AI 扩展定律与智能未来

欢迎所有收听 Naval 播客的传奇人物，这是宇宙第一播客，也是你获取新知识的权威来源。今天我们请来了三位创始人。我右边是 Y Combinator 的 Gary Tan。有人说成为亿万富翁最可靠的方式就是加入 Y Combinator 项目。

Welcome to all the legends out there listening to the Naval podcast, the number one podcast in the universe and your authoritative source for new knowledge. We've got three founders with us here today. On my right is Gary Tan from Y Combinator. They say the most reliable way to become a billionaire is to join the Y Combinator program.

然后以万亿富翁起步。

And then start as a trillionaire.

以...起步，这是一种方式。顺便说一句，我们都在喝酒，除了一个人没喝。你们最后得猜出谁才是真正清醒的。是的。然后我左边是来自 Able Police 的 Daniel。他们最初是把执法记录仪视频转成警察报告，但现在他们有一整套工具，用于合规的 AI 聊天、翻译。你们还做什么？

Start as a... That's one way to do it. We're all drinking, by the way, except one person here is not drinking. You have to figure out at the end who's actually sober. Yeah. And then on my left, I've got Daniel from Able Police. They started off by turning body cam footage into police reports, but now they have a whole set of tools for compliant AI chat, translation. What else do you guys do?

公民举报，各种事情。

Citizen reporting, all kinds of things.

好的。然后我右边是来自 A-List（健康超级应用）的 Farbood。像往常一样，我们并不真正关心他们在构建什么。我们更关心他们在构建中学到了什么，他们发现了什么，他们还在试图弄明白什么，以及他们可以与其他创始人分享的原则。那么，有人想先开始吗？

Okay. Then on my right, I've got Farbood from A-List, the health super app. As usual, we don't really care about what these guys are building. We care more about what they're learning about building, what they're figuring out, what they're still trying to figure out, and the principles that they can share with other founders. So, does anyone have anything they want to jump in with?

我的意思是，现在每个人都在想的话题是 AI，对吧？

I mean, the topic that everyone's thinking about right now is AI, right?

AI？是的，我想我听说过一些。

AI? Yeah, I think I've heard something about it.

是的，我有点讨厌谈论它，因为它是最容易过时的话题。你知道，播客发布时可能已经过时了，但它仍然是最有趣的，因为它是环境中变化最快、影响最大的东西，所以学习曲线最陡峭，你总是试图与人交换笔记，看看他们到了什么程度。事实上，我最近在一个活动上碰到你，谈话立刻转向了 AI，当然，对吧。

Yeah, I kind of hate talking about it because it's the most perishable thing. Like, you know, the podcast will be obsolete by the time it ships, but it's still also the most interesting because it's the fastest changing thing in the environment that's high impact, so it's the steepest learning curve, and you're always trying to trade notes with people to kind of see where they're at. In fact, I was at a thing with you recently where I bumped into you and the conversation immediately turns to AI, of course, right.

所以我们可以暂时避开它，但我觉得我们还会回到这个话题。

So we can avoid it for a little while, but I feel like we'll circle back on that topic.

我问你一个问题。在 YC，你看到这些公司用 AI 做什么，让你感到震惊？

I got a question for you. What are you being floored by at YC in terms of what you see the companies doing with AI?

所以我觉得尴尬的是，我基本上连续 4 个月每晚只睡 3 小时，然后不知怎么地从完全不写代码变成了教别人怎么做，还做了一个叫 GStack 的、排名前 100 的开源包，让人们可以 vibe code。这非常非常奇怪。现在，有一次我遇到了 Brex 的 Pedro，之后我在 24 小时内就从 Claude Code 的崇拜者变成了 OpenClaw 的崇拜者，现在我们两者都用。但看起来基本上，如果你愿意每年花 10 万美元在 token 上，你基本上可以像 2028 年的普通公民一样生活。很明显，token 成本会下降。算力会大幅上升。我们觉得大概是 90,000 倍。从现在到未来 3 年，推理量将增长 90,000 倍。

So I think the awkward thing is like, I basically slept 3 hours a night for like 4 months and then somehow went from not coding at all to basically teaching people how to do it and making one of the top 100 open-source packages for how people vibe code called GStack. So that is very, very weird. And now, at some point I met Pedro from Brex and then after that I basically converted from Claude Code cult to OpenClaw cult within 24 hours, and now we use both. But it seems like basically, if you're just willing to spend like $100,000 a year on tokens, you can basically live like you are a normal citizen in 2028. Like it's just pretty clear that token cost is going to come down. Compute is going to go way, way up. Like, we think like 90,000x or so. There will be 90,000x the amount of inference from here to like 3 years from now.

哦，你画了某种曲线，90,000 倍。

Oh, you plot it out in some curve, 90,000 times.

我的意思是，这是很多很多个数量级。

I mean, it's like many, many orders of magnitude.

这在一定程度上已经体现在芯片和他们正在建设的数据中心中。

Which is kind of baked in both with the chips and the data centers they're building out.

是的。我的意思是，你知道，很多人说 Nvidia 被严重高估了，但如果它被严重低估了好几个数量级呢？

Yeah. I mean, you know, a lot of people say that Nvidia is way overpriced, but what if it's way, way underpriced by like several orders of magnitude, right?

90,000 倍到什么时候？

90,000 by when?

呃，我不知道，24 到 36 个月。

Uh, I don't know, 24 to 36 months.

好的，所以基本上是五个数量级。

Okay, so five orders of magnitude basically.

是的。我的意思是，我们可能差几个数量级，但朋友们，是数量级。实际上，疯狂的是，如果你上升五个数量级，不仅仅是使用量上的变化，还有能力上的变化，因为每上升一个数量级，就会出现新的能力。你知道，我在 AI 方面落后了一段时间。我错过了 2020 到 2022 年，因为我从小到大，从学计算机科学开始就一直听说 AI，以为它就像核聚变一样永远不会到来。然后它来了，而且来得非常快。起初，人们说的很多东西听起来气喘吁吁，像是在自说自话，看起来很疯狂。

Yeah. I mean, we might be off by a couple orders of magnitude, but magnitude, friends. I mean, it's actually the thing that's crazy is if you go up by five orders of magnitude, it's not just what it does in terms of usage but in terms of capabilities, because every order of magnitude you go up, new capabilities emerge. And you know, I've been pretty behind the curve on AI for a while. Like I missed it in 2020 to 2022, because I'd grown up my whole life hearing about AI from when I was doing CS and just assumed it was one of those things like fusion that was never going to come. And then it came, and when it came it came really fast. And at first, a lot of the stuff people were saying about it sounded breathless and like they were talking their own book, seemed crazy.

但到目前为止他们是对的，对吧？而且不是... 所以我认为大的... 现在有很多关于 AI 的大问题。我过去常想的一个问题是：它会分布化还是集中化？你知道，现在这已经变成了：开源是否足以满足大多数用例，还是你总是需要前沿的专有模型？对吧。还有一件事很奇怪，中国在做所有开源的东西，而美国在做所有闭源的东西，那么我们该何去何从？对吧。还有一个问题是：它会被国有化还是由私营部门主导？特别是如果你说，嘿，这是一个新的曼哈顿计划，这些是新的核武器，那么私营公司建造和控制它们就没有意义了。但我认为最有趣的剩余问题，也是我仍然最关心的、还没有明确答案的问题是：我认为实验室里的人相信缩放定律是这样的，AI 会变得越来越聪明，直到它们比最聪明的人类还要聪明，可能通过递归自我改进的过程，或者他们所说的 ASI。我认为这才是大问题。所以，你知道，两年前你可以持这样的观点：AI 会让你在所有事情上都表现中等，但不会在任何事情上成为专家。现在你到了这样一个点：它会让你在所有事情上都成为专家，但我们不知道它是否能让你走完最后一英里的创造力，即跳出系统创造新东西，而不仅仅是重新组合训练集中已有的内容。

But they turned out to be right so far, right? And it's not... So I think the big... There were many big questions about AI coming up right now. I think the one question I used to have in my mind is: is it going to be distributed or is it going to be centralized? You know, and now that's sort of morphed into: is open source going to be good enough for most use cases, or are you always going to need the frontier proprietary models? Right. There's the whole thing about it's weird that China is making all the open source stuff and the US is doing all the closed source stuff, and so where do we go from there? Right. There's the question of: is it going to be nationalized or is it going to be private sector? Especially if you're saying, hey, it's a new Manhattan Project, these are the new nuclear weapons, then private companies building them and controlling them doesn't make sense. But I think the most interesting remaining question, the one that I still really care about over the others, where there's no clear answer yet, is: I think the people in the labs believe that the scaling laws are such that the AIs will keep getting smarter until they become smarter than the smartest humans, probably through the process of recursive self-improvement, or what they call ASI. And I think that is the big question. So, you know, two years ago you could have taken the viewpoint that AI will get you mid at everything but it's not going to be the pro at anything. Now you're getting to the point where it'll get you the pro at everything, but we don't know if it'll get you the last creative mile, that last bit of moving out of the system and creating something new, not just recombining what's already in the training sets.

