Perplexity CEO：进攻，进攻，进攻

老兄，我问每个我见过的最优秀的创始人这个问题：你更多是出于对失败的恐惧，还是对胜利的渴望？

Dude, I ask this of the best founders that I meet. Are you motivated more by the fear of failing or by the thrill of winning?

胜利的渴望。

Thrill of winning.

为什么？

Why?

因为我没什么可失去的。我出身卑微，从没想过自己会走到今天。我的人生已经远超想象。我曾在印度读本科，用实验室里别人打游戏的显卡训练神经网络，纯粹为了好玩。我的路一路走到这里。对我妈妈来说，能找到一份工作就是成功，因为我们在印度属于中低收入家庭，连英国或美国的中低收入都不如。那时我们最大的愿望就是进谷歌工作，当个谷歌工程师就算赢了。所以和那个目标相比，我已经做得非常好了。我真的没什么可失去的。所以每当我试图避免失败、采取防守姿态时，我就提醒自己，那是最愚蠢的做法。最好全力以赴，永远保持进攻。进攻，进攻，再进攻。

Because I have nothing to lose. I came from nothing. I never even imagined myself to be doing all this. So my life has already been extraordinary beyond any level of imagination. I was just in India, doing my undergrad, training neural nets with graphics cards that people in the labs were using for playing video games. It was all for fun. My path led me all the way here. For my mom, just getting a job was success because we were financially lower middle class in India, which is not even like lower middle class in the UK or the US. From there, all we wanted was to get a job at Google. Being an engineer at Google was considered a win. So I'm already doing remarkably well compared to that ambition. There's really nothing for me to lose. That's why anytime I try to act like I'm trying to avoid failure and being on the defense, I remind myself that it's the stupidest thing to do. It's better to go all in and try your best. Be on the offense all the time. Attack, attack, attack.

那你回顾一下，今天你在哪些方面还不够激进？

When you review then, what are you not being aggressive enough on today?

也许早期我们在社交媒体上大肆宣传 Perplexity 与谷歌的对比，我自己也经常这么做。有些人因此不喜欢我。现在，我谈论产品和竞争对手时更加谨慎了。但这并非缺乏攻击性，只是那样做已经没意思了，大家从我这里听得够多了。

Maybe in the early days, we'd be very loud on social media talking about Perplexity versus Google, and I used to do that myself a lot. Some people don't like me for having done that. Today, I'm a lot more measured in how I talk about our products, our competitors, and stuff like that. But it's not a lack of aggression. It's just that that is boring. People already heard that enough from me.

你后悔当初那么大胆的言论吗？

Do you regret being so bold in your messaging?

不后悔。

No.

所以不是信息变得微妙或成熟了，而是它过时了，我需要新东西。

So it's not a nuance and maturation of message. It's that it's stale and I need something new.

不仅如此。我觉得这个框架已经不再相关了。我们做搜索起家，Perplexity 最初就是搜索。我们打造了世界上第一个答案引擎，至今人们仍这样认识 Perplexity。一提到 Perplexity，人们就会想，‘哦，那是个答案引擎。’之后我们又做了很多产品：智能体、浏览器智能体、深度研究、计算机等等。我们做了这么多，但大家记住的还是第一个产品。而且印记已经打下。我们改变了谷歌的路线图。可以说，我或 Perplexity 公司对 google.com 的改变，比谷歌任何产品经理都大。

Not just that. I kind of don't think it's a relevant framing anymore. We worked on search. Perplexity started out as search. We built the first answer engine in the world that people know Perplexity even today. If you mention the name Perplexity, people will think, 'Oh, that's an answer engine.' We built a lot more things after that. We built a lot of agents, browser agents, deep research, computer. We built so many products after that, but we're still known for that first product. And the mark has already been made. We changed the road map of Google. You could argue that I or the company Perplexity changed google.com more than any product manager at Google has ever done.

给我讲讲这个论点。

Make that argument for me.

嗯，谷歌从来没人想推出答案引擎。没有。没人想动那个每年赚 2500 亿美元的界面。现在你看 AI 模式，它看起来完全就是 Perplexity。没有任何区别：字体、引用、内联文本加粗、内联超链接、建议追问。整个体验简直就像 Perplexity。只不过它还没那么好。

Well, nobody ever wanted to ship an answer engine at Google. Nobody. Nobody wanted to tinker anything on the interface that made them $250 billion a year. And now you look at AI mode. It looks exactly like Perplexity. There's not even any difference: the font, the citations, the specific bolding of inline text, inline hyperlinks, suggested follow-ups. The whole experience is literally looking like Perplexity. Except it's still not as good.

他们向你学习并适应，这对你是好是坏？

Is that bad or good for you that they learn from you and adapt?

有好有坏。显然，我知道这会在 2024 年底发生，所以一点也不意外。只是时间问题。我仍然惊讶于他们的质量还没到位，因为我经常测试所有产品。但我很高兴他们让谷歌变成了它应该成为的样子。我相信前沿才是赚钱的地方。AI 的前沿不再是回答问题，而是真正为你做事。我们仍然拥有世界上最先进的深度研究。那才是人们订阅我们 Pro 或 Max 产品的原因。他们不是为了传统方式获取答案，而是要求复杂的研究报告，要求能为你做事的智能体。如果我们 2024 年还觉得一切已定，就不可能做到这些。不，答案引擎一直是我们前沿产品的导流工具。你需要一个起点，对吧？每家公司都需要一个成功的产品来打造下一系列产品。在 AI 领域，没人能安坐不动，以为一切搞定。包括 Anthropic。如果 Anthropic 认为 Claude Code 已经赢了，那么 6 到 12 个月后，他们可能就不存在了。这就是整个领域令人不安的事实。

It's both good and bad. Obviously, I knew this was going to happen around the end of 2024, so it never caught me by surprise. It was just a matter of time. I'm still surprised that the quality is still not there because I regularly test every product out there. But I'm happy that they changed Google to be what it should be. I believe that the frontier is where the money is. The frontier in AI is not about answering questions anymore. It's about actually going and doing work for you. We still have the state-of-the-art deep research in the world. That's actually where people subscribe to pay for our pro or max products. It's not for getting answers in the traditional way. They're asking for sophisticated research reports. They're asking for agents that go and do things for you. We wouldn't have been able to do all that if we were sitting in 2024 thinking we have everything settled. No. The answer engine was always a lead gen for the frontier products we built. You need something, right? Every company needs to have one successful product to build the next set of products. In AI, nobody can sit comfortably thinking they have it all sorted out. Including Anthropic. If Anthropic thinks Claude code is already a win, in 6 or 12 months from now, they won't even be around. That's the uncomfortable fact about the whole field.

你今天告诉我——如果你不介意我引用的话——你在开始前刚说，你认为 OpenAI 还没准备好上市。两年前，当除了 ChatGPT 没人愿意碰任何产品时，你相信自己会有资格说这话吗？

Would you argue today you just told me, if you don't mind me quoting you here, you just told me before we started that you think OpenAI isn't ready for an IPO. Would you have believed you'd be in a position to say this 2 years ago when nobody had wanted to deal with any product other than ChatGPT?

所以即使是处于如此巨大优势地位的人，也可能被置于不再称王的位置，他们是在追赶，对吧？这就是现状。问题不在于 Perplexity、Anthropic 或 OpenAI 有没有护城河。

So anyone even in such a massive advantageous position can be put in a position where they're no longer the kings. They're fighting from behind, right? So that's the state of the field. It's less about Perplexity or Anthropic or OpenAI not having moats or having moats.

我能反驳你这一点吗？

Can I push back on you on that?

可以。

Yeah.

我坚持两年前的观点，即使当时他们占主导地位，并且仍然拥有主导的消费产品。但我坚持是因为我认为他们在财务上还没有准备好。当你看到他们的资产负债表时……

I would stand by 2 years ago even when they were dominant and they still had a dominant consumer product. But I would stand by it because I don't think they are financially ready. When you look at the balance sheet of that...

好吧。也许我把这两件事分开。让我们把 IPO 的财务准备与主导领导者的认知分开。

Okay. Maybe I'll decouple that. Let's decouple financial readiness for an IPO versus perception of a dominant leader.

嗯。

Yeah.

你现在认为他们是主导领导者吗？

Do you perceive them as a dominant leader right now?

是的。

Yes.

在什么方面？

In what?

消费者搜索。

Consumer search.

嗯，但那里没有钱，对吧？因为它已经被商品化了。所以它总是领先的。例如，为什么他们全力投入 Codex？因为那里有钱。我们在 Computer 上也在做同样的事。Anthropic 在 Claude Code 上也在做同样的事。谷歌还没有这个类别的产品，但我相信他们会跟进。Meta 正试图以每月 200 美元推出 Hatch。你看到发生了什么，对吧？所以没有人……

Well, except there's no money there, right? Because it's been commoditized. So it's always a lead. For example, why are they going all in on Codex? Because that's where the money is. And we're doing the same on Computer. Anthropic is doing the same on Claude Code. Google doesn't yet have a product in this category, but I'm sure they're going to come after that. Meta is trying to launch Hatch for $200 a month. You see what's happening, right? So nobody has...

但肯定有比 Codex 和 Claude Code 更多的钱。

But there has to be more money than just Codex and Claude Code.

这不是关于代码。这是关键。至少非广告收入（我不是在说广告收入），在非广告订阅或基于使用的收入中，钱在那些前沿领域。而今天的前沿是走出去为你做事。

It's not about code. That's the main thing. The money, at least in non-advertising revenue (I'm not talking about advertising revenue), in non-advertising subscription or usage-based revenue, the money is in whatever is the frontier. And today the frontier is about going out there and doing things for you.

那你认为这不会为 OpenAI 带来 1000 到 2000 亿美元的广告业务吗？

Do you not think then that that will be a 100 to 200 billion dollar advertising business for OpenAI?

还有待证明。我们来分析一下广告类别。谷歌上排名第一的广告商是谁？亚马逊。第二名呢？Booking.com。第三或第四，我想是 Expedia。那么你认为 Booking.com 在谷歌上花了多少钱？160 亿左右，大概是这个数。一个疯狂的数字。那你现在怎么订酒店或机票？是在 ChatGPT 上订还是在谷歌上订？

Yet to be proven. Let's work through the categories of advertising. Who's the number one advertiser on Google? Amazon. Who's the number two? Booking.com. Number three or four, I think is Expedia. So how much do you think Booking.com spends on Google? 16 billion, something like that. Something crazy like that. And how do you book your hotels or flights today? Do you book it on ChatGPT or do you book it on Google?

谷歌。

Google.

为什么？

Why is that?

对我来说，实际上我喜欢探索。我想看到各种选项。

For me actually, I like discovery. I would like to see the options.

正是如此。对吧。所以界面与其说是对话，不如说是探索。所以当决策更主观、更凭感觉时，你不需要一个客观的答案引擎。你再想想另一类广告，直接面向消费者的产品、时尚。大部分广告预算都流向了哪里？流向了 Meta、Instagram。因为你只是在浏览。你只是在无意识地刷屏，对吧？所以聊天界面目前没有捕捉到那种用户意图和用户行为，这就是为什么它从来都不太适合广告。而且它从根本上破坏了人们进入产品并想要准确答案时的信任，而这正是 Perplexity 所著称的。然后你却说，‘嘿，顺便说一句，你问了最好的蛋白粉，但顺便说一句，这些是你可以看看的好蛋白粉。’这有点损害人们对你的平台和产品的信任。所以这是另一个原因……如果你想想，Meta 或过去其他一些公司曾试图在消息应用和电子邮件中投放广告，但从未真正成功过。在中国，微信上成功了，因为他们没有其他方式来资助整个系统。整个经济和用户行为都是围绕游戏化优化的。在美国不是这样。所以我对广告在聊天界面中真正起飞持悲观态度。我很乐意被证明是错的，但我对此持悲观态度。

Exactly. Right. So the interface is less about conversations and more about exploration. So when the decision-making is more subjective and vibes-based, you don't need an objective answer engine. And you think about the other category of advertising, direct-to-consumer products, fashion. Where is most of that advertising budget going into? It's going to Meta, Instagram. Because you're just browsing. You're just doom scrolling or whatever you call it, right? And so the chat interface doesn't capture that user intent, that user behavior right now, which is why it was never a great fit for advertising. And it also fundamentally corrupts the trust that people have when they go into a product and they want the accurate answer, which is what Perplexity is known for. And then you're like, 'Hey, by the way, you asked for the best protein shake, but by the way, these are good protein shakes that you can check out.' It kind of hurts the trust that people have in your platform, in your product. And so that's another reason why I... If you think about it, Meta or some other companies in the past have tried to put ads inside messaging apps and emails and it's never really worked out. It works out in China in WeChat because there's no other way for them to fund the whole thing. The whole economy and user behavior has been optimized around gamifying. It's not how things work in America. So I'm bearish on advertising really taking off in the chat interface. I'm happy to be proven wrong there, but I'm bearish on that.

我想花两个小时来拆解这一点。第一个，按时间顺序和你说的方式，‘钱在前沿’。我越听这个，就越怀疑，因为我认为我们大大高估了前沿模型在做相当基础的工作时的重要性。

There are two hours that I want to unpack that. The first one, just taking them chronologically and how you said them, 'money's in the frontier.' The more I hear this, the more I question it because I think that we dramatically overestimate how important frontier models are to do quite basic work.

