扎克伯格的开源 playbook：Meta 的 Muse Glimmer 与 14 页宣言

Meta又转型了，这是第三次了，现在又回到了开源公司的路线。

Meta has pivoted again for the third time and now back to an open source company.

Meta的故事真是相当疯狂。

The story of Meta has been pretty insane.

两年前，他们完全开源，一心想要搅乱整个闭源模型市场。

Two years ago, they were totally open source, hell bent on wreaking havoc on the entire marketplace of closed source models.

他们通过他们的模型Lama来做到这一点。

They were doing this through their model called Lama.

对于那些不熟悉的人来说，Lama就像一个接近前沿的开源模型，被认为会以有意义的方式改变世界。

And for those who aren't familiar, Lama was just kinda like this close to frontier open source model that was expected to change the world in a meaningful way.

但事情并没有像扎克伯格预想的那样发展，所以他们关闭了所有模型开发，转向闭源，和其他实验室一样。

It did not go as Zuck foresaw, and therefore they went and they closed off all of their model development and went closed source just like all of the other labs.

现在他们又改回去了。

Now they're going back.

他们正在回归开源，为此他们花费了巨额资金。

They're transitioning back to open source, and as a result, they've spent a tremendous amount of money on this.

那么上次财报会议上到底发生了什么？

So what happened in the last earnings report?

在投入了这么多资本支出之后，扎克伯格直接发了一份14页的文件，讲清楚了他为什么要重返开源界。

After all the CapEx spend came in, Zuck just dropped a 14 page paper explaining exactly why they're making a reentry into the open source world.

而这个回归的核心就是一款模型，也就是我们现在要聊的 Muse Glimmer。

That entry is defined by a singular model that we're gonna talk about right now called Muse Glimmer.

Muse Glimmer 这个模型真的非常厉害，能装进笔记本，本地就能跑，我觉得它会改变整个开源AI的生态。

And Muse Glimmer is a really impressive model that fits on your laptop, runs locally, and I think it's gonna shift the world of open source AI.

我觉得这是扎克伯格和Meta的妙招，要知道我可是这档节目里过去可能是Meta最坚定的看空者。

I think this is a great move from Zuck and Meta, and this is coming from probably the Meta's biggest bear in the in the past on this show.

对。

Yeah.

绝对没错。

Definitely.

我们是专业的喷子。

We are professional haters.

我得先说明一下。

I should preface this with.

过去六个月我一直是扎克伯格和Meta的专业黑粉，但我很高兴看到这个转变。

I'm a professional hair of Zuck and Meta for the last six months, but I love to see this pivot.

我会解释为什么。

And I'll I'll explain why.

第一，Meta在AI上的整个理念一直是开源。

Number one, Meta's entire philosophy around AI has always been open source.

他这么做的简单理由是他认为每个人都应该能接触到这个东西，对吧？

And his simple reasoning behind that is he believes like everyone should get access to this thing, right?

如果每个人都能接触到这个东西，就不会有AI的集中权力核心。

If everyone has access to this thing, there'll be no centralized core power for AI.

未来不会有AI霸主。

There's no AI overlord going forwards.

然后他像你提到的那样转变了方向，现在他回来了，带来了不止一个，而是两个新的开源模型。

He then kind of pivoted on that, like you mentioned, and now he's back with not one, but two new open source models.

那这些模型是什么呢？

So what are the models?

第一个，头条是叫Muse Glimmer。

Number one, the headline is called Muse Glimmer.

它是一个开放权重、开源的300亿参数模型。

It's an open weight, open source 30,000,000,000 parameter model.

如果你不太了解这些相对模型的规模，这其实是个相当小的模型。

Now, if you're not familiar with the sizes of these relative models, this is a pretty tiny model.

它实际上可以装进你的MacBook里运行。

It can actually fit and run on your MacBook.

它只需要20GB的数据或内存，这比其他任何前沿模型都要小得多，那些模型都是几百上百GB的。

It's 20 gigabytes of data or RAM that it requires, which is significantly smaller than any other frontier model, which is like hundreds and hundreds and hundreds of gigabytes.

他们能做到这一点的方法是，把一个更大的模型缩小。

And the way that they were able to achieve this is they took a bigger model and they made it smaller.

这叫做量化。

It's something called quantization.

那他们为什么要创建这个模型呢？

Now, why did they create this model?

嗯，他认为，或者说Meta认为，本地运行的私有AI模型应该是未来AI模型普及给每个人的方式。

Well, he believes or Meta believes that locally run private AI models should be the future of how AI models are dispersed amongst everyone.

现在我们付费订阅，才能使用云端模型，这些模型托管在像Anthropic、OpenAI这样的中心化公司里。

Right now we pay subscriptions to get access to a model that lives in the cloud, that lives in a centralized company like Anthropic, like OpenAI.

有了这个模型，你可以在本地运行它。

With this model, you can kind of run it locally.

这带来了几个我们节目里多次提到的好处。

And that brings several different advantages that we've mentioned quite a few times on this show.

第一，它便宜得要命。

The first one being it's cheap as hell.

第二，它可以在你的私有数据上运行，所以它能变得更聪明，更贴合你个人。

The second one being it can run on your private data, so it can become a lot more smarter and attuned to you specifically.

第三，它在你处理不同提示词时，能非常高效地给出答案和结果。

And the third thing is it's very efficient when it serves you answers and results to your different prompts.

所以我觉得这是个不错的模型，但肯定算不上顶尖。

So I think it's a good model, but it's certainly not frontier in any way.

对。

Yeah.

离顶尖还差得远，而且我猜他们可能也不太在乎这个。

It's far from frontier, and I imagine they're probably not really concerned about that.

我们在之前的节目里提到过，大家都在争夺前沿，但现在有不同的前沿。

We mentioned this on previous episodes where there is this fight for the frontier, but now there are different frontiers.