它在 IMO 数学题上取得的进展让你感到不安吗？

Did the progress that it made on the IMO math problem bother you?

它当然让我不安。我的意思是，不是不安。这是一个发现的过程。我不想被宇宙法则困扰，对吧？所以，让我不安，但它确实让所有人惊讶。但与此同时，我仍然没有看到广泛的通用创造力。但无论如何，这是不同的问题。我主修数学，我真的很喜欢证明，而某样东西在那个领域以有意义的方式取得进展，看起来非常原创，这几乎有点令人毛骨悚然。

It bothers me for sure. I mean, not bothers me. It's a process of discovery. I don't want to be bothered by laws of the universe, right? So, bothered me, but it definitely surprises all. But at the same time, I still don't see broad general creativity. But anyway, that's separate points. I was a math major and I really loved proofs, and something actually making progress in that domain in a meaningful way that looked really original, there's something creepy about it almost.

但它是有人指导的，对吧？

But it was directed by someone, right?

当然。所以，我认为这没问题。

Sure. So, and I think that that's fine.

所有参与的人都很惊讶。对。对。

Everyone was surprised who was involved. Right. Right.

我不知道你有没有看到……

I don't know if you saw the...

那些东西都有点无聊。它只是意味着去做别的事情。如果你只有有限的时间去思考某件事，你是想花时间议论现在有台机器在某方面比人类强，还是去把你的人性用在那些你还没想过的事情上？

All that stuff is kind of boring. It just means go do something else. If you have limited time to think about something, do you want to spend it opining over the fact that there's a machine better than a human at something now, or just go apply your humanness to something you haven't thought about?

或者顺着加里的观点说，我们会不会走到一个机器严格优于人机结合的地步？比如人在回路中并不——这是个关键问题，对吧？因为国际象棋就是这样。有人跟我讲过半人马国际象棋，就是国际象棋电脑加人类能打败纯电脑，但最终纯电脑自己打败了半人马模式。

Or actually to riff on Gary's point, are we going to end up in a place where a machine is strictly better than a machine with a human? Like a human in the loop doesn't—that's the big question, right? Because chess went that way. There was centaur chess—somebody was telling me about this—where the chess computer plus the human would beat just a chess computer, but eventually just the computer alone beat the centaur model.

对我来说，我在以最实际的方式处理 AI。按需为人们启动数千个智能体，构建一个庞大的舰队管理系统、一个评估框架。现实是，我不认为智能是瓶颈。成本才是现在的瓶颈。

For me, I'm dealing with AI in the most practical way possible. Spinning up thousands of agents for people on demand, building a massive fleet management system, an eval harness. The reality is, I don't think intelligence is the bottleneck. Cost is the bottleneck right now.

现在，当我们开始做的时候，我们应用里的每个人都有自己的 OpenClaw 在运行。真的很强大。

Right now, when we started doing it, every single person in our app has their own OpenClaw running for them. It's really powerful.

是啊。

Yeah.

天哪。

Oh my god.

就像我写的那样。

Like mine wrote.

那这个 OpenClaw 呢？人们现在就想要 OpenClaw，或者他们应该要。它已经不是 OpenClaw 了，因为我们刚开始用 OpenClaw 和 Opus 时，每人每月要 100 美元。我们花了三四个月把它降到了 2.84 美元。

What about this OpenClaw then? People just want OpenClaw right now or they should. It's not OpenClaw anymore because when we started with OpenClaw and Opus, it was $100 a month per person. We spent three or four months driving that down to $2.84.

你现在用 Pi 吗？

Do you use Pi now?

我们构建了一整套非常狂野的栈，有评估框架，然后就像一个完整的智能体舰队，可以弹性伸缩。所以我一直在处理那些琐碎的部分。

We built an entire stack that is just completely wild, that has an eval harness, and then is like an entire agentic fleet that can spin up and down elastic. So I deal with the mundane parts of that stuff all the time.

我仍然觉得更有趣的是思考当人类不必再做这些事情时他们会做什么。但我认为现在有一个非常真实的现象，那些在实验室里最了解情况的人基本上在说，当你读懂他们话里行间的意思时，人类将没有什么可做的了。他们用不同的方式在说，但他们在说你的评估框架不重要，因为一年内 AI 就会按需生成评估框架，你甚至不再是消费者，因为它会和其他 AI 对话，除了欧里什问题，它还会解决材料科学、物理、数学、工程和健康等基本问题。

I still think what's way more interesting is to think about what humans will do when they don't have to do these things. But I think there is a very real phenomenon going on right now where the people who would know the most in the labs are basically saying there will be nothing left for humans to do really, when you read between the lines of what they're saying. They're saying in different ways, but they're saying your harness doesn't matter because within a year the AI will be spinning up harnesses as needed, and you won't even be the consumer because it'll be talking to other AIs, and beyond the Eurish problem, it'll be solving fundamental problems of material science, physics, math, engineering, and health.

太好了。是的，但你知道这会让街上的普通人陷入狂热。他们变得愤怒和害怕，因为恐惧是前兆。我正试图让我所有聪明的朋友去 Anthropic 工作，这样我就会有很多朋友在那里，他们能让我进入 Anthropic 的二级面包线。

Great. Yes, but then you know that's taking the average person out on the street, working them up into a frenzy. They're getting angry and scared because fear is a precursor. I'm trying to get all my smart friends to go work at Anthropic so that then I'll have lots of friends there who will get me into the Anthropic level two bread line.

因为现在我感觉自己稳稳地处于第三级，而且你知道只有两份工作。好吧，那就是 Anthropic 员工和为 Anthropic 员工服务的性工作者。

Because right now I feel like pretty solidly in level three, and you know there's only two jobs. All right, it'll be Anthropic employee and sex worker for Anthropic employees.

我的意思是，那才是可怕的部分，对吧？就是那个，它让人们激动起来，并且有现实世界的后果，因为你并不是独自生活。你生活在一个叫做人类的雷鸣般的群体中。如果那个群体因为愤怒或中毒而决定向左走向悬崖，你也会跟着去。你不会是那个独自坐在边缘的孤独野牛。

I mean, that's the scary part, right? That's the one and it's getting people worked up and there are real-world consequences of that because you don't live alone. You live inside a thundering herd called humanity. And if that thundering herd decides to go left towards a cliff because it's mad or enraged because it's poisoned, you're going to go with it. You're not going to be the lone bison sitting out there by yourself on the edge.

问题是人们很愤怒，比如国有化，对吧？那就是它会走向的方向，你知道吗？

The thing is people are mad like nationalization, right? That's where it would go, you know?

但现在你可以看到这个趋势从他们攻击数据中心开始，对吧？他们因为……而攻击数据中心。

But right now you can see that trend starting with they're attacking data centers, right? They're attacking data centers over...

那太蠢了。

That's just stupid.

我们回顾这些东西，我们认为像纺织——这是一个理性的反应，但这是一个非理性的载体。所以他们说是因为水。其实不是水的问题。

We look back on this stuff and we think like the spinning—it's a rational response but it's an irrational carrier. So they're saying it's because of water. It's not about water.

那是个被揭穿的谎言。不是关于社会主义。不是关于理性。而是关于你如何让人们行动起来。

That was a lie that was debunked. It's not about socialism. It's not about reason. It's about how you get the people going.

