Alex Wang 谈 Meta 的 AI 战略与新模型发布

好了，Kylie，这周我们又请到了一位大嘉宾。

All right, Kylie, we've got another big guest this week.

重量级的。Alex Wang，Meta 人工智能工作的负责人。那是几个月前的事了？

Huge. Alex Wang, the chief of Meta's artificial intelligence efforts. It was how many months ago?

大约 10 个月前。大约 10 个月前，他还是 Scale AI 的创始人、联合创始人兼 CEO。Meta 算是半收购了那家公司，完全收购了 Alex，之后他就一直很抢手。他一直在 AI 保护计划中。

About 10 months ago. About 10 months ago, he was the founder, co-founder, and CEO of Scale AI. Meta sort of quasi acquired the company half acquired the company and fully acquired Alex and he's been in high demand. He's been in AI protection program since.

是的，直到今天在 Core Memory Pod 上，我们都没怎么见过他。我不太确定为什么会这样，但他来了。他会告诉我们，我们会知道的。他 hopefully 会告诉我们他们刚发布了一个新模型。我确定我们会聊到一些，然后他们对我来说有点神秘，比如他们在 AI 上的哲学立场。Alex 在 Scale 的时候，他们总说自己是瑞士，他在某些事情上声音很大，但并非总是关于 AI 本身以及他对它的感受。

Yes, we haven't seen much of him whatsoever until today on the Core Memory Pod. Yes, I'm not sure exactly why this happened, but here he is. He's going to tell us well, we'll find out. He's going to hopefully tell us about they just released a new model. I'm sure we will get into some of that and then they're a little bit of a mystery to me about like where they are philosophically on AI. Alex has always been a bit of when he was at Scale, he was they always said they were Switzerland and he was loud on some things but not always AI itself and how he feels about it.

而且他们去年因为用数百万美元雇佣了很多人、组建了这个团队而上了很多新闻，每个人都在等着看他们要用这些资源做什么。所以，我们会聊聊招聘 soup 和所有这些数百万美元。所以，就是这样。Alex Wang 出现了。没错，兄弟。我是 Ashley Vance。我是 Kylie Robison。这是 Core Memory。

And they made a lot of news last year with everyone they hired for millions and millions and millions of dollars and built up this team and everyone's been waiting to see what they're going to do with all of these resources. So, we will talk about recruiting soup and all these millions of dollars. So, this is it. This is Alex Wang's he's emerged. Hell yeah, brother. I am Ashley Vance. And I'm Kylie Robison. And this is Core Memory.

Alex，谢谢你来做客。

Alex, thank you for being here.

嗯，很高兴来到这里。我觉得我们在 Meta 之前有过一些短信交流。我们有个乡村音乐短信链。

Yeah, excited to be here. I feel like we kind of texted a little bit pre meta happenings. We had this kind of country music text chain.

没错。

For sure.

而且我也认识 Nat Friedman 很久了。

And then I'm also I've known Nat Friedman for a long time.

是的，是的。然后我觉得你们两个消失了。然后什么？

Yeah, yeah. And then I feel like the two of you disappeared. And what

进了战壕，是的。

Went into the foxhole, yeah.

是的，变得非常安静。现在你带着新模型出现了。

Yes, went very very quiet there. And now here you are with the new model emerging.

嗯。

Yeah.

但你们确实安静了一段时间。我们呃，是的，我们有很多工作要做。我的意思是，我觉得呃，事实证明在 9 个月内从头构建一个前沿模型呃，是的，需要很多呃，很多艰苦的努力，但是呃，但是是的，我们，我的意思是，看到大家使用新 Spark 真的很令人兴奋。呃，你知道，我们发布的模型，呃，我们还有更好的模型在酝酿中，所以很令人兴奋。

But you guys went very quiet for a bit. We uh yeah, we had a lot of work to do. I mean, I think uh turns out um building a frontier model from scratch in 9 months is uh yeah, it takes a lot of uh takes a lot of painstaking effort, but um but yeah, we're I mean, it's been really exciting to see everyone use new spark. Um you know, the model we released and um we have better models cooking, so it's exciting.

那么你，我的意思是，你以前像个旧金山人。旧金山人。我猜。呃，然后我的意思是，我假设你在 Menlo Park 工作。

And so what you I mean, you were like a San Francisco a light. San San Franciscan. I guess. Um and then I mean, I assume you work at Menlo Park.

是的。所以你不得不，像我一样搬到南湾，是的。哇。

Yeah. So did were you did have you had to like I moved down to South Bay, yeah. Wow.

是的。我搬了。所以你完全投入了。

Yeah. I did. So you're full on committed.

我完全投入了，是的。对我来说，现在的城市是 Palo Alto。

I'm full on, yeah. And for me, the city now is Palo Alto.

好的。在 University Ave 上走走，买杯波霸奶茶。好的，我一直在想这个。那么你们之间的安排是怎样的？我的意思是，我最熟悉的人是你、Nat、Daniel Gross。我只和 Zach 一起出去过一两次。呃，是的，我的意思是，我想描绘一下你们是如何组织的。

Okay. Walk on University Ave, get a boba. Okay, I was wondering about this. So what's like the arrangement between I mean, the people I know best are you, Nat, Daniel Gross. I've only hung out with Zach once or twice. Um yeah, I mean, what's the I'm trying to pick paint a picture of how you guys are arranged.

是的，呃，所以基本上呃，整个部门叫做 Meta 超级智能实验室，呃，由我负责。然后它有几个部分。所以，呃，有一个部门叫 TBD，呃，是那种大型模型研究实验室。呃，我认为，呃，你知道，它有点名声在外，但那里有很多呃，你知道，顶尖的研究人员和基础设施工程师。呃，他们实际上在技术上都是向我汇报的。呃，所以这是一个设置。然后还有，呃，一个叫产品和应用研究的小组，简称 PAR。那是呃，Nat Friedman 领导的。所以，他们负责，呃，我们构建的所有产品，呃，以及将这些优秀模型实际部署到世界。然后，呃，在 Meta 超级智能实验室的整体框架下，还有 FAIR，呃，它继续做探索性和令人兴奋的研究。呃，我对他们的很多科学研究特别兴奋。所以，你知道，我们展示了一些非常出色的工作，呃，关于使用模型、AI 模型来理解大脑，以及使用 AI 模型来理解，你知道，计算化学。我们构建了一个原子的通用模型，呃，简称 UMA。然后，呃，这些部分构成了 Meta 超级智能实验室，我负责它，同时也在 TBD 实验室扮演非常亲力亲为的角色。然后，呃，Daniel Gross，呃，帮助领导 Meta Compute，它真正专注于我们的长期基础设施规划，以确保，你知道，我们显然可以，呃，构建所有 GPU 基础设施和数据中心基础设施，以支持这个非常大胆的努力。呃，他领导那个部门并与我们紧密合作。

Yeah, so um so basically uh the the whole unit is called Meta Superintelligence Labs, um which which I oversee. And then there's various pieces of it. So, uh um, there's a unit called TBD, which is um, the uh, sort of large model research lab. Um, that I think, uh, you know, is somewhat infamous, but that's where a lot of the, um, you know, leading researchers and infrastructure engineers are. Um, they actually all technically report to me. Um, so that's one set up. Then there's also, um, a group called, uh, product and applied research or PAR for short. That's what uh, Nat Friedman heads up. So, they're responsible for, um, all the products that we build, um, and the actual sort of deployment of these great models to the world. And then, um, also within the, uh, overall Meta Superintelligence uh, Labs umbrella is FAIR, uh, which continues to do exploratory and exciting research. Um, I'm particularly excited about a lot of their scientific research. So, you know, we've shown some pretty great work on, um, you know, using models, AI models to understand the brain, as well as using AI models to understand, you know, computational chemistry. We have built like a universal model for atoms, um, UMA for short. And then, um, and so that those pieces constitute Meta Superintelligence Labs, which uh, I oversee in addition to having a very hands-on role with, um, TBD Lab. And then, um, Daniel Gross, uh, helps lead up Meta Compute, which is really focused on our long-term infrastructure planning to ensure that, you know, we obviously can, um, build up all of the GPU infrastructure and and data center infrastructure necessary for this very bold endeavor. Um, and uh, and so he heads that up and partners closely with us.

那么，我的意思是，在加入之前，你在这个群体中最了解谁？

And like, I mean, who did you know the best out of that group before you got into this?

呃，嗯，呃，我认识 Nat 和 Daniel 其实很久了。Nat 是我在 Scale 最早的天使投资人之一，呃，我想在我完成 YC 之前，Nat 就投资了 Scale，并且多年来一直给我建议。Daniel，我想我也是在那段时间认识的。你知道，非常非常早，然后这些年逐渐了解他们。然后我们还有首席科学家 Sheng Xia，他帮助监督整个 MSL 的科学议程。他是我在开始 MSL 之前就认识的人，但你知道，自从他加入后，我们变得更亲近了。所以。

Um, well, uh, I knew I've known Da- Nat and Daniel actually for a long time. Nat was one of my very first one of my very first angel investors at Scale, Um, and I think like in before I completed YC, Nat had invested in Scale and had given me advice throughout the years. Daniel, I think I also met around that time. You know, very very early early on and have gone to know them through the years. And then we also have our chief scientist, Sheng Xia, who helps oversee the scientific agenda across all of MSL. And he is somebody who I have gotten to know, you know, I knew before starting MSL, but you know, since he's come in we've gotten a lot closer. So.

我很好奇，退一步看。自从你有点隐退，你的公司完全改变，已经 10 个月了。现在你在 Meta。那是什么样的经历？交易是怎么达成的？你是怎么最终去 Tahoe 和 Zuck 谈话的？你能带我们回顾一下第一次会面是什么样的吗？

I'm really curious like taking a huge step back. It's been 10 months since you kind of went into hiding, your company, you know, completely changed. And now you're at Meta. What was that experience like? What How was the deal made? Like How did you end up going to Tahoe and talking with Zuck? Can you just walk us through what that first meeting was like?

是的，所以是这样的。我认识 Mark 很多年了。我想即使在我经营 Scale 的时候，他也很慷慨地花时间，我能够从他那里得到很多建议。他显然是一个非常有经验的创始人，在某种意义上可以说是创始人中的创始人。所以我们认识很多年了，我们实际上在这次热潮之前就讨论过 AI，因为你知道，Scale 显然从 2016 年就开始做 AI 了。你知道，当时主要是自动驾驶，然后你知道，技术的各种转变。然后大约一年前，几乎正好是一年前，你知道，我们进行了一次对话，开始探讨是否有更紧密合作的方式。

Yeah, so it was So, I've known Mark for many years now. I think even while I was running Scale, I would He was very generous with his time and I sort of was able to get a bunch of advice from him. And he obviously is just such an experienced, you know, founder and like the the founder at this point in some in some sense. And and so we've known each other for many years and and we'd actually talked about AI before a lot of this craze because, you know, Scale obviously we'd been working on AI since 2016. And you know, back when it was mostly self-driving and then, you know, the various transitions of the technology. And then around a year ago, almost like literally a year ago, you know, we had a conversation where we sort of started exploring if there were a way to work more closely.

