扎克伯格与 Alexandr Wang：Muse Wonder、千亿美元押注与 MSL 的重启

John Collison

说说 Muse 的近况吧。看第三方报告的话，这可能是迄今增长最快的智能体或 AI 应用。你们自己最喜欢的 Muse 用法是什么？

Give me the update on Muse. If you were looking at the third-party reports, you would see that this might be the fastest-growing agent or AI app that I think has existed yet. What are your guys’ personal favorite Muse use cases?

Alexandr Wang

我最喜欢的是：我的 Muse 叫 Oiler，它不停地从越来越冷门的卖家那里给我淘越来越冷门的茶。现在它每周都会从台湾或日本的某个茶园买点茶，运回来给我。

My favorite thing is I have my Muse, whose name is Oiler, constantly source increasingly obscure teas from increasingly obscure vendors of tea. At this point, on a weekly basis, my Muse is buying some random tea from some farm in Taiwan or some farm in Japan and sourcing it over and bringing it back.

Mark Zuckerberg

我最喜欢的用法，大概也是这个产品最好的评测，是让它当我的 MMA 教练。我在健身房装了摄像头，我在里面练的时候它看着画面，在回合间隙给我反馈。这其实很难，因为回合间只有 1 分钟，它得足够快地处理完所有视频。我老是跟它说：你花了 30 秒才给我反馈，我只有 60 秒休息，你得更快，优化你自己。我觉得这很能说明未来的走向：其实每个人都会写大量代码，只是自己不知道在写代码。

My favorite both use case and probably eval for the product is having it coach me in MMA. I put cameras in my gym, it watches the cameras when I’m in there, and it gives me feedback in between the rounds, which is a very hard problem, it turns out, because it has to process all the video quickly enough; you only get 1 minute in between the rounds. And I’m constantly telling it, look, you took 30 seconds to give me that feedback, I only have a 60-second break, you gotta go faster, optimize yourself. It’s a good example of how stuff is going to go in the future, where basically everyone is actually going to be writing tons of code. They’re just not going to know that they’re writing code.

Mark Zuckerberg

我们确实用大量视觉感知数据训练了模型。这涉及我们的一个核心思路：我们是端到端地把模型训练成好的个人智能体。为什么 Meta 要自己做这些模型？首先，超级智能整体上会让我们把公司运营得更好；其次，对我们做的产品来说，某些能力组合就是更适用。Alex 从一开始就强调要把多模态做好，因为人们用这些服务做的事情里，多模态占的分量格外重。

We did train the models with a lot of visual perception data. This gets into a core approach thing that we have, which is we’re training the models end-to-end to be good personal agents. So why would Meta want to build these models? First of all, superintelligence overall is going to allow us to run the company better. But for the products that we’re building, there’s certain capabilities and certain mixes of skills that are just much more applicable. This is one of the things that Alex focused on from the beginning: we really want to make sure that our multimodal program is good, because that’s disproportionately important for the types of things that people use all these services for.

Alexandr Wang

大多数人体验世界的方式是高度多模态的：和人说话、听东西、看视频、自己拍东西。今年早些时候业界甚至有一种张力，因为 AI 行业的共识几乎是「一切都是编程，只有编程重要」。但像 Muse 这样的产品、很多消费场景能不能成功，取决于多模态是否够强。

Most people in the world experience the world in a highly multimodal way. They like talking to other people and listening to things and seeing videos and filming things of their own. And especially earlier this year, there was actually a point of tension almost in the industry, because so much of the consensus view within the AI industry was that it was just all about coding and coding was all that mattered. For the success of a product like Muse and a lot of these consumer use cases, it really depends on the ability to be strong and multimodal.

Mark Zuckerberg

我们从早期就谈到一个规律：产品越来越薄，模型越来越成为产品本身。我们内部的说法就是「模型就是产品」。做智能体时，Muse 里很多复杂的产品工作，是围绕如何处理你的数据和接入的东西的隐私与安全模型，这部分我们很自豪；但能力、个性这些，是训练进模型里的。Nat 在想怎样让产品感觉好时，他谈的是对模型做后训练。

One thing that we talked about from early on is that there is this pattern where the products are getting thinner and the model is getting to be more of the product. One of the things that we say internally is the model is the product. When you’re talking about an agent, a lot of the complex product work that went into Muse is around the privacy and security model for handling your data and the things that get connected to it, and that’s something we’re very proud of. But the capabilities, the personality, a lot of the stuff you’re training into the model. When Nat is thinking about what is the key stuff we need to make the product feel good, he’s talking about post-training the model.

Alexandr Wang

用户能很直观地感觉到用 Muse 久了的体验：一部分是它处理某些任务的可靠性，另一部分是它如何主动、在什么情境下主动，以及它跟你说话时你的感受。有意思的是，这些产品决策怎么写成规格？最终是模型设计决策，归结为训练方法、数据、评测这些东西。

There’s very viscerally, from a user, a feeling of using Muse over time. Some meaningful portion of that is just the reliability of Muse to handle certain tasks. But another meaningful portion is how is it proactive, in what context is it proactive, and when it talks to you, how do you feel about how it talks to you. The funny thing is, how do you specify these product decisions? It’s ultimately in model design decisions, which ultimately boil down to training methods or data or evals.

Alexandr Wang

Muse 没有模型选择器，也没有推理强度选择器。很多 AI 产品是由技术背景很强的小群体设计的，他们心里没有一个随机抽样的普通美国人。有些模型的命名太离谱了，有时编号倒着走，有时干脆换成另一套词。不仔细读 API 设计文档的人怎么知道该选哪个？

Somewhat infamously, a lot of the AI products have been designed by a highly technical group that do not have this sort of randomly sampled American in their mental model. The naming of some of the models is so egregious. Sometimes the numbers go backwards and sometimes it’s just a totally different word. How is somebody who’s not carefully reading API design specs supposed to know what they’re supposed to pick? It’s just far too confusing.

