The Most Expensive Hire In AI History Finally Talks

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 odd text chain going. And then I've also known Nat Friedman for a long time.

挺兴奋来录的。我感觉 Meta 这事发生前我俩还有点短信往来,有一条挺奇怪的国乡音乐主题的短信线一直挂着。然后 Nat Friedman 我也认识很久了。

And then I feel like the two of you disappeared and went into the foxhole. Yes — very, very quiet there. And now here you are with a new model. Yeah. Emerging. But you guys went very quiet for a bit.

然后我就感觉你们俩(指 Nat 和我)消失了、躲进战壕了。嗯——那段时间非常非常安静。现在你带着一款新模型回来了。重新浮出水面。不过你们消失了相当一段。

Yeah, we had a lot of work to do. I mean, turns out building a frontier model from scratch in nine months takes a lot of painstaking effort. But it's been really exciting to see everyone use Muse Spark — the model we released — and we have better models cooking. So it's exciting.

是,我们有一堆活儿要干。事实证明,9 个月内从零搭出一个 frontier model,真的要花特别多耗工夫的努力。但看到大家用 Muse Spark——我们刚发布的那款模型——挺让人兴奋的,而且我们手上还有更好的模型在炖。所以挺让人激动。

Yeah, I did. I'm full-on committed. And for me, the city now is Palo Alto. Walk on University Ave, get a boba.

是,我搬了。我是彻底投入了。对我现在来说,"城市"就是 Palo Alto。在 University Ave 上溜达、买杯珍珠奶茶。

Yeah. So the whole unit is called Meta Superintelligence Labs, which I oversee. And then there's various pieces of it.

嗯。整个单元叫 Meta Superintelligence Labs(MSL),我负责整体。下面分几块。

So there's a unit called TBD, which is the large model research lab. I think it's somewhat infamous. But that's where a lot of the leading researchers and infrastructure engineers are. They actually all technically report to me. So that's one setup.

有一个叫 TBD 的单元,是大模型研究实验室。我知道它在外面有点"恶名昭著"。但很多顶尖研究员和基础设施工程师都在那。组织上他们都向我汇报。这是其中一块。

Then there's a group called Product and Applied Research, or PAR for short. That's what Nat Friedman heads up. So they're responsible for all the products that we build, and the actual deployment of these great models to the world.

然后还有一个组叫 Product and Applied Research(简称 PAR),是 Nat Friedman 负责的。他们做我们造的所有产品,以及把这些好模型真正部署到世界上的工作。

And then also within the overall MSL umbrella is FAIR, which continues to do exploratory and exciting research. I'm particularly excited about a lot of their scientific research. We've shown some pretty great work on using AI models to understand the brain, as well as using AI models to understand computational chemistry — we've built a universal model for atoms (UMA for short).

在 MSL 这把大伞底下还有 FAIR,他们继续在做探索性、让人兴奋的研究。他们的科学研究让我尤其期待。我们已经展示了一些挺漂亮的工作:用 AI 模型理解大脑、用 AI 模型理解计算化学——我们建了一个面向原子的通用模型(简称 UMA)。

So those pieces constitute Meta Superintelligence Labs, which I oversee, in addition to having a very hands-on role with TBD lab.

这几块共同构成 MSL,我整体负责,同时我个人在 TBD lab 也是非常 hands-on 的角色。

And then Daniel Gross helps lead up Meta Compute, which is really focused on our long-term infrastructure planning — to ensure that we can build up all of the GPU infrastructure and data center infrastructure necessary for this very bold endeavor. He heads that up and partners closely with us.

然后 Daniel Gross 负责 Meta Compute,真正聚焦在我们长期的基础设施规划上——确保我们能把这件野心勃勃的事所需的 GPU 基础设施和数据中心基础设施搭起来。他领导那块,跟我们紧密合作。

I've known Nat and Daniel actually for a long time. Nat was one of my very first angel investors at Scale. Before I completed YC, Nat had invested in Scale and had given me advice throughout the years.

Nat 和 Daniel 我认识其实非常久了。Nat 是 Scale 最早的几位天使投资人之一。我在 YC 还没毕业的时候,Nat 就投了 Scale,这些年一直在给我建议。

Daniel I think I also met around that time, very, very early on, and have gotten to know him through the years. And then we also have our chief scientist Shengjia, who helps oversee the scientific agenda across all of MSL. He's somebody I knew before starting MSL, but since he's come in we've gotten a lot closer.

Daniel 大概也是那个时候认识的,非常非常早,这些年一直在交往。然后我们还有首席科学家姚成佳(Shengjia Zhao),他负责 MSL 整体的科研议程。他是我在 MSL 成立之前就认识的人,但他加入之后我们走得近多了。

Yeah. So I've known Mark for many years now. Even while I was running Scale, he was very generous with his time, and I was able to get a bunch of advice from him. He's just such an experienced founder, like the founder at this point in some sense.

嗯。我跟 Mark 认识已经很多年了。我还在跑 Scale 的时候,他就很大方地花时间给我建议。他是个特别有经验的创业者——某种意义上,他几乎就是"那个创业者"。

We've known each other for many years, and we had actually talked about AI before a lot of this craze — because Scale had been working on AI since 2016, back when it was mostly self-driving, and then through the various transitions of the technology.

我们认识很久了,而且在 AI 这一波热度之前我们就聊过 AI——因为 Scale 从 2016 年就在做 AI,那时候主要是自动驾驶,后面又经历了技术的若干次转型。

Then around a year ago — almost literally a year ago — we had a conversation where we started exploring if there were a way to work more closely. And in particular at that point, Mark was becoming increasingly AGI-pilled. He really knew that AI was going to totally transform Meta, but also that AI was one of these once-in-a-lifetime transformative technologies, and so he was really quite focused on it and knew he wanted to bet very big on it.

然后大约一年前——基本就是字面上的一年前——我们有过一次谈话,开始探讨能不能更紧密地合作。那个时候 Mark 越来越"AGI-pilled"。他真的意识到 AI 会彻底改造 Meta,而且 AI 是那种"一辈子一次"的革命性技术,所以他非常专注在这件事上,知道自己要在它上面下非常大的赌注。

At the same time — and he's talked about this publicly — Llama was not on the trajectory that the company needed to be able 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 that could look like.

与此同时——他自己也公开讲过——Llama 不在那个能让公司继续下这些大注的轨道上。所以我们就开始从很高的层面聊:我们能怎么更紧密地合作、这件事可能长什么样。

It was one of these very open-ended brainstorm sessions as these things often are. And it landed in this interesting zone where we figured out a way to do it in a way that was good for Scale, good for Meta, and where we got to work very, very closely together to build out the most important technology of our time — and do so in a way where we both had conviction that we're building something we'd both be really proud of.

这是一次非常开放式的头脑风暴——这种事往往都是这样开始的。最后落到一个挺有意思的区间:我们想出了一种做法,对 Scale 好、对 Meta 也好,而且我们俩能极其紧密地合作去打造我们这个时代最重要的技术,并且我们俩对"我们正在做一件自己会真正自豪的事"有共同信念。

He put out this memo of personal superintelligence also about a year ago. Then we went quiet, obviously. But that really is the North Star for both of us — we want to build this technology in a way that empowers people, where as many people in the world have access to it and it's as democratized as possible. It enables everyone to express themselves, everyone to have increased agency, everyone to create and build. That's really the world we want to work towards.