前沿并不意味着前沿模型。前沿只是指你现在能用 AI 达到的任何前沿成果。Greg Brockman 最近发推说模型不再是产品了，对吧？有趣的是，作为前沿实验室的领导者，他完全有动机说模型就是产品。而谷歌的人也这么说。我记得有个谷歌的人一直在发推说模型就是产品。我忘了是谁。所以 Greg 对的原因是因为如果你拿 Codex 或 Perplexity Computer、Claude Code，那是什么？那是一个编排系统，对吧？它拿一个模型，配上智能体框架。什么是智能体框架？最简单的理解：智能体循环应该如何运行的规则。有哪些技能、子智能体、连接器、工具和权限？没有这个框架，你就不一定能捕捉并将模型的内在智能转化为有价值的输出 token。如果你只是模型 token 的转售商，你就没有业务。因为模型会被商品化，所以即使你是模型构建者，你也没有业务。作为基础设施层，你在服务这些输出 token 上有一些业务。但作为应用层或模型构建者，如果你只是转售直接来自模型的 token，你实际上没有业务。如果你知道如何拿模型，将其锚定在有价值的上下文中，用非常好的智能体框架编排，连接到正确的工具和连接器，无论是个人连接器还是企业连接器，并在一个统一的系统中为用户提供体验，你就有业务。我们在 Perplexity 的差异化方式是我们不仅跨工具、文件和连接器编排，我们还跨模型编排。这是 Anthropic 和 OpenAI 无法声称的差异化，因为你不会在 Claude Code 框架中找到 GPT-5。你不会在 Codex 框架中找到 Claude Opus 4-7 或 8。它们是相互竞争的，对吧？而你在 Perplexity Computer 中可以找到这两个模型。这样我们可以提高每瓦每用户的 token 价值。

Frontier doesn't mean frontier model. Frontier just means whatever is the frontier outcome you can have right now with AI. Greg Brockman recently tweeted that the model is no longer the product, right? And it's funny because as a leader of a frontier lab, he has every incentive to say the model is the product. And that's what Google people tell. I think one of the Google people keeps tweeting that the model is the product. I forgot who. And so the reason Greg's right is because if you take Codex or Perplexity Computer, Claude Code, what is that? It's an orchestration system, right? It takes a model, pairs it with an agent harness. And what is an agent harness? Think of it the simplest way: rules for how the agent loop should run. What are all the skills and sub-agents and connectors and tools and accesses? And without the harness, you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens. The output tokens, if you're literally just a reseller of model tokens, you have no business. Because the model will get commoditized, so even if you're a model builder, you don't have a business. As an infra layer, you have some business on serving those output tokens. But as an application layer or model builder, you don't really have a business if you're just a reseller of tokens that come directly out of the model. You have a business if you know how to take the model, ground it in valuable context, orchestrate it with a really good agent harness, connected to the right set of tools and connectors, whether it's personal connectors or business connectors, and provide the experience to people in one single unified system. And the way we differentiate ourselves at Perplexity is we don't just orchestrate across tools and files and connectors, we also orchestrate across models. That is the differentiation that Anthropic and OpenAI cannot claim because you wouldn't find GPT-5 inside the Claude Code harness. You wouldn't find Claude Opus 4-7 or 8 inside the Codex harness. These are competing with each other, right? Whereas, you would find both these models inside Perplexity Computer. And that way we can increase the token value per watt per user.

如果你假设每美元的价格本质上就是功率（瓦特），那是除了政府之外没人能补贴的东西。谁用最少的电力提供最有价值的输出 token，谁就为最终用户创造了最大价值，拥有最强的定价权和最大的价值。所以这就是需要解决的编排问题。AI 中唯一最重要的指标是每个用户的 token 价值。

If you assume that the price per dollar is fundamentally power in watts. That's the thing that nobody else can subsidize other than the government. Whoever provides the most valuable output tokens with the least amount of power expended produces them generates the greatest value to the end user and has the most pricing power, has the most value. So that is the orchestration problem to solve. The single most important metric in AI is token value per user.

如果模型不是产品，而是变成一种可以随时切换的公用事业，这对 OpenAI 和 Anthropic 的价值意味着什么？

What does it mean for the value of OpenAI and Anthropic if model is not the product and it becomes a utility, something you can switch into and switch out of?

每个人都以为我们都在构建模型层竞赛。其实不是。我甚至认为构建模型是保持前沿的一种方式，但你必须拥有一个能生成有价值 AI 输出 token 的界面。最有价值的 token。它不一定是产品本身。这是大多数创始人需要重新学习的最重要的事情。我自己也不得不这样做：要在 AI 产品层成功，无论你是否构建模型，都不是要打造一个拥有十亿用户的产品。这种心态必须彻底转变。现在有一些超级用户正在推动这个 token 经济。看看那些疯狂的故事：有个工程师让亚马逊每月花费五亿美元，因为他在云代码里设置了一个愚蠢的智能体循环。好吧，也许那是个错误，但 Meta 和其他公司里确实有真正的工程师每年为这些编码工具花费每人一千万美元。Perplexity AI 计算机上有用户，我想有一个用户每月花费超过一万美元。类似这样。疯狂。而且不是浪费钱。他们的业务依靠在这些框架内运行的智能体循环运转。他们以我们构建产品时无法想象的方式复杂地使用这些产品。即使在我们公司内部，也有人设置了看起来像独立软件架构的多智能体层级和智能体循环。我经常请这些人来向公司其他人解释：‘嘿，你们在用这些工具做什么？你们显然消耗得比我们预期的公司平均水平高得多。’大量使用智能体和不使用智能体的人之间最大的区别在于他们是否运行重复的定时任务。你是否把 AI 当作一次性任务，只是委派一个任务然后完成。这就像用于深度研究之类的，对吧？单一任务。而 AI 则持续为你监控某些东西。AI 持续根据某些事件触发并执行操作，给你发出警报。就像你设置持续运行的工作流。每次收到入站邮件，它进行分类；每次出现延迟峰值，它必须识别代码库的哪个部分导致了这个，进行根本原因分析，然后找到正确的工程师。所有这些事情。这就是前沿所在。所以回到我的主要观点：这些产品不会被一亿人使用。但它们产生的收入将超过谷歌或 Meta 的广告收入。这一定会发生。

Everyone thinks we're all building the model layer of the race. We're not actually. I would even argue that building models is a way to stay at the frontier, but you have to own an interface in which valuable AI output tokens are generated. The most valuable tokens. It doesn't have to be the product. This is the single most important thing to unlearn for most founders. And I had to do it too, which is to be successful in AI product layer, whether you're a model builder or not, it's not about building something that gets a billion users. That mentality has to completely shift. There are a few power users who are propelling this token economy right now. If you look at all these crazy stories of how there's this one engineer who got Amazon spend like half a billion dollars a month because of some stupid way they set up like agent loop inside Claude Code. Okay, maybe that's a mistake, but there are real engineers in Meta in other companies spending like 10 million a year per engineer on these coding tools. There are users in Perplexity AI computer. There's one user I think who spends upwards of like $10,000 a month. Something like that. Crazy. And not like wasting it. They're not wasting money. Their business runs using agent loops that are running inside these harnesses. And they use these products in sophisticated ways that I couldn't even conceive when we were building the product ourselves. Even internally inside our own company, there are some people who set up this kind of multi-agent hierarchy and agent loops that looks like its own software architecture. And I often just ask these guys to come explain to the rest of the company, 'Hey, like what are you doing with these tools? Like you clearly are consuming it way over what we thought the average person in the company would do.' And the single biggest differentiation between those who use agents a lot and those who don't is whether they run repetitive cron jobs. Whether you use AIs as one-off tasks, you just delegate a task, and then it gets done. That's like using it for deep research or whatever, right? One single task. Whereas the AI is continuously monitoring something for you. The AI is continuously triggering based on certain events and going and doing certain things, giving you alerts. Like you set up workflows that keep running all the time. Every time you get an inbound email, it triages, or every time there's a latency spike, it has to identify which part of the code base caused that. It has to go and do the root cause analysis, and then identify the right engineer. All these things. This is where the frontier is. And so going back to my main point, these products are not going to be used by 100 million people. But they will generate revenue that's going to be higher than the advertising revenue of Google or Meta. It's going to happen.

完全理解你说的。我想聚焦一个具体点，就是你说到超级用户的时候，因为我认为一个核心数字是马克·贝尼奥夫说他们在 Anthropic 上花了 3 亿美元，这大约是……

Completely understand what you say there. I do just want to focus in on a specific element there when you were saying like the power users, because I think one of the core numbers is actually Mark Benioff said they spent 300 million on Anthropic, which works out to be about two.

是的。如果他能告诉我们这 3 亿美元是如何在员工之间分配的，那会很有趣。

Yeah. It'll be interesting to know from him if that 300 million came from, you know, what is the distribution across employees?

所以，这大约是 Salesforce 内部开发者的支出。约占开发者薪资的 3.8%。你认为 24 个月后，token 支出占开发者薪资的比例会是多少？因为这从根本上改变了 OpenAI 和 Anthropic 的价值。如果保持在 3.8%，它们就不会成为 5 万亿美元的公司。但如果像 McQuaid 的 Brandon 所说一年内达到 100%，它们就会成为 10 万亿美元的公司。

So, it works out to be So, that was on developers within Salesforce. It's about 3.8% of developer salaries. What percent of developer salaries do you think will be spent on tokens in 24 months' time? Because that fundamentally changes the value of OpenAI and Anthropic. If it stays at 3.8%, they will not be $5 trillion companies. But if it's 100% like Brandon at McQuaid said it will be in a year, they will be $10 trillion companies.

嗯，我认为它们当然可以成为 10 万亿美元的公司，无论是否达到今天开发者薪资的百分之百，因为还有很多非开发者的工作也会由智能体完成。这实际上正是我们在 Perplexity 计算机上关注的重点。我们瞄准的不是开发者市场。我们瞄准的是非开发者所做的一切。你的财务部门、企业开发部门、销售代表、数据科学团队、研究分析师。我认为那实际上是一个更大的市场。甚至不止如此。把它想象成云代码乘以 10。那就是那个市场的规模。

Well, I think they can certainly be $10 trillion companies, whether it's going to be a full percent of the developer payroll today or not, because there's a lot of non-developer work that will also be done with agents. And that's actually what we focus on for Perplexity computer. We're not going after the developer market. We're going after anything that non-developers do, basically. Your finance department or your corp dev or your sales reps or your data science teams, your research analysts. I think that's actually even bigger market. It's not even like that. Think of it as Claude Code multiplied by 10. That's the size of that market.

如果我追问开发者薪资支出，你认为 24 个月后 token 支出占薪资的比例会是多少？

If I push you on developer salary spend, what percent of token spend as a portion of a salary do you think we'll see in 24 months?

很难说。我认为成本会下降。这就是为什么很难说。

It's hard to say. I think the costs are going to go down. That's why it's hard to say.

你认为成本会下降吗？因为这是我们遇到的挑战。我们以为从聊天转向智能体时成本会下降，token 成本会下降。但它们却上升了。

Do you think the costs will go down? Because this is the kind of the challenge we've had. We thought when we went from chat to agent that costs would go down and token costs would go down. They've gone up.

是的，暂时如此。

Yeah, for now.

帮我理解一下，以及这种情况如何改变。

Help me understand that and how that changes.

我认为在软件领域，你愿意为前沿付费。这有点像如果你知道某个工程师很厉害。如果你有一个下一个 Jeff Dean，你宁愿雇佣那个人，而不雇佣五个中等水平的工程师，但达不到 Jeff Dean 的水平？用同样的预算？是的，对吧？假设你有一百万美元。你可以雇佣五个年薪 20 万的人，或者雇佣一个 Jeff Dean 并支付他一百万。你会怎么做？

I think in software, you kind of want to pay for the frontier. It's kind of like if you know some engineer is awesome. If you know you have a the next Jeff Dean, would you rather hire that person and not hire five people who are medium engineers but not Jeff Dean level? With the same amount of budget you have? Yes, right? Let's say you had a million dollars. You could hire five people worth 200k or you could hire one Jeff Dean and pay them a million. What would you do?

一个 Jeff Dean。

The one Jeff Dean.

是的。所以，我认为你会为前沿付费。但什么是前沿一直在变化。从现在起 12 个月后，假设有一个开源模型和 Opus 48 一样好。

Yeah. So, I think you would pay for the frontier. But what stays frontier keeps changing. In 12 months from now, let's say thought experiment there is an open source model as good as Opus 48.

嗯。

Mhm.

而且你仍然需要为推理付费。没有什么是真正免费的。但它可能比 Opus 48 便宜 10 倍。当你把它与正确的智能体框架以及所有连接器、GitHub 等一切配对时，你所有的开发者工作流都能正常运行。

And you still have to pay for inference. Nothing is truly free. But it's going to be like let's say 10 times cheaper than Opus 48. And when you pair it with the right agent harness and all the connectors GitHub everything and all your developer workflows work fine.

你为什么认为 token 消耗仍然会很高？它不会用于你今天做的同样的事情。但可能会有一些你目前无法想象的不同事情，用前沿模型来做。

Why would you assume that the token spend is going to be still high? It's not going to be for the same things you're doing today. But there might be a different set of things you might do with the frontier that you're not conceiving today.

我的预测是，会出现像完全自主的软件工程师那样的智能体。今天，我们都在使用像 Claude Code 或 Codex 这样的工具来写代码，但还不是真正的软件工程师。

My prediction would be agents that are like completely autonomous software engineers. Today I think we're all using tools like Claude Code or Codex to write code, but not as literal software engineers.

现在有一大波人对前沿模型持悲观态度，因为他们意识到用开源模型花很少的钱就能做很多事情。你实际上是在说，这确实是真的，但我们仍然会为前沿模型付费，所以它仍然能积累巨大价值。

There is a large wave of people that is now bearish on frontier models because they're realizing that you can actually do a lot with open models for a fraction of the price. What you're saying is actually that is true, but we will still pay for the frontier and so it still accrues great value.

没错。我认为这种区别感觉像是一个矛盾，但其实不是。感觉两件事不能同时为真，但事实并非如此。事实上，我认为前沿模型将越来越成为少数人甚至想要的东西。你可以说，到了一定程度，AI 能写软件这件事甚至都不再有趣了。我们已经习以为常了，对吧？假设情况就是这样。未来，公司不再由成千上万的软件工程师组成，而是会有更多公司拥有较小的软件团队，我们每个人都会使用大量 AI。这实际上对世界有好处。我们会看到许多不同的企业。我们会看到软件劳动力被分配到以前根本不可能的地方。而且，无论是什么，AI 将设计芯片、设计药物、想办法制造机器人、想办法治愈癌症。这些应用没有 1000 万用户，只有少数几家公司。但这些工作的成果将影响许多人的生活。我认为，这就是前沿模型的发展方向。你也可以从前沿实验室的动作中看到这一点。Anthropic 收购了一个兽医实验室。可能是为了人才，也可能是为了运行兽医实验的基础设施。但想象一下，把所有这些 token 都用于中间训练，而不仅仅是来自 GitHub 的 token。那将会产生一些有趣的东西。

That's right. And I think this distinction feels like a contradiction, but it's not. It feels like two things cannot be true simultaneously, but that's not quite the case. In fact, I would argue that the frontier is increasingly going to be a thing that very few individuals might even want. You could argue that after a point, it's not even interesting that AIs can write software. We've normalized it, right? Let's say that's going to be the case. Instead of companies being built with tens of thousands of software engineers, there'll be a lot more companies with smaller software teams and each of us will be using a lot of AIs. That's actually good for the world. We'll be seeing a lot of different businesses. We'll be seeing allocation of software labor in places that was never even possible. And whatever it is, AIs will be designing chips, designing drugs, figuring out how to build robots, figuring out how to cure cancer. These are applications where you don't have 10 million users. It's like a few companies. But the effect of that work will touch a lot of human lives. I think to me, that's where the frontier is headed. You could also see that from the moves that frontier labs are making. Anthropic bought a vet lab. Could be for the talent, could be for the infrastructure to run vet lab experiments. But imagine taking all those tokens and putting it in the mid training instead of just tokens from GitHub. So that's going to produce something interesting.