Mark Zuckerberg

Muse Wonder 是站在前沿的新一级模型。做预训练时你不会直接上最大的集群，而是沿着阶梯往上爬，每次训练大约大四五倍的模型，不断重复。第一批模型 Muse Spark 都基于去年年底那次预训练，说实话它并不大，但团队往里塞进的智能和能力让我们都很惊讶。之后我们一直在做下一代，预训练完成有一阵了，一直在做后训练，现在它来了。它大得多，你会在它思考的细腻程度上感觉到。我们重点把它做成最好的个人与职业智能体模型：很强的记忆，以及在复杂任务上坚持到底的能力。

Muse Wonder is a new class of model that is at the frontier. When you’re doing these pre-trains, you don’t just go to whatever the biggest cluster is that you have. You want to work your way up the ladder training models that are maybe 4 or 5 times bigger, and you keep repeating that process. The first set of models, Muse Spark, was all based on this initial pre-training that we did at the end of last year, and we’ve all been quite surprised and impressed with how much intelligence and capability the team was able to build into a model which is frankly not that big. The pre-training for this next model has been done for a bit, we’ve been working on the post-training, and now it is here. It is significantly bigger, so you will feel that in terms of the nuance in which it can think. A lot of the areas we focused on were making it the best personal and professional agentic model: very good memory, very good ability to persist through complicated tasks.

Mark Zuckerberg

众所周知，我们对 Llama 4 的走势不满意，于是基本重建了整个团队。团队得先做出不同层级的模型才能走到这一步。现在我们站上了前沿，往后做的一切都是推前沿，而不是补之前那些背景功课。我认为接下来每一代更大模型之间的间隔会开始缩短。

Quite famously, we were not happy with the trajectory that we were on with Llama 4. We rebuilt the whole team. The team basically had to go build up different tiers of models to get to this one. And now that we are at the frontier, everything that we do going forward will just be pushing the frontier, as opposed to the background work we needed to do before. I think the time between each subsequently bigger class of models will now start to compress.

Alexandr Wang

这个系列会有两个模型：Muse Wonder 和 Muse Wonder Full，后者是高端得多的版本。按我们的测量，它在一系列个人与职业智能体评测上是 SOTA。几个亮点：一是主动记忆，我们认为要做个人智能体，记忆是明确的前提，这方面我们的投入比很多其他团队都要多，内部有 Memory Bench 评测。记忆指的是：它能多好地记住你几天、几周、几个月前提到的事，并在之后的任务里用上。

We will have 2 models in this series that are going to be available: Muse Wonder and Muse Wonder Full, which is a much higher-end model. We’ve seen incredible performance across a slate of evaluations on personal, professional agent tasks. It’s state-of-the-art based on our measurement. Active memory is one area where we think that to build personal agents, memory is one of the clear requirements, and it’s one of the areas where we’ve invested disproportionately compared to many of the other efforts. We have our own internal evals, Memory Bench. In the experience of using the model, it’s how well does it remember things that you mentioned to it over days, weeks, or months, and then leverage those in the process of doing future tasks.

John Collison

所以这不在外挂框架里，而是在模型本身？

It’s not a harness, it’s in the model itself?

Alexandr Wang

对，在模型本身。它把记忆当成一个核心系统来意识，知道如何调用各种来源完成之后的任务，也知道怎样最有效地做上下文压缩。当然还有很大提升空间，我们的评测也显示了这一点。第二是主动专注，基本就是在很长时程里完成整件任务、遇到挫折时以对齐的方式找别的解法的能力。第三是领域专长。另外编程能力有大幅提升，它在 DeepSWE 和 SWE-Marathon 上是 SOTA，在 Terminal-Bench 4 等基准上也很好。用它写代码的用户在我们的用户里是少数，但他们会感受到很大的提升。

Yes, it’s in the model itself. It’s aware of its memory as a core system, how it’s able to leverage various sources to conduct tasks in the future, as well as how it should do compaction most effectively. To be clear, there’s still plenty of room to improve; our evaluations show this. Active focus is another, which is basically the ability to conduct complete tasks over very long horizons, and when faced with various setbacks along the way, to identify other solutions in an aligned way. The third one is domain expertise. And it’s a big step up in coding capabilities: it is state-of-the-art on DeepSWE and SWE-Marathon and does very well on other benchmarks like Terminal-Bench 4. Users who use it for coding will be a minority of all of our users, but the ones who do will see a big step up.

Mark Zuckerberg

因为它大得多、服务成本更高，在 Muse 里我们会先主要放在付费订阅里。我们希望 Muse 本身对所有人免费，用 Muse Spark 做模型有大量免费 token；但如果你想用 Wonder 做大多数难任务，至少目前要用付费版。希望不久之后能有一个和 Wonder 一样好、但是 Spark 体量的模型，能服务全球数亿人。

Because it’s a lot bigger and more expensive to serve, we’re going to start off primarily in paid subscriptions in Muse. What we want to do with Muse as a product is for it to be free for everyone, so we have this huge amount of free tokens with Muse Spark as the model. But if you want to use Wonder for most of the hard tasks, at least for now you’ll use the paid version. We hope at some point not too far from now to have a model that is as good as Wonder but in the Spark form factor, and that will be a great thing to serve hundreds of millions of people around the world.

Alexandr Wang

它也会通过 Meta Model API 和 OpenRouter 提供，Muse Code 用户也能用，Muse Wonder Full 则面向 Muse Code 的订阅用户。

It’ll also be available via API, on the Meta Model API and OpenRouter, and for Muse Code users. And Muse Wonder Full will be available to subscribers of Muse Code.