有智能的前沿，Anthropic 和 OpenAI 在搞这个。

There is the frontier of intelligence, which Anthropic and OpenAI are kind of working on.

还有每个 token 的成本和效率的前沿。

Then there's the frontier of cost and efficiency per token.

我觉得 Meta 就是把自己定位在这里。

That's kind of where I see Meta placing themselves.

实际上，他们同时在多个战线作战，而 Muse Glimmer 看起来就处在最底层，因为它是个非常小的模型。

In fact, they're fighting on multiple fronts at once, and it seems like Muse Glimmer is sitting at the very base of that because it is such a small model.

它做了非常深度的量化处理。

It's so quantized.

它真的能在本地机器上跑起来。

It's able to actually run on a local machine.

我觉得这个模型最大、功能最全的版本大概需要55GB的内存，我桌上那台MacBook就能带得动，效果真的很让人印象深刻。

I think the fully capable largest version of this takes about 55 gigabytes of memory, which runs on my MacBook that I have sitting on my desk right now, and it's really impressive.

这让它们进入了一个其他公司很少参与的赛道，就是这种边缘推理和本地计算的方向，我觉得苹果很快也会推出类似的产品。

This puts them in a race that not many other companies are competing on, which is just like that edge inference local compute thing that I think we can expect to see out of Apple fairly soon.

比如，新的 Siri 要跑本地模型了，这感觉就是针对那个情况做出的回应。现在你终于能免费运行一个能力还不错的推理模型了。

Like, with the new Siri gonna be running on local models, this is very much feels like a response to that, where now for the first time you're able to run this free inference that's fairly capable.

它的上下文窗口还挺大的。

It has a somewhat large context window.

我记得是几十万 token 吧。

I believe a couple 100,000 tokens.

具体数字我不太确定，但大概是几十万 token 的上下文。

I'm not sure the exact number, but couple 100,000 tokens of context.

它的输出速度，比如每秒处理的令牌数，表现相当不错，运行水平大概相当于十八个月前的前沿模型。

It has, you know, pretty good outputs in terms of tokens per second, and it runs at a level that you would expect a frontier model would have ran at maybe, say, eighteen months ago.

所以它做不了特别复杂困难的任务，但对于一个能在你机器上运行、随时随地可用、还能放心处理你所有数据的模型来说，已经足够了。

So it's not gonna do any crazy, difficult, complicated tasks, but for something that can run on your machine, you could use anywhere in the world at any given time, and you can trust it to manage all of your data.

比如，如果你的机器上有很多敏感信息，你想用健康数据，但不想交给公开的模型，这真的是一个以前没有过的简单方法。

So for example, if you have a machine, runs a lot of sensitive information, you wanna put health information through that, you might not want to give to a public facing model, this is a really easy way of doing that in a way that we haven't really had before.

我觉得这就是扎克伯格新战略的第一步，可以称之为“焦土2.0”，就是为什么要开源模型？

And I think that's that's kind of step one in this new strategy that Zuckerberg's going for in, I guess, could call it, like, scorched earth two point o, where why do you open source models?

嗯，就是希望大家都去基于它们开发，从而削弱那些前沿实验室的利润空间。

Well, you want everyone to go off, build on them, to remove margins away from these frontier labs.

这似乎是这个战略的第一个版本，大家还挺兴奋的。

And this seems like the first version of that that people are are kind of excited about.

而且你会发现他现在把这种逻辑用在了很多模型上。

And you see him apply this logic to quite a few of his models now.

比如，他也开放了权重，或者计划开放Muse Spark 1.2的权重，如果你一直在关注我们的节目，就知道这是一个全新的模型，Meta的前沿模型，大概一周半前发布的。

Like, he also open weighted or he plans to open weight Muse Spark 1.2, which if you have been following along with the show, was a brand new model, Meta's Frontier model that they released, I think it was like a week and a half ago.

Muse Spark 1.2 特别的地方在于，它是一个非常好的编程模型，他们通过某种方式设计了 AI 代理，让它们互相协作，拿你的代码去并行处理。

And what made Muse Spark 1.2 really special is it's a really good coding model in the way that they've kind of engineered AI agents to kind of like replicate amongst themselves, take your code and go work in parallel.

所以这是一种非常高效的工作方式。

So it's a really efficient way of working.

我觉得这有点像 Meta 发布这些模型时的核心理念。

And that's kind of like Meta's core thesis, I feel, with a lot of these models that they're releasing.

就是说，他们不是要做出最聪明的模型，而是想做出最便宜、最实惠的模型，同时还能完成大约 90% 到 95% 的工作。

It's like, they're not trying to get the most intelligent model, but they're trying to produce the most cheapest and affordable model that still gets about 90 to 95% of the work done.

扎克伯格在过去几个月里多次说过，这是 Meta 的核心哲学。

And Zuck has said in many times over the last couple of months that this is Meta's core philosophy.

他们相信，本地私有运行的高效 AI 模型最终会超越前沿模型。

They believe that locally private run effective AI models will eventually outcompete the frontier.

现在，当你比较这些模型的时候，对吧？

Now, when you kind of compare these models, right?

你听这些的时候可能会想，好吧，谁在乎呢？

You're probably listening to this and you're thinking, okay, who cares?

Meta有一款模型。

Meta has a model.

我都不知道多久没用过Meta AI了。

I haven't used Meta AI intelligence in God knows how long.

我为什么要用这个模型呢？

Why should I use this model?

如果你现在用的是Anthropic或者OpenAI的API，每个月花个几千块，哪怕是一两百美元，如果是企业用户，每年投入几百万甚至上亿美元的话，这款模型可能会让你觉得更划算。

Well, if you are using say the Anthropic or OpenAI API, and you're spending a couple thousand bucks a month, even a couple $100 a month, if you're an enterprise that's spending 10 to hundreds of millions of dollars per year, this might be a model that might be more bang for your buck.