前沿问题是否存在一个渐近线？我知道这听起来很荒谬，但如果你不断追逐下一个前沿问题，你会遇到癌症、气候变化，我希望它们都能被解决。但如果你一直在解决这些问题，是否存在一个极限？

Is there an asymptote to frontier problems to be solved? I know that sounds ridiculous, but if you are continuously on the chase for the next frontier problem, you get to cancer, you get to climate change, and I hope they solve both. But if you're on the treadmill of continuously solving, is there an asymptote to that?

没有数学论证表明，用 AGI 或 ASI 类系统创造的经济价值存在上限。埃隆对此有一个很好的论点：他说在后 AGI 经济中，金钱将失去所有意义，因为你会生产出丰富的能源和劳动力。从根本上说，经济建立在能源和劳动力之上。如果你能大量生产它们，金钱还有什么意义？所以我认为我们不会在前沿问题上耗尽可解决的问题。我认为我们总是会创造。为什么人们甚至想要理解宇宙？为什么我们想要理解亚原子粒子、量子物理、黑洞理论、宇宙的起源？目的是什么？但我们仍然去做了，因为那一直是人类的目的：理解未知。大卫·多伊奇有句名言：我们是唯一能够对已经熟悉的事物感到好奇的物种。你可以盯着一个水果，你知道它是芒果，你知道它的味道、外观、形状、生长季节，但你仍然可以看着它，问一个你以前没问过的问题。其他动物物种做不到。一旦它们在心理模型中有了它的外观和感觉，它们就会忽略它。它不再有趣了。

There's no mathematical argument to there being a cap on the amount of economic value one can create with AGI or ASI-like systems. Elon has a good argument for this: he says money loses all meaning in a post-AGI economy because you'll be producing an abundance of energy and labor. Fundamentally, the economy is grounded in energy and labor. If you can produce an abundance of them, what meaning does money have? So I don't think we run out of things to solve at the frontier. I think we're always going to create. Why would people even want to understand the universe? Why did we want to understand subatomic particles, quantum physics, black hole theory, the origins of the universe? What is the purpose? But we still went ahead and did it because that's kind of what the purpose of humanity has always been: to understand the unknown. David Deutsch is famous for saying this: we are the only species capable of being curious about what is already familiar. You can stare at a fruit, and you know it's a mango, you know exactly how it tastes, how it looks, the shape, what seasons it grows in, but you can still look at it and ask one more question about it that you haven't asked before. Other animal species cannot. Once they have it in their mental model what it looks like and feels like, they ignore it. It's no longer interesting to them.

你提到了智能体的使用，你说如果你做重复性任务与一次性任务（比如定时任务）相比。我认为 Sam Altman 说过我们将拥有全天候的 AI，他们还谈到了一个即将推出的硬件产品。你认为我们会有持续运行的智能体吗？

You mentioned about agent usage and you said if you do repetitive tasks versus one-off say cron jobs. I think Sam Altman said we're going to have 24/7 AI and they've talked about a hardware product that's going to come out. Do you think we will have continuous agents running?

是的，我认为会。

Yeah, I think so.

我认为这就是为什么我相信我提到的编排问题——最大化 token 价值。

And I think that's kind of why I believe the orchestration problem I talked about maximizing the token value.

你能帮我解释一下吗？抱歉，你说编排问题的时候……

Can you just help me? Sorry, when you say the orchestration problem...

是的。所以有四个目标：准确性、智能和准确性，然后是隐私和成本。这些目标相互竞争。你可以说，通过建造巨大的数据中心并消耗大量电力来运行它们，你可以最大化智能和准确性。但你会牺牲隐私和成本，因为一切都会集中化，而且你要支付很多。你也可以说一切都可以在本地运行。这对隐私和成本有好处，但可能不是前沿智能或前沿准确性。所以解决方案是找到一个最佳平衡点。在必要时使用本地模型。在必要时使用服务器端模型。并在有价值的个人背景下协调本地模型和服务器端模型。有时智能可能已经存在，但系统可能无法工作，因为工具集没有正确的上下文。所以构建一个世界级的工具集，它甚至可以让一个普通的模型看起来很棒，并且能够为正确的任务和任务的正确部分使用正确的模型，以及子智能体。甚至利用我们设备上的计算资源，这些资源不需要一直在服务器上。这是一个编排问题，一个路由器。一个很棒的路由器，一个主编排路由器。如果你这样做，你就可以实现全天候 AI 的愿景，而不用担心破产。因为没有人能负担得起一个全天候运行在服务器上的前沿 AI。想象一下，你把它打开，然后永远不能关掉，除非发生什么疯狂的事情。

Yeah. So there are like four objectives: accuracy, intelligence and accuracy, and then privacy and cost. These are all competing with each other. You could argue that you could max out on intelligence and accuracy by building giant data centers and spending a lot of power to run them. And you could miss out on privacy and cost because everything will be centralized and you're going to be paying a lot. You could argue that everything can run locally. That'll be good for privacy and cost, but may not be frontier intelligence or frontier accuracy. So the solution is to figure out a sweet spot. Use local models when necessary. Use server-side models when necessary. And orchestrate across local models and server-side models grounded in valuable personal context. Sometimes the intelligence might already be there, but the system might not work because the harness isn't grounded in the right set of tools. So build a world-class harness that can even make an okay-ish model appear great and be able to use the right model for the right task and the right part of the task, sub-agents. And even utilize the compute we all have in our own devices, which doesn't need to be always on a server. That is an orchestration problem, a router. An awesome router, a master orchestrator router. Now if you do that, you can realize the vision of a 24/7 AI without people freaking out about going bankrupt. Because no one's going to be able to afford a 24/7 frontier AI running on the server. Imagine you just turned it on and you could never switch it off unless something crazy happened.

大多数人担心这些 AI 的点是‘哦，要是它发疯了怎么办？’但真正的问题其实是成本。没人能负担得起一个每秒级别精度、全天候运行的定时任务。所以瓶颈在于编排和本地算力。我认为需要构建一个持续学习的本地模型，它可以通过压缩上下文窗口来节省成本，尽量在本地保留算力，只在必要时依赖服务端前沿模型。它不断学习、适应、进化。这个模型不仅仅是模型，而是模型加上框架、本地电脑芯片以及它控制的设备生态系统。这个系统将成为你自己的智能。本质上，数据中心搬到了你的本地设备上，你可以控制它、拥有它，不用担心有人监视你或查看你的所有 token——那些非常有价值的个人 token。想象一下，你有非常敏感的交易材料。假设你在做一笔交易，而某个前沿实验室拥有你写备忘录时使用的所有 token。想象一下，有人可能黑进那个服务器，偷走你的交易。你肯定不希望那样。

The thing most people worry about those AIs is, 'Oh, what if it does something crazy?' But the real concern is the cost. Nobody's going to be able to afford a cron job at the fidelity of a few seconds that runs all the time. So the bottleneck is orchestration and local compute. I believe one needs to build a continuously learning local model that can save on compaction context windows, try to preserve as much compute locally, and rely on the server-side frontier only when necessary. It keeps learning, adapting, evolving. That model is not just a model; it's a model plus the harness plus the local chip in the computer and the ecosystem of devices it controls. That system is going to be your own intelligence. Essentially, the data center moves to your local device, and you get to control it, own it, and not worry about someone spying on you or looking at all your tokens—very valuable personal tokens. Imagine you have very sensitive deal materials. Let's say you're doing a deal, and a frontier lab has all your tokens used to write a memo. Imagine someone could hack into that server and steal your deal. You wouldn't want that.

老实说，从伦敦风投那里能偷到更有价值的东西。不过，我明白你的意思。

I'm going to be honest, there's much more valuable things for people to steal from London VCs. But yes, I can see your point.

你不只是又一个伦敦风投。我记得上次看到你有 4 亿美元的基金。想象一下，你已经在为 40 亿美元的基金布局了。每个人都有一定程度的敏感信息。我相信，全天候在线的智能体将由那些想扮演编排者角色的公司实现，不是模型构建者，不是前沿模型构建者，而是编排者。这正是我们想做的。Computers 被明确定位为智能体编排者。管弦乐队里的乐手就是那些利用不同模型的子智能体。把他们想象成乐器。工具、连接器、模型——这些都是乐器。乐手是子智能体。交响乐是作品。系统是管弦乐队。而 Computers 是指挥。这就是它的定位。它编排的东西不断进化——从模型到文件到工具到芯片到设备。但这并不重要；只要它能正确编排并最大化每个用户的 token 价值，你就不在乎。如果你能解决这个问题，长期来看你将捕获 AI 领域最大的经济价值。短期来看，可能实验室的收入在指数增长，但长期来看，这才是真正重要的目标。

You're not just yet another London VC. You have a $400 million fund last time I read. So imagine you're already making moves for the $4 billion fund. Everyone has certain levels of sensitive stuff. I believe the 24/7 always-on agent will be realized by the company that wants to play the role of the orchestrator, not the model builder, not the frontier model builder, but the orchestrator. That's what we want to do. Computers has been positioned explicitly as the agent orchestrator. The musicians in the orchestra are these sub-agents that utilize different models. Think of them as the instruments. The tools, connectors, models—these are all the instruments. The musicians are the sub-agents. The symphony is the work. The system is the orchestra. And Computers is the orchestra conductor. That's how it's being positioned. What it orchestrates keeps evolving—from models to files to tools to chips to devices. But it doesn't matter; you don't care as long as it orchestrates things correctly and maximizes the token value per user. If you can solve this problem, you will capture the most economic value in AI long-term. Short-term it might look like labs' revenue growing exponentially, but long-term, this is the one objective that truly matters.

谁最有能力做到这一点？

Who is best positioned to do that?

我相信是我们。因为我们有动力不去最大化 token 消耗，而是为用户提供最大价值。每次 AI 栈的任何部分改进，我们的产品都会改进。今年年初以来，Anthropic 的模型取得了巨大进步。但同样真实的是，我们的收入自年初以来增长了两倍多。这很大程度上要归功于 Anthropic 的模型进步。同时，由于 OpenAI 与他们竞争并降低了相同能力的成本，我们也降低了烧钱速度。现在，随着开源、本地模型和本地芯片的进步，我们将把部分推理移回本地设备，进一步降低成本。所以，每次 AI 栈的任何部分——芯片、模型、框架——变得更好，我们的系统都会大幅改进。如果我们的系统改进，用户就会喜欢并支付更多。所以，回到你的问题，谁最有能力在那个世界里赢得编排者的目标？是那些产品或业务能从栈的任何一层别人的进步中受益的公司。如果 Jensen 生产出更好的芯片，对我们有好处。如果 Dario 生产出更好的模型，对我们有好处。如果 Apple 生产出更好的设备，对我们有好处。我很高兴我们能在栈的每一层都成为一个非常正和的参与者，而不必依赖任何一个人获胜。

I believe it's us. Because we have the incentive of not token maxing. We have the incentive of delivering the most value to the user. Every time any part of the AI stack improves, our product improves. Since the beginning of the year, Anthropic's models have made tremendous progress. But what's also true is that our revenue has more than tripled since the beginning of the year. A lot of thanks to model progress made by Anthropic. We also brought our burn down thanks to OpenAI competing with them and bringing down the cost of the same capability. Now, with progress in open source and local models and local chips, we're going to move some inference back to local devices and bring down the cost even more. So, every time any part of the AI stack—chips, models, harnesses—gets better, our system improves tremendously. If our system improves, our users love it and they pay more. So, to your question of who's best positioned to win in that world for that objective of being an orchestrator, it's the one whose product or business benefits from other people's progress at any layer of the stack. If Jensen produces a better chip, it's great for us. If Dario produces a better model, it's great for us. If Apple produces a better device, it's great for us. I love that we are able to be a very positive sum player at every layer of the stack and not have to rely on any one person to win.

当我们看不同的提供商——服务端对比设备端——很多人都在谈论 AI 基础设施泡沫，我觉得这很可笑、愚蠢又弱智。从你看到的来看，我们现在在多大程度上存在数据中心供应问题？

When we look at the different providers—server-side versus on-device—a lot of people talk about an AI infrastructure bubble, which I think is funny, stupid, and moronic. To what extent do we have a data center supply problem today from what you see?

我认为最大的问题其实是电力。我们来分解一下数据中心是什么。它只是从戴尔或超微买一堆芯片吗？不，那只是一部分。你得拿到土地或租用房产，购买涡轮机发电，或者与电力供应商和电网供应商合作，还要解决冷却问题。还有很多其他工作，速度要慢得多。所有这些都需要获得许可。所以前置时间很长。今天已经在使用的模型是在 Hopper 代训练的。Blackwell 代的模型——我认为第一个是 Mito——已经令人害怕。人们已经开始为之疯狂。想象一下，每个人都在数十万个 Blackwell 上预训练模型。那些模型将比今天存在的强大得多。然后 Vera Rubin 明年将全面推出。所有配备 Vera Rubin 的数据中心明年都会投入使用。那个模型将更加强大。所以，物理建设时间总是前沿能力的瓶颈。这就是为什么那一层有价值。任何知道如何把 GPU、芯片、网络、电力、冷却整合在一起，并编排软件层将其转化为前沿输出 token 的人——这种垂直整合具有巨大价值。

I think the biggest problem is actually in power. Let's break down what a data center is. Is it just buying a bunch of chips from Dell or Supermicro? No, that's just one part. You have to secure land or lease a property, buy turbines to generate power or work with power suppliers and grid suppliers, and work on cooling. There's a lot of other work that is far slower. You have to get permits for all these things. So there's a lot of lead time. The models already in use today were trained in the Hopper generation. The Blackwell generation model—I think the first one is Mito—is already scary. People are already freaking out about it. Imagine everyone pre-trains a model on hundreds of thousands of Blackwells. Those models will be far more powerful than what exists today. Then Vera Rubin is coming next year in full capacity. All the data centers with Vera Rubin will be used next year. That model will be even more powerful. So there is a certain physical build-out time that always bottlenecks frontier capabilities. That's why there's value in that layer. Whoever knows how to put together GPUs, chips, networking, power, cooling, and orchestrate the software layer to convert that into frontier output tokens—that vertical integration has a lot of value.