具体来说，我觉得马克当时越来越相信 AGI，并且深知 AI 将真正改变 Meta，而且 AI 是那种一生一次的变革性技术。所以他非常专注，并且知道要大力押注。同时，他也公开说过，Llama 4 当时的发展轨迹并不符合公司继续下注的需求。所以我们当时在高层面上讨论如何更紧密地合作，会是什么样子？这是一个非常开放的问题和头脑风暴。最后我们找到了一个对规模、对 Meta 都有利的方式，让我们能紧密合作，共同打造我们这个时代最重要的技术。我们都坚信会做出让自己骄傲的东西。他大约一年前发布了关于个人超级智能的备忘录，然后我们就沉默了，但我认为那是我们共同的北极星：我们希望以赋能人们的方式构建这项技术，让世界上尽可能多的人能够使用，尽可能民主化，让每个人都能表达自己、增强自主性、创造和建设。这就是我们想要努力实现的世界。

And in particular, I think Mark at that point was becoming increasingly AGI-pilled and really knew that AI was going to truly transform Meta, but also that AI was one of these once-in-a-lifetime transformative technologies. So he was really quite focused on it and knew he wanted to bet very big on it. At the same time, he's talked about this publicly, like Llama 4 was not on the trajectory that the company needed to continue making some of these bets. So we were talking at a very high level about how we could work together more closely, what could that look like? And it was one of these very open-ended questions and brainstorm sessions. Then it landed in this interesting zone where we figured a way to do it in a way that was good for scale, good for Meta, and where we got to work very closely together to build the most important technology of our time. We both had conviction that we were building something we'd be really proud of. He put out this memo of personal superintelligence about a year ago, and then we went quiet, but I think that really is the North Star for both of us: we want to build this technology in a way that empowers people, as many people in the world have access to it, as democratized as possible, enabling everyone to express themselves, have increased agency, create and build. That's the world we want to work towards.

我很早就报道过你。我认识你很久了。

I did a really early story on you. I've known you for a long time.

是的，是的。我 21 岁的时候？我想是的。

Yeah, yeah. When I was 21? I think so.

有很多传说：你是最年轻的白手起家亿万富翁，Scale 是一家非常知名的公司，你以预判 AI 走向而闻名。以前我和你聊天时，Scale 是你身份的一部分。从一家知名公司的创始人到一家有 8 万名员工的公司任职，即使职位很高，也是非常不同的。我真的很惊讶。虽然涉及很多钱，所以也说得通。但以我对你的了解，我觉得那一定是个很厉害的推销，因为这是一个巨大的转变。

There's all this lore: you were the youngest self-made billionaire, Scale was such a prominent company, and you had this reputation for reading the tea leaves of where AI was going. When I would talk to you, Scale was part of your identity. It's very different to be the founder of a prominent company and then taking a role at a place with 80,000 employees, even if it's a prominent role. I think I was really surprised. There's a lot of money involved, so okay. But just knowing you as much as I did, I was like, man, that must have been a hell of a sales pitch because it's a big flip.

是的，非常不同。完全不同。在整个过程中，我一直在想，进展比我长期预期的要快得多。在 AI 模型加速进步的背景下，有几件事让我印象深刻。第一，构建 AI 模型的人拥有越来越大的权利，无论是经济上还是产品上，可以在模型周围构建更多东西。早期有很多关于生态系统如何演变的争论，但由于模型改进速度和研究速度都非常快，成为构建模型的地方是生态系统中令人兴奋的位置之一。第二，下一阶段的技术很大程度上归结为算力。如果你有大量算力，你就能构建东西、下大赌注、部署产品，而没有算力就无法做到。这会导致科技生态系统出现有趣的分层。现在我们觉得所有科技公司都一样，但实际上，你应该把拥有大量算力的公司和没有算力的公司区别看待，因为前者能构建后者无法构建的东西。这创造了一个非常有趣的动态。Meta 这个机会令人兴奋的部分在于，马克全力投入 AI，下了大赌注，而且这创造了条件，让我们可以用大量算力进行构建，配合正确的研究和产品努力，我们能够真正对世界产生巨大影响。

Yeah, very different. It's super different. A lot of what I was thinking about throughout this process is that progress has happened a lot faster than I expected for a long time. In the accelerating progress of these AI models, a few things really started to stick with me. One is that those who build the AI models have greater and greater rights, so to speak, or both economic and product rights to build so much more around those models. There were early debates about how the ecosystem plays out, but because of how fast the models are improving and how fast the research pace is, building being a place that is building the models is some of the most exciting places to be in the ecosystem. The second is that so much of this next phase of technology really boils down to compute. If you have lots of compute, then you have the ability to build things and make big bets and deploy products that you just can't without that compute. This will cause an interesting stratification for the tech ecosystem. Right now we think all tech companies are the same, but in reality you should think about companies with lots of compute very differently from companies without that compute because there are things companies with compute can build that those without cannot. So it creates a very interesting dynamic. Part of what was very exciting about the opportunity at Meta is that Mark is all in on AI and has bet very big, but also that this created the conditions where we can build with huge amounts of compute and with the right research and product efforts, we have the ability to really make a huge dent in the world.

你们有大量算力，还挖来了很多优秀人才。那是当时报道的热点，前所未见。这些人已经来了 10 个月了。感觉怎么样？挑战是什么？在 Meta 拥有这个全新团队最令人兴奋的是什么？

You guys have a ton of compute and you poached a lot of amazing talent. That was a part of that reporting frenzy at the time. It was unlike anything I've ever seen before. It's been 10 months with many of these people. So what has it been like? What have the challenges been and what has been the most exciting about having this whole new team at Meta?

当我来到 Meta 时，很明显需要重新调整努力方向，重建我们的 AI 工作，以走上正确的轨道，因为 Llama 4 不在同一轨道上，我们落后于前沿。所以我们需要制定一个计划，让我们能够以非常快的速度追赶并有望超越前沿。

When I got to Meta, it was clear that there needed to be some reset of the efforts and some rebuild of our AI efforts to get on the right trajectory because Llama 4 was not on the same trajectory and we were behind the frontier. So we needed to build a plan that would enable us to have a very fast velocity to both catch up and hopefully exceed where the frontier is.

你能具体说说吗：你发现了哪些问题？

Can you be as specific: what were the problems that you found?

可能更根本的问题在于，许多领先的实验室都围绕一个前提来构建整个组织：超级智能即将到来，而且非常接近，这是一个非常现实的可信目标。然后你围绕这个基本信念来构建实验室和业务的整个计划。所以首要任务之一就是认真对待超级智能，然后围绕这个核心前提重建所有其他假设。我认为这很根本。

Probably the more fundamental ones are just around that a lot of the leading labs build the entire organizations around the premise that superintelligence is coming and it is very close, and this is a very realistic thing to believe that we can create and produce. Then you build the entire plan of the lab and the business around this fundamental belief. So that was one of the first things: to just take superintelligence seriously and then start to rebuild all of your other assumptions around that core premise. So that's something I think was somewhat fundamental.

所以你的意思是，他们在某种程度上缺乏对这种东西的宗教般的信念？

So you mean that they're like lacking this religious conviction to all this on some level?

是的，我认为这在 AI 领域其实相当普遍。很多大公司里有很多人并没有这种信念，因为仔细想想，结构有点不同。很多大公司有非常聪明的人在做 AI，但这和那些初创公司有点不同，那些初创公司是从零开始，带着超级智能即将到来的疯狂想法。所以，我不想再认为这是个问题了。显然现在，MSL，Meta 超级智能实验室，名字里就带着，围绕超级智能即将到来这个概念建立。所以我们为这个努力制定了一些原则。我想这回答了我们需要解决哪些问题。一是认真对待超级智能。二是技术声音最大。三是科学严谨，聚焦基础，并下大赌注。所以，TBD 和 MSL 的概念，当我开始的时候，我在想一个实验室的形态应该是什么样的，才能让你有极快的速度，赶上甚至可能超越前沿。我归结为三种方式。一是每个研究员拥有更高的算力。很多大实验室有大量算力，但分散在很多不同方向上。这实际上阻碍了每个研究员的研究速度。所以，如果你建立一个更专注的努力，团队更小，每个研究员算力更高，你实际上可以更快地取得研究进展。二是人才密度。我觉得人类组织总是重新学习这一课，但非常小的团队，每个人都很厉害，总是会比非常大的组织移动得更快，因为大组织责任更分散，更像大杂烩。最后一个是雄心勃勃的研究赌注。我认为这在行业里是公认的。确实存在这些非常大、非常冒险的研究赌注，但如果成功，可以完全改变范式，彻底改变我们构建现代 AI 的方式。所以，除了显然要构建非常有竞争力的前沿模型，我们还将大量资源和算力分配给这些大的雄心勃勃的赌注，因为它们会成功，给我们带来未来不可思议的模型。

Yeah, and I think this is relatively common actually for AI. I think there's a lot of people at all the large companies who don't actually have this conviction, because if you think about it, it's a bit of a different construction. A lot of the big companies have very smart people who work on AI, but it's a little bit different from these startups where it's like these new efforts start from scratch with this crazy idea that superintelligence is coming. So, I don't want to think there's a problem anymore. Obviously now, MSL, Meta Superintelligence Labs, is in the name, built around this concept that superintelligence is coming. So, there are a bunch of principles that we sort of laid out for the effort. And I think this sort of answers as to what were the things that we had to resolve. One is take superintelligence seriously. Two is technical voices are loudest. Three is scientific rigor, focus on basics, and make big bets. So, the concept of TBD and MSL broadly when I sort of got started was I thought about what would actually be the shape of a lab that would enable you to have incredibly fast velocity and catch up and potentially even overtake the frontier. And I came down to three ways that I felt like that was possible. One is to have much higher compute per researcher. So, a lot of the larger labs have lots of compute, but it gets spread so many different ways. That actually impedes the research velocity of any individual researcher. So, if you build a more focused effort with a smaller team that has higher compute per researcher, you can actually make faster research progress. Two is talent density. I feel like human organizations always relearn this lesson, but the very small team where everyone is cracked is always going to move faster than the very large organization where responsibility is more distributed and it's more of a melange. And the last one was on very ambitious research bets. I think this is very well agreed upon in the industry. There clearly exist these research bets that are very big and very risky, but if they work out, can totally change paradigms and totally shift how we build modern AI. So, in addition to obviously building towards very competitive frontier models, we're allocating a huge amount of our resources and compute towards these big ambitious bets, because they pan out and that gives us incredible models going forward.

好的，我们在 Core Memory 做什么？我们报道创新、快速、前瞻的公司，这就是为什么 Core Memory 由 Brex 赞助，因为 Brex 是许多这类公司的智能金融平台。从初创公司到全球最大的公司，有 3 万家公司依赖 Brex 的技术管理财务。他们有智能公司卡、高收益企业银行和出色的费用自动化工具。我讨厌做费用报销，Brex 的 AI 和软件直接处理这些费用，弄清楚我们在哪里花钱，为你处理很多事情，这样你就不用自己浪费时间了。访问 brex.com/corememory 了解更多，跟上节奏吧。让我们行动起来，摆脱那些过时的财务软件，走向未来。Core Memory 和 Brex。

All right, what do we do at Core Memory? We cover innovative, fast-moving, forward-thinking companies, which is why Core Memory is sponsored by Brex, because Brex is the intelligent finance platform for many of these companies. 30,000 companies from startups to the world's largest corporations rely on Brex's technology for their finances. They've got smart corporate cards, high-yield business banking, and expense automation tools that are fantastic. I hate doing my expenses, and Brex's AIs and software run right through those expenses, figure out where we're spending money, and take care of so much stuff for you, so you don't have to waste your time on it yourself. Go to brex.com/corememory to learn more, and just get with the program. Let's get going. Let's get out of this archaic finance software and move toward the future. Core Memory and Brex.

Ashley 总是谈到这些实验室如何互相超越，然后开始提供同样的东西。你在谈论冲向前沿。我也在想你们雇佣的人。据报道，薪水高得离谱，前所未有。那么，你是如何达到你所说的范式的？具体来说，你想实现什么范式？

Something Ashley always talks about is how these labs are sort of leapfrogging over each other, and they start to just serve the same thing. And you're talking about racing towards the frontier. I'm also thinking about the people that you guys hired. It was reported for these really, really wild salaries like we've never seen before. So, like how are you getting to this paradigm you speak of? Like specifically what paradigm are you trying to achieve?