而且说实话，你不可能拿它来处理所有事情。

And listen, you're not gonna use it for everything.

你也不可能指望它帮你规划整个公司的战略。

You're not gonna get it to plan your entire company strategy.

你也不能指望它从零开始开发一个全新的应用，还要求它是最好的。但你可以把它当成主力工具来用，也许再配合一些更智能的模型一起工作。

You're not gonna get it to build a brand new app from scratch and expect it to be the best thing, but you'll use it as a workhorse, maybe in conjunction with some of these more intelligent models.

所以看到扎克伯格重新扛起Meta在开源领域的角色，还是挺不错的。

So it's great to see Zuck kind of like reprise Meta's role in open source.

回想一下，以前是Llama，然后他好像转了方向，大家都挺生气的。我很高兴他现在在做这件事。

If we remember, like in the past it was Lama and then he kind of like pivoted and everyone was like kind of super angry about I'm glad that he's doing this.

但问题是，他为什么要这么做呢？

But the question is like, why is he doing this?

对吧？

Right?

Meta是一家盈利的上市公司，扎克伯格还是想赚大钱，这对Meta来说有什么意义呢？

Like, why would this make sense for Meta, which is a for profit publicly traded company, Zuck still wants to make a lot of money.

这到底怎么讲得通呢？

How does this make any sense at all?

他昨天宣布开放一批模型的时候，基本上把他之前的整套理论都给推翻了。

And he kind of dropped his entire thesis yesterday when he announced that he's open waiting a bunch of these models.

他说，我相信每个人都应该能接触到超级智能。

He goes, I believe everyone should have access to super intelligence.

我还专门写了一篇长文，讲讲Meta在构建那个世界时的理念和价值观。

And I wrote a long piece about Meta's philosophy and values for building that world.

基本上，他描述的这个未来建立在三个核心支柱上。

And essentially like this future that he describes balances on like three core pillars.

第一，他认为每个人都应该被赋予个人能力，也就是获得属于自己的私人AI模型，而不应该集中在特定的前沿AI实验室里。

Number one is he believes everyone should be individually empowered, AKA get access to their own private AI model and that it shouldn't be centered in specific frontier AI labs.

第二，他认为推动人类前进的是发明创造。

Number two, he believes the thing that pushes humanity going forwards is invention.

所以，人类发明全新事物的能力和意愿，应该决定谁能获得这些不同的模型。

So humans ability and intention to invent brand new things should drive who gets access to these different models.

那个人可以是任何人，每个人都可以。

And that could be anyone and everyone.

第三点，他相信一种核心的权力平衡，讽刺的是，这正好是中国那边的对手们一直在倡导的东西。

And the third thing is he just believes in a core balance of power, which is something ironically that adversaries over in China have been advocating for a while now.

所以能看到一个美国前沿AI实验室也提倡这一点，确实挺好的。

So it's it's nice to see like a frontier American AI lab advocate for the same.

确实，但这感觉有点像……我不觉得这是一个很有信心的表态，因为态度变化太大了。

It is, but it feels a little bit like I can't I can't feel like this is a high conviction post because the sentiment has changed so much.

他并不是说，他们一开始是这样，但后来不行了就改了。

It's not like he's like, they started this way, but then they changed this way when it wasn't working.

现在事情顺利了，他们又回过头来开始相信这套。

And now that things are working, they're gonna go back and start to believe this.

感觉他们就是没有一直坚持自己的原则。

It feels like they're just not really standing strong on morals over time.

这可能是猜错了，但我的意思是，他现在说的基本上就是大家争论了很久的那些观点。

And this this could be the wrong guess, but, I mean, he's he's here saying basically the exact arguments that everyone's been arguing for a very long time.

就是说，我们不信任单一的垄断体系来处理这件事。

It's like, we don't trust a single monoculture to handle this.

我们不相信两三家公司的未来能最好地服务人类。

We don't believe that the future of humanity is going to be best served by two to three companies.

我们不相信任何认为AI会淘汰大多数工作的公司会急着比别人更快地开发它。

We we don't believe that any company who thinks that AI will eliminate most jobs would rush to build it faster than anyone else.

这些是我们听了很久的论点，随着公司越来越符合这种模式，我们一遍又一遍地听到这些回声。

And these are the arguments that we've heard for a really long time, and we just get these echoed over and over again as the company starts to fit more of this mold.

说实话，那篇文章挺有意思的。

I mean, essay was interesting.

我更关心的是，这些新文章到底会带来什么样的实际影响。

What I will be most interested in is just seeing what the actual downstream effects are of these new essays.

看起来扎克伯格越来越有信心能搞定这件事，他们应该会推出一些真正能改变人们使用 AI 方式的模型。

And, like, hopefully, it seems like Zuck is getting more confident that they're going to figure this out, that they're going to actually be releasing models that can meaningfully shift the way people engage with AI.

这大概算是个起跑信号吧。

And this is probably, like, the starting gun.

他就说，嘿。

He's like, hey.

新时代，新公司，新风气。

New era, new company, new ethos.

我们打算先这么干着，看看效果怎么样。

We're gonna start running with this and see how it goes.

到目前为止，看起来他们好像还真走上正轨了。

And so far, it looks like they're kind of on their way.

现在情况挺好的。

Like, things are pretty good.

我看了Muse Glimmer，其实我挺想试试这个的。

I was looking at, Muse Glimmer, and I actually I want to try this one.

所以我打算今天就在本地电脑上装一下，试试看。

So my plan is to install this on my local machine today and give it a go.

因为这个模型有个特别的地方，就是对于还在用OpenCLaw的人来说，还记得那个东西吧，这个模型其实是个非常智能的智能体模型，能端到端地完成任务。

Because one of the things that is unique about this one is for people who are still using OpenCLaw, remember that thing, this actually is a very agentic model that runs end to end for task completion.