所以，这就是为什么市场给基础设施公司的市盈率比 Meta 这样的公司更高。尽管 Meta 建设了大量基础设施，但它仍被当作软件公司估值。

So, that's why the markets are pricing infrastructure companies with a higher PE ratio than companies like Meta, for example. Even though Meta builds a lot of infrastructure, it's valued as a software company.

当我们看到 Meta 最近几天想扩大资本支出区间，并考虑筹集更多资金来增加资本支出时，我理解很多 AI 提供商正在让市场大开眼界，因为他们并没有从那些 AI 产品中赚钱。对 Meta 来说，资本支出与广告精准度提升相关，这能带来 6% 到 8% 的收入增长。我理解这一点。但单就资本支出区间而言，这说不通。

When we see Meta's capex band wanting to increase in the last few days and thinking about raising more money to increase capex, I get it with a lot of AI providers that are opening eyes around the base, because they aren't making money from those AI products. For Meta, the capex correlates to increasing accuracy on ads, which is like a 6 to 8% bump in revenue. I get it. But for the capex band, it doesn't make sense.

嗯，我相信他们明白市场在说什么。他们不傻，他们看到了市场的信号。从我读到的信息来看，他们正在推出很多订阅产品。所以，他们肯定会……基本上，公司不能只是一个最大化用户参与度并将其转化为广告收入的社交平台，对吧？我认为这需要他们推出大量智能体、订阅制产品，甚至可能推出像马斯克在 SpaceX 做的那种出租服务器的 Meta Cloud。也许一旦他们这样做，市场叙事就会改变，对吧？但回到我的观点，Micron（HBM 供应商）在未来 6 到 12 个月内可能比 Meta 更有价值，这并非不可想象。它已经接近一万亿美元了，而 Meta 大约是 1.3 到 1.4 万亿美元。

Well, I believe they understand what the market is saying. They're not dumb; they see what's being said. I think they're introducing a lot of subscription products from what I'm reading. So, they're definitely going to... Basically, the company needs to not just be a social platform maximizing engagement and turning that into ad revenue, right? I think that requires them to launch a lot of agents, subscription-based products, and maybe even a cloud Meta Cloud that rents out servers like what Elon is doing at SpaceX. And maybe once they do that, the narrative might change, right? But to go back to my point, it might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next 6 to 12 months. It's already at like a trillion. And Meta is like 1.3 to 1.4 trillion.

你能帮我理解这一点吗？因为内存已经是一个巨大的瓶颈。从成本来看，它的价格已经上涨了五倍。

Can you help me understand that? Because memory is already a massive bottleneck. It's increased five times in price in terms of the COGS.

没错。

Right.

但人们会说，‘哇，Micron 现在估值已经充分了。’为什么它还没有充分定价？

But people are going, 'Wow, Micron is fully priced at this point.' Why is it not fully priced?

因为它仍然是瓶颈。无论什么成为瓶颈，都会主导价格。AMD 表现很好，因为 CPU 再次成为瓶颈。智能体循环、智能体框架都在 CPU 上运行。Token 由前沿模型在 GPU 上生成，但所有其他工作——比如 Claude 生成一个编码脚本，决定从不同网站下载 500 个文件，然后处理大量数据，以特定方式转换，生成图表，并托管在网站上供他人分享——所有这些计算都在 CPU 上运行。智能体比人类更依赖 CPU，对吧？所以企业级 CPU 的需求突然上升。受益者是英特尔和 AMD。于是它们成了瓶颈。谁成为瓶颈，谁就会赢。基础设施现在是瓶颈，因为需求巨大而供应不足。所以，任何提供内存、存储 SSD、CPU 算力的公司，突然都变得有趣了。它们比那些只建数据中心却不知道如何将其转化为有价值产出的公司更重要。

Because it's still the bottleneck. Whatever is the bottleneck will command the price. AMD is doing really well because CPUs became a bottleneck again. Agent loops, agent harnesses are all running on CPUs. The tokens are produced by the frontier models on GPUs, but whatever work, like let's say that Claude generates a coding script that decides to download 500 files from different websites and then munches a lot of data and transforms it in certain ways and generates a plot and then hosts it on a website that you can share with other people. All that compute is running on CPUs. Agents are using CPUs more than humans, right? And so suddenly there's a rise in enterprise CPUs. And the beneficiaries are like Intel and AMD. So they get to be the bottleneck. Whoever is going to be the bottleneck will win. Infrastructure is the bottleneck right now because there's a lot of demand and we just don't have the supply. So whoever supplies memory, SSDs for storage, CPU compute, suddenly these are all interesting. They're more important than companies that are just building data centers and not knowing how to turn that into a valuable output.

你认为你和 Nebius 以及你们的 CoreWeave 未来会成为可持续的、价值数千亿美元的公司，还是只是在解决短期的供应问题？

Do you believe you and Nebius and your CoreWeave will be a sustainable multi-hundred-billion-dollar company in the future, or is it solving a short-term supply problem?

我当然认为它们可以持续。我认为有些——你看，我不具体知道哪一家会赢。还有像 Crusoe、Firebird 等一堆公司。关键在于资源利用能力。你需要从自然资源丰富的地区获取电力。建设数据中心的成本相当低，时间也短。而且你的服务要可靠。如果有人承诺从你这里购买 10 万块 GPU，服务必须足够好。你需要提前确保供应，做好规划。我认为有些公司甚至在电力层面进行创新，自己发电是降低利润率的一种方式。所以我认为这一层确实有价值，因为工作难以复制。这是我的看法。你可以说 OpenAI 可以做 CoreWeave 做的所有工作，这也是他们想通过 Stargate 做的事情。但为什么 CoreWeave 在建设数据中心方面比 OpenAI 更成功？

I certainly think they can be sustainable. I think there are some—look, I don't know particularly which of those is going to win. And there are also other players like Crusoe and Firebird and a bunch of companies. It's all about being resourceful. You got to take power from areas where there's a lot of natural resources. And the cost to bring up the data center is pretty cheap. And the time to bring up the data center is cheap. And your service is reliable. If somebody commits to buying 100,000 GPUs from you, the service should be pretty good. And you should be able to secure the supply ahead of time, plan well. I think some companies are even innovating at the power layer, generating their own power is one way to bring down the margins. So I think there's certainly value in that layer because it's hard to replicate work. That's how I see it. You could argue that OpenAI can do all the work that CoreWeave is doing. And that's kind of what they wanted to do with Stargate. But why is CoreWeave more successful at building data centers than OpenAI?

这很难做。运营强度很高。

It's hard to do. It's operationally intensive.

是的，运营强度很高。你必须专注。你得花大部分时间获取许可、解决电力问题、找出供应链中的各种瓶颈，不断提前规划，仔细测试所有系统，处理数据中心运营中出现的各种物理问题。有一个概念叫 TCO（总拥有成本），你必须考虑进去。话虽如此，我认为如果你只是一个服务器租赁商，那没什么价值。如果你只是按小时费率向不同公司出租 GPU 服务器机架，价值不大。你必须在此基础上构建一些软件，就像 AWS 做的那样。它叫 Amazon Web Services，而不是 Amazon Servers，对吧？所以你需要有上层的软件编排，从而获得软件层面的利润率。我认为这就是为什么你会看到像 Nebius 这样的公司转向 AI 模型推理，采用开源模型或托管你自己的模型。这是 Fireworks、Baseten 等其他公司的商业模式，但你可以想象 Nebius 也去做这个业务。

Yeah, operationally intensive. You got to focus. You got to spend most of your time securing permits, figuring out power, figuring out bottlenecks in the supply chain here and there, constantly plan ahead, test all these systems carefully, deal with random physical issues that arise in running a data center. There's something called TCO, cost of operations. You got to factor that in. That said, I don't think there's value if you're just a server renter. If you're just a GPU server rack renter, leasing it to different companies on certain hourly pricing rates, there's not a lot of value. You have to actually build some software on top, kind of like how AWS did. It's called Amazon Web Services, not Amazon servers, right? So you have to have some software orchestration on top that allows you to get software margins on top of what you're doing. I think that's why you're seeing moves like Nebius going for AI model inference, taking open-source models or hosting your models. That's a business model of certain other companies like Fireworks and Baseten, but you could imagine Nebius just going for that business.

这正是我想问的问题。我刚刚采访了 Nebius 的联合创始人，他非常清楚地指出，他面临的挑战是：有大量资金只想要容量和算力，但他意识到，如果想要长期可持续的业务，就必须构建全栈产品。这对我来说是核心认识。当我看到推理层，比如你说的 Fireworks 或 Baseten，你认为这会如何发展？我们会看到独立的、价值千亿美元的推理公司，还是会看到它商品化？

That was exactly going to be my question. So, I just had the co-founder of Nebius on the show, and the really clear takeaway was the challenge that he has, which is there's a huge amount of money that wants just capacity and compute, with the awareness that he needs to build a full-stack product if he wants to have a long-term sustainable business. That was the core realization for me. When I look at the inference layer, like you said, Fireworks or Baseten, how do you think that plays out? Do we have standalone hundred billion-dollar companies in inference alone, or do we see that commoditize?

有可能。我的意思是，这取决于逆向思考。要打造一家千亿美元的公司需要什么？假设收入为 100 亿美元。

Possible. I mean, it's all about working backwards. What does it take to build a hundred billion-dollar company? Assume 10 billion in revenue.

没错。100 亿美元收入，30% 到 40% 的毛利率，可观的净利润和现金流。

Exactly. 10 billion revenue, 30 to 40% gross margins, good amount of net income, good cash flow.

好吧，对于一家能很好运营 AI 托管推理、服务器容量和数据中心建设的公司来说，100 亿美元的收入并非不可想象。这完全取决于一些不可控因素，比如开源模型继续表现出色。如果开源模型不再优秀，或者它们与前沿模型之间的差距超过 12 个月，比如 15 个月、18 个月，那么我认为这些公司就没有商业模式了。因为它们将无法托管，只能向 OpenAI 或 Anthropic 出租容量。

Okay, 10 billion in revenue is not that inconceivable for a company that can do AI hosted inference, server capacity, and data center build-outs very operationally well. It's all about factors beyond their control, like open-source models continuing to be awesome. If open-source models stop being good or the gap between them and the frontier is more than 12 months, like 15 months, 18 months, then I don't think these companies really have a business model. Because they're not going to be able to host; they're only going to be able to rent capacity to OpenAI or Anthropic.

这正是 Emad Mostaque 所说的。他说如果发生整合，只剩下 Anthropic 和 OpenAI，或者两三个主导供应商，那对他们来说是最大的威胁。

That's exactly what Emad Mostaque said. He said if consolidation happens and there's Anthropic and OpenAI, or two or three dominant providers, that is the biggest threat to them.

没错。但你必须做一个信仰之跃，假设来自中国或 Nvidia 的模型在其模型和 NeMo 上取得良好进展，这样市场上就会有足够多的因素阻止整合发生。但如果你是那些公司，你无法掌控自己的命运。这基本上就是问题所在。

That's correct. Yeah. But you have to make a leap of faith assumption that models from China or Nvidia are making good progress on their models and NeMo, so there will be enough factors in the market to keep consolidation from happening. But you don't control your own destiny if you're those companies. That's basically the problem.

完全理解。好吧，所以我们可以有仅推理业务就价值 1000 亿美元的独立公司。我只是在榨取你的知识。当我们看像 OpenRouter 或 Foundry AI 这样的模型选择公司时，它们刚发布了一个模型选择或模型路由产品，发布时表现很好。那是模型选择和路由领域价值 1000 亿美元的业务吗？

Totally get that. Okay, so we can have standalone companies that are $100 billion in inference alone. So I'm just pillaging you for your knowledge. When we look at model selection companies like OpenRouter or Foundry AI, which just released a model selection or model routing product that did very well on launch. Is that a $100 billion business in model selection and routing?

可能不是。我认为你只能作为路由器的提供商，你必须用路由器来产生有意义的东西。实际上，OpenRouter 的大部分商业价值不在于路由器，尽管产品叫 OpenRouter，但它并不是跨模型路由。它实际上只是跨同一模型的不同端点进行路由。所以，我们问这个问题：如果你想使用 Claude Opus 或 GPT-5，为什么你不直接用你自己的 API 密钥，而是用 OpenRouter？最简单的理由是模型回退。有时你的 API 密钥可能没有速率限制，或者即使有速率限制，OpenAI 服务器也可能出错，无法保证你需要的响应时间。OpenRouter 会用他们获得的资金提前一年购买容量，确保速率限制，并在多个不同的 OpenAI 模型提供商（如 Bedrock、Azure 或 OpenAI 本身）之间拥有多个端点。这种路由是有价值的。它本质上是一个基础设施问题：解决可靠的 token 供应。它不是通过决定这个提示应该去 GPT 还是 Claude 来降低成本。那不是他们实际向开发者销售的东西。那不是商业模式。而且对于很多中国的开源模型，你可能不希望你的 API token 流向中国。如果你没有精力与不同的推理提供商合作来验证谁好，你就只能信任 OpenRouter 来处理所有事情，并为你提供 token。所以，它的路由不是在决定哪个模型更便宜或适合任务的层面，而更像是可靠的 token 供应。我认为这一层肯定有一些价值。否则，他们不会有这么多用户，每月路由数万亿 token。但这不是高利润率的业务。他们的商业模式是：通过保证大量供应从模型提供商那里获得折扣，但仍然按 API 上的标价向用户收费。这个差价就是他们的利润。你明白吗？

Probably not. I think you can just be a provider of a router; you have to use the router to produce something meaningful. Actually, most of the business value of OpenRouter is less in the router, even though the product is called OpenRouter, it's not routing across models there. It's actually just routing across different endpoints of the same model. So, let's ask this question: if you wanted to use Claude Opus or GPT-5, why would you not just use it with your own API key versus using it inside OpenRouter? The single simplest argument is model fallbacks. Sometimes your API keys might not have the rate limits, or even if you have the rate limits, there might be an error on OpenAI servers that doesn't guarantee the response time you need. OpenRouter would pay for capacity a year ahead with the funding they have and secure rate limits and multiple endpoints across multiple different providers of OpenAI models, be it Bedrock, Azure, or OpenAI themselves. That routing is valuable. It's essentially an infra problem: solving reliable token supply. It's not about lowering cost by deciding if this prompt should go to GPT or Claude. That's not what they're actually selling to the developer. That's not the business model. And for a lot of these Chinese open-source models, you probably don't want your API tokens going to China. And if you don't have the bandwidth to work with different inference providers to verify who's good, you're just trusting OpenRouter to take care of all that and supply the tokens to you. So, it's routing not at the level of deciding which model is cheaper or task, but more like a reliable token supply. I think there's some value in that layer, definitely. Otherwise, they wouldn't have this many users and trillions of tokens routed per month. But it's not a high gross margin business. The way the business model works for them is they secure a discount from the model providers by guaranteeing a lot of supply, but they still charge the user the listing price on the API. That difference is their margins. Do you understand?