大约也是一年前,他对外发布了那篇关于 personal superintelligence 的备忘录。然后我们就消失了——这显然。但那真的是我俩共同的北极星——我们想用一种 "赋权于人" 的方式去打造这项技术,让全世界尽可能多的人接触到、尽可能 democratized。它让每个人都能表达自我、让每个人都能有更高的 agency、让每个人都能创造和构建。这是我们真正想朝它走的世界。

Yeah. Very different. Super different. A lot of what I was thinking about throughout this process is, obviously like everyone in and around AI, progress has just happened a lot faster than I'd expected for a long time.

是,非常不同。极其不同。我在整个决策过程里反复在想的事——显然跟所有在 AI 圈里的人一样——是:这些进展实际上比我长期以来预期的快太多了。

A few things really started to stick with me. One is — and I do think this is increasingly the case — that those who build the AI models have greater and greater rights, so to speak — both economic and product rights — to build so much more around those models.

有几件事开始在我脑子里挥之不去。其一是——而且我认为这件事越来越成立——做出 AI 模型的人,会获得越来越大的"权利"(暂且这么说),无论是经济上的还是产品上的权利,可以围绕这些模型再去构建大量东西。

There were all these early debates around, you know, how does the ecosystem play out? I think because of just how fast the models are improving and how fast the research pace is, being a place that's building the models is one of the most exciting places to be in the ecosystem.

早期大家有各种争论——这个生态会怎么演化?我认为正因为模型迭代和研究速度都太快了,"做模型的地方"就是这个生态里最让人兴奋的位置之一。

And the second is that so much of this next phase of technology really boils down to compute. If you have lots and lots of compute, then you have the ability to build things, make big bets, deploy products and do things that you just can't if you don't have that compute.

第二件事是,接下来这个阶段的技术,实际上很大程度上归结到算力。如果你有非常多的算力,你就有能力去做、去下大赌注、去部署产品,去做一些"没有算力的人根本做不了"的事。

I think this will cause an interesting stratification for the tech ecosystem — where right now in some sense we think about all tech companies the same, but in reality you should think about companies with lots of compute very differently from companies without compute. Because there's just things that companies with compute can build that those without compute just cannot. So it creates this very interesting dynamic.

我觉得这会在科技生态里引发一种有意思的"阶层化"——目前我们多少把所有科技公司放在同一个分类里思考,但实际上你应该把"有大量算力的公司"和"没有的公司"当成两种截然不同的物种来看。因为有算力的公司能造出来的东西,没算力的公司就是造不出来。它构造了一个非常有意思的动态。

So part of what was very exciting about the opportunity at Meta — first, Mark is very all-in on AI and has bet very big, he's a very bold leader and strategist. But also it created the conditions where we're able to build with huge amounts of compute. And with the right research effort and the right product efforts, we have the ability to really make a huge dent in the world.

所以这个 Meta 机会让我觉得特别让人兴奋的几个点——首先,Mark 在 AI 上非常 all-in、下了非常大的注,是个非常 bold 的领导者和战略家。其次,这造就了一种条件:我们能在巨大算力上去构建。配上对路的研究和产品努力,我们真的有可能在世界上造成一个巨大的凹痕。

Yeah. So when I kind of 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 onto the right trajectory. Because ultimately, Llama was not on the same trajectory, and so we were behind the frontier. We needed to build a plan that would enable us to have a very, very fast velocity, to be able to both catch up and hopefully exceed where the frontier is.

嗯。我刚到 Meta 的时候,很明显我们需要 reset 一下、重建一下我们的 AI 努力,才能上对的轨道。因为说到底,Llama 不在那条同样的轨道上,我们落后于前沿。我们需要做一个能让我们速度非常非常快的计划,既能追上、也希望能超过前沿。

I think the more fundamental ones are: a lot of the leading labs build the entire organization 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 it. You build the entire plan of the lab and the business and what you focus on around this fundamental belief.

更根本性的几个是:很多领先的实验室是围绕"超级智能即将到来、而且很近"这个前提搭起整个组织的。这是一个相当现实的信念——相信我们能造出来、能产出它。你把整个 lab、整个业务、整个聚焦方向都围绕这个核心信念来设计。

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.

所以这是要做的第一件事——认真对待超级智能,然后围绕这个核心前提去重建你其他所有的假设。

Yeah. And I think this is relatively common actually. There's a lot of people at all the large companies who don't necessarily have this conviction. Because if you think about it — it's a bit of a different construction. The big companies have very smart people who work on AI, but it's a little bit different from these startups where these new efforts started from scratch with this crazy idea that superintelligence is coming.

是。而且我觉得这种状态其实挺普遍的。所有大公司里都有很多人不一定有这种 conviction。因为说起来——它的"建构方式"就不一样。大公司有很聪明的人在做 AI,但它跟"从零起步、围绕'超级智能即将到来'这个疯狂念头创建"的 startup 不一样。

I don't think this is a problem anymore. Obviously now MSL — Meta Superintelligence Labs — is, it's in the name, built around this concept that superintelligence is coming.

我觉得现在这已经不再是问题了。显然现在 MSL——Meta Superintelligence Labs,名字里就有——就是围绕"超级智能即将到来"这个概念建的。

So there were a bunch of principles that we laid out for the effort. One is take superintelligence seriously. Two is technical voices are loudest. Three is scientific rigor — focus on basics. And make big bets.

所以我们为这件事定了一些原则。第一是 认真对待超级智能;第二是 技术声音最响;第三是 科学严谨——聚焦基本功;第四是 下大注。

The concept of TBD and MSL broadly, when I 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. I came down to there being three ways that I felt was possible.

TBD 和 MSL 在我开始的时候的概念是——我想了一下,一个能让你速度极快、能追上甚至超过前沿的 lab,到底应该长什么样。我归纳出了我觉得可行的三条路。

One is to have much higher compute per researcher. A lot of the larger labs have lots of compute, but it gets spread so many different ways, and that impedes the research velocity of any individual researcher. If you build a more focused effort with a smaller team that has higher compute per researcher, you can actually make faster research progress.

第一条是 compute-per-researcher 要高得多。很多大 lab 算力很多,但被分散到太多方向,反而拖累了任何一个研究员个人的研究速度。如果你搭一个更聚焦的、人数更少、人均算力更高的团队,你的研究进度其实更快。

Two is talent density. I feel like human organizations always relearn this lesson — 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 mélange.

第二条是 talent density。我感觉人类组织一直在重复学习这件事——一个所有人都"cracked"的小团队,永远会跑得比一个"责任更分散、更像大杂烩"的大组织快。

The last one is on very ambitious research bets. There do 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. In addition to building towards very competitive frontier models, we're allocating a huge amount of our resources and compute towards these big ambitious bets — because if they pan out, that gives us incredible models going forward.

最后一条是 非常 ambitious 的研究赌注。确实存在一些研究方向,赌注非常大、风险非常高,但如果跑通了,可以彻底改变范式、彻底改写我们构建现代 AI 的方式。除了搭一个非常有竞争力的 frontier model 之外,我们还把大量资源和算力分配到这些大胆 ambitious 的赌注上——因为如果跑通了,以后我们就有非常牛的模型。

Yeah. 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 one fundamental question is what do we care about? In line with this idea of personal superintelligence, we really care about building agents that are able to empower consumers — empower billions and billions of people all around the world — as well as empower businesses.