是的，我认为有很多大胆的研究赌注，我无法详细说明所有。但我确实认为一个基本问题是我们关心什么？我们，与个人超级智能的理念一致，真正关心构建这些能够赋能消费者的智能体，赋能全球数十亿人，以及赋能企业。我们有 Meta 这个不可思议的生态系统。我们有数十亿用户，我想大家都知道，但我们平台上还有数亿企业使用 Meta 来运营他们的业务。所以，我们非常关心构建这样一个未来：我们能够构建非常强大的智能体，赋能我们平台上的每一个消费者和企业，并构建这种新的智能体生态系统。我们思考了很多这看起来是什么样子，以及我们需要构建哪些能力来实现它。显然，在这条轨迹上，有很多子组件非常重要，比如我们必须有很好的智能体能力。我们必须有很好的编码能力，因为随着深入，很多需要构建的东西都是软件。我们需要有很好的多模态能力。所以，这决定了我们需要很多底层的东西。然后我们需要解决长期运行智能体的许多更大问题，比如如何思考记忆挑战？如何构建长期运行的智能体？我们构建能够代表用户执行越来越复杂任务的智能体。所以，这些都是我们思考的很多高层部分。

Yeah, I mean, I think there's a bunch of bold research bets, and I won't be able to go into detail on all of them. But I do think one fundamental question is what do we care about? And we, in line with this idea of personal superintelligence, really care about building these agents that are able to empower consumers, so empower billions and billions of people all around the world, as well as empower businesses. We have Meta's kind of this incredible ecosystem. We have the billions of users, which I think everyone knows about, but we also have hundreds of millions of businesses on our platforms that use Meta to run and operate their businesses. So, we really care a lot about building towards this future where we're able to build very powerful agents that empower and enable every one of both the consumers and the businesses on our platforms, and build kind of this new agentic ecosystem. We think a lot about what that looks like, and what are the capabilities we need to build towards that. Obviously on that trajectory, there's a bunch of sub components that are really important, like we really have to have great agentic capabilities. We have to have great coding capabilities because so much of what needs to be built is software as you really get into it. We need to have great multimodality. So, it informs a lot of the underlying bits and pieces that we need. And then we need to solve a lot of the bigger questions for long-running agents, like how do you think about the memory challenges? How do you think about building long-running agents? We build agents that are able to do more and more complex tasks on behalf of the users. So, those are a lot of the high-level pieces that we're thinking a lot about.

所以，如果你试图在公司内部建立这种超级智能宗教，这种安排方式与创办 OpenAI 或 Anthropic 截然不同，就像你之前说的，那些是从零开始建立的，这些公司有一种身份，并随着时间的推移逐渐形成。从外部来看，你们所做的看起来更像是雇佣兵式的。

So, if you're trying to get this superintelligence religion built within the company, there is the way this was arranged, it's quite different to starting OpenAI or Anthropic, where, like you were talking about earlier, this is all built from the ground up, and these companies have sort of an identity and it gets shaped over time. For the outsider's view, what you guys did looks far more mercenary.

你懂我的意思吗？就像我们要去，你们出局了，我们要去抓一堆高价人才，把他们带进来。这让我想起，你知道，我记得 Grok 刚起步的时候，埃隆用他那种埃隆的方式说：‘我们就是要比任何人都获得更多的算力，而且我们有这样一个核心团队要围绕它来建设。’然后感觉他们还是追上了，但从未达到那种逃逸速度，尤其是在人们心目中的品牌形象等方面。所以，我的意思是，你提到的那些东西，似乎很难用钱买到。

You know what I mean? It's like we're going to go you guys are out, we're going to go grab a bunch of high-priced people, bring them in, and it reminds me, you know, I just remember when Grok was starting up, and it was like Elon in his Elon way, he's like, 'We're just going to get way more compute than anybody else, and we have kind of this core team we're going to build around.' And then it still felt like they caught up, but then never reached that escape velocity, especially in people's minds of brand and things like that. So, I mean, it just seems like it's a hard thing to buy some of the bits that you're talking about.

是的，我认为这是较大的叙事偏差之一，或者说外部看法与内部日常之间的差异。我想很多人都有你提到的那些印象。这很大程度上是因为媒体报道以及事情的发展方式。很多报道在各种方面被夸大了，但最终还是传开了。部分原因是我们招聘得太快了。我加入时就知道，‘如果我们想打造出色的模型，我们必须昨天就拥有团队。’所以我们不得不闪电战，非常迅速地完成。但我认为实验室内部的文化实际上非常像初创公司。有很多因素造成了这种感觉。一是这是 Meta 内部一个全新的团队，但实验室的文化是，每个人都对我提到的这些事情非常着迷和兴奋。人们加入是因为每个研究员拥有很高的算力，所以他们能比在其他地方取得更多进展。人才密度很高。人们看到这是一个真正精英的小团队，我们会给他们资源和自由去进行非常大胆的研究赌注。所以认为研究人员只是受金钱驱动是错误的。对他们大多数人来说，留在原地的财务前景也非常好。所以金钱不是主要动机；而是有机会从零开始建设，拥有大量算力，追求雄心勃勃的研究方向，并且在一个不臃肿的团队中工作。因此，氛围和文化要健康得多。许多来自其他实验室的访客经常评论说，这里的氛围让他们想起早期的 OpenAI 或早期的 Anthropic，那些更初期的阶段，因为从某种意义上说，我们现在才成立 10 个月。

Yeah, I would say this is one of the larger narrative violations, or maybe differences between external perception and what the day-to-day inside is like. I think a lot of people have some of the impressions you're talking about. A lot of that is formed because of the reporting and how it sort of went down. And a lot of the reporting was overstated in various ways, but it all sort of bubbled up. Part of it was because we did the recruiting so quickly. When I got in, I knew, 'If we want to build great models, we need to have the team yesterday.' So we had to blitz it and do it very quickly. But I think the culture within the lab is actually very much a startup. There are a bunch of things that have created this feeling. One is that it was an entirely new-built team within Meta, but the culture of the lab is that everybody was very attracted and excited about these things I talked about. People joined because there was high compute per researcher, so they could make more progress than they would elsewhere. There was great talent density. People saw it was a truly cracked group that was pretty small, and we were going to give them the resource and freedom to make very bold research bets. So it's an incorrect assumption to think that the researchers are just money motivated. For most of them, the financial prospects of staying wherever they were looked very good as well. So money was not the primary motivation; it was the opportunity to build from scratch, have lots of compute, approach ambitious research directions, and do so in a group that didn't feel bloated. As a result, the vibe and culture are much healthier. Many people who visit the lab from other labs often comment that the vibe reminds them of early OpenAI or early Anthropic, the more nascent stages, because in some sense we're now 10 months old as an effort.

而且因为 Mark Chen 上过这个播客，提到了招聘大战中的汤风波。是啊，这是真的吗？Zach 做汤了吗？你们做汤来招人了吗？

And just because Mark Chen was on this podcast and brought up the soup debacle during these recruiting wars. Yeah, did you, is this true? Did Zach make soup? Did you make soup to recruit people?

我不知道我们是否做了汤，但我听说实际上是 Zach 做的，不过我不确定。我不知道我们是否做了这个汤，但我确实认为，我们必须向每个人展示我们真的在乎这项技术以及他们具体的研究方向。这是一个非常个性化的招聘过程，但也是一个让人们知道我们是认真的过程。默认情况下，很多人不知道如何看待 Meta 的 AI 努力，或者对我们了解不多。所以需要大量地去接触人，和他们交谈，解释我们在构建什么，我们关注什么，为什么我们在乎这项技术，我们想用它做什么。这非常重要。

I don't know if we made the soup, but I was told that it was actually made by Zach, but I don't know. I don't know if we made this soup, but I do think it is true that we had to show everyone that we really cared about this technology and their specific research directions. It was a very individualized recruiting process, but it also was one where people had to know we were serious. By default, a lot of people didn't know what to think about Meta's AI efforts or didn't know that much about us. So it took a lot of going to people, talking to them, explaining what we were building, what we were focused on, why we cared about the technology, what we wanted to do with it. That was very important.

是的，我之后会继续，但不仅仅是反复讲招聘的事。你知道，当你在 Scale 的时候，大家都称你为 AI 界的瑞士，然后如你所记得的，你认识所有人，你处于中心位置，然后感觉其中一些带来了个人代价。我的意思是，你和 Sam 曾经是室友，我给 Sam 发短信说你会上节目。他说的可不太好听。所以，这似乎对你个人来说一定付出了代价。

Yeah, I'll move on after this to not just belabor all the recruiting stuff, but you know, when you were at Scale, you were like the everyone called you the Switzerland of AI and then as you remember, you knew everybody and you were in the center of things and then it feels like some of this came with a personal cost. I mean, you and Sam used to be flatmates and I texted Sam about you coming on the show. He did not have flattering things to say. So, it seems like some of this must have come at a personal cost to you.

是的，我认为其中一些是不幸的。我诚实的期望是，随着我们越来越接近超级智能，我作为一个人真诚地希望，这个行业中不同人之间的所有敌意——这在当下显然非常热门，因为其他事情正在发生——会随着时间的推移而消退。人们会走到一起，意识到我们正在构建这项极其重要的技术，在构建过程中，我们所有人都必须对此非常深思熟虑。我觉得我的一个责任是确保我们开发的技术以及我们部署它的方式尽可能深思熟虑。你提到的一些事情，比如招聘方面，那是相信我们有一个使命，而不仅仅是产品。你可以追求自己的研究使命。而且你也很年轻。我们几乎同龄，这很有趣。而且我不是亿万富翁。

Yeah, I think some of this is unfortunate. My honest expectation is that as we get closer and closer to superintelligence, my hope genuinely as a human is that all the animosities between various people in this industry, which is very topical obviously right now with other things happening, subside over time. People will come together and realize we are building this incredibly important technology, and it's important for all of us to be really thoughtful about that as we build it. One of the things I feel is a responsibility of mine is to ensure that the technology we develop and the ways we deploy it are as thoughtful as possible. Some of the stuff you talked about, like in terms of recruiting, it was trusting that we have a mission and it's not just products. You can go for your own research mission. And you're also quite young. We're almost the same age, which is very funny. And I'm not a billionaire.

但 Jan 在离开后不久对媒体说，你年轻且缺乏经验，还会有更多人离开。所以我很好奇，作为这家大公司的领导者，这对你来说意味着什么？而且你相当年轻。读到那些话是什么感觉？你和他谈过吗？

But Jan had said in the press shortly after he left that you were like young and inexperienced and more people are going to leave. So, I'm curious, like, how has that boded for you as a leader at this huge company? And you're quite young. Like, what was reading that like? Have you talked to him?

是的，在那之后几周我在印度见到了他，嗯，我的意思是，Jan 是一个著名的、非常直言不讳的人，我想每个人都知道 Jan 在想什么。

Yeah, I saw him in India like a couple weeks after that and, um, I mean, Jan is a notable, very outspoken person and I think everyone always knows what Jan is thinking.

所以，但他显然说了他说的话，我在印度见到他。他祝贺我们推出了 Muse Spark。

So, but no, I think he obviously said what he said and I saw him in India. He congratulated us on the Muse Spark launch.

我看到你们在 X 上发的东西了。

I saw you guys, Patrick things up at X, yeah.

是的，我的意思是，我确实就像我之前说的那样。我认为所有的个人恩怨，随着我们越来越接近超级智能，我们会……

Yeah, I mean, it truly I do exactly what I just said before. I think all personal animosities, like, I think as we get closer and closer to superintelligence, we'll...

好像越来越糟了，不是吗？是的。

Like it's getting worse, doesn't it? Does.

嗯，是的，也许先变糟再变好，但是，嗯，但是，不，我认为，嗯，我认为，呃，你知道，我对我们如何建立 MSL、我们的研究工作以及我们取得的进展充满信心。嗯，我很兴奋地向世界展示我们研究人员所做的令人难以置信的工作和我们取得的进展。

Um, yeah, maybe it gets worse then it gets better, but, um, but yeah, no, I think that, um, I think that, uh, you know, I have a lot of conviction in how we've set up MSL and the research efforts that we have and the progress that we're making. Um, and I'm excited to show the world the incredible work that our researchers are doing and the progress we're making.

