从迷茫到规模化：Alex 的创业之旅

请大家一起欢迎 Alex Wang。我们非常高兴能邀请到你。Alice 说，SBC 非常擅长创始人旅程中“负一”的阶段，也就是在那个曲折、混沌的时期，尝试各种想法、压力测试、与人交流。那么，请带我们回到你自己旅程的“负一”阶段。在最终确定 Scale 之前，你尝试过不同的想法吗？早期的构思阶段是什么样的？

So everyone please join me in welcoming Alex Wang. We are very excited to host you. Alice says you know SBC really specializes in the minus one part of the founder journey, and that really means in that squiggle, in that kind of soup of trying out different ideas, stress testing them, talking to people about it. So take us back to the minus one days of your own journey. Did you try out different ideas before you kind of converged on Scale? What were the early days of ideation like?

是的，我们参加了 YC。在 YC 批次的前半段，可以说是我们所谓的“曲折期”。存在很多存在主义的焦虑；你真的不知道自己在做什么。我记得你会有很多创业想法的 Google 文档，不断尝试想出更多点子。我读了 Paul Graham 的文章《如何获得创业想法》，那是一个很好的聚焦框架。基本思想是，你活在未来，看看未来有什么，然后反向构建。但当时非常不清晰。在 YC，周围有很多人在做各种事情，所以很难知道什么是好主意。你还会觉得，任何你开始做的新想法，周围的人都已经做了一段时间了，所以你一开始就落后了。这带来了很多动荡。但最终我们建立了 Scale。这一切就像一场完美风暴，充满了机缘巧合。这个想法是，根据 Paul Graham 的框架，未来我们显然会更动态地编排人类算力，就像我们今天编排任何其他资源一样。所以人类计算将像 API 一样易于使用，而那个 API 当时还不存在。这就是 Scale 最初的想法。有了想法后，我花了一个晚上找域名。scaleai.com 是可用的，所以我买了它。结果证明这是一个异常好的决定。Scale 这个名字非常好。然后我们在 Product Hunt 上发布，获得了足够的关注。之后大概有 4 到 6 个月的“游荡”模式，你有几个客户，努力争取更多。我记得我会亲自和任何进入公司 Intercom 的人交谈。如果你访问 scaleai.com 并点击聊天气泡，你就会和我聊天。当时仍然非常不确定是否能成功。直到 6 个月后，我们开始有一个客户想要做得很大，然后这个想法才开始成熟。但大概有一年左右的时间，我称之为“游荡期”，我认为这对大多数初创公司来说已经算短的了。

Yeah, so we did YC, and the first half of our batch, I would say, was probably what we refer to as a squiggle. There was a lot of existential angst; you don't really know what you're doing with your life. I remember it was like you have these Google Docs of startup ideas and you keep trying to come up with more startup ideas. I read Paul Graham's essay 'How to Get Startup Ideas,' and that was a really good focusing framework. The basic idea is you live in the future, look at what exists in the future, and build backwards. But it was very unclear. When you're in YC, there are a lot of people working on many things around you, so it's hard to know what is a good idea. You also feel like any new idea you start working on, everyone around you has already been working on something for a while, so you're already behind the eightball. There's a lot of turmoil associated with all that. But ultimately we built Scale. It was kind of a perfect storm how it all came together, with a lot of serendipity. The idea was, per the Paul Graham framework, in the future we're clearly going to orchestrate human compute much more dynamically, like we orchestrate any other resource like compute today. So human computation will be as easy to use as an API, and that API didn't exist. That was the original idea for Scale. After we had the idea, I spent a night just looking for a domain. scaleai.com was available, so I bought it. That turned out to be an unusually good decision. Scale has been a really good name. Then we launched on Product Hunt, and it generated enough traction. After that, there were probably four to six months of just wandering mode, where you have a few customers and you're trying to get more. I remember I would personally talk to anyone who came into our company's Intercom. If you were a visitor on scaleai.com and clicked the chat bubble, you'd end up talking to me. It was still very unclear if it was going to work. It wasn't until six months later that we started getting one client who wanted to be really big, and then the idea started coming into its own. But there was probably a year or so of what I generally call wandering, and I think that's on the low end for most startups.

我们来聚焦一下那个焦虑的部分。你是怎么应对的？因为事后回顾，很容易把它看作一种必经之路和学习过程，但身在其中时，感觉糟透了。那么，当时你是直面焦虑，还是只是随波逐流，每天、每小时都起伏不定？

Let's focus a little bit on that angsty part. How did you deal with it? Because it's easy to look back on it and acknowledge it as a rite of passage and a learning process, but when you're in it, it sucks. So did you at that point stare the angst in the face, or were you just along for the ride, up and down every single day, every single hour?

我要说的是：我不确定它是否比公司未来其他时期更糟糕。创办公司时有很多糟糕的时刻。但让我度过这一切的主要原因是，一旦我们有了 Scale 的想法，我真心相信未来这个东西会存在，而它现在还不存在。所以你有足够的梯度去努力，然后你可以对路径有合理的信心：在某个时候我会弄清楚如何从 A 点走到 B 点，这将是一段漫长的旅程。这相当令人安慰。但在过程中，一直困扰我的是 YC 的生存几率，就像《饥饿游戏》。90%的公司都会死，但它们不会马上死。所以不像《饥饿游戏》里第一天就有人死，对于初创公司，如果你反正会在五天后死，那还不如现在就死。最可怕的想法是：你可能已经死了，但要三年后才发现。但我是一个相当焦虑的人，我把所有的不确定性都转化为焦虑地做事。如果你这样做，至少你会觉得自己在朝着某个方向前进和进步。

One thing I'll say: I don't know if it sucked more than other periods in the future of the company. There are plenty of times that suck when building a company. But the main thing that got me through all of it was that once we arrived at the idea for Scale, I really believed that in the future this thing will exist and it doesn't exist right now. So you have enough of a gradient to go off of, and then you can have reasonable confidence in the pathway: at some point I'll figure out how to go from point A to point B, and that's going to be a long journey. That was pretty comforting. But in the process itself, one thing that kept getting in my head was the survival odds in YC, like The Hunger Games. 90% of the companies just die, but they don't die for years. So unlike The Hunger Games where people die the first day, for startups, if you're going to die anyway in five days, you might as well die now. That was the most terrifying idea: you could already be dead but you might only find out in three years. But I'm a pretty anxious person, and I channeled all the uncertainty into anxiously doing things. If you do that, at the very least you feel like you're moving and making progress in a particular direction.

你担心过吗？回到你的观点，对于任何想法，你眯起眼睛都能看到大概 10 个人在做类似的事情。我常说，如果你在早期容易被竞争吓到，那你就处境艰难了。如果再配上很多创始人常有的极端偏执，那就是致命的组合，因为你永远无法克服。所以我的问题是：Mechanical Turk 当时已经存在，还有一些其他玩家。你担心过他们吗？还是说，你根本不觉得那是问题？

Do you ever worry? Coming back to your point that for any idea you have, you can kind of squint and see 10 people ostensibly doing something similar. I always say that if you tend to get easily psyched out by competition in the early days, then you're in a hard spot. If you pair that with extreme paranoia, which a lot of founders have, that's a deadly combination because you're never actually going to get over it. So my question is: Mechanical Turk kind of existed and there were some other players. Did you worry about them, or were you like, no, that's not a concern?

我认为担心竞争和对其偏执是有区别的。担心竞争在某种意义上是健康的，因为它让你保持诚实并努力工作。但对其偏执，即不断回头看、怀疑自己的决定，是没有生产力的。对我们来说，我们看到了 Mechanical Turk 和其他玩家，但我们相信未来是通过 API 以更无缝的方式让人工智能变得可访问。所以我们专注于实现那个愿景，而不是担心别人在做什么。话虽如此，我们始终关注竞争格局，但没有让它瘫痪我们。

I think there's a difference between worrying about competition and being paranoid about it. I think worrying about competition is healthy in the sense that it keeps you honest and keeps you working hard. But being paranoid about it, where you're constantly looking over your shoulder and second-guessing your decisions, is not productive. For us, we saw Mechanical Turk and other players, but we believed that the future was about making human intelligence accessible via an API in a much more seamless way. So we focused on building that vision rather than worrying about what others were doing. That said, we were always aware of the competitive landscape, but we didn't let it paralyze us.