Mark Zuckerberg

关于基准：基准必须不断更新，它们饱和得太快了。早期我们在做「爬阶梯」的科学时，更看重外部基准，确保方向没错；之后每发布一次，我们看的内部评测占比都更大。

The benchmarks have to keep moving, because they just get saturated so quickly. Early on, when we were just doing the science of trying to climb the ladder, we cared more about what the external benchmarks were, just to make sure we were on the right track. Each release that goes by, the larger share of what we’re looking at are internal ones.

Alexandr Wang

我们在 Muse Wonder 的研究博客里量化了一个指标，叫「大模型味儿」，它和「刷榜感」基本是反义词。我们的量化方式就是分布内表现与分布外表现的差距：模型在 RL 和后训练分布内的任务上表现多好，和分布外任务比差多少。从 Muse Spark 线到 Muse Wonder 线，这个指标有巨大的跃升，因为底座模型好得多。

One of the things that we quantified in our research blog post for Muse Wonder is this metric for big model smell. Big model smell and the feeling of benchmaxing are kind of inverses of one another. The way we quantify that is the gap between in-distribution performance and out-of-distribution performance: 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. We see this huge step up from the Muse Spark line to the Muse Wonder line, because it’s just a much better base model powering this new line.

Mark Zuckerberg

个人超级智能有很多不同之处，比如个性更重要，还有「分寸感」：你要让它办成事，它得非常了解你，同时在外面办事时尽量少透露信息。比如你有过敏或者怀孕了，要订一家餐厅，你不想把自己的身体状况告诉餐厅。如果你做的是编程智能体，大家共享同样的上下文，分寸感就不太重要。

There are things that are very different about it. The personality probably matters more. One of the things we talk about is this idea of discretion. If you want to 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. If you have an allergy or you’re pregnant and you’re trying to get a table at a restaurant, you don’t want to have to tell them your medical condition. If you’re building a coding agent that works in an environment where everyone has access to the same context, discretion isn’t particularly important.

Alexandr Wang

如果过度关注外部基准，你顺着供应链往下挖，会发现那往往是某个博士生读博时做的一个基准。你等于把所有产品思考外包给了一个碰巧做出爆款基准的博士生，这显然不对。所以 Meta 和 MSL 很多人的重心是打磨出色的评测。我们有一张表，详细列出我们认为对终端用户真正重要的每一项小能力和细节，也许每项只占用户体验的 2%，但加起来，这种执着最终会造就感觉神奇的模型。

If you really dig into the supply chain of external benchmarks a little bit, it’s some PhD student who made a benchmark in the course of doing their PhD. So you’re kind of outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark. Clearly that’s wrong. A huge part of what a lot of people at Meta and MSL are focused on is crafting amazing evals. These are detailed very rigorously in the spreadsheet of all the little capabilities and nuances of the product that we think really matter to the end user. Maybe each of them matters like 2% to the user, but in aggregate, that’s the obsession that will ultimately result in these models that feel magical.

Mark Zuckerberg

我第一次联系 Alex，就是问「有空聊聊吗」，聊了 44 分钟左右。两天后 Alex 发来一长篇，标题大意是「如果我是你，我会这么做」。第一条就是：你不需要很多人，人才密度才是关键。这决定了后来疯狂的招聘和极其严格的筛选。我们不把算力像抹花生酱一样摊到许多相互竞争的项目上，而是投给关键的人，让每个研究员有大量算力。

When I first reached out to Alex, my first thing was like, hey, do you have time to catch up and talk? We talked for 44 minutes or whatever. Two days later, Alex sends me this long thing, which is like, well, here’s what I think I would do if I were in your position. And the very first thing is that you don’t need a lot of people, and talent density is the thing that matters. That informed this crazy recruiting push and very intense filtering. We’re not peanut-buttering our compute across a lot of competing efforts. We’re investing in the key people, a lot of compute per researcher, and then they can do their work.

Mark Zuckerberg

我们有分层体系。内部 Avocado 是 Muse Spark 线的代号，Watermelon 是这一代的代号，水果名用完了，得换命名规则。研究团队不到 200 人，更接近 100 人，另外还有一批基础设施的人。

We have this tiering system. Internally, Avocado was the code name for the Muse Spark line, and Watermelon was the code name for this one. We’re out of fruits now, so we’re going to have to switch to a new convention. The research team is less than 200 people. It’s closer to 100, and then there’s a bunch of infra folks.

Mark Zuckerberg

预训练也要看不同类型、不同模态的数据。要擅长感知，就得有大量图片、音频、视频；要擅长某些领域，不能只在最后用 RL 灌进去，要从一开始就让它接触。人们很在乎的有购物、个人建议、个人法律、健康和税务问题。

Different modalities. If you want it to be good at perception, then you want to make sure it has a lot of images and audio and video. If you want it to be good at certain domains, you don’t just want to RL that in at the end. You want it to have exposure to that from the beginning. There are things like shopping that people care about a lot, personal advice, personal legal questions, personal health questions, personal tax questions.

Alexandr Wang

我们的理念是，最好的个人智能体应该让你感觉身边有一支专家团队：你的法律团队、财务团队、医疗团队。在不少这样的领域里，Muse Wonder 在很多金融和法律评测上是 SOTA，这从预训练就开始了。预训练是风险最高的研究环节：我们把很大比例的算力投进了 Muse Wonder 的预训练，过程中你会经历悲伤的七个阶段。它也是最根本的科学，因为样本非常少，你能多准确地推断会发生什么，取决于你能多深入地推理整个过程。

Our philosophy is that the best personal agent should feel like a team of experts in your corner: your legal team, your finance team, your medical team. On a bunch of these domains, Muse Wonder is actually state-of-the-art on a lot of the finance evals as well as a lot of the legal evals, and that starts in pre-training. Pre-training is both the highest stakes of the research areas, because we invested a huge percentage of our compute into pre-training Muse Wonder, and all the 7 stages of grief happen while you’re pre-training a model. And it is the most fundamentally scientific, because you just don’t have very many samples. Your ability to infer what will happen really depends on your ability to big-brain reason about everything that’s happening in this process.

