19 岁创办 Scale AI：亚历山大·王的创业之路

好了，让我们给 Alexander Wang 摇滚巨星级别的待遇。我们为什么不从后台聊起——我们在后台说，看待这场活动的一个酷炫方式是：这个房间里其实全是和我们一样的人，只是我们那时是 18 岁、20 岁——而且观众里还有一些 16 岁的年轻人。直接跳到你的故事吧。你一直很聪明，数学奥赛之类的。带我们回到那时的 Alex：你当时有什么感受？在想什么？是什么把你带上了这条路？

All right, full rockstar treatment for Alexander Wang, everyone. Why don't we start out—backstage we were saying that one cool way to think about this event is that this room is actually full of people who are just like us, but when we were 18 or 20—and there are some 16-year-olds in this audience too. Let's jump to your story. You were always really smart, math olympiad and all that. Jump us to the Alex of that time: what were you feeling? What were you thinking? And what drove you down this road?

是的，我在新墨西哥州的洛斯阿拉莫斯长大——现在因为奥本海默而出名——但那里真的是个偏僻的地方。我记得我参加了很多数学竞赛和计算机科学竞赛，但我知道自己想做大事情，只是不太清楚怎么做，或者具体的路径是什么。我有一个非常喜欢编程的朋友，高中毕业后他在硅谷实习过——我想他的第一份实习是在 Palantir。他对我影响挺大。所以高中毕业后，我来到硅谷的 Quora 工作。我在那里工作了一年，算是间隔年，然后去了 MIT。我在 Quora 工作时 19 岁，去 MIT 时 18 岁，创办 Scale 时 19 岁。我记得 17 到 19 岁那段时期，我感觉自己一直在变化，想做的事情也一直在变。我周围的人让我学到很多，我就像对着消防水带喝水，不断地被大量信息冲击。我强烈推荐两件非常重要的事：第一，在公司工作真的很有价值，因为从外部你完全不知道公司是怎么运作的，不知道真正做出一个东西是什么样，不知道迭代一个东西是什么样，也不知道一群人是怎么做决策的。所以我觉得那非常重要。第二，去 MIT 上学也非常重要，因为它给了我很多探索自己感兴趣的东西的机会。正是在 MIT，我开始训练第一批模型，摆弄当年刚出来的 TensorFlow。也正是在那里，我最终想出了 Scale 的创意。在 MIT 待了一年之后，我申请了 Y Combinator。那时候能进去感觉像个奇迹。YC 对我的创业旅程非常关键。YC 是那种很奇妙的组合——他们非常支持你，显然希望你成功，但他们也很真实，会直接告诉你什么时候你像个傻瓜。我觉得那是我们生活中都需要的东西。所以，那就是当时的故事：我 19 岁，创办了 Scale，剩下的就是历史了。

Yeah. I grew up in Los Alamos, New Mexico—which is now Oppenheimer famous—but it really was the middle of nowhere. I remember I did all these math competitions and computer science competitions, but I knew I wanted to do really big things. It just wasn't exactly clear how, or what the exact path would be. I had a friend who was really into programming, and after high school he got an internship in the Valley—I think his first internship was at Palantir. He was kind of an influence for me. So after I finished high school, I ended up working at Quora here in Silicon Valley. I worked there for a year, took a gap year, and then I went to MIT. I was 19 when I worked at Quora, 18 when I went to MIT, and 19 when I started Scale. I remember that period from 17 to 19—I felt like I was constantly changing. What I wanted to do was constantly changing. I was learning so much from the people around me; I felt like I was drinking from the firehose pretty constantly. I would definitely recommend two things that were really important. One: working at a company was really valuable, because from the outside you have no idea how companies work. You have no idea what it looks like to actually build something, to iterate on something, or for groups of people to make decisions. So I thought that was really important. Two: going to MIT was actually really important, because it gave me a lot of opportunity to explore what was interesting. It was at MIT that I started training my first models, playing around with TensorFlow, which had just come out that year. That's where I ultimately came up with the idea of Scale. After one year at MIT, I applied to Y Combinator. It felt like a miracle to get in at that time. YC was really critical to my entrepreneurial journey. YC is this amazing blend—they're very supportive and obviously want you to succeed, but they also give it to you very real and tell you when you're being a dumbass, which is what we all need in life. So that was the story till then: I was 19, started Scale, and the rest is history.

我猜你当时和 Jared Friedman 一起工作。你最初带来的想法，和后来真正的 Scale 其实很不一样。

I guess you worked with Jared Friedman at the time. And you came in with actually a very different idea than what ended up becoming Scale.

是的，我们想构建一个 AI 智能体——说来好笑——用来帮助人们获得医疗服务。这是一个很好的例子，我觉得这种想法最终一定会出现。我们现在甚至已经看到了：帮助人们获得医疗服务的 AI 智能体非常真实。但当时时机不对。我们大概做了一两个月，然后 Jared 把我们拉到一边说：“伙计们，我不知道这东西能不能成。”那正是我们需要听到的话。那时候我在 MIT 学过 AI，训练过模型，于是我们回到设计之初，深入思考机会在哪里，然后想出了 Scale。

Yeah, so we wanted to build an AI agent—funnily enough—to help people get medical care. It was a great example of an idea that I think will ultimately exist. We're even seeing it now: AI agents to help people get medical care are very real. But it was the wrong timing. We worked on it for about a month or two before Jared pulled us aside and said, “Guys, I don't know if this is going to go anywhere.” That's exactly what we needed to hear. At that time, I had studied AI at MIT and trained models, so we went back to the drawing board, thought deeply about where the opportunity was, and came up with Scale.