这产品品质不行，所以我知道我能做得更好。没错，当时已经有不少东西了，我觉得真正的答案在于那种虚张声势的自信。这其实挺有意思的。我见过——我和 Palantir 的一位高管一起参加一个会议，他正在跟一个政府客户谈话。他当着我的面说：‘你们没选 Palantir 没关系，但你们会后悔的。如果一开始就选 Palantir，你们会开心得多。我们做得比任何人都好。不过你们没选也没事，你们最终会回心转意的，到时候我们还是朋友。但现在……’我一点没夸张，那话术咄咄逼人得离谱。而这正是 Palantir 成功秘诀的一部分——那种根深蒂固的信念，认为他们做的任何软件都是上帝赐予地球的最佳礼物，没人能比他们做得更好。这里面有些是自我强化的循环。我提这个是因为，如果你是个理性的第三方旁观者，你可能会觉得‘这看起来很荒谬、很讨厌’。但我认为这其实是成功故事的一部分。你需要某种非理性的自信，才能相信自己能在市场上竞争。长期来看，你需要相信自己能比其他人招到更好的人才，做出更好的产品决策，付出更多努力，而别人可能没那么在意。久而久之，这些事会累积成你当初那种非理性的自信。所以我在 Scale 相信的一件事是——我参加过很多编程竞赛和数学竞赛，所以我觉得如果非要跟人竞争，我完全没问题。我至今已经参加过很多比赛了，所以放马过来吧——我们就是要做出更好的东西。我记得大概半年前，我跟一个 18 岁的 YC 创始人聊过，我问她同样的问题：‘你怎么看待竞争？’她说：‘竞争嘛，就是做得更好，干得更好，变得更好。’我觉得这有点虚张声势，但我认为这确实是答案——你只需要真心相信自己会更好。

It's just not a good quality product, so I know I can do better. Yeah, there were a bunch of things that existed, and I think the actual answer here is the right attitude is kind of this blustery confidence. And it's actually pretty interesting. I see this—I was at a conference with this guy from one of the key executives at Palantir, and he was talking to this government customer. He said this right in front of me: 'You know, it's okay that you guys didn't pick Palantir, but you're going to come to regret it. It would have been nicer if you just picked Palantir from the start, and you'd be way happier. We do a much better job for you than anybody else possibly could. But it's cool that you guys didn't, and you'll come around eventually, and eventually we'll be friends again. But for right now...' I'm not even exaggerating, it was a ridiculously aggressive talk track. And that's actually part of the success recipe for Palantir—this deep fundamental belief that any software they build is God's greatest gift to Earth, and nobody else can build software as good. There's some of that that is a self-reinforcing loop. I mention this because obviously if you're a rational third-party bystander, you can be like, 'Oh, that seems ridiculous and obnoxious.' But I think it's actually part of the success story. You need some kind of irrational self-belief to believe that you will be able to compete in the marketplace. Over the long run, you need to believe that you can recruit better than everybody else, make better product decisions, go the extra mile when others don't care as much. Over the long term, those things will ultimately accumulate to the irrational self-belief you had. So one of the things I believed, or that we believed at Scale, is that I did a lot of competitive programming and math competitions, so I felt if I have to end up competing with people, I'll do just fine. I've done plenty of competitions to date, so bring it—we're just going to build the better thing. I remember there was an 18-year-old YC founder I talked to maybe six months ago, and I asked her the same question: 'How do you think about competition?' She said, 'For competition, just be better, do better, just be better.' I think that's kind of blustery, but I think that really is the answer—you just have to really believe you're going to be better.

说到 Palantir 那个轶事，他们非常擅长让客户觉得没有他们的软件，业务会差一个数量级。这是我在 Dropbox 时纠结过的问题，因为在 Dropbox 我们只做好软件，让软件自己说话。每次你推出一个功能，两天后别人就会推出同样的东西。我们花了一段时间才明白，他们其实并没有真的推出，只是很会宣传。那么你在 Scale 遵循哪种哲学？你喜欢做产品，还是喜欢卖产品，还是找到一个平衡点？

Speaking of that Palantir anecdote, they're really good at making customers feel that without their software, their business will be literally an order of magnitude worse off. This is something I struggled with at Dropbox, because at Dropbox we would build good software and the software would speak for itself. Every time you would launch something, somebody else would launch the same thing two days later. It took us a while to understand that they weren't actually launching it, but they were very good at talking about it. So which philosophy do you follow at Scale? Do you like to build the thing, or do you like to sell the thing, or do you find a happy medium?

我认为你必须两者兼顾。在 Scale，我们的客户数量少，但都是大客户。在那个世界里，很多时候感知就是现实。可能你合作过的很多公司都是数据驱动的，真相会浮现，大家共享现实感。但大多数大公司并非如此，政府内部也不是这样，很不幸。在很多大客户那里，感知比现实更真实。现实往往太丑陋，很少有人真正面对现实；大多数时候他们只是选择相信他们所处的感知。所以这意味着，如果你做企业销售或企业产品，你的工作既包括改善现实，也包括塑造感知。我认为在某些方面，Palantir 的超能力之一就是他们比大多数科技公司更擅长塑造感知，因为他们把自己看作一个表演剧团和软件公司的结合体。我甚至不是随便说说——他们真的会给所有新员工发一本表演手册，他们这么做了很长时间。

I think you have to do both. At Scale, we have a small number of customers, very large customers. In that world, perception is reality a lot of times. Probably a lot of the companies you've worked with are very data-driven, where the truth makes its way and everybody shares a sense of reality. That's not true at most large companies, and also not true within the government, unfortunately. At a lot of large customers, perception is more real than reality. The reality is just so ugly most of the time that very rarely do people actually confront reality; most of the time they just choose to believe the perceptions they live in. So what that means is if you end up doing enterprise sales or building enterprise products, your job is just as much to improve the reality as it is to shape the perception. I think in some ways, one of Palantir's superpowers is that they shape the perception better than most other technology companies do, because they view themselves as a combination of an acting troupe and a software company. I'm not even saying that—they literally give an acting book to all new hires; they did for a very long time.

我觉得你刚才说的很有深意。如果我没理解错的话，大多数公司，包括政府，已经生活在某种感知里了，因为现实太痛苦。所以你的工作不是用现实去销售，而是用一个更好的感知去取代一个糟糕的感知。好吧，抱歉——你也应该创造价值。我只是说这是在做好产品之上的暗黑艺术。我认为从根本上说，你也需要做出好产品。

So I think there's something profound in what you just said. If I'm understanding you correctly, most companies, including the government, are already living in some perception because reality is so painful. So your job is not to outsell reality; it's to outsell a shitty perception with an even better one. Okay, all right, sorry—you also should produce value. I'm just saying this is like dark arts on top of building a good product. I think fundamentally you also need to build a good product.

也许稍微聊聊你个人——你会把自己的 10 倍优势定义为什么？你显然非常有竞争力，有极度的自信，但除此之外还有什么？

Maybe switching to you personally a little bit—what would you self-identify as your 10x spike? You're clearly very competitive and have extreme self-belief, but apart from those, what else?