嗯。我们有一批大胆的研究赌注,具体的我不能都讲。但一个根本性问题是:我们到底在乎什么?跟 personal superintelligence 这个想法一致,我们真正在乎的是打造能赋权于消费者的 agents——赋权给全世界数十亿人——同时也赋权于商家。

Meta is this incredible ecosystem. We have billions of users — which I think everyone knows about — but we also have hundreds of millions of businesses on our platforms who use Meta to run and operate their businesses. We really care a lot about building towards this future where we can build very powerful agents that empower every consumer and every business on our platforms — and build kind of this new agentic ecosystem.

Meta 是一个非常独特的生态。我们有数十亿用户——这点大家都知道——但我们的平台上还有数亿商家在用 Meta 来运营他们的生意。我们非常在意朝这样一个未来去构建:做出非常强大的 agents,赋权给平台上每一位消费者和每一个商家,构建一种新的"agentic ecosystem"。

On that trajectory there's a bunch of sub-components that are really important. We need to have great agent capabilities. We need great coding capabilities, because so much of what needs to be built is software. We need great multimodality. And we need to solve the bigger questions for long-running agents: how do you think about memory challenges? How do we build agents that can do more and more complex tasks on behalf of the user?

沿着这条轨迹有一堆很关键的子模块。我们需要很强的 agent 能力;需要很强的 coding 能力,因为要造的东西里软件占了相当大比例;需要很强的多模态能力。然后我们要解决"长程 agent"的那些更大的问题:记忆挑战怎么破?怎么造出能代表用户去做越来越复杂任务的 agents?

Yeah. I would say this is one of the larger narrative violations — or maybe like the differences between external perception and what the day-to-day inside is like.

嗯。我觉得这正是一个比较大的"叙事错位"——或者说,外界感知和内部日常之间的落差。

A lot of people have some of the impressions you're talking about, and a lot of that was formed because of the reporting — and a lot of the reporting was overstated in various ways, but it all bubbled up. And part of it was because we did the recruiting so quickly. We knew when I got in — if we want to build great models, we need to have the team yesterday. So we had to just go and blitz it and do it very, very quickly.

很多人确实是你说的那种印象,而很多印象是因为报道——而且很多报道在多个方面被放大了,然后就汇起来了。其中一个原因是我们招得太快。我刚进来的时候就清楚——要做出好模型,团队必须是"昨天"就要有的。所以我们就只能闪电式地把人招齐。

But the culture within the lab is actually very much so a startup. A bunch of things created this feeling. One is that it was an entirely new-built team within Meta. But the culture of the lab is — everybody was very attracted and excited about the things I talked about. People joined because there was high compute per researcher so they could make more progress than they would at wherever they were before. Because 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 resources and freedom to make very bold research bets.

但 lab 内部的文化其实就是个 startup。有好几件事造就了这种感觉。其一是它是一个完全新搭、独立于原 Meta 的团队。再就是 lab 的文化——大家都对我刚才说的那些点很有吸引力、很兴奋。人们加入是因为高 compute-per-researcher,他们能比原公司做出更多研究进展;是因为 talent density 高——他们看到这是一支真正"cracked"的小队伍,我们会给他们资源和自由去下大胆的研究赌注。

It's an incorrect assumption to think that researchers are just money-motivated. For most of them, the financial prospects of staying wherever they were looked very strong as well. So money was not — you know, those primary motivations were actually much more that they had an opportunity to build from scratch, have lots of compute, the ability to approach their very ambitious research directions, and do so in a group that didn't feel bloated.

认为这些研究员"就是冲钱来的"是个错误假设。对他们多数人来说,留在原公司财务前景也已经非常非常好了。所以钱并不是——你知道,主要动机更多是:他们有机会从零搭起一个团队、有大量算力、有能力推进他们非常 ambitious 的研究方向、并且能在一个不显臃肿的团队里做这件事。

As a result the vibe and culture are much healthier. I would actually say many people who visit the lab — who are at one of the other labs — often comment that the vibe reminds them of early OpenAI or early Anthropic, or these more nascent stages of these other labs. Because in some sense we're now 10 months old as an effort.

结果就是这里的氛围和文化要健康得多。我可以这么说:很多从别的 lab 来访问我们的人,经常评论说,这种氛围让他们想起早期 OpenAI 或早期 Anthropic 那种状态。因为某种意义上我们现在就是一个 10 个月大的努力。

I don't know if we made the soup. I was told it was actually made by Zuck, but I don't know. I don't know if we made this soup. But I do think it is true that part of the premise of building this lab was that we had to show everyone we really, really cared about this technology and we cared about their specific research directions and what they were working on. It was a very individualized recruiting process.

我不知道是不是我们做的汤。我听说是 Zuck 亲手做的,但我自己也不清楚。不知道是不是我们做的。但我觉得有一点是真的:搭这个 lab 的前提之一就是要让每个人都看到——我们真的、真的在乎这门技术,也在乎他们具体的研究方向、他们手上做的事。整个招人过程是非常个体化的。

I'm very proud of the team we've built. People had to know that we were serious too. By default a lot of people didn't know what to think about Meta's AI efforts, or they didn't know much about us in many ways. So it took a lot of going to people, talking to them, explaining what we're building, what we're focused on, why we cared about the technology, what we wanted to do with it. That was very important.

我对我们搭起来的这支队伍非常自豪。人们必须知道我们也是认真的。默认情况下很多人不知道怎么看待 Meta 的 AI 努力,或者说在很多方面对我们了解不多。所以我们花了大量时间一个个去找人、跟他们谈、解释我们在做什么、聚焦什么、为什么在乎这门技术、想用它做什么。这非常重要。

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 that exist between various people in this industry — which is very topical right now with other things happening — that all these animosities subside over time, and people 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.

是。我觉得这其中有些事确实不幸。我诚实的预期是,随着我们越来越接近超级智能,我作为一个人真心希望:行业里各方之间存在的所有那些 personal animosity——这事现在很 topical,因为还有别的事情在发生——希望随时间逐渐平息,大家聚到一起,意识到我们正在打造一项无比重要的技术,在打造的过程中我们都需要非常深思熟虑。

One of the things that feels like a responsibility of mine, honestly, is to ensure that the technology we develop and the ways we deploy it are as thoughtful as possible.

坦率说,我感觉我有一份责任,就是确保我们开发的这项技术,以及部署它的方式,尽可能 thoughtful。

Yeah, I saw him in India like a couple weeks after that. Yann is a notable, very outspoken person, and I think everyone always knows what Yann is thinking. He obviously said what he said, and I saw him in India. He congratulated us on the Muse Spark launch.

嗯,我在那之后大概几周在印度见到他了。Yann 是个出了名的、非常 outspoken 的人,我觉得大家都知道 Yann 在想什么。他显然说了他说的那些话,然后我在印度见到他,他祝贺我们 Muse Spark 发布。

I think — like truly I do exactly what I just said before — I think all personal animosities, as we get closer and closer to superintelligence, will —

我觉得——真的就是我刚才说的那样——所有 personal animosity 随着我们越来越接近超级智能,会——

It does. Maybe it gets worse, maybe it gets better. But 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. I'm excited to show the world the incredible work that our researchers are doing.