我也不想过多强调这一点，但这是你一直面临的挑战吗？比如人们认为你太年轻、缺乏经验，无法领导 Meta 这么大的项目？还有人批评你不是工程师。

I also don't want to belabor the point, but is that a challenge you continue to face? Like people thinking you're too young and inexperienced to lead such a huge effort at Meta? You also get the knock that you're not an engineer.

哦，是的。这绝对不是真的。比如我曾经是硅谷的一名软件工程师，但是嗯呃，这难道不让你生气吗？

Oh, yes. That is definitely not true. Like I once upon a time I was a software engineer in Silicon Valley, but um uh Like doesn't this piss you off?

是的，我的意思是，老实说，关于年龄问题，我在硅谷的整个职业生涯都有人这么说。所以，嗯，你知道，在某种程度上，呃，我几乎不再去想它了，因为它一直存在。嗯，但是，是的，我知道，我认为，我认为总是有，实际上我认为人工智能领域的很多人总是有各种对他们的错误描述，或者嗯，你知道，人们总是说，而且他们总是有点嗯，呃，你知道，外面说的从来都不完全正确，这可能会令人沮丧，但我选择把它转化为我们正在做的工作和我们发布的东西，嗯，再次，我为 Muse Spark 感到非常自豪。我对我们正在开发的模型和产品更加兴奋。所以，我认为从长远来看，嗯，从长远来看，这一切都会顺利发展。

Yeah, I mean I honestly I mean so on the age thing I mean like people have said this my whole time in Silicon Valley. So, um you know, to some extent like uh I actually I like I almost don't even think about it anymore cuz it's just it's always there. Um But yeah, I know I think that I think that there are always I think actually many people in AI always there's like various mischaracterizations about them or um there's like you know, people always say and and they're sort of always um uh you know, what's out there is never always correct and it can be frustrating, but I choose to just channel it into the work that we're doing and what we put out there and um again, I am really proud of Muse Spark. I'm even more excited about the models that we have cooking um and uh and the products that we have cooking. So, I think like in the long arc of you know, um in the long arc all of this will play out just fine.

你参加过数学奥林匹克，对吧？我的意思是，对于不了解的人来说，在我报道硅谷的经验中，在这些比赛中表现出色的人往往非常擅长编程、工程和思考这些问题。我会说，随着时间的推移，你在某些圈子里确实有了一点像大规模销售员的声誉，并且享受生活之类的。我的意思是，所以我在想，当你接受那份工作时，会不会更难指挥别人，或者只是完全不同。我不知道。

You were a math Olympiad, right? I mean you were you I mean which if for people who don't know, I mean in my experience covering Silicon Valley is like everyone who does well in these competitions tends to be quite um proficient at at coding, engineering, and and thinking about these problems. I will say, I mean, over time you did get a reputation among some circles as kind of like a salesman at scale a bit and and enjoying life and things like that. I mean, it so I I did wonder when you took that job if it was going to be um kind of like harder to boss people around or just just quite different. I don't know.

是的，顺便说一句，实际上，总的来说，我对 MSL 的管理理念不是指挥别人。我认为，呃，你知道，史蒂夫·乔布斯有句名言：大多数公司雇佣人然后告诉他们做什么，但我们雇佣人是为了让他们告诉我们做什么。我认为，你知道，这基本上是 TBD 和 MSL 整个理念的核心，以及我们如何建立它，我们将雇佣杰出的研究人员，并为他们创造最好的环境，让他们做出职业生涯和人生中最重要的工作。所以，所以，你知道，呃，长话短说，我实际上并不是想指挥任何人。我是在努力为研究人员创造最好的环境，让他们做出令人难以置信的工作。

Yeah, I By the way, I actually in general in terms of my management philosophy for MSL is not to boss people around. I think that like uh you know, there's a great Steve Jobs quote which is most companies hire people and tell them what to do, but we hire people for them to tell us what to do. And I think, you know, that is like pretty core to the entire thesis of TBD and MSL and how we built it is that we're we're going to we were going to have we're going to hire brilliant researchers and you know, create the best environment for them to do the work of their careers and the work of their lives. So, so I think you know, uh so long story short, I'm not trying to boss anyone around, actually. I'm trying to create the best environment for researchers to do incredible work.

好了，播客朋友们。我来给你们讲讲 SendCutSend。他们是一个制造奇迹。无论你是在改装自己的车，还是大型汽车公司、火箭制造商，如果你需要金属切割、加工或弯曲，你都会选择 SendCutSend。他们会为你做对，而且做得很快。访问 sendcutsend.com/corememory，无论你想做什么，都能享受 15% 的折扣。我们喜欢 SendCutSend。今天就去看看吧。

All right, pod friends. I am here to tell you about SendCutSend. They are a manufacturing phenomenon. Whether you're working on your own car or you're a giant automotive company, a rocket maker, if you need metal cut, machine, or bent, you're going to SendCutSend. They'll do it for you right and they will do it for you fast. Go to sendcutsend.com/corememory for 15% discount on whatever you're trying to make. We love SendCutSend. Go visit them today.

关于 News Park 说一分钟。我的意思是，就在我过去几天阅读所有内容并稍微玩了一下模型时，我试图，我们试图理解你们在 Meta 想要实现的目标。看起来在基准测试上，你知道，有些做得好，有些落后于其他模型。看起来你们在强调这一点，如果我把技术细节搞错了，你可以纠正我，但看起来你们在强调你们觉得你们有一些效率提升，其他模型可能没有，然后你们在做这个疯狂的事情，比如你们正在开发的 16 个智能体。我昨晚在玩那个。所以，我觉得你们认为你们选择了一些技术方向，可能领先于别人，但后来我昨晚看了你在 X 上的所有推文。有人称赞你，也有人挖苦你，你会说，你知道，等下一个东西吧。所以，是的，我想我们试图弄清楚，嗯，看起来你们并没有像插旗一样说，你知道，我们用这个模型征服了一切。

On news park for a minute. I mean, you just as I was reading everything over the last couple days and playing with the model a bit. I'm trying I'm just trying to wrap We were trying to wrap our head around exactly where you guys see it in terms of what you're trying to achieve at Meta. It seemed like on the benchmarks, you know, did well on some, was behind the models on others. It seemed like you guys were emphasizing this and apologies if I get the technical stuff right, you get wrong, you can set me right, but it seemed like you were emphasizing that you felt like you guys had some efficiency gains that some of the other models maybe maybe didn't have and then you're doing this crazy thing with the like 16 agents you're working on. I was playing with that last night. So, it felt like to me you guys think you've picked a couple technical directions where maybe you're ahead of people, but then I went through all your tweets on X last night. There were people who would compliment you. There were other ones that would take a dig at you and you would say, you know, just wait for the next thing. And so, yeah, I guess we're trying to figure out um It didn't seem like you guys were like planting a flag and being like, you know, we have conquered everything with this model.

是的，是的，不，绝不是。我认为，你知道，我们所做的是，在过去的 9 个月里，我们重建了很多技术栈和很多研究。所以，我们重建了预训练栈，重建了强化学习栈，重建了很多科学方法，并在数据上做了大量工作。所以，从很多方面来说，过去 9 个月发生的事情真的就像是对核心研究栈的一次全面翻新。而 MuSpark 是那个 Scaling 阶梯上的早期数据点，但它不是，你知道，从某种意义上说，MuSpark 就像是主菜或开胃菜，我猜是开胃菜，法语中的主菜，但它有点像我们正在构建的东西的开胃菜，但我们正在开发更大的模型，我们期望更大的模型，你知道，我们对更大的模型比 Musepark 更兴奋。嗯，但我们认为这是一个重要的数据点，可以发布到世界上，因为，嗯，你知道，我们构建的整个程序都是围绕可预测的 Scaling 开发的。所以，我们看到 Scaling，我想我们在博客文章中讨论过，是在多个轴上。我们看到预训练 Scaling 非常一致且可预测。我们看到可预测的强化学习 Scaling。嗯，我们看到强化学习的 Scaling。我们看到可预测的测试时 Scaling。嗯，呃，你刚才提到的思考模式，我们也在多智能体 Scaling 中看到了非常令人兴奋的结果。所以，嗯，我们程序的一切都是为了随着我们的发展而继续 Scaling。所以，你知道，Musepark 就像是我们 Scaling 轨迹上的早期数据点。

Yeah, yeah, no, by no means. I think that you know, the what we did is, you know, over the past 9 months, we rebuilt a lot of the stack and a lot of the research. So, we rebuilt our pre-training stack, we rebuilt our RL stack, we rebuilt a lot of the science and did a lot of work on data. So, in many ways what's been happening over the past 9 months is really like a, you know, a full-on renovation as it were for the for the sort of like core research stack. And MuSpark is kind of the early data point on that scaling ladder, but it is not, you know, in some ways like MuSpark was is kind of like the entree or the appetizer, I guess appetizer, entree in French, but it's kind of like the appetizer for what we're building, but we are in development of larger models and we expect our the larger models to be, you know, we're much more excited about the larger models than we are even about Musepark. Um, but it was an important data point we thought to to put out there into the world because it was a, um, you know, the entire program that we've built is is developed around predictable scaling. So, and we see the scaling, I think we talked about this in blog post, on many axes. We see pre you know, very consistent pre-training scaling and predictable pre-training scaling. We see predictable RL scaling. Um, we see and scaling reinforcement learning. We see predictable test time scaling. Um, and uh, a lot of what you just talked about contemplating mode is we're also seeing very exciting results in multi-agent scaling. And so, um, everything about our program is built to continue scaling as we go. And so, you know, Musepark was like this early data point on our scaling trajectory.

但下一个数据点我们更兴奋，再下一个数据点我们甚至更兴奋。所以我们很期待向大家展示我们整体 Scaling（规模扩张）努力中的下一个台阶。具体到 Muse Spark，整体端到端性能最终比我们预期的要好得多。它涌现出许多能力和行为，我们在训练中发现了这些，非常兴奋。例如，它在智能体式视觉编码方面的能力，能够生成网站或游戏。其中一些能力源于它既是一个相当强的智能体模型，又非常擅长多模态。所以我们对这个模型有很多兴奋点，因此我们把它发布了出来。我们认为对大多数消费者用例来说，它实际上是一个非常好的模型，并且与其他模型相比相当有竞争力。我们部署的 Muse Spark 在智能体编码方面还不太有竞争力，所以这些是我们正在为下一代模型努力的方向。但我预计我们生产的下一代模型总体上会比 Muse Spark 更好，这让我们很兴奋。即使是我们发布的 Muse Spark——明确地说，我们设定了预期，不认为它会是一个全面领先的前沿模型——但它是一个非常好的模型，我想很多尝试过它的用户都体验到了这一点。

But the next data point we're a lot more excited about, and the data point after that we're even more excited about. So I think we're excited to show people this next rung up on our overall scaling efforts. And for Muse Spark specifically, the overall end-to-end performance ended up being quite a bit better than we expected. It had a bunch of emergent capabilities and behaviors that we were pretty excited about, found in training. For example, some of its abilities in agentic visual coding, being able to produce websites or games. Some of these capabilities emerged from the fact that it is both a pretty strong agentic model and also pretty strong at multimodality. So there were a lot of things we were very excited about with this model, so we put it out there. We think for most consumer use cases it's actually a very good model and quite competitive with other models. Muse Spark as we deployed it is not yet competitive on agent coding, so those are capabilities we're working on for the next set of models. But I would expect the next model we produce to be better overall than Muse Spark, and that's something we're pretty excited about. Even Muse Spark as we released it—to be clear, we set expectations that we didn't think it would be a state-of-the-art model across the board—but it is a very good model, and I think a lot of users who tried it experienced that.