对我们来说，这个特质——我个人以及 Scale 很多关键人物都展现出来的——是真正的创造性解决问题，但面对的是你能想象到的最不性感的问题。我们要做的大量工作是大规模运营。我们在全球有几十万贡献者，必须协调他们为模型生产高质量数据的过程。这是一个非常不性感的问题，但我觉得我们对待这个问题的尊重程度，跟对待数学奥林匹克问题一样，并且应用同样的技巧。我们会问：‘解决这个问题最有创意的方式是什么？’

For us, the thing—my personal one, as well as I think it's exhibited in a lot of the key people at Scale—is really creative problem solving, but in the face of the least sexy problems you can imagine. A lot of what we have to do is large-scale operations. We have hundreds of thousands of contributors all around the world; we have to coordinate this process of them producing high-quality data for models. It's a very unsexy problem, but I think we treat that problem with the same level of respect that we would a math Olympiad problem, and we apply the same kinds of techniques. We ask, 'What are the most creative ways that we could solve this?'

你如何分解问题？让我们用科学方法来解决。我认为很多伟大的科技公司都会这样做，但我们并没有那种只解决特定问题的崇高使命感。我们考虑问题的影响力，然后拼命工作。这些因素的结合最终让我们能够超越所有竞争对手：我们聪明，善于解决问题，并且对最不起眼的问题也给予最大的尊重。随着时间的推移，我们设计出了一些非常聪明的运营方案、技术和产品。很多时候，当我看到其他公司或初创公司时，他们要么是优秀的团队却把自己局限在智力上有趣的深奥问题上，要么是草根团队什么活都干，但无法以可扩展的方式解决问题，或者无法有效地抽象问题。所以，强大的问题解决能力和科学流程，加上对要解决问题的谦逊态度，这才是关键。

How could you break it down? Let's apply a scientific method towards solving it. I think something a lot of great tech companies do, but we hold no sort of elevated sense of purpose like we'll only solve certain problems. We think about the impact of the problems, and then we work like hell. The combination of these things is ultimately the recipe that allows us to outcompete all the competition: we're smart, we're good at problem solving, and we treat the most lowly problems with the utmost respect. Over time, we've devised some pretty clever operational solutions, technology, and products. A lot of times when I look at other companies or startups, you have one or the other: either an exceptional team that limits themselves to esoteric problems that are intellectually interesting, or a scrappy team that works on any problem but doesn't solve them in scalable ways or abstract them effectively. So the combination of strong problem-solving and scientific process with humility about the problems to solve is key.

那些不性感且运营繁重的问题通常很难激励人们兴奋起来，尤其是在湾区或硅谷。那么，你是否尝试过在招聘或入职等环节大规模建立某种文化，让这成为你 DNA 的核心部分？

Those unsexy problems and operationally heavy ones are often hard to motivate people to get excited about, especially in the Bay Area or Silicon Valley. So have there been cultural things you've tried to institute at scale, either in hiring or onboarding, to make that a core part of your DNA?

我们鼓励员工思考的主要是他们的净经济影响，而不是技术问题的难度。在学校里，你通过解决更多技术问题来获得精通，但技术难度和经济价值之间存在相关性。许多具有经济价值的问题在技术上并不非常困难，但有很多琐碎和混乱需要梳理。这正是大多数好的创业机会所在。所以我们激励员工的方式是：如果你有效解决了这个问题，净经济影响才是相对于技术难度而言重要的。那是你想要攀登的阶梯，这意味着那些不性感但非常有价值的问题会得到有效解决。

The main thing we encourage people to think about is their net economic impact much more than the difficulty of their technical problems. In school, you gain mastery by solving more technical problems, but there's a correlation between technical difficulty and economic value. Many economically valuable problems are not super technically difficult but have a lot of hair and mess to sift through. That's where most good startup opportunities are. So we motivate our staff by saying: if you solve this problem effectively, the net economic impact is what matters relative to technical difficulty. That's the ladder you want to climb, and it means the unsexy but very valuable problems get solved effectively.

我认为 Facebook 也做了类似的事情。在 Facebook，我们费了很大力气，因为在大多数公司，尤其是在工程晋升阶梯上，你解决更难的问题就能得到晋升，这隐含着地位和尊重。但在 Facebook，我们会晋升最高效的工程师，也就是产出最多代码的人，可能是在最不性感的问题上。很多高级工程师不理解：'为什么这个人晋升了？他们只是把日志错误减少了一个数量级，但那很容易。'但总得有人去做。

I think Facebook did something pretty similar. At Facebook, we took great pains that in most companies, especially on the engineering ladder, you get promoted if you tackle harder problems. There's implicit status and respect for that. But at Facebook, we would promote the most productive engineer, the one who cranked out the most code, maybe on the least sexy problems. A lot of senior engineers didn't get it: 'Why is this person getting promoted? They worked on reducing errors in logs by an order of magnitude, but that's easy.' But somebody's got to do it.

这是来自 Facebook 的一个好教训。

That's a good lesson from Facebook.

让我们换个话题，谈谈地缘政治。你对明年有什么预测？在技术、硬件、芯片方面，美中之间的冷/热战正在升温。未来 12 个月会如何发展，又会如何影响硅谷？

Let's shift gears and talk about geopolitics. What predictions do you have for next year? On technology, hardware, chips, the cold/warm war between the US and China is heating up. How does that play out over the next 12 months and affect Silicon Valley?

首先要观察的是，几周前发生了一件在地缘政治上意义重大的事件：OpenAI 发布了 o1 预览版，而第一个复现思维循环——测试时计算缩放——来自中国的 DeepSeek R1 模型。这非常令人惊讶：第一个复现不是来自 Anthropic、Google 或任何美国公司，而是来自中国实验室的开源模型。这说明领先的美国实验室和中国实验室之间没有研究差距。它们在性能上基本持平，这具有深远的影响。需要关注几点：美国和中国之间哪种 AI 技术会被更多地出口？将会有一场竞赛，看哪个 AI 堆栈能在全球范围内被更广泛地采用。在第一波技术浪潮中，美国胜出：谷歌搜索在全球占主导地位，除了中国；美国社交网络在全球占主导地位，除了中国。

The first thing to observe is one of the most meaningful geopolitical events happened a few weeks ago: OpenAI released o1 preview, and the first replication of the thinking loop—test-time compute scaling—came from China via DeepSeek R1. This is very surprising: the first replication wasn't from Anthropic, Google, or any American company, but an open-source model from a Chinese lab. This tells us there's no research gap between leading US labs and Chinese labs. They're basically caught up in performance, with far-reaching implications. A few things to watch: which AI technology gets exported more between the US and China? There will be a race to see which AI stack becomes more globally adopted. In the first wave of technology, the US came out on top: Google search is globally dominant except China, American social networking is dominant everywhere except China.

就像这场竞赛的第二阶段，当它更多是关于硬件和电信时，中国实际上胜出了。华为技术在全球范围内得到了相当广泛的出口。它某种程度上与一带一路倡议捆绑在一起，中国很快成为了世界上大多数国家的首选合作伙伴。现在我认为人工智能是第三阶段。所以你可以看到，美国迫使阿联酋决定他们是想站在中国华为的阵营，还是美国的阵营。目前阿联酋选择了美国。我认为我们会看到更多这样的决定，各国将决定他们是站在美国 AI 阵营还是中国 AI 阵营。我不认为美国，即使从外交政策的角度，真的在乎成为全球的 AI 阵营。我认为我们最在乎的是成为我们部分合作伙伴的 AI 阵营，但不是所有合作伙伴。那将是一场战争。另一场战争是，你知道，真正值得关注的是出口管制。所以拜登政府做的最重要的事情之一就是实施了非常严格的芯片出口管制。所以中国的高端 GPU 数量可能只有美国的百分之一，因为我们不让他们买。而且他们也没有 ASML 光刻机，也没有所有那些能让他们真正建立芯片产业的前驱体。随着特朗普总统上任以及领域内新玩家的出现，我认为最大的问题之一是：我们是否会维持出口管制的强硬路线？我认为我们当然应该。这是让我们保持强大优势的最大因素之一。而那场谈判会是什么样子？因为有一个迫在眉睫的威胁：到 2027 年，习近平主席和中国共产党表示他们想要拿下台湾，并且他们已要求军队，即中国人民解放军，准备在 2027 年前拿下台湾。所以这个三年后的日期迫在眉睫，届时中国，至少是中共，表示他们将拿下台湾。而一旦特朗普总统在 1 月 20 日就职，谈判的时钟就开始滴答作响：我们能否阻止中国采取行动，或阻止中国炸毁台积电，或阻止那种灾难性情景？我们需要用什么来交换？我认为那将是，无论是头条新闻还是某种低强度的长期谈判，都是我们在未来几年需要关注的大局。最终，我认为世界上有太多的激励因素。第二次世界大战将被避免。我认为没有人有兴趣卷入热战。但我认为我们需要关注，为了阻止他们入侵和接管半导体行业，我们将不得不付出什么。