Mark Zuckerberg

扩大模型规模的一大挑战不只是 FLOPs 和算力，你还得让数据同步扩大，这就是最初预训练扩展律的结论之一。如果互联网增长没那么快，而你想训练更大、更聪明的模型，数据从哪来就是个很有意思的问题。

One of the bigger challenges of scaling bigger models is it’s not just about how many flops and how much compute. You want to be able to scale data with that. That’s what came out of the original scaling laws for pre-training. So if the internet isn’t growing that quickly and you want to scale bigger models that can be more intelligent, one of the interesting problems is where’s the data coming from.

Alexandr Wang

后训练有两个阶段。一个思路是：SFT（监督微调）负责初始化你希望模型日后展现的行为，SFT 结束时这些行为也许已经内化了，但只有 10% 的时候会出现；RL 则在各种情境下给模型强化，把这些行为的出现率大幅拉高。RL 和后训练的规模化是明确的一条轴，尤其是随着预训练变大。如果要对 AI 说一句话，那就是：什么都要扩大。

One mental model that can be helpful is that SFT, supervised fine-tuning, is where you initialize a lot of the behaviors that you want the model to be able to exhibit later on. At the end of SFT it has those behaviors baked in, but it’ll only do it like 10% of the time. Then in reinforcement learning, you give the model reinforcement in various situations to bring the elicitation of those behaviors dramatically up. The scaling of RL and post-training has been one of the clear axes, especially as the pre-trains scale up. If I were to say one broad statement about AI, it’s just scale everything.

Mark Zuckerberg

拿 Muse Spark 来说，从 1 到 1.1、1.2、1.3，我们完全没改预训练，只是在上面继续做更多 RL。所以一个新级别模型的早期版本，RL 与预训练的比例比后期版本低。Muse Spark 和 Muse Wonder 是不同的预训练。Muse Wonder 暂时没给编号，等第一个小版本时再决定。

With Muse Spark, version 1 to 1.1, 1.2, 1.3, we didn’t change anything about pre-training. We just continued doing more RL on top of it. So the early versions of a model of a new class have a lower RL-to-pre-training ratio than the later versions. Muse Spark and Muse Wonder are different pre-trains. And we decided to kick the can down the road on this question for Muse Wonder by not giving it a number yet. When we have our first dot release, we will have to decide that.

Alexandr Wang

RL 有意思的一点是，首先要生产环境和任务，这里有很多刻意设计；而且在 RL 里，数据往往少即是多，手感和科学与预训练几乎反过来。另一点是，它会极度放大给模型的指令。很多我们关心的安全与对齐问题，恰恰出现在前沿模型的大算力 RL 中：如果你在某些场景施加大量优化压力，而模型唯一的成功方式是做出不对齐或不安全的事。所以训练 Muse Wonder 时我们非常小心地挑选 RL 环境，确保它们都在引导更对齐的行为。

In the RL stage, part of what’s very interesting is first just producing the environments and producing the tasks, and that’s one of the areas where there’s a lot of intentionality. In a lot of cases, less is actually more when it comes to data in RL, so the feel and the science of it is pretty flipped from pre-training. The other is that it can give very magnified instruction to the model. A lot of safety or alignment cases that we care about are where you would see, in large high-compute RL runs on frontier models, a lot of optimization pressure applied in scenarios where the only way for the model to succeed is to do something misaligned or unsafe. So in training Muse Wonder we curated the RL environments very thoughtfully to ensure that all of them were eliciting and driving towards more aligned behavior.

Mark Zuckerberg

不同实验室对对齐的理论很不一样。对齐是个很哲学的概念：对谁对齐？我们给人的是智能体，你让它做事，就要高度确信它会做你让它做的事，而不是给你一通说教。如果你想做灾难性的事，它当然要得体地、不惹恼你地拒绝。总体上我们把对齐看成对它所服务的那个人的指令与价值对齐，这和一些其他实验室的思路很不一样。

Different labs have very different theories. Alignment is a very philosophical concept: who do you want it to be aligned to? We’re trying to give people agents. If you ask it to do something, you need to have a high confidence that it’s going to do what you ask it to do, with high reliability. You don’t want a moralizing lecture for why you’re wrong for asking. Obviously, if you’re trying to do something that’s catastrophic, it needs to gracefully, without pissing you off, not do it. But in general we view alignment as more of a question of instruction and values alignment with the person who it’s intending to serve. It’s very different from how some of the others do.

Alexandr Wang

Muse Wonder 从去年 8 月就开始规划了。我们不是做完 Muse Spark 才开始做 Wonder，Spark 一直是通往 Wonder 的垫脚石。

We’ve planned Muse Wonder since last August. It’s not that we started working on Muse Wonder when we finished Muse Spark. We really viewed Muse Spark as a stepping stone to get to Muse Wonder.

Mark Zuckerberg

衡量一个实验室的算力实力，可以看它单台最大的超级计算机有多大，能不能跑大规模训练。我们在俄亥俄建了 Prometheus 集群，规模超过一吉瓦；现在在路易斯安那建 Hyperion，会扩到 5 吉瓦。这些规划早于 MSL，大约在 2023 或 2024 年就开始了。

Part of our philosophy is that you can measure the compute strength of a lab by how big its largest single supercomputer is for doing these large training runs. We built this Prometheus cluster in Ohio that’s scaling to larger than a gigawatt, and now we’re building this Hyperion cluster in Louisiana that’s going to scale to 5 gigawatts. Those actually predate MSL. It was probably 2023 or 2024 when we started making the plans to build those clusters.