我很好奇，发布前沿模型的障碍是什么？你还需要什么才能达到所有这些基准并大获成功？

I'm curious what was the holdout for releasing a frontier model? What do you still need in order to hit all of those benchmarks and blow it out of the water?

答案只有一个词：Scaling（规模扩张）。Muse Spark 还处于早期阶段，我们有很强的可预测性。所以我们知道如果把这个模型扩大，从增加的模型规模中能期待什么样的性能。我们预计即将推出的模型将能够在各方面表现更好。

The one-word answer is just scaling. Muse Spark is early on the ladder, and we have very strong predictability. So we know if we scale this model up, what performance to expect from that increased model size. And we expect the upcoming models to just be able to perform much better across the board.

那什么时候会发生？

When does that happen?

未来几个月。我们构建了整个项目，以便能够非常快速地推进。有一段时间我们不得不重建所有基础，但现在我们正处于快速 Scaling（规模扩张）模式。

Coming months. We built the whole program so that we would be able to move very fast. There was a time period where we had to rebuild all the foundations, but now we're in fast scaling mode.

所以，整个努力大约一年了。你觉得你们在技术上做了什么与众不同的事情？

So, like a year into the whole endeavor. What do you feel like you're doing technically that's different to everybody else?

我们发现的一件事是，Muse Spark 表现非常好，在某些方面甚至比我们最初预期的还要好，尤其是一年前。当我们分析为什么它表现这么好时，我们认为很大程度上是因为从头构建了一个非常干净的栈，并且在这个重建过程中有能力以正确的方式做所有事情。我们真的拥有这种奢侈——能够构建一个非常干净的预训练栈和非常干净的强化学习栈，并由确切知道如何构建这些系统的专家以正确的方式完成一切。这能够显著加速我们的轨迹，而且我认为这在模型中真正体现出来了。

One of the things we found is that Muse Spark performed very well, in some ways even better than we originally expected, especially a year ago. When we analyzed why it performed so well, we think a lot of it comes down to having built a very clean stack from scratch and having had the ability in this rebuild process to do everything the right way. We really had this luxury—the ability to build a very clean pre-training stack and very clean RL stack, and to do everything the right way by the experts who know exactly how to build these systems. That was able to meaningfully accelerate both our trajectory, and I think it really shows in the model.

在我做这些采访之前，我把你和模型扔进所有 AI 系统，让它们四处探查。反复出现的一点是这种 Token 效率。这是你们觉得已经搞清楚了的东西，还是 Muse Spark 的一个意外之喜？在一些基准测试上，你们似乎用比其它模型少得多的努力就做到了。

Before I do these interviews, I throw you and the model into all the AI systems and get them to poke around. The thing that kept coming back was this token efficiency. Is this something you guys feel like you've figured out, or was it a happy accident with Muse Spark? It seemed like on some benchmarks you were doing it with far less effort than other models.

是的，这对我们来说是一个令人兴奋的结果。例如，在 Artificial Analysis 上，它用比其它一些实验室的模型少得多的 Token 就取得了非常相似的结果。我认为这证明了干净栈的价值。其它一些模型可能需要更多 Token 的一个原因可能是栈的其它部分存在某种根本性的低效，通过让模型思考更长时间来弥补。所以我们对我们发现的 Token 效率印象深刻且兴奋。坦率地说，随着我们继续 Scaling（规模扩张）模型并整体扩大规模，我们认为这对未来的性能是个好兆头。

Yeah, this was an exciting result for us. On Artificial Analysis, for example, it used to achieve pretty similar results with many fewer tokens than models from some other labs. I think this is a testament to the clean stack. One reason why some other models may require a lot more tokens could be that there's some level of fundamental inefficiency at another part of the stack that gets patched by enabling the models to think longer. So we were pretty impressed and excited about the token efficiency we found. Frankly, as we keep scaling the models and continue scaling overall, we think that bodes really well for future performance.

Muse Spark 在视觉基准测试上非常出色。这种效率和视觉专长似乎对你们的硬件努力非常重要。你之前谈到过一系列 AI 产品，它们能看到你所见、听到你所听。你能多谈谈这如何融入你们服务 AI 的产品愿景吗？

Muse Spark was really good at vision benchmarks. That efficiency and vision expertise seemed like it would be really important for your hardware endeavors. You've talked before about a constellation of AI products that can see what you see and hear what you hear. Can you talk a little more about how that fits into your product vision for serving AI?

百分之百。Meta 整体上非常令人兴奋的一点是，Ray-Ban Meta 眼镜已经成为一个热门产品。我们卖出了数百万副，并且有一些忠实粉丝。这是一个非常令人兴奋的方向，思考你与技术的关系会是什么样子，如果它真的能稍微淡出背景、更具上下文感知，看到你所见、听到你所听，并在你需要的时候更加智能和有用。同时还能捕捉关于你生活中正在发生的事情的所有上下文，以及哪些是真正重要、应该关注的事情。

100%. One of the things that's very exciting about Meta overall is that the Ray-Ban Meta glasses have been a hit product. We've sold millions of copies and we have some big fans. It is a very exciting direction for all these devices to think about what your relationship with technology looks like if it really can fade into the background a little bit and be more contextual, see what you see, hear what you hear, and be much more intelligent and helpful in the moments where you need it. And also capture all this context about what is happening in your life and what are the things that really matter and that it should pay attention to.

我们确实看到了一个与个人超级智能一致的未来世界，你有一系列设备来捕捉上下文，让技术淡出，帮助你从智能体那里获得智能洞察——主动的洞察，或者你提到某事，智能体就去研究或为你采取行动。它就像一个超级智能的助手，让一切变得更好。但我觉得有个问题：我喜欢这些眼镜，经常用它们看视频和打电话，而且我通过 WhatsApp 运营整个公司。我拒绝用 Slack。我旅行太多，WhatsApp 已经融入我的生活。坦白说，直到你上节目之前，我都没用过 Meta 的 AI 智能体。我总是用 Claude 或 ChatGPT。我今天第一次看到 WhatsApp 上的 AI 智能体按钮，尽管它一直在那里。所以我身处你的世界却没注意到。我不可能是唯一的。

And we really see a future world in line with personal superintelligence, where you have a constellation of devices that capture context and enable the technology to fade away, helping you get intelligent insights from agents—proactive insights, or you mention something and the agent goes off to do research or take actions for you. It becomes a superintelligent sidekick that makes everything better. But I feel like there's a problem: I love these glasses, I use them all the time for video and phone calls, and I run our entire business on WhatsApp. I refuse to use Slack. I travel so much that WhatsApp is embedded in my life. Full confession: I don't think I've ever used Meta's AI agents until you came on the show. I always use Claude or ChatGPT. I saw the AI agent button on WhatsApp for the first time today, even though it's been there. So I'm in your world and didn't even see it. I can't be unique.

是的，有一点是我们知道在推动整个生态系统更紧密整合之前，我们需要优秀的模型和产品。在很多方面，我们一直在等待拥有能够支持我们关心的多数消费者用例的优秀模型。现在我认为我们到了一个激动人心的时刻：我们的模型相当不错，还有更好的在路上。所以我们将进行大规模整合，把我们所有应用家族与 AI 结合，并把我们的商业产品与 AI 整合，将生态系统的几乎所有部分与 AI 编织在一起。某种程度上，你过去几年在 Gemini 上看到了类似的情况，我们很兴奋要经历这个过程。

Yeah, one thing is we knew we needed great models and great products before pushing for tighter integration across our ecosystem. In many ways, we were waiting to have great models that enable most consumer use cases we care about. Now I think we're at an exciting point: our models are pretty good, and we have better ones on the way. So we're going to undergo a large-scale integration of all our family of apps with our AI, and integrate our business products with our AI, knitting together almost all pieces of our ecosystem with our AI. To some extent, you've seen what that looks like for Gemini over the past few years, and we're excited to go through it.

对我来说也一样。我也用 Google 运营业务，主要玩 Gemini 看看效果。我很好奇你认为这在消费者心中会如何发展。OpenAI 和 Anthropic 在一个世界里，ChatGPT 是强大的消费品牌，Claude 在编程和商业领域占主导。Google 则让人们在使用各种服务时遇到 AI。我不认为我们见过这样的竞争。我老了，所以回到文字处理时代——人们最终选择了 Microsoft Word。在浏览器大战中，是 Internet Explorer 对 Netscape。你认为这两组面临不同的挑战吗？我觉得大多数人还是选 ChatGPT。我们太早了。

This is the same for me. I run our business on Google too, and I mostly play with Gemini to see what it does. I'm curious how you think this plays out in consumers' heads. You've got OpenAI and Anthropic in one world where ChatGPT is a strong consumer brand, and Claude is dominant in coding and business. Google is asking people to run into AI as part of all these services. I don't think we've seen a competition quite like this. I'm old, so I go back to word processing days—people settled on Microsoft Word. In the browser wars, it was Internet Explorer vs. Netscape. Do you see these two groups having different challenges? I feel like most people still pick ChatGPT. We're so early.

我觉得很有趣——如果你一年前坐在这里，我们会说，‘OpenAI 和 ChatGPT 已经在消费者领域赢了；他们会一骑绝尘。’然后一年后，Anthropic 凭借 Claude Code 取得了突破性成功，这在当时有些可预见但并非超级可预测，并在收入上超过了他们。同时，Gemini 已经大量分发，并从包括 ChatGPT 在内的其他产品中抢占了消费者市场份额。所以我认为我们处于 AI 极其动态的阶段。很难说我们到了终局，因为未来会有许多尚未发明的新产品，面向消费者、开发者和企业，每个都可能比之前的更大。ChatGPT 是增长最快的产品，然后 Claude Code 又是增长最快的业务。这说明了 AI 的一些内在特性：随着 AI 达到新的智能和能力水平，它会解锁新的形态，每个都成为一波令人难以置信的新浪潮。所以下一波会更大，我们远未到终点。未来还会有更多激动人心的产品范式。

I think it's funny—if you were sitting here a year ago, we would say, 'OpenAI and ChatGPT have won on consumer already; they're going to run away with the whole thing.' Then fast forward a year, Anthropic had this breakout success with Claude Code, which was somewhat foreseeable but not super predictable, and overtook them in revenue. At the same time, Gemini has distributed quite a lot and eaten consumer market share from the rest, including ChatGPT. So I think we're in an incredibly dynamic phase of AI. It's very hard to say we're in the end game because there will be so many new products for consumers, developers, and businesses that haven't been invented yet, each potentially bigger than before. ChatGPT was the fastest growing product, then Claude Code is again the fastest growing business. This says something intrinsic about AI: as AI gets to new levels of intelligence and capability, it unlocks new form factors that each become an incredible new wave. So the next wave will be even bigger, and we're nowhere near the end. There will be many more exciting product paradigms.

产品过剩的问题确实存在：我们有这些不可思议的模型，但能做出什么消费者真正想用的东西？我也很好奇你如何调和普通消费者对 AI 的情绪。我二十多岁，不只在科技圈，我看到 Instagram 故事上有人发疯狂的内容说多讨厌 AI。情绪似乎很低落。而你们有数十亿用户，把 AI 作为按钮提供。你如何调和这种情绪？你在消费者方面看到了什么？

The product overhang question is real: we have these incredible models, what can we make that consumers actually want to use? But I'm also curious how you square the sentiment of the average consumer toward AI. I'm in my 20s, not only in tech, and I see crazy stuff posted on Instagram stories about how much people hate AI. The sentiment seems to be in the toilet. Then you guys have billions of users and serve AI as buttons. How do you square that sentiment? What do you see on the consumer side?

是的，AI 的情绪确实非常低，至少可以这么说。

Yeah, AI sentiment is definitely very low, to say the least.