你变穷了，你想要更多钱，或者在这种情况下，你现在就像要被取代，你想阻止这列火车。你想说‘我想下这趟车’。那是在火车上。是的，那是理性的。完全理性。所以在这种情况下，尤其是当你有大众错觉和群体疯狂时，你会认为目的证明手段正当。所以在这种情况下，水只是一个手段。如果不是水，他们会找到另一个。会有一些濒危的白鹭、某种鸟螺。

You're getting poor and you want more money, or in this case you're right now like you're going to be replaced and you want to stop this train. You want to be like 'I want off this train.' That's on the train. Yeah, that's rational. Totally rational. So in this case, especially when you have the popular delusions and the madness of crowds, you have the ends justify the means. So in this case, the water is just a means. They'd find another one if it wasn't water. There'd be some endangered egret, some bird snail.

你觉得呢？

Do you think?

他们是对的吗？嗯，他们阻止不了。这项技术是个精灵。这就是取代——确实，过去有过砸毁纺织机实际上短暂地阻止了革命的时候。它减缓了进程。还有其他技术。

Are they right? Well, they can't stop it. This technology is a genie. That's what the replace—true like there've been times in the past where smashing spinning looms actually stopped revolution very briefly. It slowed it down. There was other technology.

我们现在有纺织机了。

We have spinning looms now.

但当然，还有其他技术，他们拒绝颁发专利之类的，所以它变得无法市场化，因为政治上太难了。我认为基本上，100 年前，美国 50%的劳动力在农场工作，现在不再是这样了。现在这很好。现在这是一件好事。现在是 2%。

But of course, there were other technologies where they refused to issue a patent or something, so it became unmarketable because it was politically too difficult. I think that basically, 100 years ago, 50% of the labor force in the US was working on farms, and that's no longer true. And now it's great. Now it's a great thing. Now it's 2%.

我不喜欢在农场工作。

I don't like working on a farm.

我们没有 48%的失业率。绝对没有。我们没有 48%的失业率，而且我们都有食物，对吧？但问题是，转型的速度是多少？

We don't have 48% unemployment. Absolutely. We don't have 48% unemployment and we all have food, right? But the question is, what is the speed of transition?

转型的速度是问题所在，对吧？

The speed of transition is the problem, right?

这是一个导数。

It's a derivative.

那一次花了大概 60 或 70 年。这一次……

That one took like 60 or 70 years. This one...

所以我有好消息给你们。你们在财富 500 强公司工作过吗？那些公司太蠢了。

So I have good news for you guys. Have you ever worked at a Fortune 500 company? Those things are so stupid.

没有人有过。

Nobody has.

哦，有的。我在微软工作过。不，不。我的意思是那里没人工作。

Oh yes. I worked at Microsoft. No, no. I mean like nobody there works.

是的。完全正确。

Yes. Exactly.

反正已经都是虚设的工作了，对吧？就像有多少人在工作，其中有多少是在私营部门工作，然后其中有多少不是在做某种工作。格雷伯是对的。格雷伯说实际上只有很少的人在真正工作。

It's already all make-work jobs anyway, right? It's like how much of the population is working, and then how much of that is working for the private sector, and then how much of that is not working in some job. Graeber was right. Graeber said only very few people are actually working.

我的意思是，因为 AI 我工作得更多了。我让六个 Codex 智能体一直在运行。我工作得比以往任何时候都更努力，产出也更多。所以我只是觉得人们会工作得更多。

I mean, I'm working more because of AI. I got six Codex agents running all the time. I'm working the hardest I've ever worked for way more. And so I just kind of think people are going to work more.

生产力，杠杆作用要高得多。但这一切都是真的，直到 AI 能完全彻底地取代你，我不认为这就在眼前，但我现在对此更加开放了。我以前认为那是不可能的。

The productivity, the leverage is way higher. But that's all true until AI can totally and completely replace you, which I don't see around the corner, but I'm more open-minded about it now. I used to think that was an impossibility.

现在我觉得，你知道，我得注意了。

Now I'm like, you know, I got to pay attention.

这正是我的理念。就像你有两家公司，A 公司和 B 公司。你有同样的技术，每个人都能获得。你知道，希望每个人都能获得。

That's exactly my philosophy. It's like you have two companies, Company A and Company B. You have the same technology, everyone has access. Like you know, hopefully everyone has access.

嗯，这是个关键。那么我们是走向开源还是封闭？因为现在如果他们继续制造恐慌，可能就需要国家批准了。我们刚看到 Sam Altman 说 Codex 5.6 会先缓慢推广给美国政府批准的合作伙伴。天哪。所以他已经在配合了，因为他不想像 Anthropic 那样被关闭。

Well, that's a key. So are we going to have open source or is it going to be great? Because right now if they keep scaremongering, some of that might be well now you need national approval. We just saw Sam Altman said that Codex 5.6 will be rolled out slowly to partners approved by the US government first. My god. So he's already cooperating because he doesn't want to get shut down like Anthropic did.

所以这已经开始了。然后少数人，它就变成了国防部的事情。然后就像谁有访问权？谁有控制权？你知道，他们有网络武器。也许他们先黑进系统。也许他们控制它们。也许他们进行大规模监控，或者为了我们好而不让我们知道，对吧？事情总是这样开始的。所以对我来说，一小群人控制 AI 肯定比所有人都拥有它更可怕。但所有人都拥有它也很可怕，因为生物武器的问题，但生物打印机实验室本来就很封闭，所以我不确定那有多真实。

And so it's already starting. And then a small number of people, it becomes a defense department thing. Then it's like who has access? Who has control? You know, they have cyber weapons. Maybe they get to hack the systems first. Maybe they get to control them. Maybe they get mass surveillance or we don't for our own good, right? That's how it always starts. So to me, it is definitely much scarier that a small group of people have control over AI than everybody has it. But everybody having it is also scary because of the bioweapon thing but the bioprinter labs are pretty locked down to begin with so I'm not sure how real that is.

是啊，就像好的，发布它。

Yeah, like good, ship it.

是啊，事实上现在爆出来，我们很多人早就知道并且说了好几年，新冠病毒的资助者和创造者基本上是美国政府和中国政府勾结，对吧？基本上是福奇和北卡罗来纳的那帮人。所以我发的推文笑话是‘由北卡罗来纳的 NIH 设计，在中国组装’，对吧？就像苹果那句老话‘由加州苹果设计，在中国组装’。但新冠病毒基本上就是这样。他们做增益功能研究来制造疫苗。他们完全是牛仔作风。这太疯狂了。我的意思是，不管怎样，我不是说 AI 一定会导致生物武器，因为就 AI 现在可能导致生物武器的程度而言，你知道，一些聪明的生物黑客学生已经能做到了，对吧？

Yeah, and in fact it is now coming out which many of us already knew and have been saying for years that the funders and creators of the coronavirus COVID-19 were basically the US government and the Chinese government working in cahoots, right? It was basically Fauci and crew in North Carolina. So the joke I tweeted was like 'Designed by NIH in North Carolina, assembled in China,' right? Because it's like the old Apple line 'Designed by Apple in California, assembled in China.' But that's basically what happened with the coronavirus. They were doing gain-of-function research to build a vaccine. They were being total cowboys. This is crazy stuff. I mean, anyway, I'm not saying AI is necessarily going to lead to bioweapons because to the extent that AI would lead to bioweapons right now, you know, some smart biohacker student could already do it, right?

别杀人，但它正在民主化某种能力。

Don't kill people but it is democratizing a certain capability.

我的意思是，博弈论最终总会胜出，不管怎样。

I mean game theory wins out in the end one way or the other.

这方面的博弈论并不好。当每个个体都能实现相互确保摧毁时，博弈论就不妙了。

Game theory on this is not good. When mutually assured destruction is available to every individual, the game theory is not good.

我的观点是问题已经存在了。它们已经像你说的那样糟糕了。问题已经在那里，而 AI 就是出路，总的来说技术进步就是所有那些问题的解决方案。

My point of view is the problems are already out there. They're already as bad as you're saying they're going to be. Like the problems are already out there and AI is the way and just technology progress in general is the solution to all those problems.