Alexandr Wang

因为一切都要花这么久，过去一年里有很多时候，你得在还没有任何证据证明东西能不能成之前，就做巨额投资和重大决定。这有点像高难度模式的创业，和「精益创业」正相反：精益创业是先做出能跑的东西再加倍下注，AI 里不能这么干。

Because of how long everything takes, there were many points over the past year where you have to make very large investments or very big decisions long before there’s proof points around whether or not the thing is even working. In some sense it’s entrepreneurship on hard mode. It’s the opposite of Lean Startup. Lean Startup is you see if you get something working and then you double down. You cannot do that in AI.

Mark Zuckerberg

去年年底我们的对话是这样的：我们对进展挺兴奋，但还没见到模型，我们要签一张一千亿、两千亿美元的支票，我问「Alex，你确定这能成吗？」

At the end of last year, we’re like, all right, we are quite excited about the progress, we haven’t seen the models yet, and we’re going to write this check for, I don’t know, $100 billion, $200 billion. I’m like, Alex, are you sure this is going to work?

Alexandr Wang

去年年底整个高管团队在夏威夷聚会，Susan 好几次把我拉到一边说：我只是想告诉你，如果这不成，从财务上说会非常非常糟。当时除了我做一份研究进展的「氛围汇报」、给大家看几张图，确实完全不清楚能不能成。

We were in Hawaii at the end of last year, the whole exec team got together, and Susan pulled me aside many times and was like, I just want to tell you, if this doesn’t work, it’s going to be very, very bad in financial terms. At the end of last year, other than me giving a vibe report on how our research was going and showing some plots, it was certainly not clear at all that it was going to work.

Mark Zuckerberg

可以算一笔经典账：运营一个吉瓦级集群一年，大概 150 亿美元量级；如果算建设的资本开支，因为要预先付清，大概 500 到 600 亿美元。而这只是训练的一个集群。研究不只是跑那一次训练，大概同样多的投入花在每个改动都要沿阶梯一路测到大规模上，然后还有推理服务。更复杂的问题是：产品还不存在，你却要提前好几年决定为它建多少推理算力。去年有一刻不少人说「也许该保守点」，我说不，我们就这么干，要让它成。

You can do the classic math. To operate a gigawatt a year, call it on the order of $15 billion. If you’re talking about building the CapEx, you pay for it upfront, so it’s probably closer to $50 or $60 billion. But that’s just one cluster for one part of training. Research isn’t just doing the run; probably an equal amount is because for each change you make, you’re testing all the ladders up to the big scale. And then there’s serving. Some of the more complicated questions were how much product demand we think we’re going to have, because the product doesn’t exist, and we have to make it years in advance. There was a moment where a bunch of people were like, maybe we should be a little more conservative, and it’s like, no, we’re doing this. We’re going to make it work.

Mark Zuckerberg

不过我其实不觉得风险有这么大。第一，如果 AI 整体成功，很多人都想要我们的算力，按每吉瓦 150 亿美元的成本，我们能卖得更多。所以要让它失败，不只是我们失败，而是 AI 整体失败。即便那样，我们也有很多别的用途能把算力用起来。而且我们相信自己在做的东西：先是 ChatGPT 这样的文字助手，然后 Anthropic 把编程智能体做成了，却没人做个人消费版的智能体，这看起来就是 Meta 应该做的。

I don’t actually think it’s quite as risky as this. The numbers are big, but there are 2 things. One is if AI works overall, plenty of other people want all our compute. If the cost is $15 billion a gigawatt, we could get a lot more for that. So in order for it not to work out, it’s not just that we have to fail; AI overall has to fail. And even if that happens, we still have all these other uses to productively use compute. But also we believed in what we were building. First there’s ChatGPT as the text-based assistant, then Anthropic got coding agents to work, and no one had built the personal consumer version of that. It just seemed like this is the thing that Meta should build in the world.

Mark Zuckerberg

我认为每一代科技公司里通常不会有一家通吃。互联网时代有五六家很大的公司，再加整个生态。下一个时代很可能只有少数几家公司有能力打造前沿超级智能，它们会主导议程。如果 Meta 想继续做一家为人们发明新东西的公司，我看不出不端到端自建超级智能怎么做得到。别人好像都执着于自动化知识工作，那其实并不是最有感召力的未来；我们的信念之一是，超级智能的最高用途是发明和创造，而不是自动化。

In every generation of technology companies, there’s usually not one company that wins everything, and I don’t think that’s how this is going to go either. With the internet there were 5 or 6 very big companies and a whole ecosystem. In this next era there are probably going to be a small handful of companies that have the ability to build frontier superintelligence, and those companies are going to effectively drive the agenda. If we want to be an interesting company that is inventing new things that matter to people in the future, I don’t see how you can do that if you don’t have the ability to build end-to-end your own superintelligence. Everyone else seems to be obsessed with automating knowledge work, which isn’t even that charismatic a version of the future. A big part of our belief is that the highest use of superintelligence is invention and creation, not automation.

Mark Zuckerberg

我们在模型里训练了很强的浏览器操作能力，这样谁都不需要为 Muse 做任何集成。然后我们按重要程度逐个找合作伙伴，愿意合作的就直接给 API 接入，更高效也更快。长远看，就像今天每家企业都有网站、社交账号、邮箱，每家企业也都会有智能体。现在打电话要求深度集成的企业远比试图屏蔽的多。

We trained into the models very good browser use, so no one has to do anything to integrate with Muse. Then we’re going down the list of the most important partners, and for people who want to work with us, we’ll get direct API access, which is way more efficient and faster. In the fullness of time, just like today every business has a website and social media accounts and email addresses, every business is going to have agents too. We have way more businesses calling us asking to integrate deeply, and very few who have tried to block it.