我完全明白你。我们谈到了瓶颈，你提到了 HBM 高带宽内存和美光，以及它们必须说的价值和可能是什么。三年后我们会遇到什么我们今天没有讨论的瓶颈？

I totally get you. We spoke about bottlenecks and you said about HBM high bandwidth memory and Micron and the value they have to say and what it can be. What bottleneck will we have in 3 years that we're not discussing today?

我认为电力仍将是瓶颈。对我来说感觉就是这样。除非数据中心建设的方式发生巨大变化。我实际上认为建设数据中心会遇到很多阻力。这是因为人们错误地认为数据中心消耗大量水或电力，但这并不正确。Satya 甚至说过，这些公司的效率就像一罐水之类的。

I think power will remain the bottleneck. I think it feels like that to me. Unless something dramatically changes in the way data center build-outs happen. I actually believe there'll be a lot of resistance to building data centers. It's because people incorrectly think that data centers consume a lot of water or eat up a lot of power, which isn't true. Satya even made a statement that it's like a can of water or something in terms of how efficient these companies are.

你认为这就是他们抵制的原因吗？我认为这是因为它是失业、财富不平等加剧的象征。

Do you think that's why they're putting up resistance? I think it's because it's a symbol of job losses, increasing wealth inequality.

有很多原因。有很多担忧，对将要发生的事情的恐惧，以多种不同的方式表现出来。有时表现为对财富不平等的仇恨和想对富人征税。有时表现为对环境和气候变化的担忧。有时表现为，‘哦，电网价格因为你们建了这么多数据中心而上涨’，或者‘我现在为手机和笔记本电脑付更多钱，因为你们买光了所有内存，导致内存价格上涨’。所以，我认为有很多不同的表现形式，但共同的情绪是对 AI 的相当负面的情绪。

It's a lot of things. It's a lot of apprehensions, fear about what's going to happen, channelizing in so many different ways. Sometimes it's channelized through hatred for wealth inequality and wanting to tax people. Sometimes it's channeled through concerns with the environment and climate change. Sometimes it's channelized in a way where you're like, 'Oh, the price of the grid is going up because you guys are building all these data centers,' or 'I'm paying more for my phones and laptops now because the RAM prices have gone up because you guys went and bought all of it.' So, I think there's a lot of different ways in which it's getting channelized, but the common sentiment is a pretty bad sentiment about AI.

你认为这对那些数据中心的发展有意义吗？

Do you think it would be meaningful to the development of those data centers?

我认为目前 100 个数据中心中有 40 个因为公众抵制而未能开发。这就是电力瓶颈所在。你可能会看到某些国家抓住这个机会，允许这些模型构建者在那里建设数据中心。Elon 要去太空做这件事。所以这将是一个有趣的实验。因为那里可以利用大量的太阳能。而且其他国家有大量的自然资源。法规可能更友好。所以，我们仍将看到数据中心的建设。可能不会发生在美国。

I think right now 40 out of 100 are not being developed because of public resistance. So that's where the power bottleneck is. And you could see maybe certain countries seize the opportunity for this and allow these model builders to go build data centers there. Elon's going to space to do that. So that's going to be an interesting experiment. Because there's a lot of energy from the sun that can be harnessed there. And there's a lot of natural resources in other countries. Regulations might be more friendly. So, we're still going to see data center build-out. It might not happen in the US.

那你认为这种可能性有多大？

How likely do you think that is, though?

大概有 20%到 30%的可能性。我认为存在这种可能性的原因是，由于出口管制，DeepSeek 没有使用 NVIDIA 的堆栈，而是使用华为的堆栈。而且出口管制不仅针对 NVIDIA GPU，还针对 HBM，因此 DeepSeek 构建的这些架构在内存效率上要高得多。他们在 KV 缓存上做了创新，使其足够小，可以放在 SSD 上，推理时不需要高带宽内存。他们将拥有完全不同的推理架构和存储架构，因为他们不被允许使用 3D NAND。所以他们的架构将不仅仅是模型架构——模型架构已经相当不同了。他们在注意力层做了创新，在训练算法上也做了创新，使其不消耗大量互连带宽。基本上，他们的整个堆栈正在与他们的硬件、芯片、晶圆厂等垂直整合。这与美国的赌注截然不同。

It's probably like 20%, 30% chance. The reason I think there is some possibility is that because of the export controls, DeepSeek is not building with NVIDIA's stack. They're building with the Huawei stack. And because there are export controls on not just NVIDIA GPUs, but also on HBMs, these architectures that DeepSeek is building are far more memory efficient. They made innovations on the KV cache to be really small enough that you can host it on SSDs. And you don't need high bandwidth memory for inference time. And they're going to have a completely different architecture for inference, completely different architecture for storage, because they're not allowed to use the 3D NANDs. So their architecture is going to look... It's not just a model architecture. The model architecture is already pretty different. They made innovations on the attention layer. They made innovations on the training algorithm so that it doesn't consume a lot of interconnect capacity. So they made a lot of... basically their whole stack is getting vertically integrated to their hardware and their chips and their fabs and so on. So that's a very different bet from what America's making.

你认为出口管制对我们是有利还是有害？

Do you think the export controls have helped or hurt us?

尚无定论。短期来看是有利的，因为我认为开源和前沿模型之间存在发展差距的唯一原因就是出口管制。这确实有帮助，像 Anthropic 这样的公司也为此大力游说。但有可能正因为如此，他们现在在物理层面变得非常擅长。他们的一个优势是能更快地建设数据中心。电力不是问题，许可证不是问题，人员不是问题，劳动力不是问题，专业知识也不是问题。所以，通过迫使他们去建设这一切，你正在把他们变成一个更强大的竞争对手。

Jury's still out. Short-term it's helping because the only reason, in my belief, that there is even a development gap between open source and frontier is export controls. It's definitely helped and companies like Anthropic lobbied very hard for it. But there is a chance that because of that, they now get really good at the physical layer. And one advantage they have is they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so, by forcing them to go out there and build all this, you're converting them into a far more potent competitor.

你认为我们仍然严重低估了中国的能力吗？

Do you think we still dramatically underestimate China's capabilities?

我认为是的。因为如果人工智能不仅仅是数字化的，还包括物理人工智能，你就需要建造晶圆厂、机器人、芯片，并很好地利用能源，将其封装到本地设备中。我认为他们比美国有更多优势。

I think so. Because if AI is not just digital, it's also physical AI. You've got to build fabs, robots, chips, and harness the energy really well, and package it into local devices. I think they have a lot more advantages than America.

我们在美国拥有自己的台积电有多重要？

How important is it that we have our own TSMC in the US?

台积电实际上正在亚利桑那州建造一座晶圆厂。很多人都在谈论这件事，台积电正在投资约 1500 亿美元建设美国晶圆厂。据我上次了解，他们已经投资了 400 亿或 600 亿美元。所以亚利桑那州的台积电正在建设中。还有英特尔。这就是为什么美国政府持有英特尔 10%的股份，英伟达和软银各持有 5%。所以有大量投资进入美国晶圆厂，台积电也在投资其美国晶圆厂。埃隆正在建造一座超级晶圆厂。我认为人们已经意识到建造晶圆厂的重要性。但这也是中国特别有竞争力的原因。

TSMC is actually building a fab in Arizona. A lot of people talk about this, but TSMC is investing like $150 billion into building American fabs. And they've already invested $40 or $60 billion last time I checked. So there is a TSMC in Arizona that's coming up. There's also Intel. And that's why the American government owns 10% of Intel. Nvidia and SoftBank own 5% each. So there is a lot of investment going into an American fab as well as TSMC investing into its American fabs. Elon's building a terrafab. I think people have woken up to the importance of building fabs. But this is also why China is particularly very competent.

考虑到我们刚刚提到的中国的能力，如果我对你说：‘你的工作是确保美国保持竞争力’，你会怎么做来确保在一个日益强大的中国面前保持竞争力？

Given the capabilities of China that we just mentioned, if I were to say to you, 'Your job is to make sure America stays competitive,' what would you do to ensure that you retained competitiveness in an increasingly strong China?

我认为要更认真地对待物理基础设施，并继续投入资金。不要散布关于数据中心的假新闻，比如数据中心污染水源、消耗水资源，而要基于事实。我希望我们的产品能有所帮助。你可以去 Perplexity 问任何问题，对你的假设进行事实核查。但重要的是，我们要用公众容易理解的语言教育他们，而不是制造恐慌。不要说什么‘所有工作都会消失’。将会有很多了不起的公司用更少的人建立起来，20 到 30 人就能获得数十亿、数亿美元的估值，推动数万亿美元的新 GDP。让我们讨论如何实现这一点，如何建设它，共同创造一个更积极的未来。而不是‘90%的工作会消失，你们都会被我们的模型搞砸，我们有道德义务告诉你们这些’。这对我来说毫无意义。你不可能一边这么说，一边抱怨无法快速建设数据中心。

I think take physical infrastructure a lot more seriously. And continue funding it. And not propagate fake news around data centers about how data centers are polluting and contaminating water or sucking water, and actually be fact-driven. I hope our product helps there. You can go to Perplexity and ask any question and get fact-checked on your assumptions. But it's very important that we educate the public about what's actually going on in a language they easily understand and not fear-monger. Not be like, 'Oh, all their jobs are going to go away.' There are going to be lots of amazing companies that are going to get built with far fewer people, getting multi-billion dollar, multi-100 million dollar valuations with like 20-30 people, and propelling trillions of dollars of new GDP. Let's talk about how to enable that. Let's talk about how to build that. And create a more positive future together. Instead of '90% of the jobs are going to be gone, you're all going to get screwed over by our models, and it's our moral duty to tell you all this.' That doesn't make any sense to me. You can't win by saying that and also complaining about not being able to build data centers fast.

你认为我们因为 Dario 的营销信息——所有工作都会消失，一切都是悲观绝望——而造成了极大的伤害吗？

Do you think we've done a complete disservice by having the marketing message that Dario has had, that all jobs are going and it's all doom and gloom?

是的，我认为如此。我的意思是，他们自己在不同的社交媒体上传递的信息是矛盾的。我最近听到的一个说法是‘没有证据表明人工智能正在取代工作’。所以我认为需要围绕这一点进行一致的沟通。我还认为，很少有人谈论人工智能如何以非常不同的方式帮助你建立公司。比如当前的人工智能，作为通用人工智能，已经有很多你原本需要雇佣人来做的事情，现在可以用智能体来完成。

Yeah. I think so. I mean, they have contradictory messages in their own different social engagements so far. The most recent one I heard was 'there is no evidence that AI is taking over jobs.' So I think there needs to be a consistent communication around this. And I also think that very little is being spoken about how AIs can help you build companies in a very different way. Like the current AIs, where it's generic AI, it's already true that so many things you would hire people for you can do it with agents.

但一种看法是，‘哦，所有的工作岗位会怎样？’另一种看法是，‘嘿，我从来没有机会去实现我一直以来的这个想法。也许我和一群朋友可以一起组建公司。你们能想办法给我们算力积分吗？’你知道，亚马逊给了很多初创公司算力积分。我们创办 Perplexity 时，有大约 20 万美元的亚马逊积分、GCP 积分和 Azure 积分。加起来大约价值一百万美元的算力积分。在当今世界，这将是大约一百万美元的算力积分。我们正在资助一个名为‘十亿美元建造’的项目，向任何有可信路径打造十亿美元公司的团队提供一百万美元的算力积分。我希望这样的公司能出现数千家。

But one way of looking at it is like, 'Oh, what happens to all the jobs?' But the other way of looking at it is, 'Hey, I never had the chance to go build out a company on this idea that I've been having all this time. Maybe me and a group of friends can come together and build this. Can you figure out a way to give us compute credits?' You know, Amazon gave a lot of compute credits to a lot of startups. When we started Perplexity, we had around $200,000 worth of Amazon credits, GCP credits, and Azure credits. Cumulatively, that was worth about a million dollars in compute credits. In today's world, it's going to be like a million dollars of compute credits. We're funding this thing called the Billion Dollar Build, where we're giving a million dollars of compute credits to any group of people who have a credible path to building a billion-dollar company. I want thousands of such companies to be built.

你怎么看 Sam Altman 给 YC 公司 200 万美元的代币作为交换？

What did you think of Sam Altman giving $2 million of tokens to YC companies in exchange?

我认为我们应该多做这种事。这是正确的做法。我们应该做得更多，因为你希望新公司被建立起来。即使它们价值数亿美元，也是好的。如果有数千家这样的公司，那就是大量的新 GDP。

I think we should do more of that. That's the right thing to do. We should do a lot more of this, because you want new companies to be built. Even if they're worth multi-hundred million dollars, it's good. If there are thousands of them, that's a lot of new GDP.

我在节目前和 Amba Odaszki 聊过，她说 AI 是如何为你打造团队的。现在团队有多大？

I spoke to Amba Odaszki before the show, and she said how AI built the team is for you. How big is the team today?

大约 400 人。

It's like 400 people.

400 人。两年后会多大？

400 people. How big will it be in 2 years time?

我不知道，很难说。也许 800 或 1000 人。

I don't know, it's hard to say. Maybe 800 or 1,000.

那么公司会遵循以往的人员规模轨迹，只是解决新问题，还是会以更少的人实现更高的效率？

So will companies follow the same head count trajectory that they have always followed and we will just solve new problems, or will they be dramatically more efficient with a much fewer number of people?