它用智能体。

It uses agents.

它有工具调用功能。

It has tool use.

它有多步推理能力。

It has multistep reasoning.

它支持多模态输入和输出。

It has multimodal input and output.

看起来这是一个非常强大且性能出色的模型。

It seems like a really strong and powerful model.

确实如此。

Yeah.

我们可以从这里看到它在生成图像方面的表现，以及在基准测试中的成绩，而且它是一个真正能投入使用的模型，这一点我觉得很多其他模型都还没做到。

We could see here kinda how good it is at creating visuals, how good it is in the benchmarks, and it's an actually usable model in a way that I don't think a lot of these others have been.

所以我对那篇14页的论文持保留态度。

So I'm taking this 14 page paper with a grain of salt.

我读了一部分。

I read some of it.

我没把整篇都看完。

I didn't get my way through all of it.

Ejaaz，你说你全看完了。

Ejaaz, know you said you read the entire thing.

我对Meta很期待。

I'm I'm excited for Meta.

就是说，我希望他们赢。

Like, I want them to win.

所以我希望这是他们发布更好模型的新起点，真正重新回到竞争中，因为，你知道，双头垄断总是不太理想。

So I'm hoping this is the new starting point for them to release better models to really get back in the game to compete because, I mean, yeah, duopolies are less than ideal.

我们希望有很多人在不同的领域竞争。

We want a lot of people competing for a lot of different parts of the pie.

如果Meta能站出来这样做，而且是开源的方式，那就太好了。

And if Meta's going to come out here and do it and do so open source, that's great.

而且你得想象黄仁勋坐在那里也很开心，这个开源之王，现在又有另一家美国公司加入了打造真正有竞争力的开源模型的行列。

And you gotta imagine Jensen sitting over here pretty happy too, mister open source king, that now there's another company in the running for making really compelling open source models in The United States.

不过我觉得行动比言语更重要。

I think it's actions over words, though.

扎克伯格说这些话已经有一阵子了。

Zuck has talked this talk for a while.

你说得对。

You're right.

但他也是那个在美国发布前沿开源模型的人。

But he's also the guy that's releasing, like, the frontier open source models in America.

你看看其他所有的AI实验室。

You see all the other AI labs.

这么说吧，我举个例子。

Look, I'll give you an example here.

萨提亚·纳德拉和很多人都在公开支持开源和开放权重运动，但他们自己并没有发布任何前沿的开源模型。

Sachin Nadella, bunch of people signed that they're in favor of open source and open weight movements in general, but they haven't released any frontier open models themselves.

所以我更看重实际行动。扎克伯格确实之前折腾了不少，我也一直批评他这点，但他现在回来了。接下来我会看他的行动，看看他到底拿这个做什么。

So I look at actions over words and Zuck, yeah, he's flailed around quite a bit and I've been a big critique of that, but he's come back and I'm gonna judge him by his actions going forwards and we'll see what he actually uses this for.

从最基础的层面来看，这其实是他采取的一种老派商业策略，就是让市场充斥着基本免费的人工智能服务。

Now, the bare case for this is this is like an old school business move that he's making, which is let's flood the market with intelligence that is basically free.

这样一来，其他前沿实验室那些付费订阅服务的商品属性就被强化了。

So we commoditize the paid subscriptions that other frontier labs are paying for.

这会压低它们的估值、利润率和收入，同时也给Meta争取了更多追赶的时间。

And that brings down their value, that brings down their profit margins, their revenue, and it allows Meta more time to catch up.

那么，这是一项昂贵且有风险的事业吗？

Now, is this a costly and risky endeavor?

是的。

Yes.

Meta已经烧了多少钱来着？

Meta has burned, what is it?

大概300亿美元，就为了招募150名员工来开发一个还不是前沿的模型。

Like $30,000,000,000 in acquiring 150 staff to produce a model that is sub frontier.

如果你是股东，你可能会看着这一切想，扎克伯格，你到底在干什么？

If you were a shareholder, you're probably looking at this and thinking, Zuck, what are you doing?

但如果你看看他在那之外的投入，我觉得他今年到明年上半年，在数据中心资本支出上的花费大概会是2100亿美元。

But if you look at what he's been investing in outside of that, he's I think he's on track to spend, I think, $210,000,000,000 this year through the first half of next year on data center CapEx.

他是英伟达GPU的最大持有者之一。

He is one of the largest holders of NVIDIA GPUs.

他已经囤积了英伟达下一代GPU架构Rubin的其中一批最大规模的库存。

He's secured, like, one of the largest arsenals of NVIDIA Rubens, which is their next gen GPU architecture.

仅次于埃隆·马斯克。

Second only to Elon Musk.

你肯定会纳闷，他打算拿这么多算力来干嘛呢？

You have to question, okay, what's he gonna use all of this compute for?

他可以把算力租出去，就像SpaceX在做的那样，从中赚上一笔。

Well, he could rent it out, which is what SpaceX is doing and make a bag from that.

不过，他大概率也会训练一堆前沿模型。

But he's also probably gonna train a bunch of frontier models.

判断他能不能真正做成这件事，最关键的信号就是新模型的发布节奏。

And the biggest signal for whether he's actually gonna be able to pull this off is model release cadence.

过去大概九十天里，他已经发布了两个模型，接下来的一到两个月内，还会再推出两个。

And in the last, I think ninety days, he's released two models, but he's going to release another two more over the next, I think it's like one to two months.

我记得你在某个推文中提到过，埃隆·马斯克也打算采取同样的策略。

I think you mentioned in some tweet thread, Elon Musk is planning to do the same.