我猜当时售卖数据——大语言模型还没有真正崭露头角——但自动驾驶汽车正在兴起，计算机视觉突然变得重要。所以这算是第一个市场，对吗？

I guess selling data at the time—large language models had not really come to the fore yet—but self-driving cars were sort of coming up, and computer vision suddenly became important. So that was sort of the first market. Is that right?

是的。所以这里的背景是，我在 MIT 时做了很多项目，训练各种形式的模型。和今天相比，它们就像小玩具模型。我记得训练一个模型需要三样东西：一个 GCP 账户——在某个云服务上开账户获取算力；运行训练所需的代码；以及数据，也就是数据集。其中两样你可以在线上按一下按钮就拿到，但最后一样——数据——当时没有任何有效的方式去获取用于训练这些模型的数据。所以这让人觉得未来显然就在这里：会有一种方法，可以说是按一下按钮就能拿到数据。有趣的是，在接下来的几年里——Scale 最初的许多年里——数据仍然非常不性感。每次我们去融资，尽管数据很好、营收很高，VC 和投资人们总是非常怀疑。他们会说：“哦，我不确定这是不是好生意。它有长期性吗？它可持续吗？”这让我觉得很奇怪，因为那些投资人没有一个训练过模型。所以我猜他们真的不懂。快进到今天，我们成功融到了钱，继续前进，业务持续增长。但正是当年拒绝我们、对 AI 潜力非常悲观的那些投资人，今天在写评论文章，说数据多么关键，是 AI 领域最大的商业机会之一。所以看到整个事情周而复始，真的很有意思。

Yeah. So the story here is that when I was at MIT, I did a bunch of projects where I trained models of various forms. By comparison to today, they were little toy models. I remember to train a model, I needed three things. I needed a GCP account—an account on some cloud service to get compute. I needed the code to actually run the training. And I needed data, a data set. For two out of three, you could just press a button online and get them. But for the last one—data—there was no effective way to get data for training those models. So it felt incredibly obvious that this was going to be the future: there was going to be a way to, so to speak, press a button and get data. It was funny because in the years that followed—the first many years of Scale—data was still very unsexy. Every time we went out to fundraise, even though our numbers were great and we had great revenue, VCs and investors were always very skeptical. They'd say, “Oh, I don't know if this is a good business. Does it have longevity? Is it durable?” It was really weird to me, because none of the investors had ever trained a model. So I guess they didn't really get it. Fast forward to today: we managed to raise money, keep going, and keep growing the business. But the very same investors who passed on us and were very dour on the potential of AI are writing think pieces today about how data is critical and is one of the biggest business opportunities in AI. So it's very funny to see that whole thing come full circle.

我是说，这看起来确实是一个很好的第一性原理思考案例，对吧？你不能靠打开《华尔街日报》说“这个数据很热，我们就去做这个”来创办一家公司。你根本不可能用那种方式创办 Scale。你必须从那些你知道为真的、关于世界的简单陈述出发，然后基于此去构建一些东西。

I mean, it seems like that's actually a really good case study in first principles thinking, right? You can't start a company by opening the pages of the Wall Street Journal and saying, “Well, this data is hot, we're going to go work on that.” You literally couldn't have started Scale that way. You had to start from simple statements about the world that you know to be true, and then sort of build something for that.

我觉得如果你看看世界上所有最成功的公司，它们都是在核心想法流行之前很久就创立的。它们默默耕耘多年，直到这个商业概念成为共识。而你要成功，唯一的方法就是能在别人都还没看到的时候，早早认出这些关于世界的真相。最让我吃惊的一点是，在 Scale，我们做 AI 已经做了十年。你不能根据周围其他人说的话来做商业决策。如果太随大流，你会被搞得很混乱，最后一事无成。所以你得建立起自己的指南针，想清楚你眼中的未来会是什么样子，因为其他人只会让你困惑。

I think if you look at all the most successful companies in the world, they were started at a time long before the core idea was popular. They worked on it, toiling in obscurity for years before the concept of the business became consensus. And the only way you're going to be successful is if you're able to identify these truths about the world early, long before everyone else. One of the most surprising things is that at Scale, we've been working on AI for a decade. You just can't base your business decisions on what everyone else is saying around you. If you go too much with the herd, you will get immensely confused and end up nowhere. So you have to develop your own compass of what you think the future is going to look like, because everyone else will just confuse you.

看起来你特别擅长的一件事是，你从一个核心信念出发——“我们相信 X，没有人相信它”——但接下来经营公司的实际操作是：和投资人沟通，说服他们，不让他们打击你的士气；和那些应该马上就能理解的客户沟通；尤其是说服别人来为你工作。

It seems like one of the things you got incredibly great at was you start with this kernel of 'we believe X and nobody else believes it,' but then the mechanics of building the business are talking to investors and convincing them, not letting them demoralize you; talking to customers who should just get it; and especially convincing people to come work for you.