Alexandr Wang

Muse 是第一批人们真的把购买完全委托给智能体的产品之一，这是一个阶跃。对用户的价值一是减少摩擦，二是条件能自动应用，比如「我周三前要收到」，智能体就只找周三前能送到的。长期我们最兴奋的是更深的个性化：Muse 对你这个人、你生活的各个方面有整体的上下文，能结合你朋友圈子的情况帮你挑最合适的礼物。以往的推荐系统都是割裂的。

Muse is one of the first products where people are actually fully delegating purchases to the agents, and that’s a step change. One part of the value prop is just reduced friction. Also, criteria can be automatically applied: hey, I need something by Wednesday, so the agent will only look for things that could get to me by Wednesday. The long-term thing we’re quite excited about is greater personalization. Muse has this holistic context into you as a person and all aspects of your life, so it can help you get the exact right gift for your best friend in the context of your friend group. Historically a lot of these recommendation systems are quite siloed.

Mark Zuckerberg

我 3 岁的女儿很喜欢烘焙，我们养成了一个习惯：我在用还没完全上线的 Muse 语音模式，它能区分说话人。一周里我们聊周末要烤什么，每到周六，Muse 会确保我从 Instacart 订的食材准时送到。这个周末它知道 Watermelon 要发布，就往购物袋里加了个西瓜，因为原本的订单金额不够起送。另一件有意思的事：以前我一年大概用两次 Instacart，为了给 Muse 用专门注册了账号，现在每周用一两次。所以机会不只是转化率提高 20%，而是我成了新的 Instacart 用户。那个账号到底是给我的，还是给 Muse 的？整个商业的可触达市场会不会大幅扩大，因为智能体也要有账号了？

My 3-year-old daughter loves baking. I’ve been playing with Muse voice mode, which isn’t fully rolled out yet; it segments the voices, so it knows which one is me. Throughout the week we’re talking about what we’ll want to bake that weekend, and every Saturday when it’s time to bake, Muse makes sure I get the ingredients from Instacart and they show up on time. This weekend it knew that Watermelon was launching, so it threw a watermelon in the bag; it deduced there was not enough in the order to get the delivery. Here’s another interesting thing: I didn’t really use Instacart that much before, maybe twice a year. I went and got an Instacart account to give Muse, and now I use it like once or twice a week. So the opportunity isn’t just that it’s going to be 20% more efficient on the conversion rate. I’m a new Instacart subscriber. Is that Instacart account really for me, or is it for the Muse? Is the total addressable market for all of commerce just increasing dramatically, because now agents are effectively going to have accounts?

John Collison

（节目口播）Meta 的模型，包括 Muse Wonder，都在 OpenRouter 上和数百个其他模型一起提供。Stripe 已同意收购这家领先的模型网关，OpenRouter 目前每天路由超过 10 万亿 token。

Meta is building increasingly capable models, including Muse Wonder, and their models are available on OpenRouter alongside hundreds of others. Stripe has agreed to acquire OpenRouter, who’s the leading model gateway. OpenRouter already routes more than 10 trillion tokens a day.

Mark Zuckerberg

语音为什么一直难做？一是你希望它真的去办事；二是对话动态：人说话不是「我说完、等你说」，我停 3 秒它会不会以为我说完了？它开口后我插话，它会不会当成新问题？三是语音输入带宽很高，但听别人说话来获取信息效率很低，不如视觉。所以理想界面大概不是纯语音对语音。

There are a few things. One is you actually want it to go do stuff. Then there’s a lot about conversational dynamics. It isn’t typically, I will give a prompt, and now I will wait, and you will speak. If I pause for 3 seconds, is it going to think I’m done? If it starts speaking and I cut back in, is it going to think it’s a whole new question? And then for voice overall, it’s quite high bandwidth for you to communicate, but in terms of consuming information, having someone speak to you is very inefficient compared to having it be visual. So the ideal interface is probably not just strictly voice-voice.

Alexandr Wang

语音输入已经在爆发，连编程都是，大多数研究员就是对着键盘小声说。大学生用 AI 的方式更极端：连续 10 到 15 分钟意识流地倾倒想法，让 AI 去整理、归纳、结构化。另外，Muse 之所以能有一个「替你干活的小家伙」的具象形象，是因为它同时具备持久记忆、能替你采取行动、还有一点个性，这些汇合在一起才让这种具象化成立。

There very clearly already is an explosion of voice input, even in coding. Most of the researchers just kind of whisper into their keyboard, because it’s the highest-bandwidth or lowest-friction way to get your thoughts out. The most extreme version is how college students now use AI: they will completely word vomit for 10, 15 minutes, full stream of consciousness, and have AI do all of the organization and structuring. And the fact that Muse is embodied, this little guy that you feel is working on your behalf, is only possible as a coherent representation of the product because it has persistent memory, the ability to take actions, and a little bit of a personality. It’s the confluence of all these things.

Mark Zuckerberg

后来发现眼镜才是真正的载体，很多人即使没有显示屏或只有小显示屏也想要。我们开始出货带显示屏的眼镜，虽然不是大视场的完整增强现实，小屏幕也很强大。之后会有完全多模态的交互：语音进出，它能看到你看到的，还能「画出像素」。我们刚出的一个小设备 HomeLink，硬币大小、带 USB，插上后你的 Muse 就能连上家里所有东西，比如 Sonos 和灯光。

It turns out a lot of people want the glasses even with no display or a small display. We started shipping the display glasses that aren’t the wide field of view full augmented reality, but even the small display is pretty powerful. You’ll start getting interactions that are fully multimodal: voice in and out, it’ll be able to see what you see, and it’ll be able to paint the pixels. One of the gadgets we just shipped was the little HomeLink. It’s the size of a nickel but with a USB, so you can connect it and your Muse can connect to everything in your home, like your Sonos and all your lights.