所以我观察到这样一个趋势：那些虽然在前沿智能领域稍微落后，但拥有最多算力和资金的公司，依然能够以迭代的方式陆续发布模型，并且可能比其它一些前沿实验室更快地响应市场。

So I'm seeing this trend where the companies that even though they're a little bit behind in frontier intelligence who have the most compute, who have the most capital are still able to kind of ship models iteratively and maybe respond to the market quicker than some of these other frontier labs.

我知道这听起来挺乐观的，但我确实觉得扎克伯格又回来了，我很期待看看这些模型会是什么样。

Now I appreciate that that's optimistic thinking, but I do think Zuck is back in the game and I'm excited to see what these models are like.

我想玩玩这个 Muse Spark 1.2，还有这个能在笔记本上跑的新版本，看看能捣鼓出什么好玩的东西。

I wanna play around with this Muse Spark 1.2 and this new one that I can run on my laptop and see what fun things we can come up with.

对。

Yeah.

我试着去理解他这么做的逻辑。

I was trying to unpack the reasoning.

我就想，好吧。

I'm like, okay.

他为什么又要这样转型呢？

Why did he make this pivot again?

到目前为止，Meta 的战略看起来有太多的不确定性。

It seems like there's so much uncertainty in the strategy of Meta so far.

而且这个决定的背后似乎确实有一些相当有说服力的理由。

And there there really does appear to be some pretty compelling reasons of that are kinda underlying this decision.

因为当你想到Meta这家公司时，我会想，好吧。

Because when you think about Meta, the company, I was like, okay.

Meta到底是做什么的？

What does Meta actually do?

Meta就是个卖注意力的。

Meta's just an attention salesman.

他们业务最大的部分是广告和通过广告销售。

Their biggest part of the business is advertising and sales through that.

他们所有的服务大部分都是免费使用的。

All of their services are mostly free to use.

如果你想，你可以付费买个蓝勾，但大多数用户都是免费的，而且他们拥有庞大的用户群。

You could pay for a blue check mark if you want, but most of their users are free, and they have this tremendous user base.

所以他们可以向这些人卖订阅，但这些人不习惯，也不太可能愿意每月付20美元来用一个可以免费提供的AI系统，因为你得假设普通用户并不需要前沿级别的智能。

So they can sell a subscription to those people, but those people are not used to and not very probably likely excited to pay $20 a month to use an AI system that can be offered for free because you have to assume the average user doesn't need Frontier class intelligence.

所以，如果模型是注意力，而且你能通过AI来改善核心业务，那这大概就是他们想做的，这也可能就是为什么开源是有道理的。

So if the model is attention and if you can improve your core business through AI, then that's probably what they wanna do, and that's probably why open source makes sense.

是的，我觉得也是。

I mean Yeah.

Meta已经宣布要利用你的AI聊天记录来个性化广告，还希望你用AI来建立一份详细的个人档案，就像josh.md和ejaaz.md那样，记录我们所有的偏好、我们关心的事情，并且实时更新，还会有越来越多的方式收集你的上下文信息。

Meta already announced it wants to use your AI chats to personalize the ads, and it wants you to use the AI to build a large profile on you, to build a josh.md and ejaaz.md that has all of our preferences, all the things we care about, that gets updated in real time, and that has an increasing amount of ways in which you can collect context on you.

如果你要做这件事，那你当然希望这些模型免费、易用、开放，尽可能收集更多数据，让人觉得进入这个游戏毫无阻碍，同时也能慢慢侵蚀前沿实验室的利润空间。如果有人在每月20美元的模型和直接打开Instagram用那个酷炫的新聊天机器人之间做选择，很可能很多人会选那个聊天机器人。

And if you are trying to do this, then you really just want these models to be free and accessible and open to collect as much data as possible, to feel like there is no friction to getting in the game, and then also just to kind of, like, just chip away at the the Frontier Labs margins where if there's someone who is deciding between a $20 model or just going into their Instagram app and using this cool new chatbot that they have, like, chances are a lot of people are going to choose to use the chatbot.

这让他们处于一种比较独特的境地——他们在边缘推理方面某种程度上和苹果竞争，比如苹果即将推出的新版Siri，但他们在广告业务上又在跟谷歌竞争。

And that puts them in a way I guess that's kind of unique to them, where they're they're competing on edge inference in a way somewhat with Apple and what they will be doing with the new Siri, but they're also competing with Google in the ads business.

如果他们在搜索查询方面拥有和谷歌一样多甚至更多的数据，想想看广告能有多定制化，而且不只是广告，还有整个内容创作。

If they're gonna have just as much, if not more data than Google has, when it comes to search queries and, like, think about how custom the advertising can be, and not even advertising, but content creation in general.

如果Meta知道你喜欢什么类型的内容，你如何与某些内容互动，它就能提供非常个性化的档案，为你量身生成内容，这一切都是为了卖你更多广告。

If Meta understands what type of content you like, how you engage with certain types of content, it can serve this very custom profile that allows custom generated content just for you, all in the effort of selling you more ads.

我觉得这其实是

And I think this is

Meta的现实情况是，因为他们一直试图

the reality of Meta is as they've tried

转向硬件领域，他们做了这么多探索性的尝试，但归根结底，核心业务还是卖广告。

to pivot to hardware, they've done all these other exploratory things, it still just comes down to the core business is selling people ads.

这就好比了解用户是谁，把他们和卖家匹配起来，然后从这些交易中抽取佣金。

And it's like understanding who their user is, pairing them with a seller, and collecting commission off of those sales.

我觉得这不过是顺理成章的延伸，他们会通过这种非常轻量的 AI 模型，让这些产品变得更有吸引力、功能更强大。

And I think this is probably just a natural extension of that, and they're going to continue to make these products more compelling and more capable through these very lightweight AI models.

说到底，Meta 显然是一家追求利润的公司。

At the end of the day, Meta is obviously a for profit company.

他们就是想赚钱，广告正如你所说，是他们的核心业务。

They're trying to make money and ads, as you said, is like their main business.