我觉得创办公司这些早期的操作，你可能是有点天赋的，但没有人一开始就会创办公司。我记得我最早见的一批投资人，他们很多人后来会说：“哦，你成长太快了，变化也太快了，我当时没看出来。”我想这一点对所有创业者来说可能都成立。没有人在做一件从未做过的事时是擅长的。所以对所有创业者来说，你一开始什么事都很烂，整个游戏的关键是如何让自己不断进步、快速变好、快速学习。

I think these early mechanics of building a company are things you might have some predisposition to be good at, but nobody is good at starting a company when they start a company. I remember talking to a lot of the investors I met very early on, and many of them would say, 'Oh, you just grew so quickly and changed so quickly, and I didn't see it at the time.' I think that's probably true for literally everyone who starts a company. Nobody is good at something they've never done before. So for all entrepreneurs, you start out pretty awful at everything, and the whole game is how to develop yourself to continuously improve, get better, and learn quickly.

后台休息时我们聊到，现在其实是创办公司的绝佳时机，因为你显然可以来参加 YC。这个房间里的这些人拥有彼此，这有点不可思议。但不仅如此，现在你有一个理想的个人 AI，它会告诉你，嘿，有些事情可以这样做。你觉得这会帮助你加速得更快吗？在你看来，在 AI 时代用 AI 创业是什么感觉？

Backstage we were talking about how this is actually a really lucky time to start a company, because obviously you can come do YC. You know, the people in this room have each other, which is kind of wild. But not only that, now you have an ideal personal AI that's going to tell you, hey, these are some ways to do it. Do you think that would have helped you accelerate even faster? What do you think it's like to start a company today with AI, in the age of AI?

是的，我真的觉得我们正处在一个不可思议的世界时刻——瓶颈不再是 AI 模型的进展，而是如何把它扩散到世界其他地方，帮助世界适应这种已经存在的惊人技术。就算模型从今天开始一点进步都没有，经济、世界运转方式和我们周围的一切，仍然会经历几十年的彻底颠覆和变化。因此，这大概是文明级别的机会，让你去当梦想家，有愿景、有雄心，并通过创造惊人的东西来塑造未来世界的样子。我们在后台聊到，我十年前创办 Scale 的时候，如果你创建一家公司，你是大卫对抗歌利亚。你必须很聪明，找到进入市场的角度，想办法在资源少得多的情况下竞争。现在我认为，有了智能体和 AI 的广谱力量，更像是歌利亚对歌利亚。创业公司可能是一个被智能体和 AI 力量大幅增强的机甲歌利亚，而大公司则是更传统的歌利亚。但我觉得现在的创业公司，如果你真正拥抱 AI 智能体，找到最雄心勃勃的方式来发挥它们的优势，你很容易超越现有巨头。

Yeah, I really think we're at this incredible moment in the world where the bottleneck is not the progress of the AI models. The bottleneck is diffusing that through the rest of the world and helping the world adapt to this amazing technology that already exists. Even if the models didn't improve at all from today, there would still be decades and decades of total upheaval and change in the economy, in how the world operates, and in everything around us. As a result, it's probably a once-in-a-civilization opportunity to be a dreamer, to have a vision, to have ambition, and to impose a view of how the future world should look by building something amazing. One of the things we were chatting about backstage is that when I started Scale ten years ago, if you started a company, you had to be David versus Goliath. You had to be clever, find an angle into the market, and figure out a way to compete even though you had many fewer resources. Now I actually think with the power of agents and AI broadly, it's much closer to Goliath versus Goliath. The startup is maybe a Mecha Goliath vastly enhanced by the power of agents and AI, while large companies are the more traditional Goliath. But I think startups now, if you properly embrace AI agents and figure out how to leverage their strengths in the most ambitious ways, you can easily outcompete incumbents.

那我们聊聊超级智能吧，因为你们实验室的名字里显然就有这个词。现在在 Meta 内部，超级智能操作上到底是什么意思？

So let's talk about superintelligence, because that's clearly even in the name of your lab. What does superintelligence mean operationally inside Meta right now?

一年前，马克写了一份关于个人超级智能的备忘录，我觉得这和你的个人 AGI（通用人工智能）概念其实非常相似。我们相信，世界上每个人——所有几十亿人——都会拥有一个适应他们、为他们量身定制的超级智能，能帮他们实现目标，了解他们的背景，最终扩展他们自身的行动力。我们想得很多的是行动力扩展。我们怎么帮助人们完成以前想都不敢想的事？如果一切都很容易，世界上每个人会去做什么？我们也从生态系统的角度来思考这个问题。就像帕特里克提到的，我们不相信那种极权式的、由 AI 控制世界的整体论。我们相信这些 AI 会增强这个非常广阔的生态系统。我们相信全世界几十亿人都会拥有自己的个人超级智能；我们也相信创业精神会大爆发。今天 Meta 的平台上有两亿家企业。我们认为，随着创造力的爆发和 AI 工具的普及，这个数字应该增长到数十亿。最终，这会是一个由商业智能体和个人智能体共同运作的动态生态系统，形成一个完全由 AI 强力加持的复杂生态。