当然会大幅提高效率。这就是为什么我相信要建立更高效的公司，并且我们自己也要成为这些公司的榜样。人们应该看看 Perplexity，然后说，‘哦，用 400 人就能打造一个价值数十亿美元的公司。’所以这意味着用 40 人我可能就能建立一个价值十亿或二十亿美元的公司。这完全可行。那么对我们来说，也许用 4000 人我们就能价值 2000 亿。用 10000 人我们就能价值两万亿。这并不意味着对那 10 万我们没有雇佣的、典型两万亿美元公司所需的人不好。我更希望那 10 万人分成 10 万个小组，每个小组价值几十亿美元。那太棒了。我认为需要更多人有创业精神。有些人无论在哪家公司都是糟糕的员工，因为他们难以合作，不听从指示，或者不容易协作。但另一面是，这些特质正是创始人通常具备的。

Definitely they'll be dramatically more efficient. That's why I am a believer in building a lot more efficient companies. And being an example for all these companies ourselves. People should look at Perplexity and be like, 'Oh, with 400 people you can build a multi-billion dollar company.' So that means with 40 people I could probably build a billion or two billion dollar company. That's totally doable. So for us, maybe with 4,000 people we could be worth 200 billion. We could be worth two trillion dollars with 10,000 people. That doesn't mean it's bad for the 100,000 people we did not hire for a typical two trillion dollar company. I would rather have those 100,000 people be split into groups of 100,000 groups, and each of those thousand groups are worth a few billion dollars. That's awesome. I think a lot more people need to be entrepreneurial. There are people who would be bad employees in any company because they're difficult to work with, they don't listen to instructions, or they're not easy to collaborate with. But the flip side is that those are the kind of qualities that founders typically have.

Arvin，有一大群人不是 AI 原住民，他们没有使用 AI 来改进工作流程、提高效率。你会给他们什么建议？

Arvin, there is a population and a very large population that are not AI native people, that are not using AI to improve workflows, improve efficiency. What would you advise them?

开始行动。第一步，开始行动。并引导你的好奇心。你不需要用 AI 来做你现有的工作。如果你觉得现有的工作很无聊，即使你用 AI 来做，你也不会喜欢。

Get started. First step, get started. And channelize your curiosity. You don't need to use AIs to do your existing work. If your existing work is boring to you, you probably won't enjoy it even if you use AIs to do it.

你因为说人们不喜欢那份工作而受到很多批评，所以……

You got a lot of heat for saying that people don't like that job, so...

我没说过。如果你真的听了我的采访，我没那么说。人们想要点击诱饵文章，他们把我一句话中的内容断章取义，做成标题。

I didn't say it. If you actually listen to my interview, I did not say that. People want clickbait articles and they take something I said in one sentence out of context and make it into a headline.

你说了什么？

What did you say?

我特别说过：‘嘿，有很多人不喜欢他们的工作。’顺便说一句，这件事走红并不是因为我完全错了。我认为很多人共鸣的是，我诚实地说出了很多人不喜欢自己的工作，而这与你的经济地位或社会地位无关。你甚至可能非常富有，但做着完全不喜欢的工作，浪费成年生活的黄金岁月在可怕或压抑的事情上。所以我的观点是，如果你是这样的人，而你无法离开工作的原因是你总是担心如何从零开始创业——有很多事情要弄清楚，如何雇佣很多人，要设立办公室，等等——这已经改变了。历史上第一次，你可以和一两个朋友一起开始一个想法，并且真正有机会建立一个十亿美元的公司。

I specifically said this: 'Hey, there are a lot of people who don't enjoy their jobs.' By the way, the fact that that thing went viral is not because I was completely wrong. I think a lot of people resonated with the fact that I was actually honest in saying a lot of people don't enjoy their jobs, and that has nothing to do with your economic position or standing in society. You might even be really wealthy but doing a job that you completely don't enjoy, destroying the peak years of your adult life working on something that is horrible or depressing. So my point is that if that's you, and the reason you could never leave your job is because you were always worried how you would build a company from scratch—there are all these things to figure out, how you would hire a lot of people, you have to set up an office, this and that—that's changed. For the first time in history, you can get started on an idea with one or two other friends and maybe have a real genuine shot at building a billion-dollar company.

我完全理解。我们今天讨论的一切都建立在前所未有的需求增长之上。我们需要更多内存，更多数据中心供应，需要按需和服务端。一切都在增长。我看到一些裂痕，比如 Uber 说‘我不确定我得到了我预期的生产力提升’。微软与他们合作，设定了 1500 美元的代币预算。你认为我们会持续接受生产力提升不可动摇、我们必须这样做，还是会有波折？

I totally get that. Everything that we've discussed today has been on the back of unprecedented demand up into the right. We need more memory, we need more data center supply, we need on demand and service side. Everything's like up into the right. I'm seeing some cracks in an Uber's saying, 'I'm not sure I'm getting the productivity gains that I thought.' Microsoft lining with them putting a $1,500 token budget. Do you think we will have a continuous up into the right acceptance that productivity gains are unwavering, we have to do this, or will there be falterings along the way?

我确信过程中会有波折，人们有理由对代币最大化感到恐慌，这就是为什么我认为你需要某种形式的混合智能体式推理。你需要一些本地运行的推理算力，这样你就不需要为代币付费，本质上是不计量的智能。

I'm sure there's going to be falterings along the way, and people are rightfully freaking out about token maxing, which is why I think you need some form of hybrid agentic inference. You need some amount of inference compute to run locally that you're not paying for tokens on, unmetered intelligence, essentially.

未来最好的公司将如何构建代币预算？

How will the best companies of the future structure token budgets?

我的希望是他们不需要理解这些。他们将能够与一个为他们做这些的编排器合作。你不可能持续跟踪哪些模型在哪些方面最好，以及如何分配？‘哦，这是编码的预算。这是财务的预算。’你甚至如何理解哪些模型擅长这些领域，以及每个部门花多少钱？你无法跟踪。

My hope is that they don't have to understand that. They will be able to work with an orchestrator who does it for them. It's not going to be easy for you to constantly keep track of which models are the best at what things, and how do you allocate? 'Oh, this is the budget for coding. This is the budget for finance.' How do you even understand which models are good at each of those things, and how much do you spend on each of these divisions? You're not going to be able to keep track.

我前几天在节目里有个朋友说，谷歌将成为代币之王。他们能以最低成本生产代币。

I had a friend on the show the other day say that Google will be the token king. They can produce the lowest cost tokens out of anyone.

前几天我看到 Cloudflare 的公告，说现在智能体流量已经超过了人类流量，我很震惊。

I was shocked the other day I saw the Cloudflare announcement that now agent traffic has overtaken human traffic for them.

你为什么震惊？

Why are you shocked?

比我预想的快。我个人觉得这会发生，但可能是在两年后，而不是现在。

It was quicker than I thought. Personally, I thought that would happen, but in 2 years, maybe not now.

当智能体流量远超人类流量时，世界会如何变化？

How does the world change when agent traffic far exceeds human traffic?

我认为人们将拥有更多的自主能力。就是这样。

I think people are just going to have a lot more agency. That's it.

比如网站会消失吗？设计不再重要吗？互联网的广告模式会彻底消亡吗？

Like do websites go away? Does design not matter? Does the advertising model of the internet die completely?

不，不会。因为我相信，围绕旅游、购物或时尚的广告模式不会被智能体颠覆，因为判断不是客观的。任何判断是客观的、交易基于客观判断的领域，都会被智能体颠覆。任何交易更主观的领域，比如决策更主观，像这个房间里最好的家具是什么？为什么选这张桌子？这类事情。可能对于麦克风，你会做客观决策。而对于桌子，你可能更关心房间的美学。我觉得世界会这样分裂：主观的东西仍然基于广告，客观的东西则基于智能体。

No, it doesn't. Because my belief is that the advertising model around like travel or shopping or fashion are not getting disrupted by agents because the judgment is not objective. Anything where the judgment is objective, the transaction is based on objective judgment, that's going to get disrupted by agents. Anything where the transaction is more subjective, like the decisions are more subjective, like what is the best piece of furniture inside this room? Why this particular table? Those kind of things. Probably for a microphone you would buy an objective decision. The table you probably are caring about the aesthetics of the room. I think that's kind of how I feel the world will split and subjective things will still be ad-based. Objective things will be agent-based.

我在和 Sam Altman 在 Excel 对话后看了你的毕业演讲，他说我必须看。所以我当然看了。你提出的一个观点是，AI 时代的决定性技能是提出更好的问题。

I watched your commencement speech on the back of speaking to Sam Altman at Excel and he said I have to watch it. So obviously I watched it. And one of the points you made was the defining skill of the AI era is asking better questions.

是的。

Yeah.

今天没有人问、但也许每个人都应该问的问题是什么？

What question is no one asking today that maybe everyone should be asking?

我认为人们需要更多地问：好吧，假设我拥有很多自主能力。我该做什么？想象我给你 10 万或 1 万人的团队，以及足够的算力额度来运行这些智能体。你会做什么？比如说我问你 Harry，假设你突然有 1 万个智能体供你差遣。你会做什么？我记得在你的一期节目中，你说你做播客只是因为你觉得没有套利机会去做交易。

I think people need to ask more about like okay, assuming I have a lot of agency available to me. What do I do? Imagine I gave you a head count of like 100,000 people or 10,000 people. And enough compute credits to run those agents. What would you do? Let's say I ask you Harry, like you know, let's say you have suddenly 10,000 agents at your disposal. What would you do? I remember in some episode of yours where you said you only did this podcasting because you felt like you didn't have an arbitrage to go and do deals.

百分之百。

100%

是的。

Yeah.

为什么我还在做。我是说我喜欢我所做的，但确实。

Why I still do it. I mean I love what I do but yeah.

好的，所以你已经有了一定的分发能力。现在假设你可以花 1 亿美元在推理上，并用所有连接器等把它落地，而且一切正常。你会用这种能力做什么来推进你的目标？你的目标甚至应该是什么？我认为这就是我会问的问题。假设在未来 3 到 5 年内，你将能够通过合适的工具和智能体委派任何数字任务，并且能够做到这一点。

Okay, so you've gotten some amount of distribution. So now assuming that you could spend $100 million on inference and ground it with all the connectors and stuff and it's all working. What would you do with that capability to further your goals? What should your goals even be then? I think that's the question I would ask. Assuming that in the next 3 to 5 years you're going to be able to delegate whatever digital task you want with the right harness and agents and be able to delegate that.

从根本上说，就是建立一个巨大的基础设施，能够发现、识别、接触、建立、赢得伟大的投资，并让媒体在上面驱动这一切。这非常困难，将是投资的圣杯。但这将推动我最终的目标和抱负。

Fundamentally it would be to build a gigantic infrastructure to be able to find, identify, outreach, set up, win great investments, and have the media sit on top and power that. That is intensely difficult to do and would be the holy grail to investing. But that would power what my end goal ambition is.

是的，所以你的目标是运营一个规模大 10 到 100 倍的基金，对吧？这就是我从你那里听到的。那么假设你有一个 400 亿美元的基金，从 4 亿美元起步。那么你只需要问：假设我有所有需要的人力，我能多快完成？我认为这就是我构建这个问题的方式。我觉得 Elon 也说过类似的话：好吧，假设有人告诉你一项任务需要 10 年。问一问：怎样才能在 10 个月内完成？也许 10 个月是不可能的，但相比于那些理所当然认为需要 10 年的人，你问这些问题可能会走得更远。

Yeah, so your goal is to run like a 10 to 100X larger fund, right? That's basically what I'm hearing from you. So let's assume you have a $40 billion fund from $400 million. Then all you got to ask is assuming I have all the head count I need to do this, how much faster can I do it? I think that's how I would frame this question. I think Elon has a similar thing he spoke about once: okay, assume that a task somebody tells you is going to take 10 years. Ask the question what would it take to do it in 10 months? Maybe it's impossible to do it in 10 months, but you'll probably get pretty far asking those questions compared to somebody who takes it for granted that it's going to take 10 years.

好吧，采访者，让我把问题抛给你。你的 10 年目标是什么？在 10 个月的时间框架内看起来怎么样？

All right, interviewer, let me put it on you. What's your 10-year and how does that look in a 10-month time frame?

我认为我们的使命，超越任何资本主义层面，是让地球更充满好奇心。产品始终旨在帮助人们提出下一个问题。我的目标是真正实现这个世界需要存在的自主能力水平，但目前还远远不够。

I think our mission beyond any level of capitalism is to make the planet more curious. The product is always intended to help people ask the next question. And my goal is to truly realize that the level of agency that needs to exist in this world is quite not there.

我觉得这需要以数字为基础，老兄，才能让它可行。就像我说“哦，我想要最好的投资。”所以一个 400 亿美元的基金是有帮助的。

I think that needs to be grounded in numbers, dude, to make it possible. Like it's like me saying, 'Oh, I want the best investments.' Well, which is why a $40 billion fund is helpful.

当然，我也可以说同样的话。比如 2 万亿美元。你知道，这没关系，对吧？比如 100 倍、10 倍、1000 倍。这些都像是激励性的里程碑。

Sure, like I could say the same thing. Like $2 trillion. You know, it doesn't matter, right? Like 100X, 10X, 1,000X. These are all like motivational milestones.

你认为 Publicity 会成为万亿美元公司吗？

Do you think Publicity will be a trillion-dollar company?

是的。任何人都可以成为万亿美元公司。SK 海力士和三星在过去几周市值达到了万亿美元。你知道三星最初是一家杂货店吗？你知道吗？你不知道？好吧，是真的，他们一开始卖干鱼。认真的。海力士是 SK 集团，SK 集团最初是一家纺织公司。所以，任何人都可以成为万亿美元公司。你只需要朝着那个方向努力。我的意思是，你之前阐述的公司如何达到 1000 亿美元的逻辑完全一样。你说你需要 100 亿美元的收入。万亿美元不也一样吗？你需要 1000 亿美元的收入？

Yeah. Anyone can be a trillion-dollar company. SK Hynix and Samsung are worth a trillion last couple of weeks. Did you know Samsung started off as a grocery store? Did you know that? You didn't know that? Okay, so it's true they started selling dried fish. Seriously. Hynix was the SK Group started off as a textiles company. So, anyone can be worth a trillion-dollar company. And you just have to work your way towards that. I mean, the exact same logic for you that you laid out for how can a company be worth $100 billion. Okay, you said you need to make $10 billion in revenue. Isn't that the same for a trillion? Like you need to make $100 billion in revenue?

实际上，Coatue 最近披露了一些非常有趣的数据。我不知道你是否看到了，基本上是关于达到下一个价值水平的概率。

And there was actually some really interesting data that Coatue revealed. I don't know if you saw it recently, which basically says about the probability of reaching the next level of value.

是的。

Yeah.

概率高得多。所以，当你达到 10 亿时，达到 100 亿的可能性大得多，100 亿达到 1000 亿的可能性也大得多。

It's much higher. So, when you're at a billion, it's like much more likely to reach 10 billion, 10 billion much more likely.