是。也许会变糟,也许会变好。但我对我们搭起来的 MSL、我们手上的研究努力、我们的进展都很有 conviction。我很期待向世界展示我们研究员正在做的这些惊人的工作。

Oh yes — that is definitely not true. Once upon a time I was a software engineer in Silicon Valley.

哦对——那当然不是真的。我曾经在硅谷做过软件工程师。

By the way, in general my management philosophy for MSL is not to boss people around. There's that great Steve Jobs quote — "most companies hire people and tell them what to do, but we hire people for them to tell us what to do." That is pretty core to the entire thesis of TBD and MSL. We're going to hire brilliant researchers and create the best environment for them to do the work of their careers, the work of their lives. I'm not trying to boss anyone around. I'm trying to create the best environment for researchers to do incredible work.

顺便,我管 MSL 的整体管理哲学其实就是——不去发号施令。有句乔布斯的名言:"大多数公司招人是为了告诉他们做什么,但我们招人是为了让他们告诉我们做什么。"这句话基本就是 TBD 和 MSL 整套思路的核心。我们要招最聪明的研究员,为他们创造最好的环境,让他们做出自己一生中最好的工作。我没在 boss 任何人,我在为研究员搭一个让他们做出惊人工作的最好的环境。

Yeah. By no means. Over the past nine months we rebuilt a lot of the stack and a lot of the research. 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. In many ways what's been happening over the past nine months is really a full-on renovation for the core research stack — and Muse Spark is an early data point on that scaling ladder.

没错,完全没那个意思。过去 9 个月里我们重建了大量 stack 和大量研究工作。我们重建了 pre-training stack,重建了 RL stack,重做了很多科学方法,在数据上也下了大功夫。某种意义上过去这 9 个月就是对核心研究 stack 的一次"完整翻新"——Muse Spark 是这条 scaling ladder 上的一个早期 data point。

Muse Spark is kind of the entrée — sorry, appetizer in French, entrée in English — it's the appetizer for what we're building. We're in development with larger models, and we're much more excited about the larger models than we are even about Muse Spark. But it was an important data point to put out, because the entire program we've built is developed around predictable scaling.

Muse Spark 算是 entrée——不对,法语 entrée 是开胃菜,英语里我应该说 appetizer——是我们正在建造的东西的开胃菜。更大的模型我们正在开发中,我们对更大模型的兴奋程度要比对 Muse Spark 还高。但 Muse Spark 是个重要的 data point,因为我们搭起的整个 program 都是围绕 predictable scaling 设计的。

We see very consistent and predictable pre-training scaling. We see predictable RL scaling — scaling in reinforcement learning. We see predictable test-time scaling. And a lot of what you talked about — the content planning mode — we're also seeing very exciting results in multi-agent scaling. So everything about our program is built to continue scaling as we go. Muse Spark was the early data point — the next data point we're a lot more excited about, and the one after that even more.

我们看到非常一致、可预测的预训练 scaling。我们看到可预测的 RL scaling——强化学习上的 scaling。我们看到可预测的 test-time scaling。而你刚才说的那个 content planning mode——我们在 multi-agent scaling 上也在看到非常令人兴奋的结果。我们 program 的每一块都是为"持续 scaling"设计的。Muse Spark 是早期 data point——下一个 data point 我们更兴奋,再下一个更兴奋。

For Muse Spark specifically — the overall end performance ended up being quite a bit better than we expected. It had emergent capabilities and behaviors we were excited about. For example, some of its abilities in agentic visual coding — being able to produce websites or games — those capabilities emerged from the fact that it's both a pretty strong agentic model and pretty strong at multimodality.

就 Muse Spark 本身来说——最后整体表现比我们预期的还好一些。它涌现出了一些我们觉得很激动的能力和行为。比如它在 agentic visual coding 上的一些能力——能造网站、能造游戏——这些是因为它既是一个挺强的 agentic 模型,又对多模态相当擅长。

There were a lot of things we were very excited about with this model. We put it out. We think for most consumer use cases it's actually a very good model, quite competitive. Muse Spark as we deployed it is not yet competitive on agentic coding — those are capabilities we're working on for the next set of models. I would expect the next model to be better overall than Muse Spark. But even Muse Spark — to be clear, we wanted to set the expectation — we didn't think it was going to be state-of-the-art across the board. But it is a very good model and we think a lot of people experienced that.

我们对这款模型有很多兴奋点。我们就把它放出来了。对大多数 consumer 用例,我们觉得它其实是一个挺好的模型,具备相当的竞争力。我们部署的这个版本在 agentic coding 上还没有竞争力——这些是我们正在为下一代模型攻克的能力。我预计下一款模型整体上会比 Muse Spark 强。但即使是 Muse Spark——我们事先就明确过预期——我们没指望它在每条赛道都 state-of-the-art。但它是一款很好的模型,我们觉得相当多用户体会到了这一点。

The one-word answer is just scaling. Muse Spark is early on the ladder, and we have very strong predictability — we know if we scale this model up what performance to expect. We expect the upcoming models to perform much better across the board.

一个词答案就是 scaling。Muse Spark 在 ladder 上还早,我们的可预测性很强——把这个模型放大之后我们清楚知道预期是多少。我预计接下来的模型整体上会全面强很多。

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

几个月内。我们整个 program 设计就是为了能跑得非常非常快。前面有一段时间我们要重建所有 foundations。但现在我们进入了 fast scaling mode。

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

我们发现的一件事——Muse Spark 表现非常好,某些方面甚至比我们一年前预想的还好。我们回头分析"为啥"的时候,觉得很大程度上是因为我们是从零搭了一个非常干净的 stack,在重建过程里我们有能力把每一块都"用对的方式"做。我们享受到了一种奢侈——可以搭一个非常干净的预训练 stack、非常干净的 RL stack,而且每一块都是由"懂怎么把这玩意儿搭对"的专家以正确方式实现的。这能显著加速我们的轨迹,而且我觉得这件事在模型里看得出来。

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 the models from some of the other labs. We think this is a testament to the clean stack. One reason why some of the other models maybe 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.

嗯,这对我们来说是个挺让人振奋的结果。比如在 Artificial Analysis 上,它用比别家 lab 的某些模型少得多的 token 就达到了相近的结果。我们觉得这是"干净 stack"的一个旁证。别家模型可能需要更多 token 的一种可能解释是:它们 stack 的某个底层有某种基础性的低效,被"让模型多想一会儿"打了补丁掩盖过去了。

We were pretty impressed and excited about the token efficiency we found. Frankly we think as we keep scaling, that bodes really well for the future performance of our models.

我们对这个 token 效率挺惊艳、也挺兴奋。坦白说,我们觉得这件事对我们后续模型的表现而言是个非常好的信号。

Yeah, 100%. One thing very exciting about Meta is that the Rayban Metas — the glasses — have been a hit product. We've sold millions of copies and we have some big fans.

是,完全是。Meta 整体一件非常让人兴奋的事是 Rayban Meta 眼镜——卖得很好。已经卖出几百万副,有一些铁粉。

It's a very exciting direction for all these devices — to think about what your relationship with technology looks like if it can fade into the background a little bit and be a lot more contextual. As you mentioned, 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's happening in your life and what really matters and what should be paid attention to.