A year ago Mark wrote a memo about personal superintelligence, which I think is actually very similar to your concept of personal AGI. We believe that everybody in the world—all the billions of people—are going to have a superintelligence adapted and tailored to them, that enables them to accomplish their goals, knows their context, and ultimately is an expander of their own agency. We think a lot about agency expansion. How do we help people accomplish things they couldn't have ever dreamed of before? And what would everyone in the world do if everything was just easy? We also think about this in an ecosystem way. As Patrick mentioned, we don't believe in this totalizing, totalitarian view of AIs that control the world. We believe these are going to enhance this very broad ecosystem. We believe in billions of people all around the world having their own personal superintelligence, and we also believe in an explosion of entrepreneurship. There are 200 million businesses on Meta's platforms today. We think that number should go to billions with this explosion of creativity and use of AI tools. Ultimately, we think it's going to be a dynamic ecosystem of business agents working with personal agents, developing a complex ecosystem that is fully AI-supercharged.

我看到 Meta Spark 1.1 真的很兴奋。我的 Claude 也特别喜欢它。运营一个前沿实验室的感觉怎么样？你知道，Meta Spark 的级别大概相当于 Opus 的级别。接下来会有什么新东西？还有，我觉得你越来越关注开源，我想在场的观众会特别喜欢。

I was really psyched to see Meta Spark 1.1. My Claude absolutely loved it. How has running a frontier lab been? You know, the Meta Spark level is sort of the Opus level. What's coming down the pipe? And also, I think you're increasingly looking at open source, which I think this audience really loves.

是的，已经——我在 Meta 大概一年了，这一年真的很不平凡。

Yeah, it's been—I've been at Meta for about a year now, and it's been quite a year.

我觉得，你知道吗，进入 Meta——我们公开聊过——Llama 4 没有走上 Meta 需要的轨道，所以我进去之后，我们做了一次从零开始的重建，思考如何构建一个完整的前沿实验室，在某种程度上是从零开始，当然也用了很多已有的东西，并且要尽可能快地推进。那时起九个月内，我们发布了 Muse Spark 1；两个月后，我们又发布了 Muse Image 和 Muse Spark 1.1。这其中有几件事让我印象非常深刻。第一是人才密度极其重要，那是我们押注的核心。人才密度是自然复利的：你拥有的优秀人才越多，就越多的顶尖人才想加入你。从内部看到这一切非常神奇。但前沿 AI 工作是研究，是科学工作。我们在探索能用这些模型做什么，如何推动它们，以及它们能够达成的边界在哪里。这需要一套完全不同于互联网公司或互联网产品的心智模式和运营模式。它更多关乎实验、科学和 Scaling（规模扩张）。归根结底，一切都关于你如何打造一个实验室、一套运营模式、一个系统，能够与生态中将发生的所有指数级增长一起复利——包括能力、算力、采用和使用量的指数级增长。我们正处于整个生态每一个维度都极其陡峭的指数曲线上。重要的是去发展一个有机体——我这样看待实验室——它能随那个曲线一起成长。这非常令人兴奋，我们还会发布更多东西。我们刚刚发布了 Muse Spark 1.1，这是一个很棒的模型。我们会继续更新 Muse Spark 产品线。我们还有更大的模型正在路上，我认为它们将比当今市面上最好的模型更有竞争力。我们很快也会推出一个 harness，并且一直在开发一个 harness，来赋能所有开发者和智能体式开发者。另外，正如你提到的，我们也在做开源模型。我们相信一个去中心化的 AI 能力、进步和发展的世界。我们希望赋能更广泛的生态和世界上每一个人，让他们用这项技术去构建和发展。我们有很多令人兴奋的东西在路上，我们希望尽可能大地赋能生态和开发者。

I think, you know, getting into Meta — we’ve talked about it publicly — Llama 4 wasn’t on the trajectory that was needed for Meta, and so I got in there and we did a zero-based build of how to build an entire frontier lab, in some ways kind of from scratch, obviously using a lot of what we had, and move as quickly as possible. Within nine months of that moment, we launched Muse Spark 1, and then two months later we launched Muse Image and Muse Spark 1.1. There are a few things about this that have really struck me. The first is that talent density was incredibly important. That was the core thing to bet on. Talent density is something that compounds naturally: the more talented people you have, the more of the most talented people want to join you. It’s amazing to see from the inside. But frontier AI work is research — it is scientific work. We’re exploring what you can do with these models, how you can push them, and what the reaches are of what can be accomplished with them. This requires a totally different mindset and operating model than existed for internet companies or internet products. It’s much more about experimentation, about science, about Scaling. Ultimately, everything is about how you develop a lab, an operating model, a system that can compound with all the exponential growth that will happen in the ecosystem — the exponential growth in capabilities, in compute, in adoption and usage. We are on a very, very steep exponent across every dimension of the ecosystem. It’s important to develop, like an organism, the lab — that’s how I think about it — able to grow with that. It’s been very exciting, and we’re going to be shipping a lot more. We just launched Muse Spark 1.1, which was a great model. We’ll continue to have updates on the Muse Spark line. We also have bigger models on the way that I think will be much more competitive with even the very best models out there today. We’re going to be launching a harness soon, and we’ve been working on a harness to help empower all the developers and agentic developers out there. And, as you mentioned, we’re also working on open-source models. We believe in a decentralized world of AI capability, progress, and development. We want to empower the broader ecosystem and everyone in the world to build and develop using this technology. We have a lot of exciting things on the way, and we want to empower the ecosystem and developers as much as humanly possible.