是的，实际上对人来说也是如此。一个拥有 1 亿美元流动净资产的人成为亿万富翁的可能性，比拥有 1000 万美元的人大得多。

Yeah, that's true actually for even people. It's more way more likely for a person with a hundred million dollars in liquid net worth to become a billionaire than someone with 10 million dollars.

你不担心财富不平等吗？我的意思是，老实说，坦率地讲，我们俩现在都很幸运，生活在相当不错的世界和稀薄的空气中。

Are you not worried about the wealth inequality? I mean, honestly, if we were being blunt, we both are very lucky now to live in kind of nice worlds and rarified airs.

你不担心极少数人拥有那么多钱，而其他人却那么艰难，而且差距还在扩大吗？

Are you not worried by just how much money a very small number of people have and how hard it is for everyone else and that gap is getting bigger?

我认为确保这种情况不持续下去的方法就是更广泛地分配收益。顺便说一句，使用我们工具的人——我遇到过一位优步司机，我没编故事，实话实说——旧金山的一位优步司机曾告诉我，他看了我的一段 YouTube 采访，我在里面解释了如何从零开始用 AI 构建产品或网页应用。他真的去做了，还用 AI 添加了计费等功能，这给他带来的被动收入比开优步还多。所以他实际上减少了开优步的时间，因为他喜欢用编码开发新应用。这已经告诉你，对于有能动性和积极展望未来的人来说，一切皆有可能。如果你一直传播关于 AI 和财富不平等的负面信息，而新闻和媒体只写这些，我认为这会延续下去，人们只会想到坏的一面。所以，如果你觉得自己已经做得不错，谈论哪些事情可以变好，给那些处于低谷的人希望，这非常重要。就像你，你开始做播客时也是一无所有，对吧？

I think the way to ensure that doesn't remain the case is to distribute the benefits more widely. By the way, the people who are using our tools—I've had an Uber driver, I'm not making this up, as honest as I can get—an Uber driver in San Francisco once told me that he watched one of my YouTube interviews where I explained how you can build a product or a web app with AI from scratch. He went on to do it, and used AIs to add billing and all that, and that makes more passive income for him than driving Ubers. So he actually reduced the amount of time he's driving Uber because he loves coding new apps. That already tells you that for the person with agency and a positive outlook for the future, anything is possible. If you keep communicating all the negative things about AI and wealth inequality all the time, and that's the only thing news and press writes about, I think it'll perpetuate and people will only think the bad things. So it's very essential that if you think you're already doing well, you talk about what are all the things that can go well and give hope to people who are down. Like you, you started this podcasting circuit when you had nothing, right?

一无所有。

Nothing.

没错。所以这是可能的。你应该多谈谈这个，而不是说‘哦，我成功了感到很内疚，现在我知道那些还没成功的人’，你也可以成功。

Exactly. So it's possible. You got to talk more about that than be like, 'Oh, I feel so guilty that I made it and now I know about all these people who haven't made it.' You can also make it.

我对普通大众的看法更悲观。我不认为很多人有能动性。我觉得很多人有自己的心态问题。

I think I have a more pessimistic view of the general public. I don't think that many people have agency. I think a lot of people have their own mentality.

去帮助他们。我认为那是最重要的。

To help them. I think that's the most important thing.

我认为他们必须自助。

I think they've got to help themselves.

当然，但人们一旦看到‘好吧，我有点想像这个人一样，让我努力吧’，他们就会自助。你需要一个榜样，对吧？并不是说没人能塑形。这需要自律。你得改掉坏习惯。现在是 12 个月内改变人生的最佳时机。从一无所有到亿万富翁在 12 个月内实现，在某些方面现在是可能的。听着，我不是说每个人都能成功，每个人都能值十亿美元。

Sure, but people will help themselves once they see that, okay, I kind of want to be like this guy. Let me work hard. You need an example, right? It's not like nobody can get in shape. It takes discipline. You got to get rid of bad habits. And now is the best time ever to change your life in 12 months. The ability to go from nothing to billionaire in 12 months is now possible in some respects. Look, I'm not saying everyone's going to make it and everyone's going to be worth a billion dollars.

这不是这个节目的标题吗？Aravind，每个人都会成功。

Isn't that the caption from this show? Aravind, everyone's going to make it.

每个人都有成功的潜力。所以 Perplexity 价值 2 万亿美元和一位尚未获得融资的创始人价值十亿美元的可能性是一样的。同样困难。我认为你只需要给自己射门的机会。保持好奇。这就是毕业演讲传达的信息。保持好奇。

Anyone has the potential to make it. So it's as likely for Perplexity to become worth $2 trillion as for a founder who's yet to secure funding to be worth a billion dollars. It's equally hard. And I think you just have to give yourself shots at the goal. And be curious. That's the message from the commencement speech. Be curious.

我们有 Space Eyes，有 Anthropic，有 OpenAI 要上市。感觉就像有人开了枪，比赛开始了。有足够的钱来支持三个这么大的 IPO 吗？

We have Space Eyes, we have Anthropic, we have OpenAI going public. Feels like someone's kind of shot the gun and the race is on. Is there enough money to fund three such large IPOs?

肯定会有一些重新配置。可能有一些 SaaS 股票的持有者会把它投入 Anthropic 之类的。假设你相信企业 AI 会起飞。你可能想在持有大量微软和 Salesforce 股票与将其中一部分投入 Anthropic 之间进行对冲。所以假设 Vanguard 或 BlackRock 总共持有 2000 亿美元的微软和 Salesforce 股票。他们可能会说，‘好吧，我要从中拿出 300-400 亿美元投入 Anthropic。’没问题，这不是一个糟糕的赌注。

There will be some reallocation for sure. There might be some holders of SaaS stocks who would put it into Anthropic or something. Let's say you believe that enterprise AI is going to take off. You might want to hedge between having a lot of Microsoft stock and Salesforce stock versus putting some of that into Anthropic. So let's say Vanguard or BlackRock own cumulatively $200 billion of Microsoft and Salesforce. They might be like, 'Okay, I'm going to take $30-40 billion of that and put it into Anthropic.' Fine, not a bad bet to make.

那些上市的 SaaS 企业会怎么样？‘不，没事。’

What happens to all the enterprise SaaS companies that are public going, 'Nah, fine.'

他们必须经受住风暴。

They have to weather the storm.

是风暴还是持续降水？

Is it a storm or is it a continuous precipitation?

我认为你必须降低成本并创造新价值。Salesforce 做得很好，因为他们总是去收购下一个东西。如果你只是卖同样的软件，你可能就活不下去了。IBM 还在，因为他们收购了 Red Hat 和 HashiCorp。现在他们正在收购 Confluence。所以这些公司有办法存活并延长寿命。显然，保持一个品牌的相关性很难。我认为 IBM 品牌在激发人们使用其产品的情感方面已经不那么相关了。但作为一家企业，它会很棒，会没事的。

I think you have to bring down the costs and produce new value. Salesforce has done well because they always went and bought the next thing. If you're just selling the same software, you're probably not going to be around. IBM is still around because they went and bought Red Hat and HashiCorp. And now they're buying Confluence. So there are ways for these companies to stay alive and extend their lifespans. It's obviously going to be hard to preserve a brand that's as relevant. I don't think the IBM brand is that relevant anymore in terms of evoking an emotion in people to go use their products. But as a business, it's going to be awesome, it's going to be fine.

我必须以此结束。你说过 2028 年 IPO。我不得不问这个。我在团队 WhatsApp 里看到这个，就像‘Oscar Evan，2028 年 IPO。’

I have to finish on this. You said IPO in 2028. I had to ask this. I woke up to this in my team WhatsApp, and it's like, 'Oscar Evan, IPO 2028.'

我希望能比那更早。

I hope it can be sooner than that.

你怎么知道什么时候准备好了？是达到 10 亿 ARR 吗？你现在是 5 亿 ARR？

When do you know when you're ready? Is there like a billion in ARR? You're at 500 million ARR now?

比那更多。实际上远远超过。

More than that. Far more than that, actually.

真的吗？什么——

Really? What—

我们还没准备好分享，但增长非常快。

We're not yet ready to share it, but growing really fast.

营收增长对你来说比盈利能力重要得多。

Revenue growth matters much more to you than profitability.

今天。顺便说一句，我认为总的来说，你可以看看公开市场。现在人们更想要营收增长，而不是利润效率。因为这很难，很罕见。

Today. I think in general, by the way, you can look at public markets. People want top-line growth more than bottom-line efficiency right now. Because it's very hard. It's rare.

但你肯定需要其中一个。

But you definitely need one.

当然，可持续的业务。你需要有一个模型来实现利润效率，当这成为目标时。而且你还需要有一条通往那里的路径。

Of course, sustainable businesses. You need to have a model in place to get the bottom-line efficiency when that becomes the objective. And you need to also have a path to getting there.

你今天在哪些方面成本效率低下，预计 2 到 3 年内会显著改善？

Where are you cost inefficient today, where you expect to be significantly better in 2 to 3 years?

我们正在训练自己的模型。我们在优秀的开源模型基础上进行训练。这将降低我们目前在前沿模型 token 上的支出。我们预计将继续使用前沿模型来设计新产品中目前不存在的新体验和新能力。但我们现在产品中已有的东西，我们预计将完全依赖我们自己拥有和服务的模型。这将是降低成本和提高利润率的最佳方式。

We're training our own models. We're training it on top of amazing open-source models. And that will bring down the cost that we currently spend on frontier model tokens. We expect to continue to use frontier models for designing new experiences and new capabilities that do not exist today in our products. But whatever exists today in our products right now, we expect it to completely rely on models we own and serve ourselves. And that's all going to be the best way to bring down the costs and increase our margins.

世界上最大的企业会不会微调开源模型，以获得更适合它们的定制模型？

Will the largest enterprises in the world be fine-tuning open models to have tailored models that are much more specific to them?

绝对会。因为降低成本符合你的利益。

Absolutely. Because it's in your incentives to bring down the costs.

这难道不是为大型前沿模型提供商提供了另一个不利案例吗？

Does that not provide another bad case for the large frontier model providers?

前沿模型提供商只有保持在最前沿才能保持相关性。如果六个月没有新能力出现，对他们来说就很糟糕。这就是这个领域令人不安的本质。没有人能处于舒适的位置。就像我一开始说的，没人能放松。这是一匹马，它很忙，跑得更快。它还会变得更难。这就是本质。代价太大了。比如 Anthropic，我认为它价值大约 1 到 1.5 万亿美元，差不多就是 Meta 的估值。而这一切是在大约 6 年内创造的。Meta 花了 20 年才建成。所以代价如此之大。没有人能感到舒适。今天赢的人明天可能输，包括模型提供商。

Frontier model providers will only remain relevant if they remain at the frontier. If for 6 months you're not seeing a new capability, it's bad for them. And so that's the uncomfortable nature of this field. No one's ever in a comfortable position. Like I said at the start, no one can relax. This is a horse. It's busy, it's going harder. It's going to get even harder. And that's the nature. The price is too big. Like take Anthropic. I think it's worth like 1 to 1.5 trillion, something in that range. That's basically the valuation of Meta. And all this was created in like 6 years. Meta took like 20 years to build. So the price is so big. And so no one can be comfortable. And anyone who's winning today can lose tomorrow, including the model providers.

今年之前，有大约三个月的时间，人们会说：“哦，但是 Perplexity，Perplexity 怎么了？”你关注吗？你在意吗？

Pre-this year, there was like a 3-month period where people like, "Oh, but Perplexity, what's happening with Perplexity?" Do you pay attention? Do you give a care?

我当然关注所有这些。

Of course I pay attention to all of that.

在旧金山有一个特别的活动。你还记得他们问：“哦，你会做空哪家公司？”

There was one in particular in San Francisco. Do you remember where they were like, "Oh, what's the company you would short?"

是的，我们被评为最可能失败的公司。Cursor 被评为第二最可能失败。OpenAI 被评为第三之类的。

Yeah, we were voted the most likely to fail. Cursor was voted the second most likely to fail. OpenAI was voted the third or something.

你不在乎……

You didn't give a...

我觉得我们都做得很好。

I feel like we're all doing well.

Cursor 我觉得它要被卖了。SpaceX，OpenAI 很快要上市了。

Cursor I think it's getting sold. SpaceX, OpenAI going public soon.

上市了，宝贝。自从那个判断做出以来，我们的收入翻了三倍。所以，我们的亏损减少了超过 50%。所以，我不知道。我的感觉是，那些参加这些聚会的人大多数实际上并没有构建任何有用的东西。

Going public, baby. We tripled our revenue since that judgment was made. So, we're out of the burn by more than 50%. So, I don't know. My sense is that most of those people who sit on these meetups don't actually build anything useful.

嗯。

Well.

是的。

Yeah.

好了，我们来个快速问答环节，因为我可以跟你聊一整天。第一个，你认为哪个广泛持有的观点是完全错误的？

Okay, we're going to do a quick fire round because I could talk to you all day. First one, what's one widely held belief that you think is completely wrong?

我认为很多人痴迷于，你知道，在公司头一两年就确定一个模式。但我认为你唯一的机会就是快速行动。在我看来，速度，快速行动是一种表达谦逊的方式，因为你不断与世界接触，并一直试图质疑你的假设。

I think a lot of people are obsessed about, you know, from identifying a model in the first year or two of their company. But I think the only shot you have is to move fast. Velocity in my mind, moving fast is a way of expressing humility because you're constantly making contact with the world and trying to question your assumptions all the time.

今天你在内部哪些方面仍然行动太慢？

Where are you still moving too slow internally today?

我认为我们可以更加主动。我这么说很疯狂，因为我们正在构建一些最有趣的 AI 产品，而内部对我们自己产品以及竞争对手产品的采用率本可以更高。尽管我们内部已经极度智能体化，并试图尽可能多地委托给智能体。是的，这是一个很大的领域。我的希望是，我们能把这家公司几乎变成一个 AGI。这并不意味着没有人类在这里工作。但会有一个 AGI，它拥有所有必要的上下文，以半自主的方式运行公司的不同部门，人类在这里那里提供一些脚手架。这根本不会让人感到害怕。我们会很快让这种感觉正常化。它只会感觉像 10 倍工程师在管理公司的某些方面。

I think we can be even more reactive. It's insane I'm saying this because we are building some of the most interesting AI products and internal adoption of our own products, our competitors' products can be even higher. And this is despite us being extremely agent-built internally and trying to delegate as much to agents. Yeah, that's a big area. My hope is that we can turn this company almost into an AGI. That doesn't mean no humans work here. But there will be an AGI that has all the context it needs to run different divisions of the company in a semi-autonomous way with some scaffolding provided by humans here and there. And that's not going to feel scary at all. We'll normalize that feeling very fast. It's just going to feel like 10x engineers running certain aspects of the company.