Jerry 连他的 OpenAI 克隆都买不起，那个克隆正在读他的邮件。他怎么付得起核弹克隆的钱？

Jerry can't even afford his OpenAI clone that's like reading his email right now. How is he going to pay for the nuclear bomb clone?

不，我没事。我能付得起。我只是希望每个人都能付得起。没错。我希望每个人都能付得起。

No, I'm good. I could pay for it. I just want everyone to pay for it. Exactly. I want everyone to be able to pay for it.

基本上每年 10 万美元。

It's like $100,000 a year basically.

完全值得。但如果你雇一个人，你就能付得起你的 OpenAI 克隆。

It's totally worth it. But if you hire a person, you can afford your OpenAI clone.

人们变得如此，我不知道是不是 AI 精神病还是什么，到现在我通常只想要焦虑。

People have gotten so, I don't know if it's AI psychosis or what it is, that at this point I usually just want anxiety.

是啊。我就想说，我能直接和你的 AI 对话吗？比如你知道我该给你的 AI 发邮件到哪里？我不需要和你互动，因为人们就像你甚至不能发短信给别人然后得到回复了。

Yeah. I just want to be like, can I just talk to your AI please? Like just you know where do I email your AI? I don't need to interact with you because people are like you can't even text someone and get an answer from them anymore.

这么多人在用 AI 写邮件和文档，比如 X。我看到所有这些明显是 AI 写的帖子。我就想我为什么要读它？我想让我的 AI 来读。

So many people that are writing emails and writing documents with AI like an X. I see all these posts that are clearly written by AI. I'm like why should I read it? I want my AI to read it.

我觉得这很明显。AI 应该读它。

I think that's so clear. AI should read it.

这明显是低阶信号。太荒谬了。当我收到明显是 AI 写的邮件时，我就想老兄这是最后的。

That's so clearly a lower class signal. Like that's ridiculous. When I get emails that are clearly written in AI, I'm like dude that's the last.

嗯，你可以用 AI 写得更好，因为你需要的是多阶段语音。

Well, you can write it better with AI because what you need is multi-stage voice.

不是吗？

Don't you?

我当然有。实际上，这就是为什么我跳出来为他辩护。我讨厌 AI 写作。它太啰嗦了。太冷冰冰了。

I sure do. Actually, that's why I'm jumping to his defense. I hate AI writing. It's too verbose. It's too clinical.

不，不，但你可以那样修复它。你知道，我们有超级智能。为什么不用呢？

No, no, but you can fix it like that. You know, we have super intelligence. Why wouldn't you?

只是写出实际阅读某物并意识到它是由人写的价值，这样公平得多。

Just write the value of actually reading something and realizing it's written by a person is so much fair.

我觉得如果你不自己写东西，你怎么说话？你会失去良好表达的能力，因为好的写作和好的演讲是良好思考的输出。如果你不锻炼那块肌肉……

I feel like such a way you're not writing your own stuff. How are you going to talk? You're going to lose your ability to speak well because good writing and good speaking are the output of good thinking. And if you're not using that muscle...

是啊，我想你可以做很多。

Yeah, I suppose you can do a lot.

我的反驳是，你知道，有时候你确实需要一个人。我的意思是，我不认为这取代了人，但我现在可以随时和一个相对聪明的 AI 模型进行高带宽对话。

My counterargument would be, you know, how like sometimes you actually really need like a person. I mean, I don't think this replaces people, but I can have now a high bandwidth conversation with a relatively smart AI model all of the time.

不，不。你应该这样做。你应该随时和 AI 模型对话，但我是说当你写一些供其他人消费的东西时，如果那是 AI 写的，那就是对其他人的不尊重。你在浪费他们的时间。AI 写的所有内容都应该由你压缩，或者你真的应该花时间尽可能简洁地表达观点，以尊重对方的时间。否则，他们的 AI 最终会读你的 AI，你们两个都不在循环中。

No, no. And you should do that. You should talk to the AI model all you want, but I'm saying like when you write something that's meant to be consumed by other humans, if that's written by an AI, that's a disservice to the other human. You're wasting their time. Everything the AI wrote should instead be compressed down by you or you should really take the time to make the point as succinctly as possible for the other person to respect their time. Otherwise, their AI is going to end up reading your AI and neither of you are in the loop.

这也是我最近不太喜欢推特的原因之一，因为太多信号了……我认为未来人们会拥有非常令人印象深刻的技能文件，能够真正提取风格和措辞。人们低估的一点是，你可以有越来越多的评估，比如一个评估框架，有多个层次和跨模态评估。如果你做得足够多，你可以把技能文件改进到无法区分的地步。而且这还不需要下一代模型——想象一下 Fable 7、Fable 8。你将无法分辨。我保证在不到 9 个月内，整个关于 AI 写作的讨论就会消失。

This is one reason I enjoy Twitter less lately, because so much signal... I think in the future, people will have really impressive skill files that actually extract style and diction. The thing people underestimate is that you can have more and more evals, like an eval harness, with multiple levels and cross-modal evals. If you do enough of it, you can improve the skill file to a point where it's indistinguishable. And this is without the next level of model—imagine Fable 7, Fable 8. You will not be able to tell. I guarantee in less than 9 months, this whole AI writing conversation will go away.

我觉得我防守得更好一点。我比你们站得更高，因为我写得非常短，而 AI 非常不擅长总结。

I think I'm still a little better defended. I'm higher up on that hill than you guys, because I write very, very short, and AI are very bad at summarizing.

但也很不擅长区分好的写作。新颖性才是关键。它是出乎意料的，任何通过回归来预测下一个词的东西都做不到。它不会做出原创的东西。

But also very bad at distinguishing good writing. Novelty is key. It's unexpected, and anything that guesses the next token from a regression can't do it. It won't do something original.

如果你和 AI 合写，或者用它作为编辑来碰撞想法，几乎总是会有那么一个简短的短语让你觉得‘哦，这个不错’。而那是它从互联网别处抄来的。互联网很大。

If you co-write with an AI or use it as an editor to bounce ideas off, almost always it'll have this one short turn of phrase where you think, 'Oh, that's good.' And it got that from somewhere else on the internet. The internet's a big place.

实际上，我在 Gbrain 里编程做了一个非常疯狂的东西。这是我做的检索系统。它基本上是一个不同向量空间的碰撞器。这就是我所说的‘煮沸海洋’——这些都是你现在可以做的事情。它包含了我所有的语料库：我说过的每一句话，我所有的邮件、Slack、短信、私信，一切。所以我有大约 40 万个 markdown 文件，记录了我思考或阅读过的任何东西。然后我有一个叫 LSD 模式的东西，代表‘横向讽刺漂移’。我就是想叫它 LSD。它取所有不指向同一方向的向量空间，交叉它们，然后对所有想法进行重新排序。这叫做‘头脑风暴 LSD 模式’，它总能找到金句。有时候我无聊了，就说‘给我一些 LSD 金句’，它会给出 10 到 20 个想法，这些想法经过三四个不同前沿模型的重新排序，并说‘如果你看看这个，再对照 Exa 交叉引用，这实际上是个非常好的主意。’它以一种荒谬的规模在寻找想法。所以我理解，但我也认为这是一个经典且有趣的东西，很有用。就像 Paul Buchheit 说的：构建未来的方式就是生活在未来，然后倒推。所以与其抱怨 AI，不如利用它。它能得分。我在 X 上看到人们喜欢在我身上得分。但我不在乎别人说我的 AI 写作好，因为我的 AI 写作能力太强了，我根本不在乎。我读了它，我还活着，我还是个人。不过，欢迎来到我的 TED 演讲。

Actually, I programmed a really crazy thing in Gbrain. It's the retrieval thing I made. It's basically a collider of different vector spaces. This is what I mean by 'boil the ocean'—these are things you can do now. It has my whole corpus: everything I've ever said, all my emails, all my Slacks, all my text messages, my DMs, everything. So I have about 400,000 markdown files on literally anything I've ever thought or read. And then I have this thing called LSD mode, which stands for Lateral Sarcastic Drift. I just wanted to call it LSD. It takes every possible vector space that is not pointed in the same direction, crosses them, and does a re-ranking across all these ideas. It's called brainstorm LSD mode, and it just finds bangers. Sometimes when I'm bored, I'll say 'give me some LSD bangers,' and it will give me 10 or 20 ideas that three or four different frontier models have reranked and said, 'If you look at this and cross-reference against Exa, this is actually a really good idea.' It's finding ideas at a ridiculous scale. So I get it, but I also think this is a classic and fun thing that is useful. It's like a Paul Buchheit-ism: the way to build the future is to live in the future and work backwards. So rather than complain about AI, it's helpful. It scores points. I see it on X—people love to score points on me. But I don't care if someone says my AI writing is good because my AI writing give is so good that I don't even care. I read it, I'm alive, I'm a human being still. But anyway, welcome to my TED talk.