Alexandr Wang

我们还推出了 Muse Gadgets，让生态去尝试各种设备形态。发布以来，已经有好几家手机厂商来找我们，说要把 Muse Gadgets 集成进手机试试。

The other exciting thing we’ve done here is Muse Gadgets, which is unleashing the ecosystem to try cool device formats and form factors. Since releasing that, we’ve gotten a number of phone manufacturers talking to us about integrating Muse Gadgets into their phones.

Mark Zuckerberg

我们希望 Muse 被广泛使用，因为规模上会有网络效应，合作伙伴也会随着用户增多更愿意集成，最好的办法就是让它负担得起、免费可用。但收入有很多来源：订阅会是一大块；电商方面，我们相信 Muse 会帮人赚钱、省钱，长期可以从中抽一小部分；还有 API；我们成立了 Meta 企业平台，请 CJ Desai 来负责。Meta 传统上不是企业服务公司，但我们认为最难的是先做出领先的通用模型，之后投入建一支企业团队，很多企业会想用 Muse Wonder。再加上我们本来就是很赚钱、信用很好的公司，能为这些大投资兜底。

What we’re trying to do with Muse is make it widely used, because there are going to be valuable network effects at scale, and partners are going to want to integrate more the more people are using it. The best way to do that is to have it be affordable and free. But there are going to be a lot of ways to scale revenue. Subscriptions is going to be a big one. Around commerce, we believe Muse is going to make people money and save people money, and over time there are lots of ways to take a small cut of that. We’ll also have the API. A big part of why we started the Meta Enterprise Platform and had CJ Desai join us is that, traditionally Meta hasn’t been an enterprise company, but a huge part of the challenge is building the leading models in the first place, and they’re very general. Coupled with the fact that we were already starting off as a very profitable business with good credit that can underwrite these large investments, that puts us in a pretty good position as a lab.

Mark Zuckerberg

另外，我们用模型加速核心业务。核心业务规模是数千亿美元而且仍在快速增长，用这些模型让它再快几个百分点，就能为很大一部分投入买单。我们自建数据中心也是一个财务优势：别人向云付年费，而 GPU 折旧完后我们大概还会继续用，现在一块 H100 的价值甚至比去年还高，因为算力依然紧张。

The other thing is that we use the models to accelerate our core business. You have this core business that’s hundreds of billions of dollars and still growing extremely quickly, and if you can grow it a few percent faster by using these models, that ends up funding a huge amount of this too. And building out the data centers is a financial advantage for us compared to companies that are paying clouds the annual rate. At some point the GPUs will be fully depreciated, but we’re probably still going to be using them, because the value of even an H100 at this point is higher than it was last year, if compute continues being constrained.

Alexandr Wang

原因很多。一是执行：每家前沿实验室都有执行强和执行弱的时期，时间一长就平均掉了。如果每家领先时都执行完美，大概能一直领先。二是蒸馏的威力，这特别帮了开源生态。没人计划让蒸馏这么好用，但事实就是可以通过某种蒸馏自然获得大量能力，这对整个生态是好事，意味着能把很强的智能装进越来越小的模型。三是很多人本来就想要一个不是「资本主义最后只剩一家公司」的未来，这也是很多人来 Meta 和 MSL 的原因：他们想去一个以把这项技术交给数十亿人为目标的地方。

One certainly is just execution. For any given frontier lab, there are periods of strong execution and weak execution, and that starts to average itself out over time. If all players, when they were in the lead, executed perfectly, they would probably be able to stay in the lead. The power and strength of distillation is certainly another factor that’s benefiting the open source ecosystem quite a bit. It’s not like anyone planned for distillation to work super duper well, but it turns out you can get a lot of capabilities naturally through some form of distillation, which means you can pack very strong intelligence into smaller and smaller models. And then a lot of people in the ecosystem want this future where it is not just one final company in all of capitalism. That’s one of the reasons why a lot of people were excited to join Meta and MSL: I want to work at a place that is focused on giving this technology to billions of people.

Mark Zuckerberg

我们认为好的未来是通过赋能人们来实现繁荣。对于很多重要的安全问题，我们的看法是：靠多样性和制衡，让很多不同的东西都对人们可用，而不是只有几个好模型掌握在几家公司手里。在我看来，后者对未来是更有风险的格局。

We think a good future is one where you get prosperity by empowering people. For a lot of the important safety topics, our view is that the way to get there is having diversity and checks and balances, by having a lot of different things out there and accessible to people, not just having a couple of good models that are only in a couple of companies’ hands. From my perspective, that is probably a riskier configuration for the future than lots of people having access to different things.

Mark Zuckerberg

我们是 2025 年 4 月 4 日左右发布 Llama 4 的，大概 4 月 20 或 21 日，我开始打电话向人求教该怎么办。一开始就像超新星一样，把一群人硬凑在一起，头几个月我们俩都没怎么睡。我几乎清空了一两个月的日程，只见研究员、跟他们聊。我和所有实验室的负责人、所有研究负责人都谈过，然后我很确定：Alex 就是那个人。

We launched Llama 4 on April 4th, I think it was, 2025. And then I think it was April 20th or 21st when I started calling people for advice on what to do. In the beginning it was sort of like the supernova thing, where we just jammed a bunch of people together, and I don’t think either of us slept that much for the first couple of months. I kind of cleared my calendar for a month or 2 and was just meeting researchers and talking to them. I talked to all the leaders of all the labs, all the heads of research, and I was like, no, Alex is the guy.

