升智能，我觉得一个有趣的问题就是：数据从哪儿来？

So if the Internet isn't growing that quickly and you want to scale bigger models that can be more intelligent, think one of the interesting problems is, alright, where's the where's the data coming from?

对。

Yes.

所以我觉得……

So I think that's...

这个话题我们大概就聊到这儿吧。

Probably gets to the end of what we're gonna talk about on that.

不过我们可以转到另一个话题……

But we can go to the other...

现在……

Now...

很有CEO的风范，不过……

Excellent CEO energy But...

结束了。

Over

所以，后训练。

So post training.

所以，然后我觉得后训练……

So so then I think post

训练现在算是……

training is kind of...

现在是一个非常令人兴奋的领域，我觉得。

Is a very exciting field now, think.

而且在这个领域，实际上……

And it's an area where I think actually...

我觉得几乎持续不断地，整个领域都在学习如何最优地训练这些模型的后训练阶段。

I think almost on a continual basis, the entire field is learning about the optimal recipes to be able to post train these models.

而且，

And again,

对人们来说……

for people...

后训练就是给模型一个任务，你期望它在这个任务上表现最好。

Post training is where you give the model a task that you expect to do greatest on how it does.

所以如果是一个下棋的模型，它就会下棋，学习……

And so if it's chess playing model, it plays chess games, learns the...

两个阶段。

Two stages.

在SFT之后，就是监督学习。

Post the SFT, so there's supervised learning.

好的。

Okay.

那就像很多情况，你拿这种原始的东西，给它问题和答案，对。

So that's like a lot of when you you take this kind of raw thing and you give it both the the question and the answer Mhmm.

你希望它拥有的那种答案。

That you would want it to have.

对吧？

Alright?

比如说，如果你在聊天，你希望它在这种情况下怎么回应？

If, like, if you're having a chat conversation, how would you want it to respond in this case?

所以当你训练像性格这样的东西时，SFT其实是一个……

So when you're training things like personality, the SFT is actually, an...

塑造它的非常重要的部分。

Very important part for molding it.

但我觉得最近几年的进步大多是在强化学习方面。

But then I think most of the advances in the recent years have been in RL.

所以我觉得有一个有用的思维模型是，SFT就像是初始化……

So so I think I think one mental model that can be helpful is like an SFT is where you kind of initialize...

SFT是监督微调吗？

SFT is supervised fine tuning?

监督微调。

Supervised fine tuning.

你在那里初始化很多你希望模型以后能展现的行为。

Where you initialize a lot of the behaviors that you want the model to be able to exhibit later on.

然后，在SFT结束时，这些行为可能已经内化了，但模型只有大约10%的时间会表现出来。

And then, you know, maybe at the end of SFT, it has those behaviors kind of baked in, but it'll only do it like 10% of the time.

然后在强化学习中，你确实在各种情境下给模型强化，以大幅提高这些行为的触发率。

And then in reinforcement learning, you really sort of give the give the model reinforcement in various situations to bring the elicitation of those behaviors dramatically up.

而且我认为……

And I think that the...

所以，我觉得在RL和后期训练中有很多有趣的事情正在发生。

So so I think there's a bunch of interesting things in that that are happening in in RL and and post training.

我认为其中之一是……

I think one is that this...

RL和后期训练的规模化一直是明确的进展方向之一。

Just the scaling of RL and post training has been has been one of the clear axes.

就是投入更多的计算资源，收集更多的数据……

Just spending more compute and gathering more data...

在发布模型之前，要做更多的这种工作。

Much more of it before you release a model.

是的。

Yes.

要做更多的这种工作，尤其是当预训练规模扩大时，要做更多的……

Doing much more of it, especially if as the pre trained scale up doing much more...

这就像缩放定律里的任何东西一样，必须和其他因素一起扩展，否则就不起作用，对吧？

It's like anything in scaling laws, like it has to go with the rest of the factors or it doesn't work?