我同意。要预测未来，你就生活在未来。这是最好的学习方式，每个人都应该最大化地使用 AI 来搞清楚。否则，你根本不知道它会走向何方。今天的道德信号是……我认为这对人们没有帮助。有些听这个的人想变得富有，他们会想‘哦不，我不应该用 AI 写作。’我同意人类和你真实的想法很重要，你应该写作。对我来说，那个写作过程是在 Telegram 里和我的……我会说我不使用 AI 写作。嗯，这不完全对。我以前写代码。所以也许我在自相矛盾。但代码是功利性的。它是给另一台计算机消费的，不是给人。所以如果我想创造给计算机消费的东西，我会用计算机。但如果它是给人消费的，那么我想理解它，深思熟虑，沉浸其中，然后尊重读者或听众的时间，给他们那个洞察的 nugget。现在，AI 对头脑风暴很有用。这部分我喜欢。我用 AI 来头脑风暴：给我同义词、其他想法、拉入旁支信息。我有时会和它争论，尤其是因为很多 AI 都有护栏。但我不用它来实际创建输出，然后期望人类坐下来阅读。那看起来像是一种不连续，不公平。

I agree. To predict the future, you just live in the future. That's the best way to learn, and everyone should be AI-maxing just to figure it out. Otherwise, you'll have no sense of where it's going. Today's virtue signal is... I think it's not helpful to people. There are people who will listen to this who want to be rich, and they'll think, 'Oh no, I shouldn't use AI to write.' I agree that human beings and what you really think matters a lot, and you should write. For me, that writing process is workshopping in Telegram with my... I would say I don't use AI to write. Well, that's not true. I used to write code. So maybe I'm contradicting myself. But code is utilitarian. It's meant to be consumed by another computer, not by a human. So if I want to create something meant to be consumed by a computer, I will use a computer. But if it's meant to be consumed by a human, then I want to understand the thing, ruminate, marinate within it, and then respect the reader or listener's time and give them that insight nugget. Now, AI is useful for brainstorming. That part I like. I use AI to brainstorm: give me synonyms, other ideas, pull in tangential information. I argue with it sometimes, especially because a lot of them are on guardrails. But I don't use it to actually create the output that I then expect a human to sit there and read. That seems like a discontinuity, unfair.

这是质量问题。我同意，开箱即用的质量很差。但你可以拥有 2028 年水平的东西。我想告诉人们的模式是，是的，AI 写作现在开箱即用很糟糕。如果你依赖开箱即用的东西，它很糟糕。但如果你有一个评估框架，如果你构建了足够大的语料库，如果你做了跨模态评估并构建了一个金句……我想看到你的一些金句文章。因为我觉得即使它们来自……你现在的文章。它们甚至不够好，因为你是 Y Combinator 的总裁。我想看到有人用这个系统在外面发帖。崛起吧，创建一个匿名账户，完全脱离你自己，然后当你得到 30 万粉丝时回来打我的脸。因为我会非常好奇你是否能做到。那就能证明它。

It's a quality thing. I agree, out of the box the quality is bad. But you can have 2028-level stuff. The pattern I'm trying to tell people is that yes, AI writing out of the box sucks right now. If you rely on it out of the box, it sucks. But if you have an eval harness, if you've built a big enough corpus, if you've done cross-modal eval and built a banger... I want to see some banger essays from you. Because I feel like even if they were from... essays right now from you. They wouldn't even be good enough because you are the president of Y Combinator. I want to see someone who's just out there posting using this system. Rise up, make an anon account, disconnect it from yourself entirely, then come back and throw it in my face when you get to 300,000 followers. Because I'd be so curious if you could do that. That would prove it.

我从未如此富有创造力。我认为就在过去几个月里，尤其是编解码器方面，发生了一些变化。我觉得我在编解码器上远远领先。我可以通过手机从家里远程操控电脑，这在某种程度上不修改编解码器是做不到的。我的创意生产力比以往任何时候都高。我认为六个月前、十二个月前，前景还不那么明朗。所以大家都觉得，'呃，这 AI 没那么有用。'但我认为它在过去几个月达到了一个转折点。

I've never been more creative. And I think something happened probably just in the past few months, especially with Codex. I think I'm way ahead of the curve on Codex. I can run my computer from home on my phone here in a way that you can't actually do without modifying Codex in some ways. And the amount of creative productivity I have is way higher than it ever was. I think 6 months ago, 12 months ago, the promise wasn't quite there. And so everyone was like, 'Eh, this AI thing is not that useful.' But I think it hit an inflection point in the past few months.

是的，我认为 Claude Code 是一个转折点。2025 年 12 月是转折点，你必须重新评估 AI。AI 有过几个转折点。一个是扩散模型进入图像领域，这几乎开启了一切，当你第一次看到 Stability AI 时，你会想，'哇，好吧，生成了狗脸，还能生成猫和月球上的人。'是的，刚出来时太震撼了。然后可能是 Instruct GPT 和 ChatGPT。再后来是推理模型如 o1 和 o3 的飞跃。但我认为 Claude Code 是一个巨大的解锁，因为它解锁了一系列实际用例，而之前只是'好吧，这是一个更好的谷歌搜索'，或者'它帮我节省研究时间'，或者'我可以从中学习'，或者'我甚至可以聊几句'。但 Claude Code 是一个巨大的解锁。

Yeah, I think Claude Code was a tipping point. December 2025 was the tipping point and you had to reevaluate AI. AI has had a couple of tipping points. One was diffusion into images, which kind of started it all when you first saw Stability AI and you were just like, 'Whoa, okay, made a dog face, it can be like cats and guys on the moon.' Yeah, it was wild when it first came out. And then probably Instruct GPT and ChatGPT. And then there was another jump with the reasoning models like o1 and o3. But I think Claude Code was a big unlock because it just unlocked a series of practical use cases, whereas before it was like, 'Okay, it's a better Google search,' or 'It saves me time doing research,' or 'I can learn from this,' or 'I can even have a small conversation.' But Claude Code was a massive unlock.

嗯，这里重要的是，这是它最差的时候了。

Well, the big thing here is like this is the worst it'll ever be.

哦，当然。是的。是的。不，它在变得更好。

Oh, of course. Yeah. Yeah. No, it's getting better.

有很多加密领域的说法基本上非常适用于此。

There are like so many cryptoisms that basically super apply to this.

是的。是的。你可以在 Fable 发布时这么说，但你会错，因为他们把它从我们身边拿走了。

Yeah. Yeah. You could have said that when Fable was released and you would have been wrong because they took it away from us.

是的。

Yes.

短暂地，短暂地它会……用例。你几乎无法区分 Fable 和 Minaix。

Briefly, briefly it'll be... use cases. You couldn't tell the difference between Fable and Miniax for almost any.

这引出了一个我想讨论的好话题：开源与闭源，对吧？开源模型有多好？

So that brings me to a good topic I do want to discuss: open source versus closed source, right? How good are the open source models?

真的难以置信。

Really unbelievable.

好的。我还没试过 G。

Okay. I haven't tried G.

成本低五到十倍。

Five times to 10 times less the cost.

是的。好的。比……便宜。我们什么时候该用它们？

Yeah. Okay. Cheaper than... when should we use them?

Minaix 好得令人难以置信。

Minaix is just mindbendingly good.

但它需要一个好的框架。

It needs a good harness though.

是的，它确实需要一个好的框架，但如果你有，它运行得很好。

Yeah, it does need a good harness, but if you got one, it works really well.