Like the second leg of this race, when it was more about hardware and telecom, China actually came out on top. Huawei technology became pretty widely exported globally. It was sort of packaged in with the Belt and Road initiatives, where China sort of pretty quickly became the partner of choice for the majority of countries around the world, actually. And now I think AI is kind of the third phase. So you can see that the US forced the UAE to decide if they want to be on the sort of like China Huawei stack or they want to be on the US stack. The UAE for now is picking the US. I think we're going to see a lot more decisions where countries are going to decide: are they on the US AI stack or on the Chinese AI stack? I don't think the US, even the US from a foreign policy standpoint, doesn't even really care about being the AI stack globally. I think we most care about being the AI stack to some of our partners, but not all the partners. That'll be one war that happens. And the other one that happens is, you know, the thing to really watch is the export controls. So one of the most important things that the Biden administration did is they launched very restrictive export controls on chips. So there's maybe like one hundredth the number of high-end GPUs in China versus the United States, because we don't let them buy them. And they also don't have ASML machines and they don't have all the sort of precursors to enable them to actually build the chip industry. With President Trump coming in and with sort of new players in the field, I think one of the biggest questions is: are we going to maintain a hard line on the export controls? I think we certainly should. That's one of the biggest things that's enabling us to maintain a strong advantage. And what is that negotiation going to look like? Because there's the looming threat: by 2027, President Xi and the CCP have said they want to take Taiwan, and they've asked the military to prepare, the People's Liberation Army, to prepare to take Taiwan by 2027. So there's this looming date three years from now, by which China, at least the CCP, is saying they'll take Taiwan. And as soon as President Trump gets into office and gets inaugurated January 20th, that's going to start the clock on the negotiations: can we prevent China from taking that action, or prevent China from blowing out TSMC, or prevent the sort of catastrophic scenario? And what do we need to give in exchange? I think that will be, whether it's the headline or it's like a low-grade kind of negotiation over time, that is the big picture thing that we'll watch over the next few years. And ultimately, I think there's too much incentive in the world. World War II will be averted. I don't think there's any interest to get into a hot war. But I think we need to watch what we're going to have to give in exchange for them not invading and not taking over the semiconductor industry.

也许回到你关于 AI 堆栈的观点，你知道，本质上主权国家或世界各地的公司是会选择西方的 AI 堆栈还是中国的堆栈。我很好奇，你认为我们除了在实验室里真正创新之外，还有什么办法能影响这一点吗？不过我得承认，你说得对，有点令人惊讶的是，在推理时计算之后，大概已经过了八周，8 到 10 周，我们还没有看到其他一些主要实验室发布类似的东西，这真的让我很惊讶，因为我最初的想法是，这感觉有点容易复制，就像，嘿，只要在各种技术上投入更多推理时间循环，然后做点树搜索，找出最终结果。但我还没看到有人复制出来，这很奇怪。而你关于 DeepSeek 的观点很好。所以你认为我们只能靠创新，还是说我们可以通过政策来强制执行？

Maybe coming back to your point about the AI stack and what, you know, whether essentially like sovereign states or companies around the world will pick a western kind of AI stack versus kind of a Chinese stack. Curious, like, do you think it's actually possible for us to affect that by anything other than actually innovating in our labs? I will admit though, to your point, it has been a little surprising that after the inference time compute, it's kind of been what, like eight weeks, 8-10 weeks, that we haven't seen an equivalent release by some of the other major labs, which has actually really surprised me because my initial take was that this feels somewhat somewhat like easy to replicate, just like hey, just throw a bunch of like more inference time essentially like loops on kind of like the various techniques and essentially like you know do some tree climbing and figure out the end result. But I haven't seen that be replicated, which is odd. And your DeepSeek point is a good one. So do you think we can we have to just innovate right, or do you think we can actually enforce it through policy?

嗯，我认为第一步是不浪费我们的开源产业，我认为现在我们已经度过了那个阶段，但确实有一段时间有很多讨论，关于美国是否会更积极地监管开源模型。既然我们已经度过了那个阶段，我认为，你知道，我们美国确实有一些有趣的杠杆可以使用，最显著的是出口管制。所以我认为对于很多国家，我们可以谈判说，嘿，你们想要英伟达 GPU 集群吗？好吧，如果你们想要，你们可能得在我们的堆栈上构建，而这个堆栈可以包括 GPU，可以包括开源模型，包括一个广泛的包。我不知道这会不会是新政府的优先事项。我不知道这是否是外交政策的优先事项，但我认为我们有杠杆可用，就像另一方面中国也有他们的杠杆。你知道，他们可以提供大型基础设施建设，他们可以提供债务，他们可以提供大量免费技术，而这些是我们无法匹配的。所以这是一种给予和索取。几十年前，中国就像，或者说美国无可争议地是世界上基础设施和技术的提供者。希望未来十年这种情况能继续。

Well, I think the first step is to not squander our open source industry, which I think at this point we're through that, but definitely there was a lot of conversation at one point of whether or not the US is going to more actively regulate open source models. Now that we're through that, I think that there's, you know, we do have United States has interesting levers at our disposal, most notably it's the export control. So I think for a lot of countries, we can negotiate and say, hey, do you want clusters of Nvidia GPUs? Well, if you want those, you probably have to build on top of our stack, and that stack can include the GPUs, it can include the open source models, it includes sort of a broad package. I don't know if that's going to be a priority for the new administration. I don't know if this is a priority from a foreign policy standpoint, but I think we have levers at our disposal, just in the same way that on the flip side China has their levers. You know, they can offer large infrastructure build-outs, they can offer debt, they can offer a lot of free technology, and that's not stuff that we can match. So it's a kind of give and take. A few decades ago, China was like, or the US was indisputably the sort of provider of infrastructure and technology of the world. Hopefully that continues being the case in the next decade.

是的，我的意思是，关于你提到的开源模型，我认为我们会回顾扎克伯格的决定，即实际上拥有一个前沿规模的开源模型，这在那种演变中相当关键。这绝对是一个非常美国的决定。我知道，绝对，非常爱国。绝对。回到你可能的另一个想法，我的意思是，显然我们已经看到了很多关于智能体式 AI 的兴奋。智能体现在意味着一切，也意味着什么都不是。你怎么看？你认为在 2025 年这会如何发展，从消费者的角度，但也许从公司的角度更清楚一些，但对于消费者，我仍然认为我还没有使用过一个智能体为我做任何有趣的事情。嗯，我在公司和 AI 中还没有用过，但我很好奇你怎么看。

Yeah, I mean, I think to your point about the open source models, I think we're going to look back on the decision by Zuck to actually have like a frontier-scale open source model as being pretty critical in that kind of evolution. Very American decision for sure. I know, absolutely, very patriotic. Absolutely. Coming back to one of your maybe your other thoughts, I mean obviously we have seen a lot of excitement about agentic AI. Agents kind of means everything and also means nothing right now. What do you make of it? Like, how do you think in 2025 that kind of plays out, both from a consumer perspective but then maybe also I think it's a little bit clearer what it means maybe for companies, but like for consumers, I would still argue that I haven't used an agent to kind of do anything interesting for me yet. Well, I haven't used one in the company and the AI, but I'm curious what you make of that.