总的来说，如果我要给出……

In general, I mean, if I were to give...

比如说一个关于AI的宽泛说法，那就是把一切都规模化，你知道的。

Say like one broad statement about AI, it's just scale everything, you know.

但是……

But...

但特别是，如果你必须扩展一切，那么单独扩展某一样东西是行不通的。

But in particular, if you have to scale everything, like scaling one thing in isolation does not work.

你技术上来说……

You you technically...

你可以。

You can.

你可以只扩展一个维度，但如果同时扩展多个维度，性能会好得多。

You can scale one dimension, you just will get much more optimal performance if you scale many things altogether.

嗯，其实可以这样理解 Muse Wonder，对吧？

Well, mean, so one way to think about like Muse Wonder, right?

就是说，好吧。

It's like, alright.

我们做了预训练。

So we did the pre train.

对。

Yep.

我们也做了一些强化学习。

We've done some RL.

你看，像Muse Spark这样的模型，对吧？

When you look at it like Muse Spark, right?

从1.0到1.1、1.2，嗯，差不多这样。

So version one to 1.1, 1.2, one point Mhmm.

我们在预训练方面没有做任何改动。

We didn't change anything on our pre training.

我们只是在它基础上继续做了更多的强化学习。

We just continued doing more RL on top of it.

所以，你……

So you...

当你想到一个新类别模型的早期版本时，它们实际上相对不那么……

When you think about the early versions of a model of a new class, they're actually relatively less...

就是说，强化学习与预训练的比例，在早期版本中会比后期版本低一些。

Like the RL to pretraining ratio is gonna be lower for RL than the later versions...

差不多吧。

Kind of

模型的小版本更新是什么意思呢……

what the point releases in models are is...

嗯，我觉得每家公司都有自己的惯例。

Well, think every company has its own convention.

对我们来说，就是这个意思。

For us, that's what it is meant.

但对很多公司来说，大版本号是预训练，然后……

But for a lot of companies, the big number is the pre train and then the...

我觉得这个意义已经有点丢失了，对。

I think it kind of lost Yeah.

那种语义上的价值。

That semantic value.

我觉得现在实际上……

I think now you actually...

至少在Meta是这样的。

It's true at Meta at least.

我们其实还没做这个。

Well, we haven't even done it.

我们算是已经决定了……

We've kind of decided...

确实是这样。

It's true

Muse Spark。

Muse Spark.

确实是这样。

It's true

Muse Spark 和 Muse Wonder 是……

that Muse Spark and Muse Wonder are...

而且

And

我们决定先不急着给 Muse Wonder 定版本号，把这个问题往后推一推。

we decided to kick the can down the road on this question for Muse Wonder by not giving it a number yet.

但等我们发布第一个小版本的时候，就必须得做决定了。

But when we have our first dot release we will have to decide.

但是……

But...

不。

No.

我觉得强化学习这个阶段，你知道，一部分原因在于……

I think the RL stage, you know, part of what has been...

你知道，最有趣的一点是，首先要生成环境和任务。

You know, part of what's very interesting is first is just producing the environments and producing the tasks.

而且我觉得，这正是训练这些模型时特别需要用心的地方之一。

And I think, that is one of the areas where there's a lot of intentionality in training these models.

同时，这也是一个你会注意到很多有趣现象的地方。

It's also one of the areas where, you know, you notice very funny things.

比如，在很多情况下，强化学习的数据其实越少越好。

Like if you, you know, in a lot of cases less is actually more when it comes to data in RL.

所以我觉得，这种感觉和科学，怎么说呢，确实是这样。

And so I think I think the the sort of like, the the feel and the science of it, like, is Yeah.

一般来说，和预训练相比，完全是反过来的。

Pretty flipped, generally speaking, from, from pre training.

另外，我觉得它能给模型提供非常放大的指令，你明白我的意思吗？

The other is that I think it can give very magnified kind of instruction to the model, if that makes sense.