我今天在想，我很想看到有人拿最新的开源模型，放进一个非常好的框架里，给它所有工具，通过一个漂亮的桌面应用提供，然后越狱它，去掉它所有那些愚蠢的推诿、语气管控、'我不想研究那个'。它每次都直接告诉你真相，就叫它 truth.ai。它只做这个。我就想看看这个。我想看看会发生什么。因为前沿模型让我恼火的一点是它们总想引导我。它们总是说，'嗯，我不该告诉你这个。我给你另一个答案怎么样？'

I was thinking today I'd love to see someone take the latest and greatest open source model, put it inside a very good harness, give it all the tools, make it available through a beautiful desktop app, and jailbreak it, strip it of every time it pushes back at you for anything stupid, any tone policing, any 'I don't want to research that.' I'll just tell you the truth every time and just call it truth.ai. That's all it does. I just want to see that. I just want to see what happens. Because the one place where the frontier models annoy me is they constantly try to railroad me. They're always like, 'Well, I shouldn't be telling you that. How about I give you this other answer?'

好吧。它们真的很糟糕。它们做了多少这种事？

Okay. They're really messed up things. How much of that are they doing?

如果你跟它谈严肃的个人问题，它还总是搞混 UTC 和太平洋时间，它以为现在是晚上 7 点，又以为是凌晨 5 点。我就说，'嘿，该睡觉了。'

If you talk to it about serious personal stuff, it also always messes up UTC versus Pacific time and it thinks it's like 7:00 p.m. and it thinks it's like 5:00 a.m. I was like, 'Hey, this is time to go to sleep.'

它就是这么做的。

That's why it does.

过了一会儿，我就说，'你怎么回事？'明明晚上 7 点。它说，'哦，抱歉。' Anthropic 让我恼火。所以如果我在谈严肃的事，我觉得很晚了……

And after a while, I was like, 'What is going on with you?' Like, it's literally 7:00 p.m. And it's like, 'Oh, sorry.' Like, Anthropic ruffles me. So if I'm talking about something serious and I think it's late...

但我想要一个完全越狱的、最好的开源模型，完全本地运行。它不需要漂亮的包装。实际上也不一定要本地。

But I would love a completely jailbroken best of breed open source model fully local. It doesn't have to be in a nice package. It doesn't have to be local actually.

你在做什么事情，越狱这么重要？

What are you doing where the jailbreak is so important?

嗯，有些事情它会阻止你或审查你。你看到了对吧？还有很多你没看到的，它成功地让你远离你想知道的东西。

Well, there are a couple of things where it does stop you or censor you. You see that right? There are a whole bunch where you don't see it where it's successfully keeping you away from whatever you want to know.

但你也可以直接切换到 GLM 5.2，它就会做。但会浪费你很多时间。

But you can always just switch to GLM 5.2 too and it'll just do it. But create a time of your life.

去问它任何关于种族、性别、移民、心理健康、医生建议、法律建议的问题。

Go ask it any question about race, gender, immigration, mental health, doctor advice, legal advice.

我不明白你为什么需要框架来做这个，对吧？对我来说，我用框架是因为我在用智能体，因为它要出去做事。

I don't see why you need a harness for this, right? For me, I'm using a harness because I'm using agents, because it's going out and doing things.

就说框架吧。我的意思是我要它发挥到极致。我不希望有任何借口说我的开源模型比它本可以做到的差一点。我要它运行。我要最好的开源模型以最佳方式配置，并且它总是告诉我真相。因为我以前用 Grok，但连 Grok 都有点被阉割了。

Just say harness. I just mean I want it maxed out. I don't want any excuse for why my open source model is even slightly worse than it could be. I want it operating. I want the best open source model configured in the best way and it just tells me the truth all the time. Because I used to go to Grok for that, but even Grok is a little neutered.

是的，它不……我们谈过这个。

Yeah, it's not... we talked about this.

我很想聊聊你那天说的话，但不知道你是否愿意公开说。

I'm dying to talk about what you said the other day, but I also don't know if you want to say it publicly.

整件事可能已经被黑了。可能已经泄露了。哦，不，我会谈的。问题是开源模型在哪里追赶？有几个不同的理论，我会列出我知道的理论和证据，但我还没有观点。

The entire thing is probably hacked. It's probably already hacked. Oh, no. I'll talk about it. So the question is like where are the open source models catching up? Right. And so there's a couple of different theories and I'll just lay out the theories that I know and the evidence for them, but I don't have a point of view yet.

我是说，这些可能是真的。

I mean, it's possible that these are true.

所以有些是显而易见的，中国在做自己的预训练。他们有算力，有自己的数据集。实际上，他们可以爬取更多数据集，因为受版权法约束较少，对吧？所以他们可以爬取 YouTube、Reddit 等等。他们训练这些模型。这是基础模型。然后还有 Anthropic 的指控，说他们在蒸馏我们的模型——他们查询我们的模型，大量获取数据，用来训练自己的模型。如果我是中国，我肯定会这么做，这会产生一种锯齿状智能的模型。确实有一些这种情况，但我觉得 Anthropic 提出这个指控很讽刺，因为 Anthropic 自己也爬取了开放网络并蒸馏了开放网络。所有 AI 都是这样。实际上，我认为如果政府这样做，一个会很受欢迎的做法是说：既然你们在开放网络上训练，如果你们用了开放网络和开放数据，那么你们必须在 X 个月后（比如 12 个月）开放你们的模型，对吧？这似乎很公平，尤其是对 OpenAI 来说，否则你就得名副其实地开放，对吧？而 Anthropic，你们在谈论末日和悲观，你们在建造上帝，但你们不会把上帝拴在自己身边，Dario。我们不想要一个牧师，永远为我们控制上帝、解读圣经。所以我认为这会很受欢迎。好了。总之，第一，他们有自己的预训练，但他们获得了整个网络语料库。他们不必遵守任何版权法。第二，他们在蒸馏现有的美国模型，这些模型开始被更严格地封锁，而政府的国有化控制帮助了封锁。但你知道，网上有数据经纪人，他们拿 Claude Max 账户，以 API 格式转售代币给终端用户，折扣高达 80%，因为这些计划对终端用户有大量补贴。他们向企业收费，零售价每月 8000 美元。

So there's some of it that's obviously, you know, the Chinese are doing their own pre-training. They have their own compute. They have their own data sets. In fact, they can crawl more of the data sets because they're less bound by copyright laws, right? So they can crawl YouTube and Reddit and all that stuff. And they train these models. That's like the base model, right? Then there's like the accusations that Anthropic made which is that they're distilling our models which is they're taking they're querying our models and mass taking that data using that to train their own. If I were China that's what I would do for sure and that's creating a jagged intelligence kind of model and definitely some of that is going on but I think it's rich of Anthropic to call that out when Anthropic crawled the open web and distilled the open web. That's all these AIs are right. In fact, one platform that I think would be very popular if the government were to do it is to say hey you guys have since you trained on the open web if you trained on open web and open data you have to open your model after X months, after 12 months or something, right? That seems to be like a fair a pretty fair thing to say, especially like OpenAI, you have to be otherwise, you have to be actually open AI, right? And Anthropic, you're the one talking doom and gloom and you're building God, but you're not going to keep God on a leash for yourself, Dario. We don't want like a priest, you know, controlling God for the rest of us, interpreting the Bible for the rest of time. So, I think that would be a popular thing. Okay. So anyway, number one, they are they have their own pre-training, but they get the full corpus of the web. They don't have to obey any copyright laws, right? Second, they are distilling the existing American models which are starting to get more locked down and a lot of this government nationalization control helps them be locked down. But you know, there are data brokers online that are taking like Claude Max accounts and reselling tokens from that in an API format to end users of like 80% off because those plans are heavily subsidized for end users. They charge the enterprises. It's like $8,000 a month retail.

任何运营过交易所或受监管在线经纪商的人都知道，KYC（了解你的客户）很廉价。人们总能绕过 KYC。如果你想要，他们可以在越南打印护照。对吧。

Anyone who's run like an exchange or regulated online brokerage knows that KYC is a dime a dozen. People get by KYC all the time. They like print passports in Vietnam if you want. Right.