所以，毫无疑问，他们会把重心放在这上面。

So, you know, they're definitely gonna focus on that.

不过我也觉得，在更广阔的AI领域里，最近发生了一些变化，是我们不能忽视的。

But I also think in the wider market of AI, a bunch of things have shifted recently that we just can't ignore.

第一点就是，你并不需要最聪明的模型来包办所有事情。

The first one being you don't need the most intelligent model to do anything and everything for you.

我最近和一个朋友聊天，他在一家顶级公司，大概前100名吧，职位挺高的。

I was catching up with a friend who is pretty high up at a top, I would say, 100 company.

我问他，你现在怎么用AI？

And I asked him like, you know, how are you using AI these days?

你用私有模型吗？

Like, do you use private models?

还是用Anthropic和OpenAI的标准模型？

Do you use the standard models from Anthropic and OpenAI?

他说，我现在完全转向了一些开源模型，还有更便宜的模型。

And he said, I've shifted completely to some of these open source models as well as just cheaper models in general.

我用它们处理大部分工作，然后用Claude或者GPT 5.6来头脑风暴，构思我想构建的东西的架构。

And I get it to do the bulk of my work whilst I use a Claude or GPT 5.6 to brainstorm kind of like the architecture of what I want to build.

我在很多企业里都看到了这种模式。

And I'm seeing this pattern reflected across a number of different enterprises.

社交媒体上每天都在报道这个。

It's certainly been reported across social media, like every single day.

我们看到公司正在转变想法，不再盲目地在尖端智能上花大笔钱。

We're seeing the shift of companies thinking, I'm not gonna just needlessly spend a bunch of money on frontier intelligence.

有些东西我肯定还是会用，但我也想探索一些更便宜的选择。

I certainly will for some things, but I wanna explore some of the cheaper options.

对吧？

Right?

而且你可以看看那些前沿实验室自己的商业模式，对吧？

And you can look at the business models of the frontier Labs themselves, right?

我觉得在过去九十天里，已经发布了八个模型。

Like I think over the last ninety days, eight models have been released.

光OpenAI就发布了三个，对吧？

Three of them alone from open AI, right?

他们为什么不直接发布GBT 5.6 Sol呢？

Like why didn't they just release GBT 5.6 Sol?

为什么要发布Luna？

Why did they release Luna?

他们为什么还要发布 Terra？

Why did they release Terra?

一个月前，他们为什么要把 Luna 的价格砍掉八成？

Why did they slash the price of Terra, sorry, of Luna by 80 percent a month ago?

还有昨天，Anthropic 好像宣布说，他们会把 Claude Sonnet 维持在有史以来最低的价格，而且长期不变。

Why did Anthropic yesterday, I think announced that they're gonna keep Claude Sonnet, I believe at like the cheapest price ever indefinitely.

我觉得这些公司已经意识到，更便宜的模型确实实用，而且对于他们想要推销的企业客户或普通消费者来说，这能带来明显的优势。

I think these companies are realizing that cheaper models are definitely applicable and have an advantageous incentive for the customers that they're trying to sell to, whether it's enterprises or consumer customers.

而且你需要花得更值。

And you need more bang for your buck.

扎克伯格意识到了这一点，他正在抓住这个机会，但在开源方面采取的手段更加激进。

Zuck realizes this and he's seizing the opportunity, but he's being more aggressive with the open source side of things.

我觉得Meta有足够的资金，能让这列火车继续跑上一阵子。

And I think Zuck has enough capital at Meta to be able to keep this train going for a while now.

所以我挺好奇接下来几个月Meta，特别是SpaceX会是什么样子。

So I'm interested to see how, how, what Meta and specifically SpaceX looks like over the next couple of months.

我觉得SpaceX也在预告他们的下一个模型，过去三周里已经发布了两个，希望今天能正式推出。

I think SpaceX is also teasing their next model and they've released like two over the last like three weeks today, which should hopefully be releasing.

不过我觉得他们肯定会搞出点什么厉害的东西来。

But yeah, I think there's something that they're gonna pull off something cool.

对。

Yeah.

我也很期待。

I'm looking forward to it.

我很期待这些新模型会衍生出多少智能体，因为如果你在用智能体的话，肯定知道我们得聊聊谁，那就是本集赞助商 Ledger。他们为智能体提供了一套基础设施，主要分三个部分。

I'm looking forward to all the agents that they're going to spawn from these new models because if you are working with agents, you know you know who we gotta talk about, and that's Ledger, our sponsor of this episode, because they have this infrastructure for agents in which it's kind of three parts.

智能体先提出请求，然后由人类批准，最后由 Ledger 签名器来执行。

Agents propose, then the human approves, and then the ledger signer enforces.

Ledger 签名器通常以硬件钱包的形式出现，你可以用它来批准或拒绝操作。

The ledger signer is this mostly, it comes in the form of a hardware wallet in which you can approve or deny.

我们经常谈到安全问题，以及一个

We talk a lot about security and how a

很多这样的代理可能会

lot of these agents can kind

失控，做一些你可能不希望它们做的事情。

of go rogue and do things you may not want.

Ledger 提供了一个代理栈，这是一系列开源工具，可以让你评估这些代理在整个过程中的行为。

Ledger offers an agent stack, which is a series of open source tools that allows you to evaluate what these agents are doing across the way.

你可以让它做一些事情，比如重新平衡你的钱包，或者调整你想要调整的任何参数，然后它会提出建议，你批准，然后 Ledger 签名器强制执行。

You could tell it to do things like rebalance your wallet or adjust whatever parameters you want to adjust, and then it will propose you approve, and the ledger signer enforces.

它支持 Cloud Code、Codex，以及你日常使用 AI 的所有地方。

It works with Cloud Code, Codex, all of the places that you work with AI on a daily basis.