是的，我认为我们在智能体方面的现状，也就是模型整体的现状，是它们在单轮交互中相当不错。你知道，一次提示响应，它们表现很好，然后随着轮次增加，性能就像悬崖一样下降。这有点像，我的意思是，没有人会真正公开说出来。

Yeah, I think where we are with agents, so where we are with models overall, is that they're quite good within one turn. You know, with one prompt response, they perform pretty well, and then the performance just goes down a cliff as you increase the number of turns. And this is kind of like, I mean, nobody will really go out and say it.

但这就是我们目前的现实。模型在谷歌那样的用例中表现很好，你问个问题然后看答案。但在需要与模型协作完成更复杂任务时，它们表现得很差，直接掉链子。所以我认为有两件事。首先，行业和模型公司必须大力提高多轮交互的可靠性，最终让模型拥有更高层次的内在一致性，变得更像实体。现在，模型甚至不知道自己不知道什么。总的来说，在多轮交互中，它就像一台统计上更可能正确而非错误的机器，但并不是一个有心理理论的实体。所以这必须改变。但我实际上认为，对于这个智能体的分水岭时刻，最大的障碍就是产品设计。模型已经足够好了，可以做出一些令人惊叹的智能体产品或体验，对很多人来说会很棒。在座的各位可能不觉得，因为我们经常玩模型，知道它们很多方面确实很厉害，但大多数人甚至不知道模型已经这么好了。Cursor 就是个好例子：大多数工程师之前并不知道模型这么强，但把它集成到 Cursor 后，突然就成了他们工作流程的一部分，用起来方便多了。所以我认为消费者领域也会迎来这样的时刻。关键就是把模型从聊天范式里解放出来，更深入地嵌入到基本工作流程中。我认为这是 2025 年最大的创业机会：不断迭代，找到正确的智能体形态。

But this is the reality of where we are. So where they work really well is kind of like the Google use case, where you ask a query and look at the answer. Where they work really poorly is if you actually need to work with the model and do something more complicated. They just fall off a cliff. So I think there are two things. First, the industry and the model companies are going to have to work very hard to improve reliability with greater numbers of turns, and ultimately get these models to have some greater level of internal coherence and ultimately be more like entities. Right now, models don't even know what they don't know. In general, when interacting with the model on multiple turns, it's like this machine that is statistically more likely to be correct than incorrect, but it's not like an entity with any theory of mind that you're interacting with. So I think that'll have to change. But I actually think that for this watershed agents moment, the biggest blocker is just product design. The models are good enough already for there to be some agent product or agent product experience that will be pretty mind-blowing and great for a lot of people. For the people in this room, it's probably not obvious because maybe we play with models all the time and know they're really good at a lot of things, but most people don't even know that the models are that good. Cursor is a good example: most engineers didn't actually know the models were super good, and you put it into Cursor and all of a sudden it's part of their workflow, much easier to use. So I think that kind of moment will happen for consumers. It really is just breaking the models out of the chat paradigm into something that's a bit more baked into the fundamental workflows. I think this is the biggest startup opportunity in 2025: iterating to find the right agent modality.

那我们聊聊构建下一代前沿模型的最大限制之一：数据的可获取性。你认为数据墙有多真实？你对使用合成数据训练模型怎么看？在整个讨论中，Scaling 是个重要话题，所以想听听你的想法。

Maybe let's move on to talking about one of the biggest limitations to building the next generation of frontier models: accessibility of data. How real do you think the data wall is? What's your take on synthetic data for training models? And in this overall conversation, scaling is a huge part, so curious to get your thoughts.

是的，我认为过去 12 到 18 个月的狂热就是关于谁有更多算力、更多芯片。芯片越多，你就赢了。那场关于谁拥有最大集群的讨论非常缺乏深度。但逐渐清晰的是，即使拥有更大的集群，我们也遇到了一些数据限制。我们已经用尽了所有公开可用的数据。我们需要一种并行的方法，既要有专门的数据库，也要有计算能力，才能获得更好的性能。所以数据墙和我们目前遇到的进展瓶颈确实是真实存在的。我认为我们在预训练方面已经遇到了一些限制，现在很多进展来自后训练。后训练更多地受限于专门的、高质量的数据集，这些数据集与互联网上的不同。所以数据非常关键。曾经有一种研究信念认为，模型会自己生成大规模的推理轨迹或其他数据，用来训练预训练模型，但这并没有实现。许多关于合成数据的实验表明，使用合成数据会丢失数据分布中很多真正的丰富性。所以未来的现实是，我们需要依赖新形式的人类生成数据才能达到目标。但如果我们同时扩展数据和算力，我们就能继续取得进展。我们可能不需要一年前谈论的万亿美元或五万亿美元的集群，但我们可能仍然需要千亿美元的集群。所以我认为讨论正在趋于正常化。

Yeah, I think the frenzy of the last 12 to 18 months was just about who has more compute, who has more chips. The more chips you have, you're going to win. It was a very unnuanced conversation around who has the biggest cluster. But what's coming to light is that even with a much bigger cluster, we're hitting some data limits. We've hit the limits of all publicly available data. We need a tandem approach of specialized datasets in addition to computational power to yield much greater performance. So the data wall and the wall on progress we've hit right now is certainly real. I think we've hit some pre-training limits, and a lot of progress now is coming from post-training. Post-training is much more bottlenecked by specialized, high-quality datasets that don't look like the ones on the internet. So data is really critical. There was a research belief that models will just generate large-scale reasoning traces or other data that we'll train pre-training on, but that hasn't really panned out. Many experiments on synthetic data show that you lose so much of the real richness in the data distribution if you use synthetic data. So the reality going forward is that we're going to need to rely on new forms of human-generated data to get where we want. But if we scale data in addition to scaling compute, we can keep making progress. We may not need the trillion-dollar or five-trillion-dollar clusters we were talking about a year ago, but we still probably need the hundred-billion-dollar clusters. So I think there's some normalization happening in the conversation.

这有点像，之前完全是关于谁拥有最大的集群，然后资本主义会起作用吗？实际上我们有了更多的芯片，这很好。但我确实认为新的瓶颈在数据方面。所以我想知道，你认为 Scaling 在其中扮演什么角色？显然你们在后训练领域扮演着重要角色。你认为 Scale 有机会在这方面进一步加大投入吗？

It's kind of like it was actually all about who had the biggest cluster, and then will capitalism work? And we actually have a lot more chips, which is great. But I do think the new bottleneck is on the data side. So I'm curious, what do you think scaling's role is in that? Obviously you guys play a big role in the post-training part of the world. Do you think there are opportunities for Scale to even ramp up its presence on that front?

是的，我们必须在各个方面加大数据生产。坦白说，如果你从全局视角看通往 AGI 的道路，我们需要算力呈指数级增长，同时数据生产也需要一个较小但同样呈指数级增长的曲线。所以我们只需要在那条曲线上持续扩展，并且它需要与其他一切同步扩展。你不能让数据领先于算力，或者两者差距过大。我们的关键角色就是生产所有数据，以实现数据扩展和算力扩展的并行。

Yeah, we're having to ramp up data production across the board. I think frankly, if you zoom all the way out on the path to AGI, we're going to need compute to scale exponentially, and then you'll have a smaller but equivalent curve of data production needing to scale up exponentially. So we just need to keep scaling up on that curve, and it needs to scale with everything else. You can't scale data ahead of compute, or one can't get too far ahead of the other. Our key role to play is in producing all the data to enable data scaling in addition to compute scaling.

换个话题：在座很多是 AI 公司的创始人。你认为在 AI 领域创业与上一代公司相比有什么不同？有哪些显著差异？

Maybe shifting a bit: a lot of folks in the audience are founders of AI companies. What do you think is different about starting a company in the AI space versus all the companies created in the prior generation? What are some notable differences?