Mark Zuckerberg

一开始我们并不确定要「完全重启」。是在和很多顶尖的人谈的过程中意识到，为了最有吸引力，它应该给人一种「从零重新开始」的感觉；当然我们也会把 Llama 和 FAIR 的很多优秀的人拉进来。这个决定对实验室的文化影响巨大。我想的是，这个领域会不断演变，能把实验室管好的人，重要的不是他是不是最新某种训练技巧的专家，而是他是否思路清晰、能像创业者一样动态更新。

We weren’t sure upfront that it was going to have to be a full reboot. That started when we were out talking to a lot of the best people, and we realized that in order to make this as attractive as possible, it should feel like we’re starting this thing again from scratch. We’ll pull in a lot of the great people from the Llama effort and FAIR too. That ended up being really smart and made a huge difference in the culture of the lab. The field is going to keep evolving, so the important thing for someone who’s going to run a lab well isn’t whether they’re an expert in the latest modality or training technique. It’s whether the person is a clear thinker and able to dynamically update, as an entrepreneur would.

Alexandr Wang

我刚加入时开的第一批会里，我说第一件事是在某个时间点前训练出一个很好的推理模型，这显然是最重要的里程碑。结果原团队一位关键负责人（现已不在 Meta）说：我们的消费者需要推理模型吗？我觉得不需要，做个普通模型就行。我心里一沉。那件事说明公司领导层需要对 AI 更「上头」、更前倾。后来扎克伯格整个人转向 AI，大家看到后很快跟上，比如负责 Instagram 的 Adam Mosseri，他自己的 vibe coding 和个人小应用配置是我见过最讲究也最高效的之一。

One of the first meetings I had when I joined, I was in this meeting with some of the leads of the existing effort, and I was like, the first thing we need to do is train a really good reasoning model by some timeframe; this is clearly the most important milestone. And one of the key leads in the group, who no longer works at Meta, was like, do any of our consumers need a reasoning model? I feel like we don’t need to do a reasoning model. And I had this heart-sinking moment. It was an indicator that we needed leadership in the company to be much more AI-pilled and forward-leaning on how quickly the technology’s moving. As Mark started pivoting the whole company, a lot of the existing leaders very quickly picked it up. Adam Mosseri, who runs Instagram, has one of the most aesthetic but also effective personal vibe code and personal mini app setups of anyone that I’ve seen.

Alexandr Wang

组建 MSL 时，公司投入巨大，我们却明确告诉大家：没有截止日期，按正确的节奏来。但因为聚集的都是很厉害的人，每个人都很想交付、在领域里留下印记，自我驱动很强。这是一群「没安全感的超级优等生」，我每次问他们三个月后我们会在哪，得到的都是最悲观的回答；我会记录下来，三个月后发布了模型，再把他们当时的话拿给他们看。

Meta invested a huge amount to stand up MSL, and then we very clearly told everyone, there’s no deadlines, we are going to take things at the right pace. But focusing on talent density and hiring all these brilliant people, everybody had a lot of motivation to ship and wanted to leave their mark on the field, so everybody was highly self-motivated. It is a team of a lot of insecure overachievers. If I ask them where we’ll be in 3 months, I always get these very pessimistic answers. Then I record it, and 3 months later, we just shipped this model, let me show you what all of you said.

Mark Zuckerberg

什么时候知道 Muse Wonder 是好的？预训练做完后我们讨论过：RL 要多久才能超过 Muse Spark？预训练团队在给大家打预防针，说有时要花一阵。结果几乎是立刻就超过了，那时我觉得我们走对了。

We had this debate after it was done pre-training: how long will it take in RL before it is better than Muse Spark? I remember the folks on the pre-training team were trying to set expectations: look, sometimes this takes a while. And then it was immediately. So we were like, okay, I think we’re on a good path.

Mark Zuckerberg

那么多直接汇报的意义不在于直接汇报本身，而在于没有非技术的管理者。研究员都知道一个常见模式：从很技术的岗位转到不再接触细节的角色后，你会比自己意识到的更快变得不懂技术，这很危险。所以要保持技术性，大家要用智能体，要写代码。

The point of having all these direct reports isn’t that you have the direct reports; it’s the absence of non-technical managers. It’s well known among researchers that there’s this common pattern: you go from being very technical to a role where you’re not in the details anymore, and you very quickly become not technical, much faster than you probably realize. It’s quite dangerous. You have to stay technical, people have to use the agents, people should be writing code.

Alexandr Wang

过去一年和他共事最震撼的，是看到他真正的胆量，以及一旦内化了某个信念，就以多快的速度、多大的幅度改变之前的做法，转向与信念一致的世界观。即使在创始人群体里，这也很少见。很多 CEO 去年夏天都会说 AI 是对自己业务最重要的事，但看他们实际做了什么、行动与信念有多一致，这正是 Meta 和他特别的地方。

One of the very cool things about working with Mark over the past year has been watching the true degree of his boldness, and this trait of, once he internalizes certain beliefs, the speed and magnitude of willingness to change prior actions and pivot towards the worldview that is consistent with those beliefs. These are pretty unusual traits even among founders. A lot of CEOs would have told you last summer that AI is the most important thing for their business, but then you look at what they’re actually doing and how consistent their actions are with that belief. That’s a quite special thing about Meta and Mark.

Mark Zuckerberg

关于 Alex 在 X 上那些「发疯」帖子：我觉得沟通最重要的是真实。你想真实时，有些话出来不会完全如你所愿，但宁可往这个方向取舍，也不要过度照稿而失去真实。不过大部分我只觉得好笑。

With communication, I think it’s most important to just be authentic. When you’re trying to be authentic, some of the things don’t come out the way you want, and it is better to make the trade in that direction than to be overly scripted and not be authentic. But most of it I just feel is hilarious.

Mark Zuckerberg

保持在细节里，保持技术性，留在前沿。

Stay in the details. Stay technical and stay at the frontier.