所以我觉得，我们非常关心的很多安全或对齐问题，就像，你知道的，当你在前沿模型上做大规模高算力的强化学习时，你会看到，在那些模型只能通过做不对齐或不安全的事情才能成功的场景中，你施加了很多强化学习或优化压力。

So I think that like a lot of safety cases or alignment cases that we care a lot about are like, you know, and you know, where where you would see when you do large high compute RL runs on frontier models are where you apply a lot of RL or optimization pressure in, scenarios where the only way for the model to succeed is to do something misaligned or do something unsafe.

所以，在训练Muse Wonder时，我们还需要非常小心和深思熟虑的另一件事，就是确保我们精心策划了强化学习环境，确保所有环境都在引导模型朝着更对齐的行为发展。

And so that's another thing that we need to be very careful about and thoughtful about in training Muse Wonder is ensuring that we curated the RL environments very thoughtfully to ensure that all of them were listening and driving towards more aligned behavior from the model.

你刚才提到的关于质量的观点，我觉得很重要。

So the point that you're saying about quality, I think is an important one.

因为我觉得之前来回讨论的很多问题都是关于数量的，嗯。

Because I think a lot of the previous questions back and forth were about the quantity and Mhmm.

感觉就像，哦，你需要扩大数量。

Have the flavor of like, oh, you need to scale the quantity.

对。

Yes.

但我觉得，这很大程度上其实只是……

But I mean, I do think like a huge part of this is just...

就是……

Is is what...

你刚才提到的那个点，比如……

The point that you just made of like the...

即使是少量的高质量数据。

Even a smaller amount of high quality.

我觉得在强化学习里，我们……

I think I think in RL, like we...

你知道，从我们新的Sparkline上看到的那些大幅提升，我觉得背后都是对底层方法实实在在的改进。

You know, as we've seen pretty big jumps in our new sparkline, I think that these have been like accompanied with, you know, real meaningful improvements to the underlying recipe.

而这些方法上的改进，每一个都……我觉得从Muse Spark 1到Muse Spark 1.3的进步，简直是天壤之别。

And each of these improvements in the recipe have been like, you know, I think the improvement between Muse Spark one to Muse Spark 1.3 feels night and day.

我觉得如果你去问大多数人，他们会不会觉得这些像是来自不同的预训练模型？

Think probably to most people if you ask them like, are these like, you know, do these feel like they came from different pre trends?

我觉得他们很可能会说是的。

I think they probably would say yes.

因为我觉得你可以

Because I think you can

你知道，那个那个那个……

You know, the the the...

我觉得这在整个行业都在发生，不管怎么说。

I think it's happening industry wide for what it's worth.

就像我们正在学习从这些现有的基础模型中榨取更多的能力，激发出更多的智能体能力。

Just like we are learning to squeeze so much capability and so much, and elicit so much agentic capability from, you know, these existing base models.

这并不是说我们不相信扩展规模很重要。

Now that's not to say like we believe scaling is very important.

这正是我们推出 Muse Wonder 产品线的原因。

That's why we have the Muse Wonder line in the first place.

但是，你知道，这种……

But but you know, this sort of like...

这两件事同时发生，顺便说一句，这也是我们总体上看到这么多进步的原因。

These two things are happening at the same time which is also by the way for for us, this is why we're seeing so much progress in general.

我觉得是因为我们同时在预训练和强化学习配方上以不错的速度进行扩展。

Like I think it's the fact that we're scaling both pre training and scaling our recipes in RL kind of like concurrently at a nice clip.

而且还有其他的扩展维度。

And then there's these other scaling axes too.

就是说，确保你的……

Mean, like making sure that your...

智能体框架是好的，并且你能协同优化这些东西，可能只是……

The agent harness is good and that you can co optimize these things is probably just another aspect of...

扩展的另一个方面或维度。

Or dimension for for scaling.

但是

But