我认为有几个不同点。首先，变化速度更快。你必须不断适应，因为底层技术发展非常迅速。其次，资金需求通常更高，尤其是如果你在构建基础模型。但同时也更有杠杆效应，因为你可以在现有模型之上构建。第三，人才竞争激烈；每个人都在争夺同一小批 AI 研究者和工程师。第四，监管正变得越来越重要，你需要从一开始就考虑对齐和安全。最后，商业模式仍在探索中；目前还不清楚主导的变现策略会是什么。

So I think there are a few. First, the pace of change is much faster. You have to be constantly adapting because the underlying technology is evolving so quickly. Second, the capital requirements are often higher, especially if you're building foundation models. But there's also more opportunity for leverage because you can build on top of existing models. Third, the talent competition is intense; everyone is fighting for the same small pool of AI researchers and engineers. Fourth, regulation is becoming a bigger factor, and you need to think about alignment and safety from day one. Finally, the business models are still being figured out; it's not clear yet what the dominant monetization strategies will be.

在任何新技术领域，最大的问题就是存在大量不确定性，而另一面则是大量炒作。AI 也不例外。很久以前，SaaS 可能也是如此，炒作很多。但 AI 现在到了一个地步，我觉得市面上 80% 到 90% 的东西都是炒作。很多人说的话、投资者的信念、或者你在派对上听到的，其实没人真正知道到底发生了什么。但有很多人自信满满却错了。所以我觉得 AI 的一个难点……我讲个趣事。Scale 在 2018 年初融 A 轮时，我们的 Deck 里有一页写着‘数据是 AI 系统的生命线’之类的话。我们去红杉 pitch，到了那一页，一位合伙人（我不点名）非常大声且自信地说‘这不是真的’，还告诉我 AI 不再需要更多数据了，没有数据也没问题。这发生在演讲开始两分钟。我当时 20 岁，有点懵，就说‘我不知道他在说什么，我可以跟他讨论，但显然 AI 需要更多数据。’他说‘我不这么认为。’然后他就不听后面的演讲了，只有初级员工在听。我就走完了流程。结果当然是被拒了。这就是一个例子：有些人对自己坚信的事情非常自信，但后来证明是错的。这在任何早期行业都是如此，现在尤其如此，因为科技生态系统里很多人是靠‘观点’吃饭的。如果你是投资人，你需要有观点去跟 LP 讲；如果你是产品负责人，也需要有观点。很多人持有的信念并没有真正基于现实。所以如果你在这个领域创业，能弄清楚自己真正相信什么，并逆流而行去实现它，这有很大的溢价。我觉得这是最重要的事。如果你随大流，你甚至能融很多钱。AI 历史上有无数公司融了 5000 万甚至 1 亿美金的 A 轮，做的都是当时很时髦的东西，但后来发现那只是个愚蠢的想法。所以我认为这是标志性的。AI 经历了很多起起落落，未来还会有更多。但创始人最大的挑战是培养对自己真正信念的感知，同时有一个机制让你不断从生态中学习，而不是固守一个信念然后被浪潮冲走。这是最重要的。

One big thing in any new technologically new space is there's just a lot of uncertainty, and the flip side of that is a lot of hype. So AI is no different as an industry. I think a long time ago, probably Software as a Service was the same where there was a lot of hype. But AI is certainly at a point right now where I don't know, 80 to 90% of what is out there is hype. And a lot of what people will say, or what investors believe, or what other people if you go to a party will tell you, nobody knows what is actually going on truly. Nobody knows what is actually going on. But there's a lot of people who are confident and wrong. So I think one of the things that's hard about AI... I'll tell this funny anecdote. When we were raising our Series A at Scale, it was like 2017, no, 2018, early 2018. One of our first slides in our deck said something like 'Data is the lifeblood of AI systems' or something like that. And then we were pitching Sequoia, it was a whole partner pitch. I get to this slide and one of the Sequoia partners, I'm not going to say exactly who, very loudly and confidently says 'That's not true' and told me that we no longer need more data for AI, and we're going to be fine without more data. This is like two minutes into my presentation. I was 20 at the time, so I was a little shell-shocked. I said 'I don't know what he's talking about, I'm sure I could talk to him about it, but you obviously need way more data for AI.' And he was like 'No, I don't think so.' Then he didn't pay attention to the rest of the presentation, only the junior people paid attention. So I just went through the motions. And obviously they said no. That's an example of the kind of thing where there are people who are very confident about some things which will prove to be wrong. That's true in any early industry, it's certainly true right now because so many people in the tech ecosystem are paid to have a thesis. If you're an investor, you're paid to have a thesis and a point of view you can tell your LPs about. If you're a product leader in a big tech company, you're paid to have a thesis and a point of view. So many people have these beliefs that are not actually truly grounded in that much reality. So if you're building in this space, there's a large premium in having the ability to figure out what you actually believe and then executing against the flow to accomplish that. I think that's probably the biggest thing. If you follow the trends, you could even raise a lot of money following trends. There are tons of companies in the history of AI that raised like $50 million Series A or $100 million Series A by doing something that was really vogue and trendy, but then it turns out that was just a stupid idea. So I think that's the hallmark. AI has gone through many ups and downs, and there will be many more ups and downs in the future. But the biggest challenge for founders is to develop a sense of what you actually believe and also having a mechanism by which you're constantly learning from the ecosystem versus just digging your heels in a belief and then getting washed out in the tide. That would be the most important thing.

是的，你说得对，正确很重要。但有一个非常有趣的启示，特别是对 SPBC 的创始人来说：我认为目前很少有类别或垂直领域有真正的在位者，甚至还没有出现明显的赢家。大部分领域仍然有待争夺。这跟融了多少钱甚至使用指标都没有关系，因为现在所有人的使用指标都挺低的。所以我觉得你不需要被吓到。你当然需要正确，这是一个很好的过滤器。但我也不认为……我不会在这个阶段被竞争吓到，因为机会仍然巨大。而且，如果有一大群人都在用同一种方式炒作同一个想法，那么在这个子领域采取一点逆向思维会非常有趣。

Yeah, you know, I'd say that it helps to be right. But there's actually a very interesting takeaway, particularly for the founders at SPBC, which is that I think there are very few categories or verticals that actually have any kind of incumbent or even an emerging winner that's actually winning right now. I think a lot of it is still up for grabs. And it's not actually an indicator of how much money has been raised, or even honestly their usage metrics, because everyone's usage metrics are kind of low right now. So I think you just have to not be psyched out. You also have to be right, which is a good filter. But I also just don't think that there's... I would not get psyched out about the competition at this stage because there's still a ton of opportunity. And if anything, if you have a bunch of people all kind of harping about the same approach towards a particular idea, it's pretty interesting to take a slightly contrarian take on that particular subdomain.

是的，总的来说，旧金山大多数人其实并不真正相信什么，他们的想法都来自 Twitter 和派对。所以我认为你的任务就是不要成为那些人中的一员，要独立思考。这对创业至关重要，因为无论你做什么，总会有受欢迎和不受欢迎的时候。你的工作就是经受住这场疯狂的风暴。这在最高层面也是如此。比如 Nvidia，它曾经非常不性感，然后突然变得非常性感，未来还会经历这样的波动。所以这就是任务。

Yeah, I think in general, most people in San Francisco don't really believe anything and they get all their ideas off Twitter and parties. So I think your mandate is to not be one of those people and to think independently. I think that's critical for being able to start a company because at various points whatever you build will be popular or unpopular. Your job is to weather that crazy storm. That's true at the highest levels. Nvidia, for example, has been at various points very unsexy and then all of a sudden very sexy, and they'll go through those waves again. So that's kind of the mandate.