所以非常感谢 Ledger 赞助本集的这一部分。

So thank you so much to Ledger for sponsoring this part of the episode.

你可以在下方描述中的链接获取他们的产品。

You can get their products at the link in the description below.

Ejaaz，你提到了资本支出。

Ejaaz, you mentioned CapEx spend.

或者你其实没提到资本支出。

Or you actually didn't mention CapEx spend.

你提到了更强的盈利，但我要提一下

You mentioned stronger earnings, but I'm gonna mention

我提资本支出。

I CapEx spend.

对。

Yeah.

也就是他们的自由现金流。

Which is, their free cash flow.

比如，我们有

Like, we have

要谈谈

to talk about

他们的盈利，因为，我们确实想稍微谈谈经济方面。

their earnings quickly because, like, we do wanna get a little bit economic here.

比如，这个策略到底靠不靠谱？

Like, how viable is the strategy?

他们今年的自由现金流同比暴跌了91%，说白了就是进账和出账的差额大幅缩水。

Their free cash flow this year, which means basically the money that comes in and out, is down 91% year over year.

他们在花钱。

They're spending.

而且是大手笔地砸钱在几个关键支柱上。

They're spending big on a few key pillars.

其中一个是押注开源，这不会立刻带来投资回报，毕竟他们把这些模型免费开放给大家用。

One of them is this open source bet, which does not have an immediate ROI because, I mean, again, they're offering these waits for free.

另一个则是他们的硬件平台。

The other is on their hardware platform.

有一件事让我至今都觉得挺惊讶的，而且我不确定很多人知道这一点：Meta占据了整个AI眼镜市场70%到80%的份额。

One of the things that is, like, still shocking to me, and I'm not sure that many people know about this, Meta owns 70 to 80% of the entire AI glasses market.

虽然这不是一个特别大的市场，但确实是个有意义的市场，很可能不该被忽视。

And this is not a big market, but this is a market that that is meaningful and that probably shouldn't be overlooked.

这完全是和一家叫陆逊梯卡的公司合作完成的。

This is done exclusively with a company named Luxottica.

他们两家合作，在2025年这一年里，就把Meta的AI眼镜销量翻了三倍多。

Together, they have more than tripled Meta's AI glasses sales just in the year of 2025.

我们还没拿到2026年的数据，但这一点还是挺值得注意的。

We haven't gotten twenty twenty six numbers yet, but, like, that is kind of noteworthy.

有意思的是，资本支出的指导下限也提高了。

It's interesting to see the floor of the CapEx guidance being raised too.

2026年，Meta预计要花1450亿美元，这个数字真是大得让人难以想象。

In 2026, Meta is guiding for a $145,000,000,000 of spend, which is just this, like, unfathomably large number.

当你想想这些钱会花在哪里，那肯定是开源。

And when you think about the places that it's gonna land, well, it's certainly open source.

肯定是硬件。

It's certainly hardware.

但，还有什么呢？

But, like, what else?

他们真的打算全力投入建设数据中心和GPU，来训练更大的模型吗？

Are they really just gonna go all in on building out data centers and GPUs to make larger models?

这就是最终目标吗？

Is that is that the end game?

因为我觉得他们并不打算在尖端模型上竞争。

Because, like, I don't see them competing on the frontier.

那他们真的需要那么多算力来竞争开源、开放权重的模型吗？

So do they really need that much compute in order to compete for open source, open weight models?

我觉得是的。

I think so.

我认为决定你AI产品好坏的第一要素就是算力。归根结底，马斯克会用SpaceX证明这一点，而扎克伯格也打算这么做。

I think the number one factor that determines how good your AI offering is is compute At the end of the day, Elon's gonna prove that with SpaceX, and I think Zuck is planning to do the same thing.

我们昨天看到英伟达或黄仁勋宣布，他为未来的算力锁定了额外的5000亿美元收入和资本投资。

We saw NVIDIA or Jensen Huang announce yesterday that he secured an additional $500,000,000,000 of revenue and capital investment for future compute.

我们看到OpenAI与英伟达合作，金额也差不多，5000亿美元。

We see OpenAI partnering with NVIDIA for a similar amount, $500,000,000,000.

你看到Anthropic昨天跟Riot还有一家老式的密码学挖矿公司达成协议，就是为了拿到更多的算力。

You see Anthropic securing deals with Riot yesterday with an old cryptographic mining company to secure even more compute.

算力是你能不能做出更好模型的最终决定因素。

Compute is the ultimate indicator on whether you can create a better model.

这是AI竞赛里所有竞争者都绕不开的一件事。

It is the one thing that is consistent across all competitors in the AI race.

我觉得扎克伯格相信，如果他囤够算力，拿到足够多前沿NVIDIA GPU的配额，他就能搞出更好的模型。

And I think Zuck believes if he amasses enough of it, if he secures enough allocation of frontier NVIDIA GPUs, he'll be able to pull off and build better models.

至于开源策略能不能奏效，还是我之前提过的那个策略：用商品化的智能淹没市场，削减竞争对手的收入和利润。

Now, whether the open source strategy plays out, again, it's that strategy that I think I mentioned earlier, which is flood the market with commoditized intelligence, cut down or cut into revenue and margins of your competitors.

对。

Yeah.

Meta的资本支出短期内会很难看，但如果长期来看，当所有精准代理都跑在Meta的基础设施和消费者应用上，那这笔账就划算了。

Meta's CapEx takes a beating in the short term, but hopefully that plays better in the long term when you have precise agents that all run on Meta's infrastructure and consumer apps.

别忘了，Meta现在有大约40亿用户。

Also remember, Meta has like 4,000,000,000 consumers right now.

我昨天刷Reels了，虽然不太想承认，但现在好多都是Meta眼镜的内容。

I was scrolling on reels yesterday, which I don't like to admit, but a ton of them are now Meta glasses.