我的意思是，这听起来是其中一种方式——当然，当我在 OpenClaw 里使用 Muse Spark 时，它明显和 Opus 一样好，尤其是在那种带技能文件的智能体式流程中，但价格实际上便宜了 8 倍。

I mean, it sounds like one of the ways — certainly when I was using Muse Spark with my OpenClaw, it became clear that it was as good as Opus, especially for that sort of agentic flow with skill files, but it was like eight times cheaper, actually.

是的。嗯，这就说到了重点——我们不相信模型会贵到只配给最富有的开发者和公司使用。让每个人都能使用这项技术、并且用它构建他们想构建的东西，这一点很重要。我们还认为，最好的 AI 产品甚至还没有被开发出来。如果你看看 AI 生态里发生的一切，每一波浪潮都比上一波大 10 倍。当我创办 Scale 的时候，第一波可能是自动驾驶汽车。自动驾驶汽车非常棒，非常酷，但与大型语言模型和聊天机器人相比就相形见绌了。聊天机器人变得可能比自动驾驶汽车还要大 10 倍。然后几年后出现了编码智能体，编码智能体可能比聊天机器人大 10 倍。我们正处于这条陡峭的曲线上。我们将不断看到 AI 范式的新模态、新形态和新发展，每一个都会比上一个显著更大。所以我们的观点是：释放生态，去探索，去一起建设世界的未来。

Yes. Well, this goes to it — we don’t believe in a world where these models are so expensive that they get rationed only for the wealthiest developers and companies. It’s important for everyone to be able to use the technology and build whatever they want with it. We also take the view that the best AI products haven’t even been developed yet. If you look at the AI ecosystem and everything that’s happened, every wave is 10 times bigger than the past wave. When I started Scale, the first wave was maybe self-driving cars. Self-driving cars are really awesome, really cool, but that pales in comparison to large language models and chatbots. Chatbots became something probably 10 times bigger still than self-driving cars. Then there were coding agents a few years later, and coding agents are probably 10 times bigger than chatbots. We’re just on this steep curve. We’re going to keep seeing new modalities, form factors, and developments of the AI paradigm, each dramatically bigger than the last. So our point of view is: unleash the ecosystem, explore, and build the future of the world together.

那么，实际利用 Muse Spark 编码模型的最佳方式是什么？它是 OpenClaw，对吧？

So, what’s the best way to actually take advantage of the coding model from Muse Spark? It’s OpenClaw, right?

是的。目前最简单的方式是使用 OpenClaw。我们网站上有入门引导。很快我们也会有自己的 harness。最终，我们希望优秀的模型能接入所有可用的 harness，并尽可能激发生态中的组合式创新。

Yeah. Today the easiest way is to use OpenClaw. We have onboarding on the website. And soon we’ll have a harness of our own. Ultimately, we want great models that plug into all the available harnesses and empower as much combinatorial innovation in the ecosystem as possible.

是的，我知道这个 harness 还在保密中。但你能不能透露一点？我是说，我还在用 OpenClaw，我还在用 Hermes agent。这些东西，我称它们为总是在路边抛锚的法拉利。这是一个不会抛锚的法拉利吗？给我们点暗示吧。

Yeah, I know the harness is still under wraps. But can you tease us? I mean, I still use OpenClaw, I still use Hermes agent. These things are, I call them Ferraris that break down on the side of the road all the time. Is this a Ferrari that won’t break down? Tease us a little bit.

是的，希望它不会抛锚。我们确实非常专注于速度。对于任何使用这些工具的人来说，速度可能是最关键的因素之一。还有可靠性——就像你提到的，我们希望极其可靠。我们希望非常可扩展，能扩展到任何你想要的复杂而有趣的多智能体设置。坦白说，在 harness 之上还会有大量的创新，比如如何编排和建立循环，以及开发这些智能体协同工作的复杂生态。我们真的希望可扩展性很强，最终我们希望赋能人们去驾驭这项技术——无意双关。但我真心相信这些模型已经极其强大了。它们应该强大到足以推动 GDP 增长的许多增长点，而实现这一切要靠有远见、有雄心的聪明人。

Yeah, hopefully it doesn’t break down. I think we’re really focused on speed. For anyone who uses these tools, speed is probably one of the most critical things. Also reliability — like you mentioned, we want to be extremely reliable. We want to be very extensible and scale to as complex and interesting a multi-agent setup as you want. There’s so much innovation that will occur even above the harness, frankly, in terms of how to orchestrate and set up loops and develop very complex ecosystems of agents working together. We want to be really extensible, and ultimately we want to empower people to harness this technology — no pun intended. But I truly believe these models are already incredibly powerful. They should be powerful enough to fuel many points of GDP growth, and it’s up to smart people with vision and ambition to make that happen.