是的，马克在 Facebook 时经常告诉我们：你永远不会像世界说的那么糟，但也永远不会像世界说的那么好。大家总是说‘当我挣扎时，我需要相信自己’，但另一面是：当世界告诉你你做的每件事都对时，这也可能意味着你不够批判性思考，不够有主见，可能正在向平庸均值收敛。所以我觉得这很有道理。说起来容易做起来难；回音壁是真实存在的，我们都容易受影响。亚历克斯，你同意吗？

Yeah, something that Mark used to tell us at Facebook is that you are never as bad as the world tells you, but you are never as good as the world tells you. Everybody always takes the 'oh, when I'm struggling, I need to have belief in myself,' but there's a flip side: when the world is telling you that everything you're doing is right, that's also a sign that maybe you're not thinking critically enough, maybe you're not being opinionated enough, maybe you're converging to the midwit mean. So I do think there's something to be said about that. It's easier said than done; the echo chambers are real, we're all pretty susceptible to it. I'm curious if you agree with this, Alex.

我认为硅谷的一个迷思是：要在 AI 领域竞争，你需要很多钱。我可能持相反观点。我认为有更多的机会，即使你是在做基础模型领域，我觉得……

I think that one of the mythologies in Silicon Valley has been that in order to compete in AI you need a lot of money. I would probably take the opposite point of view on it. I think there's a ton more opportunity, either if you're actually doing on the foundation model space, I think the T of...

在算法变化甚至数据来源上发挥创造力，如果你在构建 AI 栈的上下游，我认为采用逆向创新策略比单纯筹集大量资金能走得更远。我很好奇，你认同这一点吗？

Opportunity in terms of being creative with algorithmic changes, maybe even in your data sources, and certainly if you're building up and down the AI stack, I think you can get a lot further with having contrarian creative strategies than simply going out and raising boatloads of money. I'm curious, does that resonate?

如果你的商业计划是需要筹集大量资金，因为你需要花很多钱，这在商业上是一个糟糕的生意。因为你显然想建立一个盈利的公司，你想做一些能赚钱的东西。所以如果你必须花很多钱，你只会陷得更深。我认为很多时候硅谷会陷入某些宏伟目标的竞赛中。所以如果你想通过控制基础模型成为世界之主，那么是的，你可能需要很多钱。但这可能不是这里大多数人、大多数公司的目标。大多数公司的目标是建立一个能够长期复利的盈利业务。我们密切关注自动驾驶汽车的惨败，那里筹集了数千亿美元，我认为接近一千亿美元用于构建大规模自动驾驶汽车。现在随着 Cruise 的事情，我认为实际上这些公司中没有一个成功。所以所有投入自动驾驶汽车的资金，只有一个赢家是 Waymo，而事实证明他们有一个无限的银行账户，所以他们不是一个你能竞争的公司。我认为这与我们今天的情况非常相似：如果你想踏上训练基础模型的跑步机，你是在与比除 10 个国家外所有国家都有更多钱的公司竞争。确实，是人类历史上最富有的实体。微软应该属于八国集团。没错。这不是一个好策略。我认为你不应该试图与一个无限的资产负债表竞争。所以我认为现在，当务之急是：什么是创造性的方式，什么是你能开发出在生态系统中真正差异化并且能长期投资的东西？我认为至少现在在 AI 领域，溢价实际上就是选择一个特定问题并长期关注。OpenAI、Anthropic 和 Google 这些公司非常可怕，但他们也有无数的事情要关心、担心和优化。他们很难专注于某个特定领域。Perplexity 我认为是最好的例子：其他产品在简单的基于搜索的 LLM 输出方面仍然不如 Perplexity。所以他们并没有花那么多钱，他们只是更专注于那个问题。所以专注总是有很大的溢价。

So if your business plan is that you need to raise a lot of money because you need to spend a lot of money, that is, in business terms, a bad business. Because obviously you want to build a profitable company, you want to build something that makes money. So if you have to spend a lot of money, you're just deeper in the hole. I think that often times Silicon Valley gets caught up in certain races for these grandiose ambitious objectives. So if you want to be emperor of the world by controlling the foundation model that is God, then I think yeah, you probably need lots of money. But that's probably not the goal of most people here, most companies. The goal of most companies is to build a profitable business that has the ability to keep compounding for a long time period. And so we watch this pretty closely in the self-driving car debacle, where there were hundreds of billions of dollars raised, I think close to a hundred billion dollars raised to build large-scale autonomous vehicles. I think now with the whole Cruise thing, I think actually exactly zero of those companies ended up succeeding. So all the money that went into self-driving cars, you have one winner which is Waymo, and it turns out they have an infinite bank account, so they're not really a company you can compete with. And I think there's a real analogy to where we are today: if you want to get on the treadmill and compete on training foundation models, you're competing with companies with more money than all but 10 countries. Truly, the richest entities that humanity has ever known. Microsoft should be in the G8. Exactly. And that's not a good strategy. I don't think you should try to compete with an infinite balance sheet. So I think that for sure right now, the imperative is: what is a creative way, what is a way in which you can develop something that is genuinely differentiated in the ecosystem, and also something that you can keep investing in for a long time horizon? I think for at least right now in AI, the premium is actually just picking a certain problem to care about for a really long time. OpenAI and Anthropic and Google, these are very fearsome companies, but they also have literally a bajillion things to care about and worry about and optimize for. And it's very hard for any of them to focus on a particular area. Perplexity, I think, is the best example of this: it's still none of the other products are as good as Perplexity in just simple search-based LLM output. And so they haven't spent that much money, they're just way more focused on that problem. So focus always has a big premium.

好的，我们几分钟后会开放给观众提问，所以开始想你们的问题。但在开放之前，你还写代码吗？

Okay, we'll open it up to the audience in a couple of minutes, so start thinking of your questions. But before I open it up, do you still write any code?

不，不写了。我前几天用了一下 Cursor，只是想看看它是什么。我认为这是一个很棒的产品。但我不确定的是，很难完全垄断一个工作流程。所以我认为如果你看看计算业务的历史，仅仅拥有一个工作流程很难长期保持，因为开发者和其他人一样，他们善变，会尝试新东西，最终新东西会有一个更酷的功能。所以我认为看看他们如何扎根并巩固自己的位置会很有趣。但我们都同意这是一个好东西。我知道开发者非常挑剔和善变。我安装 Cursor 后的感受也是：哦，我不喜欢这三个小地方，而忽略了十个很棒的地方。

No, no. I used Cursor a little bit the other day just to kind of see what the hub is about. I think it's a great product. I think what I don't know is it's pretty hard to fully monopolize a workflow in this way. So I think that if you look at the history of computing businesses, just owning a workflow is very hard to hold on to over long periods of time because developers, like anybody, they're fickle and they're going to try something new, and eventually something new will have one feature that's cooler. So I think it'll be interesting to see how they choose to sink roots and how they try to lay roots to solidify their position. But we both agree it's a great thing. I know developers are super picky and fickle. My take on installing Cursor was also like, oh I don't like these three small things, and ignore the ten great things.

你认为硅谷应该更关注哪一个人？

Who is one person that you think Silicon Valley should pay more attention to?

我认为总的来说，硅谷在政治上非常不聪明，或者说政治上的探索不足。所以大体上：有一封很久以前泄露的邮件，里面是关于婴儿潮一代的人口统计信息，那实际上是值得关注的主要线索。我认为在很多方面，如果你把握住了大的政治趋势，这些可以成为巨大的增长趋势。所以显然现在主要的政治威胁是民粹主义和特朗普主义之类的东西。我认为科技公司大多是非政治的，不应该关注这些东西。但如果你从全球角度思考政治上的变化，比如更大的孤立主义、民族主义、民粹主义等，那么对于什么是可能嵌入我们所见一切的 20 年顺风，会有很好的启示。所以我认为硅谷更大的政治素养被低估了。

I think in general, Silicon Valley is very politically unintelligent, so to speak, or politically uninvestigated. And so the broad strokes: I mean there's this great email that leaked a long time ago where it was an email thread with all this demographic information of the Baby Boomers, and that's actually the major thread to pay attention to. And I think that in many ways, if you get the broad political threads right, then those can be huge trends for growth. So obviously the major political threat right now is populism and Trumpism and all that kind of stuff. And it's one of these things where I think tech companies mostly are apolitical and shouldn't focus on this stuff. But if you were to think globally what's happening politically in terms of greater isolationism, greater nationalism, greater populism, etc., there's good takeaways for what are the 20-year tailwinds that might be embedded in all the things we see. And so I think the sort of greater political literacy in Silicon Valley is underrated.