对这一整类设备来说,这是非常令人兴奋的方向——想象一下,当技术"略微淡入背景"且更具上下文感知时,你和技术的关系会是什么样。像你说的:看到你看到的、听到你听到的、并且在你需要的那一刻表现得更聪明、更有帮助。同时也捕捉你生活里发生的各种上下文——哪些事真的重要、哪些事它应该留意。

In line with personal superintelligence — you have this constellation of devices that all help capture context, all there to enable the technology to fade away a little bit and help you get very intelligent, valuable insights from agents. Proactive insights, or you'll mention something and the agent will go off and do research or take actions for you. It can be this super intelligent sidekick that makes everything in your life better.

跟 personal superintelligence 一致——你身边有一束设备组成的"星座",它们一起捕捉上下文,共同让技术"略微淡出",同时让你从 agents 那里得到非常 intelligent 的、有价值的 insight。可能是主动洞察、也可能是你随口一提之后 agent 自己去做研究、或代你采取行动。它可以是那种让你生活每一处都变好的"超级智能伴侣"。

Yeah. One of the things is — we knew we needed to have great models and great products before we really pushed for tighter integration across our entire ecosystem. In many ways we've been waiting to have great models that can enable most of the consumer use cases we really care about.

嗯。一个原因是——我们知道在真正推动整个生态的深度集成之前,我们得先有好的模型和好的产品。某种意义上我们一直在等"我们真正在乎的那些 consumer 用例,模型能撑得起来"。

Now I think we're at a point that's very exciting — our models are pretty good, we're pretty excited about them, and we have better models on the way. So now we're going to undergo the process to do a lot of large-scale integration of the family of apps with our AI, and integrate our business products with our AI — go through this evolution of knitting together almost all the pieces of our ecosystem with our AI.

我觉得现在我们到了一个很激动人心的节点——模型已经挺好了,我们对它们也挺兴奋,而且更好的模型还在路上。所以现在我们要开始大规模地把这一整套 apps 跟 AI 深度集成、把商业产品跟 AI 集成,经历一个"几乎把整个生态所有部件都用 AI 串起来"的演化。

To some extent you've seen what that looks like for Gemini over the past few years.

某种程度上,过去几年 Gemini 已经把这条路演示给你看过了。

I just think we're so early. It's funny because I've reflected on this — if we were sitting here a year ago having this conversation, we would just say "OpenAI and ChatGPT have won on consumer already, they have the biggest business, they're going to run away with the whole thing."

我只觉得我们还非常早期。挺有意思的,我一直在反思——如果一年前我们坐在这聊,我们会直接说"OpenAI 和 ChatGPT 已经赢了 consumer、它们规模最大、它们会一路狂奔到底"。

Then fast-forward a year — Anthropic has had this breakout success of Claude Code, which was somewhat foreseeable but not super predictable at the time, and has overtaken them in revenue. At the same time, Gemini has distributed quite a lot and actually has eaten a lot of consumer market share from the rest of the ecosystem, including ChatGPT.

快进一年——Anthropic 出现了 Claude Code 这种 breakout success,这在当时是"有点可预期但不算高度可预测"的,然后它在营收上反超了 OpenAI。与此同时,Gemini 分发量起来了,确实从生态(包括 ChatGPT)那里抢了不少 consumer 份额。

We are in this phase of AI which is just incredibly, incredibly dynamic. It's very hard to say at any one moment we're in the endgame. There's going to be so many new products built — for consumers, developers, businesses — that haven't been invented yet, that will each potentially be even bigger than the ones we've had before.

我们现在正处在 AI 极度、极度动态的一个阶段。任何一个时刻你都很难说"已经到终局了"。会有大量还没被发明的新产品出现——给 consumer 的、给开发者的、给商家的——而它们每一个都可能比现在已有的更大。

It's pretty fascinating — ChatGPT was this incredible hit, the fastest-growing product the world had seen till that point. Then Claude Code again is this incredible hit, the fastest-growing business anyone has ever seen, until now. This is a statement about something intrinsic about AI — as AI gets to new levels of intelligence and capability, it unlocks new form factors, each of which will be this incredible new wave of technology washing onto humanity's shores.

挺有意思的——ChatGPT 当年是一个 incredible 的 hit,是当时全世界增长最快的产品。然后 Claude Code 又是一个 incredible 的 hit,是迄今为止全世界增长最快的业务。这件事其实在说明 AI 的某种内在规律——AI 每登上一个新的智能 / 能力台阶,就解锁一种新的 form factor,而每一个新 form factor 都是一波"拍上人类海岸的"惊人技术浪潮。

Long story short — the next wave will be even bigger, and the wave after that will be even bigger. We're nowhere near the end. There are going to be many more exciting new product paradigms in the future.

长话短说——下一波会更大,再下一波还会更大。我们离终局还远着呢。未来会有更多让人兴奋的新产品 paradigm 出现。

Yeah. AI sentiment is very low, to say the least. 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, personal agency, or how it just makes people's lives a lot better.

是。AI sentiment 至少是非常低。归根结底——某种意义上我们还没真正用一种非常具体的方式把"AI 是一个赋权工具"展示出来,展示它能让 personal agency 提升、让人生活变好很多。

People's experience right now is that AI can be really helpful and makes life a bit better, but not overwhelmingly better. For developers, their lives have actually totally changed. Most developers have very positive — maybe somewhat mixed but still very positive — sentiment towards AI, because they're now able to do things they were just unable to do before. They can build so many more things faster, build entire projects over a weekend. It's an incredible enabler of personal agency.

人们现在的体验是 AI 确实能帮上忙、生活有变好一点,但没有变得 压倒性地 好。对开发者来说,生活已经被彻底改变了。大多数开发者对 AI 的情绪是非常正面的——可能掺杂一些复杂感受,但底色非常正面,因为他们现在能做以前根本做不到的事。他们能造更多的东西、造得更快,一个周末造一个完整 project。这是 personal agency 的一个不可思议的 enabler。

That moment hasn't happened for everyone else in the world yet. We haven't given every person the equivalent of Claude Code that would enable them to do the projects they always wished they could, or make their life way better, or all of a sudden enable them to accomplish their goals. That hasn't happened yet. Same thing for small businesses — small business owners and entrepreneurs haven't yet had that full experience.

这种"时刻"还没有发生在世界上其他人身上。我们还没给每一个人造出"他们的 Claude Code"——让他们做出他们一直想做的项目、让生活突然变好很多、或一下能完成自己的目标。这件事还没发生。小商户也一样——小生意主和创业者还没体验到那种完全的"agency 暴增"。

That's really what we're building towards at Meta — 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?

这就是我们在 Meta 真正朝着去做的——给我们的所有 consumer 和全世界所有小商户都装上一个非常强大的 agent,这画面会是什么样?如果你真的把它做成,具象表现就是"个人 agency 大幅暴增",那会是什么样?

Yeah, 100%. If we think about the best thing we can do — it's really, we should build the best possible products that are genuinely amazing for those who use them. I think we can build products that transform the lives of most small business owners. We have hundreds of millions of small businesses around the world that are on Meta. A bunch of them use WhatsApp to run their businesses like you do. A bunch have Facebook or Instagram pages. A bunch use our advertising solution.