我认为这归根结底是，在某种程度上，我们还没有真正展示出这如何成为个人赋能或个人能动性的工具，或者它如何让人们的生活变得更好。我觉得人们现在的体验是，它确实很有帮助，让生活好了一些，但并没有好到翻天覆地的程度。相比之下，对很多开发者来说，他们的生活已经完全改变了。大多数开发者对 AI 持非常积极的态度——也许有些复杂，但总体非常积极——因为他们现在能做以前做不到的事，能更快地构建更多东西，能在一个周末内完成整个项目。这真是个人能动性的惊人催化剂，但这个世界上的其他人还没有迎来那个时刻。到目前为止，我们还没有给每个人一个像 Claude Code 那样的工具，让他们能完成一直藏在心底的项目，让生活变得更好，或者突然实现他们的目标。这还没有发生。小企业也是如此。小企业主和创业者还没有完全体验到这一点。

And I think this comes down to on some fundamental level we haven't yet demonstrated in a very real way how this is actually a tool for personal empowerment or personal agency or how it just makes people's lives a lot better. Like I think people's experience right now is that it can be really helpful and it makes your life quite a bit better, but it's not overwhelmingly better. Versus I think for a lot of developers, I think their lives have actually totally changed and I think most developers have very positive sentiment, maybe somewhat mixed but very positive sentiment towards AI because they're now able to do things that they were just unable to do before and they can build so many more things faster and they can build entire projects over a weekend and it's just this incredible enabler of personal agency and that moment hasn't happened for everyone else in the world yet. So far we haven't yet given every person what the equivalent of Claude Code is that would enable them to do the projects they always wished they had in the back of their mind or make their life way better or all of a sudden enable them to accomplish their goals. That hasn't happened yet. Same thing even for small businesses. Small business owners and entrepreneurs haven't yet had that full experience yet.

所以，这正是我们在 Meta 努力的方向：给所有消费者和全世界的小企业提供非常强大的智能体，如果真能实现，会带来个人能动性的巨大提升，那会是什么样子？这将是疯狂的事情，因为如果你去美国任何一个小镇的餐馆网站，你会发现它自 2002 年以来就没更新过。所以，给每个人一个多智能体架构的产品听起来是一个巨大的飞跃。

So, that's really what we're building towards at Meta is what does it look like to give very powerful agents to all of our consumers and all the small businesses in the world and what does that look like if you're actually able to nail it in the form of a huge increase in individual agency? That would be a crazy thing to nail because if you go to a small town in anywhere in America and go to that restaurant's website, I mean, it hasn't been updated since 2002. So, giving everyone a multi-agent architecture product sounds like a huge leap.

而且，回到 Kylie 之前的问题，你看，Meta 做的事情有些我喜欢，有些我不喜欢。我觉得有很大一部分公众对这家公司相当不信任。就像你说的，AI 目前并不总是最受欢迎的东西。我确实觉得你们要赢得人们的信任，门槛更高。

And I mean, to Kylie's earlier question, I mean, look, there's things that I like that Meta does and then there's things I don't like. I think there's huge swaths of the public that view the company quite cynically. It just feels like you guys do have a... It's like you said, AI in general not always the most beloved thing at the moment. I do feel like the bar is higher for you guys to get people to trust you.

是的，100%。但我认为，如果我们思考我们能做的最好的事情，那就是我们应该构建最好的产品，让用户觉得真正惊艳。我认为我们可以构建改变大多数小企业主生活的产品。而且，我们在 Meta 上有数亿小企业。很多像你一样用 WhatsApp 运营业务，很多有 Facebook 页面或 Instagram 页面，很多使用我们的广告解决方案。所以，我认为那里存在一个机会，在某种程度上只有我们有，因为只有我们有数十亿用户、数十亿使用我们产品的人，以及数亿小企业。让我个人非常兴奋的一个想法是，如果你能为这个生态系统的双方——所有消费者和所有小企业——构建智能体，然后让这些智能体能够相互协作，那会是什么样子？Dario 总是谈论数据中心里的天才之国。我们则对在数据中心里构建一个智能体经济体感到兴奋。如果你从根本上改变经济中供需的运作方式，由智能体来中介，那会是什么样子？我认为我们可以朝着非常令人兴奋的方向发展。你完全正确，这必须与确保我们获得社会许可同步进行，让人们看到我们关心这些技术的部署方式，并真正让他们的生活变得更好。

Yeah, 100%. But I think that again, if we think about what are the best things that we can do, it's really we should build the best possible products that we think are genuinely amazing for those who use them. I think we can build products that can transform the lives of most small business owners. And we have again, hundreds of millions of small businesses all around the world that are on Meta. A bunch of them use WhatsApp to run their businesses like you do. A bunch of them have Facebook pages or Instagram pages. A bunch of them use our advertising solutions. So, I think there's an opportunity that exists there that on some level only we have because only we have billions of users, billions of people around the world who use our products and hundreds of millions of small businesses. And one of the ideas that gets me personally really excited is, if you're able to build agents for both sides of this ecosystem, for all the consumers as well as all the small businesses, then what does that look like when you enable the mechanism for those agents to work together and collaborate with one another. And so, Dario always talks about a country of geniuses in a data center. I think we're excited about building an economy of agents in a data center. What is that actually if you fundamentally change how supply and demand work in the economy and it's mediated by agents? I think there are very exciting things that we can build towards. And you're totally right that that has to be done in lockstep with ensuring that we have social permission and that people really care about that people see that we care about how these things are deployed and that we're genuinely making people's lives better as a result.

你们赢得人心和思想的一个明确方式就是开源这些东西，我是一个老开源粉丝，在哲学上相信它。那么，自从 Musepark 之后，我们在这方面进展如何？

Do you... I mean one place you guys had one clear hearts and minds was by making these things open source and I'm an old open source fan and kind of believe in it philosophically. So where are we going with that since Musepark?

是的，是的。模型比 Llama 时代强大得多，尽管那只是不久前的事。对我来说非常重要的一点是这些模型的安全性。所以，我们在高级 AI Scaling 框架中设立的一项内容是，当我们的模型触发各种安全护栏时，我们必须非常严肃对待，尤其是在生物化学、网络能力和失控方面。Musepark 在我们的测试中确实触发了一些安全检查，我们在发布的 Musepark 准备报告中详细说明了这一点。因此，Musepark 目前的形式不适合开源，但我们正在开发适合开源的模型版本。我今天早些时候的一个会议就是审查这方面的进展。所以我们很兴奋能继续支持开源生态系统，开发开源模型，我预计在未来几个月内会有更多分享，这对我们来说也是一个激动人心的里程碑。

Yeah, yeah, yeah. So models are a lot more powerful than they were even back in the Llama days, even though it's so recent. And one of the things that is very important to me is safety for these models. So one of the things that we instituted as part of our advanced AI scaling framework was we have to take very seriously when the models that we develop trigger various safety guardrails, especially around bio-chem, cyber capabilities, and loss of control. So Musepark in our testing did trigger some of those safety checks that we did, and we detailed all this in the preparedness report of Musepark that we published. So as a result, Musepark in its current form is not suitable for open sourcing, but we are working on developing versions of the model that are suitable to be open sourced. Literally a meeting I had earlier today was actually to review the progress on this. So we're excited to actually continue supporting the open source ecosystem and developing open source models, and I expect that we'll have more to share on that in the coming months, but that's an exciting milestone for us as well.

那么好吧，你们真的要坚持下去，因为你们做了开放计算项目，你们又在 Sun Microsystems 的老楼里了。我只是个历史迷。他们是开源软件的坚定拥护者，一直是微软所建世界的对立面。我认为这很重要。所以听起来你们是在承诺，这仍然是 Meta 会做的事情，这与大多数竞争对手截然不同。

So okay, you're really going to stick with it because you know, you did the Open Compute Project, you're in Sun Microsystems' old building again. I'm just a history nerd. They were such a champion of open source software and always kind of like this foil to what Microsoft had built in the world. And I kind of think it's important. So it sounds like you're saying for you guys, you're committing that that's still going to be something that Meta does, that's quite different to most of your competitors.

是的，我说过很多次了。我们会继续开源模型。但我们也必须认真对待安全性，所以我们会考虑我们最强大的模型是否足够安全以开源。

Yeah, I mean, I've said this a bunch of times. We will continue open sourcing models. But we also have to take safety seriously, and so we will consider whether our most powerful models are safe enough to be open sourced.

好的。还有一件事……如果你读过关于你在 Meta 任期的报道，有一件事很突出……我记得是《纽约时报》还是谁写了一篇关于 Alex 和 Zuck 看待世界的方式的文章。

Okay. What's... and then I mean, there's another... If you read the stories about your tenure at Meta, I mean, one thing that drops out is this... I think it was the New York Times or someone did this story on Alex and Zuck see the world one way.

他们非常注重研究，想要世界上最好的模型，而博兹和克里斯·考克斯更关注产品。Meta 是一家必须服务数十亿用户并尽可能降低成本的的公司。你可以说，它目前并不对其模型收费。我相信你可能预料到我们会问这类问题。那么你们在理念上是什么立场？在 AI 战略方向上是否存在分歧？

They're very research forward and want the best model in the world, and Boz and Chris Cox are more focused on products, and Meta is this company that has to serve billions of users and do so as cheaply as possible. You could argue that, and it doesn't charge for its models today. I'm sure you probably expected we would ask some question along these lines. But where are you guys philosophically? Is there division about what direction to take your AI strategy?

是的，首先，这份工作教会我的一件事是，主流媒体的新闻报道标准——你知道，八卦和报道之间的界限非常模糊。所以，你们并没有疯狂争吵？不，我不这么认为。我认为我们在重要问题上非常一致。我们都知道我们需要拥有非常先进的模型，既要支持我们的核心业务，也要为我们现有的用户和小企业打造最好的应用、产品和服务。在我加入 Meta 之前很久，我们就在开发商业智能体，这些都需要最好的模型。但我们也知道，我们需要构建最好的模型，然后将它们整合到我们的业务中，利用这些模型为消费者和平台上的企业打造出色的产品和服务。所以没有真正的分歧。像任何公司一样，我们会深入讨论问题，思考其影响，并确保每个人都能发表意见，但并没有大的矛盾。

Yeah, first off, the one thing this job has taught me is the bar for journalistic reporting at major outlets is, you know, the line between gossip and reporting is remarkably thin. So, you guys weren't fighting like crazy? No, I don't think so. I think we're all very aligned on what is important. We all know we need to have very advanced models both to support our core business and to build the existing apps, products, and services that we have for our users and our small businesses to be the best version they can be in the world. We've been working on business agents since long before I got to Meta, and those required the best models possible. But we also know that we need to build the best possible models and then integrate those into our business and utilize these models to build products and services that are incredible for our consumers and for the businesses on our platform. So there's no real disagreement. Like any company, we debate the points deeply, talk about them, think through the implications, and make sure everyone can chime in, but there's no major beef.

所以你认为那完全是……我也这么认为。关于 Meta 的事，就在你从 Scale 转到 Meta 之前，你在华盛顿做了很多事，警告中国在 AI 竞赛中的危险。当我看到你们做那笔交易时，我试图在脑海中理清这一点，因为我知道他们在新加坡设立办公室，保持一定距离。在我看来，与一家中国初创公司以及像 Meta 这样资源丰富的公司走得更近，似乎与你之前言论中的说法有些不同。你明白我的意思吗？

So you think that was just total... I think so, yeah. On the Meta stuff, right before you made this transition from Scale to Meta, you were doing a lot in DC, flagging the danger of China in the AI race. When I saw you guys do that deal, I was trying to square that in my head because I know they were putting offices in Singapore and creating some distance. It seemed like getting much closer with a Chinese startup and a company with resources like Meta seemed a little different to me from what you'd been saying rhetorically. Does that make sense?