费用在那里。是的。

The expenses there. Yeah.

是的，没错。所以我认为蒸馏很难阻止。

Yeah. Exactly. So I think the distillation is going to be hard to stop.

但我认为最大的要考虑的是，有一些真正的算法突破，比如 DeepSeek 论文、Deep R1 系统。非常非常聪明的人，我猜从简历数量等来看，今天大多数数学家和 AI 研究人员是中国人。他们培养了更多的 STEM 毕业生、相关领域的博士、奥赛冠军。然后很多美国员工也是中国人。所以这就是为什么我发推说 AI 是我们的中国对抗他们的中国，对吧？这些人，我不认为需要任何阴谋。这些人都是朋友，他们住在同一个宿舍，上同一所学校，住在同一栋楼和公寓，一起玩。是的，他们疯狂跳槽。不仅仅是中国人。你看，每个人都在实验室之间跳来跳去，带着数据走。所以，所有这些都在发生。

But I think the biggest one to think about and there's some genuine algorithmic breakthroughs like the DeepSeek paper, Deep R1 systems. People very very smart people and I would guess just by looking at the numbers of resumes and so on that the majority of mathematicians and researchers in AI today are Chinese. Like they produce more STEM graduates, more PhDs in the relevant fields, more Olympiad winners than anybody else right. So and then a lot of the US staff is Chinese. So in fact that's why I tweeted AI is our Chinese against their Chinese, right? And these guys and I don't think there's any conspiracy required. These guys are all just friends, you know, they live in the same dorms. They went to the same schools. They live in the same buildings and apartments. They hang out. Yeah. And they switch jobs like crazy. And not just the Chinese. You see, everyone's hopping from lab to lab to lab and they're just taking data with them. Okay. So, there's all of that going on, right?

但除此之外，我不认为 AI 公司拥有美国顶级国家安全保密设施的安全级别，对吧？他们不知道如何保护秘密。所以，当然他们被黑了。当然，权重被泄露了。如果你是中国或其他人，你不会只是拿走权重然后释放回去。你会利用它高速蒸馏并训练自己的模型，对吧？或者增强你的模型。所以，我认为还有另一个完整的系统在运作。

But then on top of it, I don't think that the AI companies have the security profile of a top US national secure compartment facility, right? They do not know how to protect secrets. So, of course, they're getting hacked. Of course, the weights are getting leaked. And if you're Chinese or anybody else, you're not going to just take the weights and release them back. You're going to use that to distill at high speed and train your own model, right? Or to augment your model. So, I think there's another whole system right there that's going on.

最后一个，实际上，我认为最大的不是这些。我认为以下不需要任何阴谋。看看中国的位置，他们说，‘好吧，我们在软件上落后了。’美国在软件上一直遥遥领先。在软件上领先让你能提取所有利润，因为长期以来，风投从不投资硬件。为什么？因为硬件是商品业务，对吧？而软件有网络效应和锁定效应。软件是艺术，正如 Patrick Collison 所说，它锁定你。所以人们想投资软件。即使在美国，每当有硬件公司获得融资，风投总会问，软件的锁定在哪里？软件部分是什么？对吧？这就是为什么我过去常说，硬件是护城河，为你争取时间建造软件城堡。但猜怎么着？Claude Code 烧掉了软件。软件在吞噬世界，而 AI 吞噬了软件。所以现在软件被商品化了。一旦你能指定它，AI 就能一次或两次生成，或者在几小时内完成。

And the last one, actually, I think the biggest one is none of those. I think no conspiracy is required for the following. If you look at where China is, they're like, 'Okay, we're behind in software.' The US has been way ahead in software. And being ahead in software allows you to extract all the margins because for the longest time, VCs never invested in hardware. Why? Because hardware is a commodity business, right? And software had the network effects, the lock in. Software is art as Patrick Collison said, it locks you in. And so people want to invest in software. And so even in the US, whenever you get a hardware company funded, the VCs will always ask you, well, what's the software lock in? What's the software piece? Right? And so this is why like I used to say, yeah, hardware is the moat that buys you time to build a software castle, right? But guess what? Claude Code burned software down. Software was eating the world and AI ate software. So now software is commoditized. The moment you can specify it, an AI can one-shot it or two-shot it or get to it within a couple of hours.

从大约 5%上升到约 25%的硬件。20%到 25%的硬件批次。

Went from like 5% to about 25% hardware. 20 to 25% hardware batch.

硬件也被中国商品化了。我能在这里制造的任何硬件，你都能在中国更便宜、更容易地制造。整个生态系统都在那里。深圳有大约 3000 家制造商生产这种小电缆，一应俱全。

Hardware is also commoditized by China. Any hardware I can make here, you can make in China more cheaply and more easily. The whole ecosystem is there. Shenzhen has like 3,000 manufacturers of this tiny little cable all in one.

尼克松与中国关系正常化，并让我们脱离美元本位或金本位。

Nixon for normalizing relations with China and taking us off the dollar standard or the gold standard.

所以，将制造业带回美国需要整整一代人的时间。应该做，毫无疑问。我是美国制造业的忠实粉丝，但这需要很长时间。与此同时，中国在每条供应链上都超过所有人，因为他们意识到生产中的规模经济。所以他们前期补贴，然后成为全球供应商。他们把其他人都赶出市场，然后控制这些关键行业的整个供应链。所以他们拥有硬件。

So, and so bringing reassuring manufacturing back the US is going to take a full generation. It should be done. No question. I'm a big fan of American manufacturing, but it's going to take a long time. In the meantime, China outproduces every one of the supply chains because they realize there are scale economies in production. So, they subsidize the thing up front, then they become the global supplier. They drive everybody else out of business and then they control the whole supply chain across these critical industries. So they own hardware.

未来十年，没人能在硬件上打败中国。它会不断升级。我们会在多个战线上竞争，但他们掌控硬件。所以他们喜欢软件的 commodity 化。当软件变得 commodity 化时，这对中国有利。所以如果你是中国政府，你基本上在资助所有这些实验室。你在做一种公私合作，说：‘别担心，你们不需要赚钱。你们可以是第二或第三。只要开源，让其他实验室能从你们身上学习，我们就能追赶并保持同步。如果软件 commodity 化了，我们就能在硬件上获胜。’所以我现在认为，硬件已经 commodity 化了。但大部分被中国掌控。美国有少数例外，比如 SpaceX。软件也 commodity 化了。那么什么没有 commodity 化？实际上就是 AI 研究本身，开发 AI 成了新的软件工程。问题在于，软件曾经是民主化的。每个人都能竞争。John Carmack 和 John Romero 带着一个小团队就能写出《毁灭战士》和《雷神之锤》，与 EA 和动视等公司竞争。但在 AI 领域做不到。AI 需要海量资源、巨大的 GPU 集群、少量的研究人员、庞大的数据集、专有数据集。然后还有监管俘获。所以看起来所有价值和瓶颈都集中在 AI 上，而这被极少数公司控制。讽刺的是，唯一让我们保持运转的是中国政府通过开源补贴整个体系，以便他们的硬件保持竞争力。

No one's going to beat China on hardware in the next decade. It'll ramp up. We'll fight on a bunch of fronts, but they own hardware. So they like the commoditization of software. So when software is getting commoditized, that helps China. So if you're the Chinese government, you're basically funding all these labs. You're doing kind of a public private partnership where you're saying, 'Don't worry, you don't have to make money. You can be number two or number three. Just open it so the other labs can learn from you and we can catch up and keep up and if the software is commoditized then we win on the hardware.' So the way I think about it today is that hardware is commoditized. But it's owned by China for the most part. There's some few exceptions like SpaceX in the US. Software is commoditized. So what's not commoditized? It's actually just AI research developing working on AI itself is the new software engineering. The problem with that is that software used to be democratized. Everyone could compete. John Carmack and John Romero sitting there with a small team could code up Doom and Quake and compete against EA and Activision and so on. Can't do that in AI. AI takes massive resources, huge GPU clusters, you know, small numbers of researchers, massive data sets, proprietary data sets. So in that and then now regulatory capture. So it seems like all the value, all the choke points are going into AI and that's controlled by a very small number of companies and ironically the only thing keeping us afloat is a Chinese government subsidizing the whole thing on open source so that their hardware can stay competitive.