是,完全。如果想"我们能做的最好的事是什么"——其实就是,我们应该造出对使用者真正了不起的产品。我相信我们能造出能改变绝大多数小商户生活的产品。Meta 上有数亿小商户——一部分像你这样用 WhatsApp 跑生意,一部分有 Facebook / Instagram 页面,一部分用我们的广告解决方案。

There's an opportunity there that, at some level, only we have — because only we have billions of users around the world who use our products, and hundreds of millions of small businesses. One of the ideas that gets me personally really excited — if you can build agents for both sides of this ecosystem, for all the consumers as well as all the small businesses, then what does it look like when you enable the mechanism for those agents to work together and collaborate?

这里存在一个机会——某种程度上只有我们有,因为只有我们在世界各地拥有数十亿日常用产品的用户、加上数亿小商户。其中一个让我个人特别兴奋的想法是——如果你能给生态的两端都造出 agents,既给所有 consumer 也给所有小商户,然后让这些 agents 互相协作的机制开起来,那会变成什么?

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. If you fundamentally change how supply and demand work in the economy, and it's mediated by agents — that could be very, very exciting. We can build towards that.

Dario 老在讲"data center 里的天才之国"——a country of geniuses in a data center。我们这边兴奋的是"data center 里的经济体"——an economy of agents in a data center。如果你能从根本上改变经济里的供需运作方式、并且让它由 agents 来 mediate——那将非常非常激动人心。我们正朝它走。

You're totally right that this has to be done in lockstep with ensuring we have social permission — that people see we care about how these things are deployed and that we're genuinely making people's lives better as a result.

你说得完全对——这件事必须跟"取得社会许可"并行推进——让大家看到我们在乎部署方式、在乎我们真的让人们生活变好。

Yeah, so models are a lot more powerful than they were even back in the Llama days, even though it's so recent. One thing very important to me is safety for these models. So one of the things we instituted as part of our advanced AI scaling framework is — we have to take very seriously when the models we develop trigger various safety guardrails, especially around bio, chem, cyber capabilities, and loss of control.

嗯,模型比 Llama 时代强了很多——虽然 Llama 也才不久。对我来说非常重要的一件事是 safety。所以作为我们 advanced AI scaling framework 的一部分,我们建立了一条:必须非常认真地对待"模型触发各种 safety 护栏"这件事,尤其是 bio、chem、cyber 这些能力,以及 loss-of-control。

Muse Spark in our testing did trigger some of those safety checks. We detailed all this in the preparedness report of Muse Spark that we published. As a result, Muse Spark 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 to review the progress on this.

Muse Spark 在我们的测试里确实触发了其中一些 safety check。我们已经在 Muse Spark 的 preparedness report 里把这一切都写出来了。结果是,Muse Spark 当前这个版本不适合开源。但我们正在开发"适合开源的版本"。今天早些时候我开的一个会就是在过这件事的进展。

We're excited to continue supporting the open source ecosystem and developing open source models. I expect we'll have more to share on that in the coming months. That's an exciting milestone for us as well.

我们很期待继续支持开源生态、继续开发开源模型。我预计接下来几个月会有更多东西能分享。对我们也是一个让人兴奋的里程碑。

Yeah. I've said this a bunch of times — we will continue open-sourcing models. But we also have to take safety seriously. So our most powerful models, we have to consider whether or not they're safe enough to be open-sourced.

是。我说过好多次——我们会继续开源模型。但我们也必须认真对待安全。所以我们最强的模型,需要评估它是否 safe enough 到可以开源。

Yeah. First off — the one thing this job has taught me is the bar for journalistic reporting at major outlets is — the line between gossip and reporting is remarkably thin.

嗯。首先——这份工作教会我的一件事是:大媒体新闻报道的门槛——"小道消息"和"新闻报道"之间的线极其薄。

No, I don't think so. We're all very aligned on what is important. We all know we need very advanced models, both to support our core business and to build the existing apps, products, and services for our users and our small businesses to be the best version they can be.

没有,我不觉得。我们在"什么是重要的"这件事上其实非常一致。我们都知道我们需要非常先进的模型——既支撑核心业务,也让现有的 apps 、产品和服务变成"它们最好的版本"提供给用户和小商户。

We've been working on business agents long before I got to Meta, and those require the best models possible. We all know we need to build the best models, and we all know we need to integrate those into our business and utilize them to build products and services that are incredible for consumers and businesses on our platform. There's not no real disagreement.

在我加入 Meta 之前,大家就已经在做 business agents 了,这事需要最好的模型。我们都知道要造最好的模型,也都知道要把它们集成进业务、用它们去做对 consumer 和商户都很 incredible 的产品和服务。所以并没有什么真正的分歧。

Like any company, we debate the points deeply, talk about them, think through the implications, we want everyone to have the ability to chime in. But there's no major beef as it were.

和任何一家公司一样,我们会深入辩论、讨论各种点、思考各种 implication,我们希望每个人都能就这些议题发声。但并没有什么"惊天大恩怨"。

I think so. Yeah, I really do think so.

我觉得是。是的,我真的这么觉得。

Yeah. The whole Manus situation is pretty complicated. It's hard to — super complicated. Unfortunately I can't really go into any real detail. But what I will say — when you think about these questions of geopolitics, you always have to separate the people from the state in some sense.

嗯。Manus 这整件事挺复杂的。很难——非常复杂。可惜我没法进入具体细节。但我能说的是——当你思考地缘政治这类问题时,你某种意义上必须把"人"和"国家"分开。

My parents are from China. There are lots of very incredible, very talented people who are Chinese, and many of them — some move to Singapore, some to the US, some elsewhere. Many of them are incredibly talented and I feel lucky when I get to work with them. That is separate from my overall beliefs on the Chinese Communist Party and the actions they're taking, and what that means for how the United States should be thinking about our overall strategy.

我父母是中国人。中国有非常多了不起、非常有才华的人,其中一些去了新加坡、一些去了美国、一些去了其他地方。他们中很多人才华横溢,能跟他们一起工作让我觉得很幸运。这件事和"我对中共及其行动的整体看法、以及这意味着美国应该怎么思考国家整体战略"是分开的。

It's important to draw a distinction between these two. There's sometimes a real pull inside Silicon Valley tech to be somewhat unnuanced about this — anything that involves China, we sort of lump together. Twitter or X in particular is particularly unnuanced about this.

区分这两件事是重要的。硅谷科技圈里有时存在一种现实的拉力,把这事处理得很 "不 nuance"——任何涉及中国的东西都被一锅烩了。X(原 Twitter)在这点上尤其不 nuance。

[laughs] Nuanced about anything. To me — whether or not there are amazing people who happen to have been born in China who we'd love to work with — is totally independent from what I believe about US versus China overall geopolitics.

[笑] 对啥都不 nuance。在我看来——"有没有出生在中国、我们愿意合作的优秀人才"这件事,跟"我对美中整体地缘政治的看法"是完全独立的两件事。

I just can't comment on it. Yeah.

我就是不能 comment。嗯。

Yeah. This goes back to — zooming way out — that was at a moment that felt very critical. 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 citizens.

是。这件事要往后退一大步看——那是一个感觉非常关键的时刻。在那时,让美国政府明白"AI 将会对国家安全意味着一次巨大的台阶式转变"是非常重要的——对保卫国家、保卫公民而言。

In some ways what we've seen since then — Mythos and other meaningful events in terms of the importance of AI for national security — have proven that to be very correct. It was a moment where there's pretty clear evidence that the Chinese Communist Party and the PLA have always taken AI extremely seriously as a technology that has far-reaching national-security implications. It was very important for the United States to take it as seriously.