我不知道你有没有看过这个，Josh，就像是从不同人视角拍的。

I don't know if you've seen this Josh, but like it's just POV from like a bunch of different people.

我觉得他们的硬件不是最好的，但他们的分发能力绝对是数一数二的。

And I think that they don't have the best hardware, but they have one of the most amazing distributions ever.

想想40亿用户。

Think 4,000,000,000.

再对比一下苹果，现在大概有35亿台活跃的苹果设备。

And then you compare that to Apple, which has, I think, 3,500,000,000 live Apple devices right now.

可能还要多得多。

It's probably a lot more.

我觉得那个数据已经过时了。

I think that data is outdated.

他们在这里确实有一个真正的护城河。

They have a real competing moat here.

再看看谷歌，最近已经出现了负现金流。

And I look at like Google who has recently gone into negative cash flow.

Meta 还没到那一步。

Meta isn't quite there yet.

所以可以说他们有点保守。

So you could argue that they're being somewhat conservative.

他们还有其他业务在赚很多钱，可以支撑一段时间，但他们的跑道不是无限的。

They have other businesses that are earning a ton of money that can support this for a while, but they don't have infinite runway.

而且，你知道，也许我们会看到 Meta 像英特尔和谷歌在过去几个月那样筹集外部资金。

And, you know, maybe we'll see Meta raising external capital similar to how Intel and Google has over the last couple of months.

我不知道。

I don't know.

我们拭目以待。

We'll see.

对。

Yeah.

我刚想问的是，这笔资本支出是不是生死攸关的。

And I was gonna ask the question whether or not this CapEx spend is existential.

比如说，如果他们又失败了会怎么样？

Like, what happens if they they fail again?

因为这不是他们第一次尝试做出真正好用的前沿模型，而第一次的结果并不理想。

Because this isn't their first time trying to get a frontier model that works really well, and the first time did not go as planned.

然后我又想，你知道吗？

And then I was thinking, well, you know what?

至少他们有GPU，可以直接转卖出去，做个好生意。

At least they have the GPUs, and they could just go resell those and make a great business out

靠这个。

of it.

所以现在市场上有个很奇怪的现象，就是砸钱建数据中心非常划算。

So there's this really weird part of the market now where it's very advantageous to just spend money on building data centers.

就算你自己不知道怎么用，也可以把它们卖给下一个最合适的人，靠这种基础设施建设能赚一大笔钱。

And even if you can't figure out how to use them yourself, you just sell them off to the next best guy, and you can make a tremendous amount of money from that infrastructure build out.

所以我觉得至少，这个想法能给你很大的安慰。

So I think at the very least, you could find a lot of solace in that idea.

关于用户规模，那确实是个很好的观点。

In terms of users, like, that's a great point.

我的意思是，40亿用户就是半个地球的人了。

That I mean, 4,000,000,000 users is half of the planet.

全球没几家公司——我不确定有没有公司——用户数量能跟Meta一样多，最多也就那么寥寥几家。

There's not many companies in the world I'm not sure if there's any companies in the world with that quite as many users as Meta, or there's a very small handful of them.

所以如果他们能想办法——我意思是，用“武器化”这个词不太好听，但就是能利用他们的人工智能工具，在社交媒体平台上推出更有吸引力的产品，那会释放出巨大的价值。

So if they're able to figure out how to I mean, weaponize is like a terrible word for this, but, like, able to leverage their AI tools to further create compelling offerings on their social media platforms, like, that's a pretty big value unlock.

所以我很期待。

So I'm I'm excited.

我对Meta持谨慎乐观的态度。

I'm cautiously optimistic on Meta.

不过我也还在观察，看他们是不是真能找到出路。

Like, I'm still looking for signals that they're going to figure it out.

但就像你提到的，看Reels的时候，很多内容都来自那副眼镜。

But like you mentioned, when you're watching Reels, a lot of it are coming from the glasses.

这是一种全新的内容创作方式。

It's this novel form of creating content.

这是其他所有公司都在努力追赶的新模式。

It's this cool new paradigm that every other company is working for.

而现在，怎么说呢，随便你怎么评价。

And right now, I mean, say what you want.

反正他们是领先的。

Like, they are in the lead.

也许不是世界上最好的硬件，但他们有目前市面上最好的。

It's maybe not the best piece of hardware in the world, but they have the best one available.

至于公司未来会怎样，我们拭目以待，但这就是目前的进展。

And we'll see what happens with the company, but that's kind of the update.

Meta又要开源了。

Meta is going open source again.

那帮人又回来了。

The boys are back.

而且他们花了大把的钱，显然铁了心要做最好的开源模型，并且把它们应用到整个产品线里，同时还在继续打磨他们的硬件产品和那些奇怪的科学项目。

And they're spending a lot of money, and they are very clearly hell bent on making the best open source models that they can and implementing them across the entire product stack while continuing to iterate on their hardware products and weird science projects.

比如，还记得那个瞳孔扩张追踪器和那个脑电波扫描仪吗？他们用这些来了解内容对你的影响，Meta的未来会挺疯狂的。

Like, remember the the pupil dilation tracker and, like, the this brainwave scanner that they're using to, like, understand how content affects you, and it's gonna be a weird future with Meta.

但他们依然在全力推进，毫不减速。

But they're working on it nonetheless, full speed ahead.

是的，我们会继续跟着这段旅程报道下去。

And, yeah, we'll just continue to cover it as we go along this journey here.

另外，我就直说吧。

Also, I will just say this.

如果你在听这期节目，而且你是个开发者，那现在真是最好的时代。

If you're listening to this episode and you are a builder, there has never been a better time to be alive.

过去几个月里，这些工具、模型、产品接连发布，频率越来越快，全都是给你去试用和体验的。

Like, all these tools, all these models, all these products that are being released over the last couple of months that are being increased in frequency are all for you to trial test out.