如果我给你无限的钱，你今天会做什么你现在没在做的事？

If I gave you unlimited money, what would you do today that you're not doing?

我会建数据中心。

I would build data centers.

你会？在太空？

You would? In space?

我没有这方面的专业知识，但我会从地球上的土地开始。我认为有很多土地，也许你可以在不同国家巧妙地获得许可和电力，但我会从那里开始。我认为物理基础设施的建设就像是工业时代的回归。那些建设工业革命的前辈们，石油管道、钢桥、生产汽车的工厂，所有这些我们今天认为理所当然的东西，都是由那些花费大量时间思考如何以成本高效的方式扩展这些事物的人建造的。所以我们需要为 AI 做很多这样的事情。是的，这就是我会做的。当然，你不能只建设基础设施。你需要能够利用所有这些基础设施为用户生产有价值的输出 token。但我们已经很擅长这个了，所以基础设施是我会专注的事情。

I don't have expertise to do that, but I would start with land on Earth. I think there's a lot of land and maybe you can be resourceful in securing permits and power in different countries, but I would start there. I think physical infrastructure build-outs is like the return of the industrial age again. The forefathers who built the industrial revolution, oil pipelines, steel bridges, factories producing cars, all these things that we take for granted today were built by people who spent a lot of time thinking about how to scale these things in a cost-efficient way. So we need to do that a lot for AI. And yeah, that's what I would do. Of course, you cannot just be building infra. You need to be able to utilize all that infra to produce valuable output tokens to the user. But we're already good at doing that, so infra is the thing I would focus on.

你可以买入并持有 10 年，SpaceX、Anthropic 或 OpenAI。这三家即将在未来几个月上市。你会买入并持有哪一家 10 年，为什么？

You can buy and hold for 10 years, SpaceX, Anthropic, or OpenAI. The three IPOs coming in the next few months. Which would you buy and hold for 10 years and why?

SpaceX。

SpaceX.

为什么？

Why?

它是一家独一无二的公司。Anthropic 和 OpenAI 可以声称他们做对方做的任何事情。但 SpaceX 是唯一一家为连接性建设太空基础设施的公司。你坐过有 Starlink 的航班吗？

It's an enough one company. Anthropic and OpenAI can claim they do whatever each other does. But SpaceX is the only company building space infrastructure for connectivity. Have you been on a flight with Starlink?

没有。

No.

你应该试试。之后你会讨厌没有 Starlink 的航班。想象一下，我们可以录制这个，我可以在飞机上飞行时观看这个播客。Starlink 让你做到这一点。这只是业务的一个方面。只是……

You should. You will hate being on a flight without Starlink after that. Imagine we can record this, I can watch this podcast while flying on a plane. Starlink lets you do that. That's just one aspect of the business. It's just one aspect of the...

业务的一个小方面。

One small aspect of the business.

是的，有很多。我很兴奋有可能在 30 分钟内从澳大利亚旅行到旧金山。这一切都感觉像科幻小说，但我对这些可能性感到兴奋。

Yeah, there's a lot. I'm excited about possibilities to travel from Australia to San Francisco in like 30 minutes. All this feels like sci-fi, but I'm excited about all these possibilities.

什么工作今天不存在，但在 5 年内会变得非常普遍？

What job does not exist today that will be incredibly common in 5 years time?

我认为它已经存在了。前向部署工程师肯定在兴起。我猜是那些有很好质量控制意识的人。也许更好的答案是，大多数存在的工作，比如有价值的工作，通常是已有事物的化身。所以我不认为我们会看到全新的东西。它会以不同的方式化身。

I think it already exists. The forward deployed engineer is definitely on the rise. I guess people with a really good sense of quality control. Maybe a better answer is that most jobs that exist, like valuable jobs that exist, are usually reincarnations of something that already existed. So I don't think we're going to see completely new things. This is going to reincarnate in different ways.

你可以给你今天刚大学毕业、刚读完计算机科学学位的弟弟妹妹一个建议。一件事，你会建议他们什么？

You can advise your little sibling who's finishing university today and just done a computer science degree. One thing, what would you advise them?

保持好奇心。不要屈服于 FOMO，试图在短期内最大化某件事。不要上 Twitter，觉得自己是个失败者，因为前沿实验室的人变得如此富有，一切对你来说都无望。还有更多东西要建。我们才刚刚开始。应用层时代或基础设施建设。有很多机会。

Stay curious. Don't give in to FOMO and try to max out on something here in the short term. Don't go to Twitter and feel like a loser that people on Frontier Labs are getting so rich and everything feels hopeless to you. There is so much more to build. We're just getting started. The application layer era or infrastructure build-outs. There's a lot of opportunities.

我们每天看到越来越多的从 OpenAI、Anthropic 等公司分拆出来的公司。我们会有数百个这样的 Neo Labs 和垂直模型吗？

We are seeing more spin-outs from OpenAI, Anthropic, you name it, every single day. Do we have hundreds of these Neo Labs and vertical models?

不。我不太相信会有太多。

No. Not a big believer in too many of them.

我觉得你得做出一些差异化。这是最重要的。比如，你会把 DeepSea 称为新实验室吗？

I think you got to produce some differentiation. That's the most important thing. Like, would you call DeepSea a neo lab?

不会。

No.

为什么？

Why?

我觉得很傻的是，我不叫它新实验室，因为我认为新实验室是大实验室的衍生品。

I think very stupidly for me, I don't call it a neo lab because I attribute neo labs to spin-outs from larger labs.

我明白了。而且有点垂直化，这可能两个维度都错了。

I see. And kind of verticalized, which is probably wrong on both axes.

但它是横向的，也不是衍生品。

But it's horizontal and it's not a spin-out.

是的，我有点喜欢实验室下差异化赌注的想法。如果有人真的质疑 Transformer 架构本身，或者有人真的质疑需要在 Nvidia GPU 上构建，诸如此类的基础赌注，或者有人质疑并去构建机器人模型。我认为这有点不相关且不同，这对实验室来说是有意义的，但我觉得有些实验室只是为了存在而存在，我不认为它们能成功。

Yeah, I mean I kind of like the idea of labs taking a differentiated bet. If somebody really questions the Transformer architecture itself or somebody really questions needing to build on Nvidia GPUs or something like that, like foundational bets, or somebody questions or goes out and builds for robotics models. I think that's somewhat uncorrelated and different and that makes sense for a lab, but I feel like there are just labs for the sake of being a lab and I don't think they're going to make it.

你能给我描绘一下——我确实喜欢这个——Perplexity 成为万亿美元公司的最合理故事是什么？你那时会做什么？编排层……

Can you paint for me—I do like this one—what's the most plausible story where Perplexity becomes a trillion-dollar company? What do you do then? The orchestration layer of...

我的意思是，准确性和编排是我们公司从一开始就持续选择的两个目标。所以我认为我们会继续这样做。我们将在设备、芯片、模型、工具、文件、连接器等各种东西上进行编排，对吧？那么一旦发生这种情况，我会做什么？我不知道。我们会规划通往 10 万亿美元的道路。

I mean accuracy and orchestration are like two goals that have been consistently chosen since the beginning of our company. So I think we'll continue to do that. We'll be orchestrating across devices, chips, models, tools, files, connectors, everything, right? So what would I do once that happens? I don't know. We'll chart our path to 10 trillion.

你现在开心吗？但你享受这个过程吗？

Are you happy now? But are you enjoying this?

当然。我的意思是，如果不是为了这个，我有很多事情可以做。我认为过程才是激励你的东西。所以你问我——我觉得你需要在某个地方给我一个你想要的数字。实际上我不是那样工作的。比如，这些数字，比如达到 2 万亿或 20 万亿，令人兴奋，但这并不能激励我。很难被财富激励。你想被影响力激励。

Of course. I mean there are so many things I could be doing if not for this. And I think the process is what motivates you. So you asked me—I think somewhere in between you need to give me a number of where you want. I don't work like that actually. I like, for example, these numbers like getting to two trillion or 20 trillion are exciting, but that doesn't motivate me. It's hard to get motivated by wealth. You want to get motivated by impact.

你见过的最聪明的人是谁？最后一个问题。你见过黄仁勋，你见过最顶尖的人。我很幸运——谁最聪明？

Who's the smartest person you've met? Final one. You've met Jensen Huang, you've met the best of the best. I've been fortunate enough to—who's the smartest?

人们各有各的聪明之处。很难比较。我见过黄仁勋、埃隆，所有这些人和像贝佐斯这样的人。

People are smart in their own ways. It's hard to compare. I've met Jensen, Elon, all these guys and like Bezos.

见到埃隆是什么感觉？

What was it like meeting Elon?

太棒了。我的意思是，埃隆是一个非常专注的人。他在 Twitter 上可能不是那样，发很多随机的推文，但他对当下所做的任何事情都极其专注。实际上，作为一个创业者，我非常想从他这样的人身上学到并拥有的一个技能，就是那种能力：把业务或其他业务中发生的所有其他事情都屏蔽掉，只专注于当下的限制性问题，即瓶颈问题，忽略其他一切。这非常难做到。即使在 Perplexity 内部，我也不能只专注于业务的某一部分。这非常困难。我总是同时关注其他事情。而他的风格就是始终只关注限制性问题，忽略其他一切。这非常难做到，因为你必须非常擅长集中注意力。你必须非常擅长忽略即使是重要的事情，只要它们对当前的核心目标来说是干扰。

Amazing. I mean Elon is a very focused person. He might not appear that way on Twitter, with a lot of random tweets, but he's extremely laser-sharp focused on whatever he's doing at that moment. Actually, the one skill as an entrepreneur that I would really like to take from somebody like him and have for myself is that ability to just zone out of all the other things happening in your business or other businesses and just focus on that limiting problem right now, the bottleneck problem, and ignore everything else. It's very hard to do. Even within Perplexity, I cannot just focus on one part of the business alone. It's very difficult. I'm always looking at other things simultaneously. And his style is to just always look at the limiting problem and just ignore everything else. And that's very hard to do because you actually have to be really good at concentration. You have to be really good at ignoring even important things which are distractions to your core objective right now.

黄仁勋是谁？或者你想象中他是什么样？

Who's Jensen Huang? Or who you thought he'd be?

好得多。

Far better.

真的吗？

Really?

是的。黄仁勋非常追求真理，这简直不可思议。我想是他或别人告诉我的，或者我在一本书里读到，他非常紧张，每天醒来都告诉自己他很差劲，然后去工作，他紧张到告诉周围的每个人，他们距离破产只有 30 天。想想看。对吧？价值 5 万亿美元的公司，未来两年保证有 5 亿美元的收入。拥有世界上最先进的芯片，他却以可能 30 天后就破产的心态运作。这就是成为黄仁勋所需要的。从这些人身上可以学到很多东西。有很多东西要学。我认为有一个方面是安于现状，认为自己成功了。到目前为止，走到这里感觉很好。但这些人不会停止。我不认为埃隆想停止。如果你看看他在 SpaceX 的薪酬方案，它是围绕在火星上建立一个有百万居民的殖民地并在太空中建立足够的算力而构建的。所以，并不是为了拥有 10 万亿美元的净资产之类的。如果他做了这些事情，我相信他会达到那个目标。但更多的是围绕让不可能的事情发生，并拥有那种长远眼光。我认为这是从这两个人身上学到的最重要的东西：很多人把创业看作，“哦，如果我赢了，取得了好结果，卖掉了公司，我就有了世代财富，再也不用工作了。”然后呢？你最终待在家里，你的孩子显然会有信托基金，他们不会看着他们的爸爸……

Yeah. Jensen is so truth-seeking, it's insane. I think he or somebody else told me or I read in a book that he is so intense that he wakes every day and tells himself that he sucks and goes and like he's so intense that he tells everybody around him that they're 30 days away from going out of business. Think about it. Right? $5 trillion guaranteed to make $500 million in revenue in the next 2 years. And has the most advanced chips in the world and he operates with that mentality that he could be 30 days away from going out of business. That is what it takes to be Jensen Huang. And there's so much to learn from these guys. There's so much to learn. I think there's one aspect of being comfortable where you are, thinking you made it. That feels good to get here so far. But these guys are not stopping. I don't think Elon wants to stop. If you look at his pay package with SpaceX, it's structured around creating a colony on Mars with a million inhabitants and building enough compute in space. So, it's not like motivating to be worth $10 trillion in net worth or something. If he does these things, I'm sure he's going to get there. But it's more motivated around making the impossible things happen and having that long-term outlook. I think that has been the biggest thing to learn from maybe these two individuals in particular: a lot of people view entrepreneurship as, "Oh, if I win and I have a great outcome and I sell my company, I would have generational money and never have to work again." And then what? You end up staying at home and your kids will obviously have trust funds and they're not going to get inspired watching their dad...

玩 Papole。

Play Papole.

是的，你知道。你不会为他们树立正确的榜样。他们无法继承你的财富并使其增值，因为他们没有看到真正这样做的人。你在他们成年之前就做到了。所以我认为你总是需要做点什么。就像黄仁勋最近说的，他希望在工作中死去之类的。这就是你需要有的态度。你需要永远工作。

Yeah, you know. You're not going to set the right example for them. They're not going to be able to take your wealth and multiply it because they didn't watch somebody who actually did that. You did it before they were adults. And so I think you always need to be doing something. Like Jensen said recently that he hopes to die on the job or something like that. That's the attitude you need to have. You need to work forever.

当黄仁勋说“如果我知道会这么难，我就不会做了”的时候，我非常沮丧。他当时——我不知道你是否看过那个采访。我当时想，“哦。”我的天。

I was so upset when Jensen said, "If I'd known how hard it was going to be, I wouldn't have done it." When he did—I don't know if you saw that interview. I was like, "Oh." My hair.

是的。我认为这非常难，但你做不是因为容易，而是尽管难你还是要做。我认为这就是它的运作方式。

Yeah. I think it's pretty hard, but you don't do it because you do it despite that. I think that's how it works.

Arvin，今天真是太棒了。非常感谢你在伦敦期间抽出时间。非常感谢你加入我。

Arvin, this has been so fantastic today. I so appreciate you taking the time while you're in London. So thank you so much for joining me.

谢谢。

Appreciate it.