让我想想。那么，有一个问题。

Let’s see. So, one question.

回顾这十年，你觉得事后人们会说什么在 AI 上是显而易见的，而人们当下却正在错过？

When you look back on the decade, what do you think they'll say was obvious in hindsight about AI that people are missing in real time right now?

如今很多讨论都围绕这些模型到底变得多强、它们能否真正弥合这个差距、我们什么时候会迎来超级智能——是两年还是五年？我们会不会撞墙？在某种程度上，我觉得这类争论有点浪费时间，因为拥有非常强大的模型是不可避免的。我认为回望时我们会说，所有这些关于具体何时发生的争论都是短视的，因为现实是，我们整个文明正处在这条令人难以置信的指数曲线上。你不可能看着过去十年 AI 的进步而不被它已经走到的地步深深震撼。十年前，最好的 AI 模型只能识别 YouTube 视频里的猫。而现在，我们在和一个数字神明对话，它能——我的意思是，我们都见过这些系统的一些破解效果，见过它们能够做到的一些事情。你只能感到敬畏。我觉得这个趋势只会继续。这些模型会变得越来越强大。所以十年后回望，显而易见的是，智能变得充裕，能动性也变得充裕。我们目前所处的趋势只会一直延续下去。这会非常奇怪。我认为在人类历史上，一群聪明人为了共同目标聚在一起，一直就是进步的瓶颈。从某种意义上说，美利坚合众国就是这样一个例子：一群非常聪明的人聚在一起，对他们想要实现的未来有一个愿景。这几乎是美国每一家公司的故事，也是每一家 YC 公司的故事。这种情况将会改变。突然之间，稀缺的资源不再是智能或能动性。我真的认为稀缺的是愿景和雄心。你对自己想让世界在未来变成什么样有清晰的看法吗？你想在影响世界未来发展方向的天平上按下自己的手指吗？世界在 5 到 10 年后会以怎样的方式变得和今天不一样？你有没有雄心、有没有驱动力去熬过所有那些糟心事来实现它？AI 会让这一切更容易。AI 智能体比十年前让这变得容易了也许 10 倍或 100 倍。但另一面是，你突然可以梦想得更远。世界需要演化很多东西，我们才能真正拥抱这项技术。今天这个世界其实还几乎没有准备好迎接这项技术。作为建设者，我们有责任为世界做好准备。我们必须帮助企业和政府适应这项新技术。我们必须帮助想清楚如何从生物安全或网络安全的角度来保卫这个世界。我们必须想清楚如何管理这项新技术带来的所有风险。但另一方面，这也是人类史无前例的机遇时刻。我们可以发展新的科学。我们可以解决健康与生物学中那些长久以来悬而未决的问题。我们可以建立以前根本无法想象的新业务。还有以前不可能存在的新的创意机会。所以，这是一个充满机遇与风险的惊人摇篮，使得没有比现在更好的时代，去成为一个建设者，并拥有关于世界应该怎样改变的坚定观点。

So much of the debate these days is around how good the models are actually getting, whether they can actually bridge the gap, when we're going to get superintelligence — is it in two years or five years? Are we going to hit a wall? In some ways, I think that debate is a bit of a waste of time, because it's inevitable that we're going to have very powerful models. I think we'll look back and say all this arguing around exactly when it was going to happen was short-sighted, because the reality is that we as an entire human civilization are on this incredible exponential. You can't look at the progress of AI over the past decade and not be totally awestruck by how far it's come. A decade ago, the best AI models could recognize cats in YouTube videos. And now we're talking to a digital god that can do — I mean, we've all seen some of the hacks and some of the things these systems are capable of. You can't help but be awestruck. I think this trend will just continue. These models are going to become more and more powerful. So a decade from now, looking back, it'll be obvious that intelligence became abundant and that agency became abundant. The current trends we're on are just going to keep continuing. This will be very strange. I think for the history of humanity, groups of smart people getting together toward a shared goal was the bottleneck of progress. The United States of America, in some sense, was an example of this: a group of very smart people got together and had a vision for the future they wanted to enact. That's the story of nearly every company in America, and every YC company. That's going to change. All of a sudden, the scarce resource isn't going to be intelligence or agency. I really think it's going to be vision and ambition. Do you have a clear view of what you want the world to look like in the future? What is the one way you want to put your finger on the scale for how the future of the world will develop? How will the world look in five to ten years that it does not look like today? And do you have the ambition and drive to go through all the crap to make that happen? AI will make that easier. Agents in AI make that maybe ten times or a hundred times easier than it was a decade ago. But the flip side is that all of a sudden you can dream bigger. The world needs so many things to evolve for us to fully embrace this technology. The world is barely ready for this technology today. As a builder, we have a responsibility to prepare the world. We have to help enterprises and governments adapt to this new technology. We have to help figure out how we secure the world from a biosecurity perspective or cybersecurity perspective. We have to figure out how to manage all the risks we see with this new technology. But on the flip side, this is also a time of unprecedented opportunity for humans. We can develop new sciences. We can solve problems in health and biology that have been forever unsolved. We can build new businesses you couldn't have imagined before. There are new creative opportunities that couldn't have existed before. So it's this incredible cradle of opportunity and risk that makes it no better time to be someone who is a builder and has a strong view of how the world should change.