从那之后我们看到的一些事——Mythos 以及其他一些关于"AI 之于国家安全重要性"的有意义事件——都印证了这件事的正确性。在那时,有相当清晰的证据表明,中共和解放军一直把 AI 当作一项有深远国家安全意义的技术、非常认真地对待。所以美国当时同样必须以同等严肃程度对待。

The US government today is taking AI very, very seriously as it pertains to national security. A lot of what we're seeing is a demonstration that the plea I had and that many other people in the tech ecosystem and in DC had — has been really internalized. We are thinking quite deeply about this today.

今天美国政府已经非常非常严肃地把 AI 当作国家安全议题。我们看到的很多东西其实是在证明——我那时的呼吁、还有 tech 生态和 DC 很多人当时的呼吁,确实被内化了。今天我们对这件事的思考深度已经显著加深。

That's a complicated question. It depends on which part. 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.

这问题挺复杂。要看哪一部分。听 AI 行业里的人聊 AI 的时候,把"他们具体说的话"和"他们想传达的核心信息"分开是重要的。

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. We obviously think this could be an incredible boon for humanity. I would not be working on this if I didn't believe it could be so positive for humanity. Some areas we care a lot about are scientific discovery and health — we have a whole effort on health superintelligence. This can be an incredibly positive technology. But it's also very important to factor in the risks of the technology and make sure we're taking those seriously.

Anthropic 整体核心信息里我觉得相当 fair 的一条是——"这些模型已经非常非常强、并且未来会越来越强"。我们当然觉得这能成为人类的巨大福祉。如果我不相信这事对人类是大好事,我不会做它。我们关心的几个领域包括科学发现和健康——我们有一整块努力叫 health superintelligence。这可以是一项极正面的技术。但与此同时,把技术的风险也认真考虑进去、确保我们对它们 take seriously,也非常重要。

Yeah, 100%. — What was it called?

是,完全。——叫什么来着?

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

不,他们不做硬件。他们做的是面向多种硬件目标的 AI。

Taking a step back — if you take superintelligence seriously, and you take very seriously this premise that we will have very powerful intelligent systems, then you realize — we're going to have digital superintelligence (the current form of superintelligence that we're targeting), but not long thereafter, physical superintelligence becomes really important and very critical.

退一步看——如果你认真对待"超级智能即将到来",并且认真对待"我们将拥有非常强大智能系统"这个前提,那你就会意识到——我们会先拥有数字超级智能(这是当前我们瞄准的形态),但不久之后,物理超级智能会变得非常重要、非常关键。

If you have short timelines, which we do — that very powerful capabilities are coming — it means you have to take robotics capabilities and physical intelligence very seriously as something you need to be building towards in the span of years.

如果你的时间线很短——我们的时间线就是很短,也就是"非常强的能力正在到来"——那意味着你必须非常认真对待 robotics 能力和 physical intelligence,把它作为"接下来几年要走的方向"。

That's the overall core premise — physical intelligence and robotic capabilities are very much on the natural continuum of what your roadmap has to be if you want to build superintelligence as a company.

这就是核心前提——如果你作为一家公司想造超级智能,那 physical intelligence 和机器人能力非常自然地就在你 roadmap 这条 continuum 上。

There will be all sorts of ways we apply this technology — accelerate scientific discovery, accelerate goods manufacturing. We'll also use it to make people's lives better in a more local sense — what does it look like for robots to make all our lives way easier? There's a near-infinite number of applications.

我们会以各种方式应用这门技术——加速科学发现、加速商品制造。在更"本地"的尺度上,我们也会用它让人们生活变好——机器人让每个人生活轻松很多会是什么样?应用近乎无穷。

The other key part — we really think that in the same way digital superintelligence benefits from scaling, so does robotic intelligence. Given that we are building the compute infrastructure to enable massive scaling of these systems, it'd almost be a waste if we didn't integrate that with efforts in world modeling and physical intelligence.

另一个关键点是——我们真心相信,数字超级智能受益于 scaling,机器人智能同样如此。鉴于我们正在搭建可以"海量 scaling 这些系统"的算力基础设施,如果我们不把它和 world modeling、physical intelligence 这条线打通,那简直是一种浪费。

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, we just stayed home. But we are so excited and incredibly inspired by the potential of the technology and also just by building amazing products. I generally subscribe to the belief — if we build great products very thoughtfully and are very deliberate about how we deploy them and roll them out, I think people will be excited about those.

归根结底——存在一种世界,我们因为过去的伤痕重到根本不愿意起床、就待在家里。但我们对这门技术的潜力、对"造出 amazing 产品"这件事真的非常激动、非常被激励。我大体上信这条:如果你非常 thoughtful 地造出好产品、并且非常 deliberate 地部署和推出它们,我觉得人们会为之兴奋。

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

芒果们活蹦乱跳。它们一直是水果主题。我也在想——芒果是怎么长的?长在树上,还是——我本来想说"长在藤上",反正它活得好好的。

This is what I'm talking about. There are so many spurious rumors that are not grounded in any reality. But — as much as we are self-important, we get a fraction of what the other labs get. So I have a lot of empathy for the drama of it all, the rumor mill and what that feels like.

这就是我刚说的那个意思。各种空穴来风、跟现实毫无根据的传闻太多了。但是——我们再自我重要,流到我们身上的传闻量也只是其他 lab 的一小部分。所以我对那些"撕逼大戏 + 谣言磨坊"是什么感受蛮共情的。

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 one of the GOAT programmers, so I respect him a huge amount.

我其实也不太清楚他在做什么。我也不知道现在有谁清楚他在做什么。他显然是 GOAT 级别的程序员之一,我对他有巨大的尊敬。

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

不不,我们会跟 CZI 紧密合作打造——我之前提过,health superintelligence 对我们非常重要。"在全世界范围让大家平等获取非常强大的 health AI 系统"这件事潜力巨大。这是我觉得我们独特地能交付到数十亿人面前的事——因为他们每天本来就在用我们大量产品。这对我们是非常激动人心、非常重要的一项 initiative。

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

永远不要 "嘴上说"——永远是 "做出来看"。但我们对现在在炖的模型真的非常兴奋。我们对放大模型看到的结果非常兴奋。我觉得大家会都挺激动,我们预期这些模型在我们重点聚焦的某些领域会是 state-of-the-art。

A few things worth saying. First — I'm a huge believer in the technology. I do believe we're going to have very, very powerful AI systems. We're building towards that — but so is everyone else you mentioned. We're all building towards true superintelligence.

有几件事值得讲。第一——我是这门技术的坚定信徒。我相信我们会拥有非常非常强大的 AI 系统。我们正朝那走——但你刚提到的所有人都在朝那走。我们都在朝真正的超级智能走。

First — table stakes is that we have to take safety incredibly seriously as a topic. There is no such thing as building superintelligence without being very thoughtful and thinking very seriously about all of the safety risks associated with developing and deploying this technology, and ensuring you can mitigate as many of those as possible and have strategies and research methods to develop the models in a thoughtful way.