我们开放给观众提问。我是 Ryan。我想问一下你对合成数据的看法。你似乎对后训练中的合成数据不太看好。我想听听更多关于这个立场的看法。你认为这反映了我们目前的状况，只需要找到正确的技巧，还是更根本性的问题？同时也理解，就像你说的，人们靠持有观点赚钱，你也靠持有观点赚钱，这个观点对规模化很重要。

We'll open it up to the audience. I'm Ryan. I just wanted to ask you about your position on synthetic data. So you didn't seem very bullish on synthetic data for post-training. I wanted to hear more about that position. And do you think that is indicative of where we are right now and just need to find the right tricks, or is this more of a fundamental thing? Also understanding that like you said, people get paid to have theses, you're also paid to have a thesis, and this thesis is important for scale.

我认为合成数据显然对后训练有效，但我想说，合成数据不是点金石。它不是那种你按一下按钮就能不断得到更多数据的东西。你需要想出各种技巧，每个技巧能让你得到多一点合成数据，你需要从根本上理解合成数据是什么：你是在利用数据的结构，或者利用你能得到的所有先验知识，或者这些结构性的先验，你是在利用数据中的底层结构来挤出一些合成数据来改进你的模型。

I think synthetic data works for post-training obviously, but it's like, I guess synthetic data is not a philosopher's stone. It's not like this thing that you just press a button and keep getting more data out of it. It's like you need to come up with all these tricks to get each trick a little more synthetic data, and you need to fundamentally understand what synthetic data is: you're leveraging the structure of the data, or you're leveraging all these priors that you can, or these structural priors, you're leveraging underlying structures in the data to squeeze out some amount of synthetic data to improve your model.

这显然作为一种范式是有效的，但它不是一种无限的追求，不是能无限扩展的东西。所以我的总体想法是：它显然有效，也显然是世界级后训练工作的一部分，但它并不是解决过度 Scaling 的方案。

That obviously works as a paradigm, but it's not like an infinite pursuit. It's not something that works to infinity. So I guess my overall thought is: it obviously works and it's obviously part of a world-class post-training effort, it's just not the solution to over-scaling.

嗨，你好。我是 Lorena。我想了解：大多数创业者在初期被告知应该瞄准中小企业，不应该做大企业，很可能会失败。所以你怎么看，尤其是我理解你早期就专注于企业客户？为什么你选择了企业客户，尽管连 Y Combinator 都给出了相反的建议？对于想直接服务企业而非中小企业的创业者，你有什么建议？

Hi, how are you? My name is Lorena. I would like to understand: most entrepreneurs at the beginning are told you should target small to medium-sized businesses, you shouldn't go for enterprise, you're likely to fail. So what would you say, especially if I understand correctly from an early stage you focused on enterprise? Why did you choose enterprise despite the advice that said even from Y Combinator? And just any advice for entrepreneurs who want to go straight to enterprise and not small to medium businesses?

大多数企业的问题在于，你甚至不能相信他们说的话，因为大多数大型企业就像这些毒瘤，你接触的多数人与公司的成功毫无关系，但他们有份工作，负责某件事。所以这就像一个老鼠窝，你必须学会在其中周旋，弄清楚如何利用它来运作。但如果成功了，显然可以非常成功。世界上所有最大的公司要么是企业服务公司，要么是消费者公司。中小企业到某个点就会触顶。这不是针对初创公司的说法。对于初创公司，如果你想专注企业客户，你必须知道你得花大约 30%的时间来应对组织的官僚主义。而如果你与中小企业合作，他们甚至没时间对你撒谎，所以无论他们说什么，你都可以按字面意思理解，我认为这对初创公司来说是一个巨大的优势。一般的建议可能是，如果你想作为一家公司起飞并取得成功，专注于中小企业是明显的路径。但取决于生态系统，如果你在一个 99%的公司都专注于中小企业的创业生态中，那么可能机会就少了。从长期来看，初创公司最终会发生的情况是：一批公司起步，其中一半试图模仿当时成功的公司，它们都陷入完全竞争，没有一家获得 traction。然后某家专注于边缘想法的公司最终没有竞争，但那个想法成功了。所以两年后，那家公司成了热门，所有初创公司都试图模仿它。你永远不想成为所有其他公司都聚焦的地方；你想尝试做一些稍微边缘的事情，因为这避免了完全竞争。

The issue with most enterprises is that you can't even believe what they tell you, because most large enterprises are these cancers, and most of the people you talk to have nothing to do with the success of the company, but they have some job where they're responsible for a thing. So it's like this whole rats nest that you have to learn to navigate and figure out exactly how you leverage it to work. But if it works, then obviously it can work incredibly well. All the largest companies in the world are either enterprise companies or consumer companies. The small to medium businesses, you tap out at some point. This is not really a statement for startups. For a startup, if you want to focus on enterprise, you have to know that you'll have to spend, let's say, 30% of your time navigating the bureaucracy of an organization. Whereas if you work with small to medium businesses, they don't even have time to lie to you, so whatever they tell you, you can take at face value, which I think is a great benefit for startups. The general advice probably is that focusing on small to medium businesses is the obvious gradient if you want to get lift-off and success as a company. But depending on the ecosystem, if you're in a startup ecosystem where 99% of companies focus on small to medium businesses, then maybe there's less opportunity there. If you look in longer themes with startups, what ends up happening is: a bunch of companies start, half of them try to model themselves against the companies that are successful at that time, they all end up in perfect competition with one another, so none of them get traction. Then some company that focuses on some fringe idea ends up having no competition, but that idea ends up working. So two years later, that's the hot company, and then all the startups try to copy that. You never want to be exactly where all the other companies are focused; you want to try doing something that's a little bit fringe, because it avoids perfect competition.

嘿，Alex，快速问一下你对人们观点和持有特定观点的看法。我认为人们的观点通常是他们的生活经历、第一性原理思考和你的输入的综合。现在如果你的大部分输入都是这样，很难正确，那么你的输入是什么？你如何筛选你的输入？

Hey Alex, just a quick question on your thought on people's opinions and having a specific opinion. I think people's opinions are often a combination of their life experiences, first principles thinking, and your own inputs. Now it's difficult to be right if most of your inputs are so what are your inputs and how do you filter for your inputs?

我选择关注的是：我努力专注于从互联网上的文章或与特定专家交谈中获取输入。我努力与那些真正不在乎别人想法的人交谈。你会找到这些人，因为他们真的我行我素；他们看起来会非常奇怪和古怪，但试着去听取那些真正独立思考的人的建议。同时，也要阅读独立思考者的文章。现代互联网生态的一大好处是，Substack 上有大量高质量的文章，你可以从在那里发文的专家那里学到很多关于任何特定事物的知识。所以这些是关键输入。然后是你的客户。我采取一种心态，即客户永远是对的。有些人在不需要这样做的情况下也能成功，但你必须对客户非常顺从。

What I choose to focus on: I try really hard to focus on getting inputs from writing on the internet or talking to specific experts. And I try really hard to talk to people who truly do not care about what other people think. You'll find these people because they really seem to beat to the beat of their own drum; they'll seem very strange and weird, but try to get the advice of people who are genuinely independent thinkers. Also, look at the writings of people who are independent thinkers. One of the great benefits of the modern internet ecosystem is there's so much high quality writing on Substack that you can actually learn so much about any individual thing from experts who post there. So those are the key inputs. And then your customers. I adopt a mindset where customers are always right. Some people are successful without needing this, but you have to be super subservient to your customers.

酷，我想时间到了。非常感谢 Alex 今天来做客分享。

Cool, I think that's all the time that we have. Thank you so much Alex for coming by and sharing today.

嗯，谢谢。谢谢你。

Yeah, thanks. Thank you.