你觉得路径变了吗？我看到一个情况——我记得是斯坦福——计算机科学专业的入学人数实际上下降了两位数百分比。人们有点担心，这在我看来有点不可思议。即使要创造出那么好的智能体，你还是需要那些技能。也许将来不是这样。我不太确定。你自己改变了吗？你想对现在在座的这些人说什么？这是一个人们真正面对的问题：他们应该变得更 wordcel（语言型）、更少 shape rotator（空间型）吗？怎么做才对？还有，这是否改变了你在 Meta 现在想招的人，以及你管理团队的方式？

Do you think the path has changed? One of the things I saw — I think it was Stanford — the number of computer science majors actually dropped by a double-digit percentage. People are sort of worried, which is kind of insane to me. You still kind of need those skills to even create agents that are that good. Maybe that won't be true. I'm not really sure. How have you changed? What would you say to people in this audience right now? It's a real question people are facing: should they become more wordcel and less shape rotator? What's the move? And has that changed the kind of people you're looking to hire, and how you manage your teams right now at Meta?

我觉得系统性、严谨的思维仍然极其重要，因为抽象层在持续变化。我以前并不相信事情会这样发展，但确实就是这样。我在我那个年代创办公司的时候，我们写代码。而现在——我相信在座没有人再写代码了，这很荒唐——但现在的问题是，你如何把智能体编排在一起。然后是，你如何发展这些智能体的组织：怎么让一百万个智能体协作好？再然后会是怎么让一万亿个智能体协作好？我认为，在我们将要运作的抽象层上，如何结构化工作流，会一直有这种需求。那种严谨的系统性思维——传统上，在我创办公司的年代，你会从写代码开始，然后你会有由人类组成的组织，你需要弄清楚如何组织这些人。这需要系统思维。现在也许更接近的是：先编排智能体，再想办法编排这些智能体大军。但我认为系统思维永远不会过时。所以，全力押注 wordcel 肯定是错误的。你需要会 shape rotate。但我认为，未来更需要的，是在世界应该如何发展上有一个更深的罗盘和哲学观，因为人类历史上有非常多的经验教训，关于文明如何度过这样的时期。所以，未来十年人类的变化很可能会超过过去 100 年的变化。

I think systematic and rigorous thinking are still incredibly important, because the abstraction layer keeps changing. I didn't used to believe that this is how it was going to play out, but it really has. When I started a company back in my day, we wrote code. And now, I'm sure nobody here writes code anymore — that's ridiculous. But now it's about how you orchestrate the agents together. Then it's how you develop these organizations of agents: how do you get a million agents to work together well? And then it'll be how do you get a trillion agents to work together well? I think there's going to be this continued need to figure out how you structure workflows at the abstraction layer we're going to be operating at. That form of rigorous systematic thinking — traditionally, in my era of starting companies, you would start by writing code, then you would have organizations of humans, and figure out how to organize those humans. That requires systems thinking. Now maybe it's much closer to: first you orchestrate the agent, then you figure out how to orchestrate these armies of agents. But I think systems thinking is never going to go out of style. So it's definitely a mistake to go all in on wordcel. You need to shape rotate. But much more of what I think is necessary going into the future is having a deeper compass and philosophical view on how the world should develop, because there are many lessons from human history around how civilization gets through this period. So, humanity will change more in the next decade than it has in the past 100 years, probably.

因此，我认为，我们必须对未来抱有积极的愿景，并清晰连贯地阐述其发展方式，这比以往任何时候都更加重要。

And so, I think the imperative for us to have positive visions for that and coherent articulations of how that should develop are more important than ever.

我们来谈得更具体一些。我好奇的是，你在朋友中间或 Meta 内部看到的哪些 AI 应用可以与我们分享——这些应用在近期显然会很重要，但可能人们还没意识到？给我们一些内幕消息吧。

Let's get a little more concrete. One of the things I'm curious about is, are there applications of AI that you're seeing among your friends or internal to Meta that you can talk about—that are obvious near-term, but maybe people haven't figured out yet? I mean, give us some alpha.

我认为在智能体式循环中仍然存在巨大的机会——也就是如何开发系统，让你能在持续反馈循环中把 Token 花费提升 1000 倍甚至 100 万倍来推动结果。想想大多数公司，它们其实就是大型反馈循环，每个环节都有人类在操作。公司获得客户，然后想办法让客户更满意。客户越满意，消费就越多。消费越多，就能雇佣更多员工，这些员工再去获取更多客户并让客户更满意。从某种意义上说，这就是每家创业公司或企业的反馈循环。在循环内部还有微小的反馈循环。我认为开发能够运转并优化这些反馈循环的智能体系统，那里有巨大的 Alpha（超额收益）。我们在 Meta 内部见过这样的案例：如果你构建了正确的智能体循环，并为智能体设定了正确的评估或优化指标，一群智能体可以轻松胜过一支 100 名工程师的团队。所以，弄清楚充满各种智能体协调问题的世界会是什么样，是当今最有趣的问题之一。