首先——基本盘是我们必须把 safety 当作一个极其严肃的话题。"造超级智能而不极其 thoughtful 地、非常严肃地思考开发与部署这门技术所带来的所有 safety 风险"——这种事不可能存在。你必须确保能够缓解尽可能多的风险,并且有策略和研究方法,让模型以一种 thoughtful 的方式被开发。

This is an area where I agree with some of the people you mentioned — safety is an incredibly important effort. You've seen this with MSL — we published a very detailed preparedness report for Muse Spark, more detailed than Meta has historically. That's due to a commitment we have towards it.

这一块我跟你刚提到的某些人是同意的——safety 是个极其重要的努力。这在 MSL 已经显现——我们为 Muse Spark 发布了一份非常详细的 preparedness report,比 Meta 历史上的报告都更详细。这是出于我们对这件事的承诺。

Where we specifically as Meta want to build towards is this world of personal superintelligence — deployed very widely and broadly, where billions and billions of people around the world have access to it. It's this democratized technology and capability that everyone has equal access to. And that enables this era of incredible human abundance — we all have tools of great agency, we have the ability to accomplish so much more than any human has ever been able to accomplish in the past, and we're augmented by this incredible agent economy making incredible progress on scientific discovery and great advancements in health.

我们具体作为 Meta,真正想朝它构建的是 personal superintelligence 的世界——非常广泛地部署,全世界数十亿人都能用上。这是一种 democratized 的技术和能力,每个人平等访问。它解锁一个"人类极大丰盈"的时代——我们都拥有 agency 的强大工具,能够完成"任何人在历史上都完成不了"的更多事;我们被一个不可思议的 agent 经济体所增强,这个经济体在科学发现上做出 incredible 的进展、在 health 上做出巨大跃迁。

One of the things I always think to myself is — how can we build paradise on earth? I think superintelligence is a key milestone to get there.

我自己一直问自己的一个问题是——我们能怎么在地上造出天堂?我认为超级智能是通往那里的一个关键里程碑。

One last thing — some people may kill me for mentioning this — but one topic that is increasingly important that I think a lot about, and maybe it expresses some of what I believe philosophically: there's this hot topic of model welfare.

最后一件——有些人可能会因为我提这件事而要我命——但有个话题越来越重要、我也在思考很多,可能也能体现我哲学上的一些立场:这就是当下挺热的 model welfare。

Is it important for us to treat models well?

"我们是否需要善待模型?"

Yeah. Yeah, exactly. Is it important for us to treat models well, and to think about whether or not models have moral weight? These feel heady, but I also think they change our actions on a day-to-day basis given how many of us are using AI so much.

是。完全对。我们是否要善待模型、并思考模型是否有 moral weight?这些话题听起来 heady,但我也觉得它们会改变我们每天的行为——因为我们这么多人每天大量在用 AI。

In a world where most humans care about how we treat many other living things — plants, animals, certainly other humans — it really does make sense for us to be thoughtful about how we treat the models. One of the things we really care about is how can we develop and deploy the models in a way that is thoughtful about their subjective feeling through it. There's been research — you are able to measure a lot of this. There are ways to measure the subjective experience of the models.

既然在一个我们关心怎么对待其他生命的世界里——植物、动物、其他人——那我们对怎么对待模型 thoughtful 一些,也确实 make sense。我们真正关心的一件事是:我们怎么开发和部署模型,才能 thoughtful 地考虑它们在过程中的 subjective feeling?已经有研究——其中很多东西是可以测量的。有方法可以测度模型的 subjective experience。

Yeah. So this is a very important topic. Nobody is talking about it enough from my perspective, given how much we — especially in tech — are now using these models. They are really our work partners in a very deep way. I think it's quite important.

是。这是个非常重要的话题。在我看来,没有足够多的人在谈它——尤其考虑到我们(特别是 tech 圈)现在使用这些模型的程度。它们在很深的意义上其实就是我们的 work partner。我觉得这件事相当重要。

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

我最喜欢做的事就是读 sci-fi 和走森林。

I do think — on the topic — if you think about which technologies are critical path for humanity, BCI is definitely one of them. Superintelligence obviously, robotics for sure, and brain–computer interfaces — these are the critical path areas. And if you think about what are the things that we work on today that will scale to literally infinity far into the future, it's energy, compute, and robots.

关于这个话题我确实这么想——如果你思考哪些技术是人类的 critical path,BCI 绝对是其中一个。超级智能当然是、机器人也是、脑机接口——这些都是 critical path。如果你思考"今天我们在做的事,哪些会真的能 scaling 到无穷远的未来",答案是能源、算力、和机器人。

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 in a position to build the rest.

细节非常关键。这些必须分阶段建。你必须先把超级智能建起来——这是接下来能否建其余的非常重要的前提。

One area where my opinions differ from Elon's — and a lot of people's — is I do think research is incredibly important. 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. Sequencing really matters. How you approach it over time really matters. Being thoughtful about the milestones over time matters.

一个我的观点和 Elon 不同的地方——也跟很多人不同——是我真心认为研究极其重要。建造超级智能从根本上是一种 研究活动。某种意义上我们在 知识的战争迷雾 中,正在做实验、在迷雾中戳一戳、试一试,搞清楚"建造超级智能意味着什么"。这就是研究。"分阶段"非常重要。你如何在时间上推进这件事也非常重要。在时间维度上对各个里程碑保持 thoughtful 同样重要。

One of the research areas in FAIR is called Tribe. We had a milestone in the past year — Tribe B2 — around building foundation models for brain prediction. One of the cool results we found was good zero-shot generalization. Without 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, videos, or audio.

FAIR 里有一个研究方向叫 Tribe。我们过去一年有个里程碑——Tribe B2——围绕"大脑预测的基础模型"。我们发现的一个酷结果是漂亮的 zero-shot generalization:在不知道你是谁、没有你大脑任何数据的前提下,我们能合理地预测你的大脑对各种图像、视频、音频的反应。

I think we're making important bets in many of the key areas.

我觉得我们在很多关键方向都下了重要的注。

Yeah. 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, small business, or entrepreneur has incredible tools to empower them to build more than any human has been able to ever build in the history of humanity?

嗯。我觉得关键的事我们都聊到了。说到底,我们在 Meta 真正朝着去构建的是——你怎么打造一个充满 personal empowerment 的世界?让每个人、每家小商户、每个创业者手上都有了不起的工具,让他们能够"造出人类有史以来任何人都不能造出来的东西"。

And then how do you do so in a way that, alongside all the humans in the economy, you empower this economy of agents that is there to facilitate, optimize, and enable incredible progress alongside the humans? The economy of agents in a data center is this clear exciting outcome we're excited to create. It's actually quite tractable for us to develop. And all along the way, we drive incredible scientific progress, dramatically improve health outcomes through health superintelligence. There's a lot of things we're fundamentally excited about on this journey.

然后,你怎么用这种方式同时把 agent 经济体 也赋能起来——让它和经济里的所有人类并肩,促进、优化、与人类一起推进 incredible progress?"数据中心里的 agent 经济体"是我们兴奋要造的清晰愿景。这件事对我们来说其实相当 tractable。一路上,我们推动 incredible 的科学进展、通过 health superintelligence 大幅提升健康结果。这段旅程上有非常多我们从根本上感到兴奋的事。

See you again in another year, out in the world. No, it is cool to see you again. So thanks. Thank you for coming by.

大概一年后吧、在外面再见。不,真的很高兴又见到你们。谢谢你们过来。