是的，显然整个 Manus 的情况非常复杂。很难深入任何细节。但我要说的是，当你思考这些地缘政治问题时，你总是需要在某种程度上将人民与国家分开。我的父母来自中国。有很多非常优秀、有才华的中国人，他们中的许多人搬到了新加坡、美国或其他地方。我很幸运能和他们一起工作。这与我对中国共产党及其行动的整体看法，以及这对美国应如何制定整体战略的意义是分开的。我认为区分这两者很重要。硅谷科技界有时会倾向于不细致地看待这个问题，把所有涉及中国的事情混为一谈。Twitter 或 X 在这方面尤其不细致。

Yeah, obviously the whole Manus situation is pretty complicated. It's hard to go into any real detail. But what I will say is that when you think about these questions of geopolitics, you always have to separate the people from the state in some sense. My parents are from China. There are lots of incredible, talented people who are Chinese, and many of them moved to Singapore, the US, or elsewhere. I feel lucky when I get to work with them. That is separate from my overall beliefs on the Chinese Communist Party and their actions, and what that means for how the United States should think about its overall strategy. I think it's important to draw a distinction between these two. There's sometimes a pull inside Silicon Valley tech to be un-nuanced about this, to lump everything involving China together. Twitter or X in particular is un-nuanced about this.

Twitter 对任何事情都不细致。不是对任何事情都不细致。但对我来说，是否有幸在中国出生的优秀人才我们愿意与之合作，这完全独立于我关于美中整体地缘政治的看法。你不能评论，因为如果看起来是中国叫停了交易，如果你不能评论，那就意味着还有暗箱操作，可能还会发生什么。

Twitter is un-nuanced about anything. Not un-nuanced about anything. But I think that to me, whether there are amazing people who happened to be born in China whom we would love to work with is totally independent from what I believe about US versus China overall geopolitics. And you can't comment because if it looks like China shut the deal down, if you can't comment, that means there's still machinations at play, something could still happen.

我无法对此发表评论。

I just can't comment on it.

说到这种情绪，你之前登的那个关于 AI 战争的报纸广告是怎么回事？是《纽约时报》上那个整版广告吗？关于 AI 和战争，我们需要认真对待。你还记得你在 Scale 的时候吗？

Touching on that sentiment, what was that newspaper ad you put out about AI war? Was that in the New York Times, that full-page ad about AI and war and we need to take this quite seriously. Do you remember while you were at Scale?

是的，从更宏观的角度看，我认为那是一个感觉非常关键的时刻。我当时觉得，让美国政府理解 AI 将给国家安全、保卫国家和公民带来巨大变革非常重要。从某种意义上说，我们之后看到的事情，比如 Mythos 和其他重要事件，都证明了这一点非常正确。那是一个我们必须认真对待的时刻。有相当明确的证据表明，中国共产党和解放军一直非常重视 AI，认为它是一项对国家安全具有深远影响的技术。而那时，我们美国也必须同样严肃对待。我认为今天的美国政府非常非常重视 AI 在国家安全方面的意义。我们看到的大量事实证明，我和科技界及华盛顿的许多人所发出的呼吁已经被真正内化，我们今天正在非常深入地思考这个问题。

Yeah, zooming way out, I think that was at a moment that felt very critical. I felt it was very important at that time for the United States government to understand that AI was going to enable a large step change in what it meant for national security and defending our country and our citizens. In some ways, what we've seen since then, like Mythos and other meaningful events, have proven that to be very correct. That was a moment where it was important for us to take that seriously. There's pretty clear evidence that the Chinese Communist Party and the PLA have always taken AI extremely seriously as a technology with far-reaching implications for national security. And that was a moment where it was very important for us in the United States to take that as seriously. I think the US government today is taking AI very, very seriously as it pertains to national security. A lot of what we're seeing demonstrates that the plea I had, and many others in the tech ecosystem and in DC, have been really internalized, and we are thinking quite deeply about this today.

那么，你不认为 Anthropic 是过度悲观者吗？

So, you don't think Anthropic are over-doomers?

嗯，这是个复杂的问题。我认为取决于哪一部分。Anthropic 是……嗯，是的，取决于哪一部分。但总的来说，我认为当你听 AI 行业的人谈论 AI 时，重要的是要把他们具体说的话和他们想传达的核心信息区分开来。

I mean, that's a complicated question. I think it depends on which part. Anthropic are... Well, yeah, it depends on which part. But on the whole, I think whenever you listen to people in the AI industry talk about AI, it's important to separate the exact things they're saying from the core message they're trying to get across.

我认为 Anthropic 传达的一个核心信息——我觉得相当公允——是这些模型已经非常强大、非常能干，而且未来只会更强大、更能干。我们显然认为这可能是人类社会的巨大福音。如果我不相信这对人类有如此积极的影响，我就不会从事这项工作。我们非常关注的领域包括科学发现和健康。我们有一个专门的团队在做健康超级智能。我认为这可以是一项极其积极的技术。但同时，也必须考虑技术的风险，并确保我们认真对待这些风险。

And I think that some of the overall message from Anthropic, which I think is quite fair, is that these models already are very, very capable and very, very powerful. And they're only going to be more capable and more powerful going into the future. And we obviously think that this could be this incredible boon for human society. Like, I would not be working on this if I didn't believe that this could be so, so positive for humanity. Some of the areas that we care a lot about are scientific discovery and health. One of the things that we have a whole effort on is health superintelligence. Like, I think this can be an incredibly positive technology. But, it's also very important to factor in what are the risks of the technology and make sure that we're taking those seriously.

我想趁剩下的时间聊聊 Ashley 最喜欢的话题——你们刚刚收购了一家仿人机器人初创公司。能多谈谈你们的雄心吗？以及你们希望构建什么、如何利用这些模型将其带入现实世界？

I want to jump into Ashley's favorite topic with the time we have left, which is you guys just bought a humanoid robotics startup. Can you tell us more about those ambitions and like whatever you can tell us about what you're hoping to build and use I imagine these models to bring into the real world.

当然。我记得叫 Assured Robot Intelligence，简称 ARI。

Yeah, 100%. I mean, I think it was called Assured Robot Intelligence, ARI.

他们做硬件吗？

And they made hardware?

不，他们不做硬件。他们为各种硬件目标开发 AI。

No, they did not make hardware. They made AI for various hardware targets.

好的。

Okay.

是的。我认为，如果你认真对待超级智能，并认真对待我们将拥有非常强大的智能系统这一前提，那么你就会意识到我们将拥有数字超级智能。我们正在追求当前形式的超级智能，但不久之后，物理超级智能将变得极其重要和关键。因此，如果时间线很短——我们确实如此——强大的 AGI 能力即将到来，这意味着你必须认真对待机器人能力和物理智能，并将其作为未来几年需要构建的方向。所以核心前提是：如果你想作为一家公司构建超级智能，物理智能和机器人能力自然是你路线图上的连续部分。我认为随着时间的推移，我们将以各种方式应用这项技术。比如，我们将用它加速科学发现，加速商品制造，也会用它让人们在更本地化的层面上生活得更好——机器人如何让我们的生活更轻松。机器人技术的应用几乎是无限的，但另一个关键点是：我们确实认为，就像数字超级智能受益于 Scaling（规模扩张）一样，机器人智能也是如此。既然我们正在构建算力基础设施以实现这些系统和模型的大规模扩展，如果不将其与世界模型和物理智能的工作结合起来，那几乎是一种浪费。

Yeah. I think that if you take superintelligence seriously and you take very seriously this premise that we'll have very, very powerful intelligent systems, then you kind of realize we're going to have digital superintelligence. Like we're going to have the current form of superintelligence that we're targeting, but then not long thereafter, physical superintelligence becomes really, really important and very critical. And so, if you have short timelines, which we do, that very powerful AGI capabilities are coming, it just means that you have to take robotics capabilities and physical intelligence very seriously as something that you need to be building towards in the span of years. And so that's kind of the overall core premise is that physical intelligence and robotic capabilities are very much so on the natural continuum of what your road map has to be if you want to build superintelligence as a company. And I think there will be all sorts of ways that we apply this technology over time. Like I think that we will use the technology to accelerate scientific discovery. I think we'll use the technology to figure out how to accelerate goods manufacturing. I think we'll also use it to figure out how to make people's lives better in a more local sense. Like what does it look like for robots to make all of our lives way easier. So, I think there's obviously a near infinite number of applications of robotic technology but the other key part here is we really think that in the same way that digital superintelligence benefits from scaling, so does robotic intelligence. And so given that we are building the compute infrastructure to enable just massive scaling of these systems and these models, it's sort of like it almost be a waste if we didn't integrate that with efforts in world modeling and physical intelligence.

感觉你们真的想拥有硬件，把模型带入现实世界。但整个过程中，我不幸地想到了元宇宙没有腿的情况，以及批评者会如何看待从 Meta 引入仿人机器人。你们凭什么认为自己做对了？你们学到了什么，让你们觉得可以做到并改变这种声誉？

It feels like something you guys are really trying to own the hardware bringing the models into the real world but this whole time I am unfortunately thinking about the metaverse no legs situation and what critics might think about bringing humanoids from meta into the world like what makes you guys right to do this what have you learned that makes you feel like we can do this and change that sort of reputation.

我认为，最终存在一种可能性：我们可能被过去发生的事情吓到，以至于早上不起床，就待在家里。但我认为，我们对技术的潜力以及构建出色产品感到无比兴奋和深受启发。我通常相信，如果我们非常深思熟虑地构建优秀产品，并非常谨慎地部署和推向世界，我认为人们会对此感到兴奋。

I think that ultimately there's a world where we could be so scarred by what has happened in the past that we just didn't get out of bed in the morning and we just sort of stayed home. But I think that we are so excited and incredibly inspired by the potential of the technology and also just building amazing products. And I generally subscribe to the belief that if we build great products very thoughtfully and take a lot of care in how we deploy them and how we roll them out to the world, I think that people will be excited about those.

好的，我看时间快到了，我们马上要失去你了。能快速问答吗？Mango 模型是活的还是死的？

All right I'm just looking at the time we're going to lose you in a second can we go rapid fire real quick? Mango model live dead?

芒果们活蹦乱跳的。它们总是水果主题的。

The mangoes are alive and kicking. They're always through fruit themed.

我知道。我在想芒果是怎么长的，是长在树上还是藤上？不过算了，它们活得好好的。

I know I'm wondering how do mangoes grow do they grow on trees or I was going to say on the vine but anyway alive and well.

好吧，因为我在 AI 圈的朋友告诉我 mega bottle 那边有动静。又一个 AI 应用？这就是我说的，有太多虚假谣言，毫无现实依据。但我觉得，尽管我们自视甚高，我们受到的关注只是其他实验室的一小部分，所以我对这些闹剧深表同情。

Okay, what because my nerds in AI land were telling me there's things afoot with the mega bottle. Another AI app? This is what I'm talking about there's just so many spurious rumors that are not grounded in any reality. But I mean I feel like as much as we are self-important, we get a fraction that the other labs get so I have a lot of empathy for the drama of it all.

Nat Friedman 和 Daniel 是 John Carmack AI 项目的最大投资者。他一直很安静。他显然曾在 Meta 工作。你和他谈过吗？有没有可能重组团队？你知道他在做什么吗？

So Nat Friedman and Daniel were two of the biggest investors in John Carmack's AI effort. He's been very quiet. He obviously used to work at Meta. Do you talk to him? Is there any chance of getting the band back together? Do you know what he's doing?

我其实不太知道他在做什么。我不知道是否有人真的知道他在做什么。他显然是史上最伟大的程序员之一。所以我非常尊重他。

I actually don't really know what he's doing. I don't know if anyone really knows what he's doing. He's obviously like one of the GOAT programmers. So I respect him a huge amount.