I think there's still astronomical opportunity in agentic looping—figuring out how to develop systems that enable you to spend 1,000x or 1 millionx more on tokens to drive an outcome in a continuous feedback loop. If you think about most companies, companies are just large-scale feedback loops where humans are operating each of the edges. They get customers, they figure out how to make those customers happier. If customers are happier, they spend more. If they spend more, they can hire more people who then go figure out how to get more customers and make those customers happier. In some sense, that's the feedback loop of every startup or business. And within that, there are micro feedback loops. I think developing agentic systems that can operate and optimize these feedback loops is where there's huge amounts of alpha. We've seen internally at Meta cases where if you develop the right agentic loop and you have the right eval or metric for the agents to optimize, you can have a swarm of agents accomplish more than a team of 100 engineers, very handily. So figuring out what the world looks like with lots of agentic coordination problems is one of the most interesting problems today.

那从机制上讲——Markdown 文件、Cron 任务——是不是基本上就是把足够多的数据交给智能体，让它能找出一些可能不属于现有分布的东西？

So mechanically speaking—markdown files, cron jobs—is it basically pointing the agent at enough data so that it can figure something out that maybe isn't in distribution?

对，从机制上讲，先弄清楚指标是什么，然后就归结为技能、Markdown 文件、Cron 任务。

Yeah, mechanically figuring out what the metric is, and then it just comes down to skills, markdown files, cron jobs.

就是 /goal 吗？

/goal?

对，/goal。一旦你真正深入下去，就会发现一切都是那么平凡，这总是很有趣。

Yeah, /goal. It's always funny how mundane everything is once you really dig into it.

所以这并不神奇，你懂我的意思吗？有些人把它讲得很玄乎——LinkedIn 上有些帖子就在说那些神奇的方法。

So it's not magic, you know what I mean? Some people put a lot of magic—there are LinkedIn threads about some magic stuff.

我的建议是：别管 LinkedIn。LinkedIn 是获取客户的地方。

My advice: ignore LinkedIn. LinkedIn is where you get customers.

我想以这个问题收尾：你有一封电报要发给 18 岁的自己。你现在会对那个人说什么？也感谢你回来与观众分享你的智慧。你会用漂流瓶给 18 岁的自己寄去什么话？

I'd like to end on this: you get a telegram to send to the 18-year-old version of yourself. What do you say to that person right now? And thank you for coming back and sharing your wisdom with this audience. What would you send in a message in a bottle to the 18-year-old version of yourself?

这归根结底是要建立你自己的内心罗盘，用来判断未来会如何发展，并对此有坚定的信念。因为你会被各种噪音和人言淹没——这会令人困惑，也很艰难。尤其是年轻时，缺乏经验，你可能很难对自己相信的东西和想做的事抱持真正的信念。但我认为那是最重要的。正如我们所说，正是对我们所构建之物的坚定信念，才让我们能够经受住多年市场、行业和周围人混乱的风暴。我的另一条建议是：试着找出世界上那条既最陡峭又能持续最久的指数曲线。几十年前，那条曲线是摩尔定律——在当时显然是正确的投资方向。现在，那是 AI 的进步，但未来还会有更多这样陡峭的曲线。如果这些曲线起步时平淡无奇，甚至看似无趣，那也没关系。当我开始研究 Scaling（规模扩张）时，我们在 YouTube 视频里做猫的检测器。很难把这个故事讲成我们时代最重要的技术，但它确实处在一个难以置信的指数曲线上。

It really boils down to developing your own internal compass for how you think the future will develop, and having strong conviction in it. Because you'll get inundated with noise and people telling you things—it can be very confusing and very hard. Especially when you're young and you don't have experience, it can feel difficult to have true conviction in what you believe and what you want to do. But I think that's the most important thing. As we talked about, it took a deep conviction in what we were building to weather the storms of many years of chaos in the market, in the industry, and in the people around us. The other piece of advice I'd have is to try to identify what is the exponential in the world that has both the steepest curve and will go the longest. Many decades ago, that curve was Moore's law—clearly the right thing to invest in at the time. Right now, it's AI progress, but there will be more of these very steep curves in the future. It's fine if these curves start out very boring or not that interesting. When I started working on scaling, we had cat detectors in YouTube videos. It's hard to explain that story as the most important technology of our time, but it was on this unbelievable exponential.

我还有最后一件事要说。我们在 Meta 很自豪地给在座每一位提供 1,000 美元的新 Spark API 免费额度。太棒了。我们会继续改进模型。现在新 Spark 的价格我认为比 Opus 便宜 8 倍，如果你把它换算成 Opus 美元……

I have one last thing to say. We at Meta are proud to offer everyone in this room a thousand dollars of free credits for the new Spark API. Fantastic. We're going to keep making the models better. Right now, new Spark is, I think, 8x cheaper than Opus, so if you convert that to Opus dollars...

那可就多多了。但我们会让大家都能了解如何获得这些额度。我们非常期待看到大家构建的东西。让我们感谢 Alexander Wang！

It's a lot more. But we'll work to get everyone the details on how to get these credits. We're really excited to see what everyone builds. Alexander Wang, everyone!