我采访过 Priscilla Chan。CZI 正在向科学和生物技术投资数十亿美元。我不知道。你们在健康基准测试中得分很高，扎克显然也有兴趣。所以感觉你们有其他人没有的资源可以挖掘。这是否在计划之中？我不知道这些事情是否必须分开。

I interviewed Priscilla Chan. CZI is investing billions and billions of dollars into science and biotech. I don't know. You guys were scoring really high on these like health benchmarks and Zuck obviously has an interest there as well. So it just seems like man you guys would have resources to tap into that other folks wouldn't. Is that like a is that in the cards is that I don't know if those things have to be separate or

不不不，是的，我们将与 CZI 密切合作，构建最好的健康超级智能。正如我提到的，健康超级智能对我们非常重要。我们认为，让世界各地的人们平等获得强大的健康 AI 系统潜力巨大。这是我们能够独特地提供给全球数十亿人的东西，因为他们每天已经在使用我们的许多产品。所以，这对我们来说是一个非常令人兴奋且重要的举措。

No no no yeah we're going to be collaborating closely with CZI to build the best health superintelligence. As I mentioned, health superintelligence is so important for us. We think that there's just so much potential in enabling equal access all around the world to very powerful health AI systems. And that's one of the things I think we uniquely can deliver actually to billions of people all around the world because they use a lot of our products already every day. So, yeah, it's a very exciting and important initiative for us.

我知道你不想谈新模型，但告诉我们一件事，你们觉得自己在哪些方面做得与众不同或领先于其他人，你们认为已经搞明白了。

I know you don't want to speak about the new models, but tell us one thing that you guys feel like you're really doing different or ahead of everybody else on something that you think you figured out.

我们总是想用行动而非言语来证明。不过，我们对正在研发的模型感到非常兴奋。我们从 Scaling（规模扩张）中看到的结果让我们很激动，相信大家也会很兴奋。我们预期这些模型会在我们重点关注的某些领域达到最先进水平。

Well, we never want to show not tell. But I think we are really excited about the models that are cooking right now. We are really excited about the results we're seeing from scaling our models, and we think everyone's going to be pretty excited. We expect them to be state of the art in some of the areas we're really focused on.

从哲学上讲，你觉得你的方法与其他前沿实验室不同吗？你运营着这么大的实验室，我不确定我了解你对这项技术被释放到世界上的看法。

Philosophically, do you feel like you have a different approach to all the other frontier labs? You're running this massive lab and I'm not sure I know what you think about this technology being unleashed on the world.

有几点值得说。首先，我非常相信这项技术。我确实相信我们会拥有非常强大的 AI 系统，我们正在朝着这个方向努力，但其他人也一样。我们都在朝着超级智能迈进。首先，我们必须极其认真地对待安全。构建超级智能不可能不考虑所有安全风险，并确保尽可能减轻风险。这是我和你提到的一部分人达成共识的领域：安全极其重要。从我们为 MS Spark 准备的详细报告就能看出，这比 Meta 以往的报告更详细，这源于我们的承诺。作为 Meta，我们特别想构建的是一个个人超级智能的世界，广泛部署，让数十亿人都能使用。这是一项民主化的技术，将开启人类富足的时代。我们都拥有强大的工具，能够完成比以往更多的事情。我们被一个智能体经济所增强，在科学发现和健康方面取得进展。我一直在想：我们如何在地球上建造天堂？超级智能是实现这一目标的关键里程碑。最后一点：模型福利是一个越来越重要的话题。善待模型重要吗？我认为是的。在一个我们关心如何对待其他生物的世界里，认真思考如何对待模型是有道理的。我们关心以考虑模型主观体验的方式开发和部署模型。有研究可以测量模型的主观体验，Helios 就在做这件事。我认为没有人足够谈论这个话题，鉴于我们现在如此频繁地使用这些模型作为工作伙伴。

A few things worth saying. First, I am a huge believer in the technology. I do believe we're going to have very powerful AI systems, and we're building towards that, but so is everyone else. We're all building towards superintelligence. First, we have to take safety incredibly seriously. There is no such thing as building superintelligence without being very thoughtful about all the safety risks and ensuring we mitigate as many as possible. This is an area where I agree with a subset of the people you mentioned: safety is incredibly important. You've seen this in terms of our preparedness report for MS Spark, which was more detailed than Meta has historically, due to our commitment. Where we specifically as Meta want to build towards is a world of personal superintelligence, deployed very widely, with billions of people having access. It's a democratized technology enabling an era of human abundance. We all have tools of great agency, able to accomplish so much more. We're augmented by an agent economy making progress on scientific discovery and health. I always think: how can we build paradise on earth? Superintelligence is a key milestone. One last thing: model welfare is an increasingly important topic. Is it important to treat models well? I think it is. In a world where we care about how we treat living things, it makes sense to be thoughtful about how we treat models. We care about developing and deploying models in a way that is thoughtful about their subjective experience. There's research measuring subjective experience of models, and Helios does that. I think nobody is talking about it enough, given how much we use these models as work partners.

你有点像科幻迷。我听你其他访谈时，你谈到对 Neuralink 很投入，以及脑机接口对人类未来的意义。我感觉你有点……

You're kind of a sci-fi guy. I've listened to your other interviews where you talk about being dialed in on Neuralink and what BCIs could mean for the future of humanity. I'm getting the sense you're kind of...

我最喜欢的事情就是读科幻小说和在树林里散步。

My favorite things to do are read sci-fi and walk in the woods.

所以你融合了这两个世界：自然和我们的超人类主义未来。

So you're mixing these two worlds: nature and our transhumanist future.

如果你思考哪些技术是人类的关键路径，脑机接口绝对是其中之一。超级智能、机器人技术和脑机接口是关键路径领域。如果你思考我们今天的工作中哪些会扩展到无限远的未来，那就是能源、算力和机器人。

If you think about which technologies are critical path for humanity, BCI is definitely one of them. Superintelligence, robotics, and brain-computer interfaces are the critical path areas. If you think about what we work on today that will scale to infinity far into the future, it's energy, compute, and robots.

然后我觉得 Meta 在这些方面，尤其是脑机接口（BCI），比如运动神经元之类的东西，比我看过的其他一些 AI 公司下了更多赌注。但我的意思是，如果你相信这一点，我会说埃隆在机器人、能源、脑机接口方面比任何人都更全力以赴。这是否意味着你们是个人行为，还是 Meta 正在全面布局这些？

And then I feel like Meta on some of these fronts, especially around BCI, like the motor neuron stuff and everything, is making more bets than I see some of the other AI companies. But yeah, I mean, if that's what you believe, I would say Elon is a little more all-in on robotics, energy, BCI than anyone else. Does that mean you guys are personal or Meta is ratcheting all these?

我认为细节确实很重要。你必须分阶段构建这些。你必须先构建超级智能。这是能够构建其余部分的一个非常重要的前提。我的观点与埃隆不同的一个方面——我认为很多人的观点与埃隆不同——是我确实认为研究极其重要，构建超级智能从根本上来说是一项研究活动。在某种程度上，我们处于知识的战争迷雾中，我们试图通过实验在这片迷雾中探索，以理解构建超级智能意味着什么。那就是研究。所以我认为顺序很重要。我认为随着时间的推移，你如何推进很重要。我认为仔细考虑各个阶段的里程碑很重要。是的，我们在 FAIR 中的一个研究领域叫做 TRIBE。过去一年我们有一个里程碑，TRIBE V2，围绕构建用于大脑预测的基础模型。我们发现的一个很酷的结果是良好的零样本泛化。所以即使不知道你是谁，也没有任何关于你大脑的数据，我们也能合理预测你的大脑对各种图像、视频或音频的反应。我认为我们在许多关键领域都在下重要的赌注。

I think the details really matter here. You have to build these in stages. You do have to build superintelligence. That is a very important prerequisite to being able to be in a position to build the rest. And one area in which my opinions differ from Elon's—I think a lot of people's opinions differ from Elon's—is I do think research is incredibly important and that building superintelligence is fundamentally a research activity. On some level, we are in the fog of war of knowledge, and we are trying to do experiments to poke and prod in this fog of war to understand what it would mean to build superintelligence. That is research. So I think sequencing really matters. I think how you approach it over time really matters. I think being thoughtful about the milestones over time matters. But yeah, we're doing one of the research areas in FAIR called TRIBE. We had a milestone in the past year, TRIBE V2, around building foundation models for brain prediction. One of the cool results we found was good zero-shot generalization. So without even knowing who you are or having any data about your brain, we can do a reasonable job of predicting how your brain would respond to various images or videos or audio. I think we're making important bets in many of the key areas.

好了，我想我们要放你走了。你自从接受这份工作以来还没这么说过话。我们把你拖来拖去。我一般不这么做，但现在是开放时间——如果有什么我们没谈到，你觉得你还没机会告诉世界的，或者你想从这次经历中传达的任何东西，或者也许我们已经谈到了所有。我不知道。

Okay, I think we're going to set you free. You haven't talked like this since you took this job. We dragged you around. I never really do this, but open floor—if there's something we didn't hit that you feel like you haven't had a chance to tell the world, or whatever you want coming out of this experience so far, or maybe we hit everything. I don't know.

是的，不，我认为我们谈到了很多关键的事情。最终，我们在 Meta 真正要构建的是一个拥有大量个人赋权的世界。所以每个个人、小企业或企业家都能拥有不可思议的工具，使他们能够构建比人类历史上任何一个人都更多的东西。这对我们来说是一个非常令人兴奋的概念。然后，如何以一种方式做到这一点，同时与经济中的所有人类一起，赋能这个智能体经济，它在那里促进、优化并实现与人类并肩的惊人进步。而数据中心中的智能体经济有一个明确令人兴奋的结果，我们很高兴能在世界上创造它，而且我认为这对我们来说实际上是相当容易开发的。然后一路走来，推动不可思议的科学进步，通过健康超级智能大幅改善健康结果。在这段旅程中，有很多事情我们从根本上感到兴奋。

Yeah, no, I think we talked about a lot of the key things. Ultimately, what we really are building towards at Meta is how do you build a world that has a massive amount of personal empowerment. So each individual person or small business or entrepreneur has just incredible tools to empower them to build more than any human has been able to build in the history of humanity. That is an incredibly exciting concept for us. And then how do you do so in a way that also, alongside all the humans in the economy, you empower this economy of agents that is there to facilitate and optimize and enable incredible progress alongside the humans. And the economy of agents in a data center just has a clear exciting outcome that we're excited to be able to create in the world, and I think it's actually quite tractable for us to develop. And then all along the way, drive incredible scientific progress, drive dramatically improved health outcomes through health superintelligence. There's a lot of things that we're really fundamentally excited about on this journey.

好的。那么，谢谢你。非常感谢你抽出时间。我们时间到了。很高兴再次见到你。明年再见。太棒了。哦，不，再次见到你真的很酷。所以谢谢你过来。

Okay. Well, thank you. Thanks so much for making time. We're out of time. It's nice to see you again. See you again in another year. Out of the world. Oh, no, it is cool to see you again. So thanks for coming by.

是的，很高兴见到你们。

Yeah, good to see you guys.

Cold Fusion 播客由我 Ashley Vance 和/或 Kylie Robison 主持，或者我们两人一起主持。它由我和 David Nicholson 制作。我们的主题曲由 James Mercer 和 John Sortland 创作，节目始终由 John Sortland 编辑。非常感谢 Brex 和 SunCut Sun 的所有支持，最重要的是感谢所有收听或观看的人。我们爱你们。请给我们点赞、评论、订阅，所有这些了不起的事情。谢谢，我们下次再见。

The Cold Fusion podcast is hosted by me, Ashley Vance, and or Kylie Robison or both of us together. It is produced by me and David Nicholson. Our theme song is by James Mercer and John Sortland, and the show is edited always by John Sortland. Thank you so much to Brex and SunCut Sun for all your support, and thank you most of all to everybody for listening or watching. We love you. Please leave us a like, a review, a subscribe, all those tremendous things. Thank you and we'll see you again.