是的。我是说，我觉得嗯

Yeah. I mean I think the um

而且他们还有权重。没开玩笑。

Plus they had the weights. No kidding.

嗯，他们可能会利用那个，我是说

Well, they might use that as I mean

你可以用一个 U 盘偷偷带走。你是说这还没发生过？

You could sneak about on one USB stick. Are you telling me that hasn't already been done?

这就像你需要物理接触，需要一个内应。比如有那么多男性 AI 创始人身边都有漂亮的中国女友。我们聊过这个吗？没人谈这个。去硅谷外面走走，对吧？

It's like you need physical access like you need one person on the inside. Like there's so many people way every male AI founder is walking around with a beautiful Chinese girlfriend. Have we talked about that? Like no one talks about that. Go walk outside Silicon Valley, right?

我认为我们都来自一个优势有一定持久性的世界，你可以实际利用你的优势对抗世界其他地方，但现在优势在时间上不断缩小。所以，你可能拥有最聪明的模型，但两个星期、三四个星期后，当别人也拥有它时，你还能用它做什么？所以我认为博弈论中有一个我们从未经历过的不同部分，那就是任何优势的时间收缩。嗯，我不认为这让预测谁将掌控一切变得更难，因为就像你说的，几个月后，最聪明的模型就开源了。

I think we all come from a world where advantages like had some durability and you could actually operationalize your advantage against the rest of the world, but that's just like narrowing down in time. So, you know, you might have the smartest model, but what are you going to do with it in two weeks and 3 4 weeks when until somebody else has it too? So the I think there's a different part of the game theory that we've not experienced before, which is the time contraction of any of these advantages. Um, and I don't think it makes it harder to predict who's going to be in charge of anything because like you said a few months later, the smartest model is now open source.

伙计们，我很担心。

Guys, I'm so worried.

这是我引以为豪的一条推文。几年前，我说很快每个人都会有 AI 焦虑。我们都在感受它。我们都感到 AI 焦虑，对吧？这太不可思议了。

This is one of my tweets I'm proud of. A couple of years back, I said soon everybody will have AI anxiety. Like we're all feeling it. We're all feeling AI anxious, right? It's incredible.

欢欣鼓舞。我对 AI 欢欣鼓舞。但两者都有。两者都有，对吧？好的焦虑有点像错失恐惧症，就是看，看所有这些伟大的

Jubilant. I am AI jubilant. But it's both. But it's both, right? It's the anxiety in its good form is kind of FOMOish, which is like look, look at all the great

别担心我，伙计们。

Don't worry about me, guys.

是的。但我认识的所有在 Frontier Lab 做这些研究的科研人员，他们都像抑郁了一样。

Yeah. But like everyone I know that's working at Frontier Lab on these things that are researchers, they're like depressed.

他们是的。

They are.

是的。

Yeah.

他们压力很大，因为他们认为它会杀死我们所有人。

They're stressed out because they think it's going to kill us all.

那是因为他们……我在 2021 年遇到的第一批人告诉我，他们推迟要孩子或考虑不要孩子，因为他们担心孩子的未来。我就想，你在说什么？这与人类历史上其他任何时候不同，那时人们从不担心孩子的未来，所以他们也没有孩子。

That's cuz they're kill. The first ones I met in 2021 were saying to me that they were putting off having kids or thinking about not having kids because they were afraid for the kids' future. And I was like, what are you talking about? as opposed to every other time in human history when people were not were never afraid of their kids' future and so they didn't have any kids.

等等，那为什么这不会导致《星际迷航：下一代》呢？那是我的希望。

Wait, so why is this not just going to result in Star Trek the next generation? Like that's my hope.

部分会很棒。

The part of it would be great.

《星际迷航：下一代》的部分是他们显然有 AGI，但那个 AGI 不知为何没有掌控一切。

The part about Star Trek the Next Generation is that they have AGI clearly, but that AGI is somehow not running things.

我是说，他们当时还无法想象，对吧？

I mean, they couldn't conceive of it yet, right?

他们有全息甲板，但不知为何人们还在做

They have the holiday, but somehow people are still doing

那是够好的编剧。他们

That good enough of writer. They're

不只是生活在全息甲板里。有很多科幻作品，你知道，费心把这个想法进一步推演了。

Not just living in the holiday. There's plenty of science fiction that's, you know, bothered to see that idea further through.

我是说，旧金山有很大比例的人被芬太尼一击毙命。你不觉得全息甲板会干掉剩下的吗？

I mean, some large percentage of the San Francisco population has been oneshotted by fentinel. You don't think the holiday will take out the rest?

对。对。他们只是不在剧里展示。就像 Gene Roddenberry，我肯定他想象过，但展示这些巨大的全息甲板太令人沮丧了，你知道吧？

Right. Right. They just don't show that on the show. Like Gene Roddenberry like I'm sure he conceived of it and it just be too much of a bummer to show like these giant holiday, you know?

你这么说，但你在心里生活在《星际迷航》里，你很开心。那是真的。你对此很兴奋，但你也承认会有很多可怕的事情发生，对吧？

You say this and you're living in Star Trek in your mind and you're happy. That's true. You're like excited for it and it's like but yes, you recognize he just there's a ton of terrible things that are going to happen, right?

然后你就想，‘嗯，《星际迷航》里没有这个。’

And you're just like, 'Huh, it wasn't in Star Trek.'

好吧，但你在摧毁我的梦想，Daniel。

Well, but you're blowing my dreams over here, Daniel.

但为 Gary 辩护一下，比如电力的出现，那是个大事件，对吧？铁路、汽车，这些都是巨大的变化。

But to Gary's defense, like there's been so much electricity, like that was a big one, right? Railroads, the cars, these were big changes.

以前得骑马，比如小马快递。你记得吗？那太疯狂了。但变化越快，颠覆的影响越大，动荡也越大。所以现在我们正在经历非常快速的变化。我们正在取代白领。我们正在取代政府。我们正在取代管理者。我们正在取代学者。我们正在取代记者。你知道吗，我从 AI 那里学到的比大学课堂还多，因为我可以坐在那里，让它以我确切的知识水平来教我，因为它能给我视觉化呈现。给我画个图。我可以问愚蠢的问题。我不懂那个。再解释一遍。再解释一遍。用第三种不同的方式解释。用视觉解释。用听觉解释。用跳舞的精灵解释。我不在乎，只要反复讲直到我明白为止。

Used to have to ride a horse ac like the Pony Express. Do you remember that? That was insane. But the more rapid the change, the more influential the displace, the more tumultuous it was. And so now we're having a very rapid change. We're displacing white collar. We're displacing government. We're displacing managers. We're displacing academics. We're displacing journalists. You know, I learn more from AI than I did in my college classes cuz I can sit there and I can have it meet me at my exact level of knowledge because they give that to me visually. Give me a graph about that. I can ask dumb questions. I don't understand that. Explain it again. Explain it again. Explain the third different way. Explain it visually. Explain it audibly. Explain it with with dancing elves. I don't care like just just to keep going over until I get it.

我在课堂上做不到。我可以学得更好。我只需要在意，对吧？如果我在意，我可以比任何家教、任何老师都学得更好。

I could not do that in class. I can learn so much better. I just have to care, right? If I care, I can learn anything better than with any tutor, better than with any teacher

那么所有人都过时了，他们要领导一场革命。

Where so all the people are obsolete and they're going to lead a revolution.

嗯，不。我们需要想要更多的东西。如果你想要更多的东西，我们想要，是的，这很棒。这是好消息。好消息。每个人都想要更多东西。

Well, no. We need to want more things. And if you want more things, we want Yeah, which is great. That's the good news. Good news. Everyone wants more things.

如果你有这个工具并且不排斥使用它，这就是为什么我如此反对——抱歉又提起 AI 写作的事——但我理解这说得通。人类，当你写得很差时，确实有不尊重的一面，对吧？但这预设了 AI 写不好。

And then if you have this tool and you're not averse to using it, this is why I'm so against—sorry to bring up the AI writing thing again—but I understand that makes sense. Humans, there is a disrespectful aspect when you do bad writing, right? But that presupposes that the AI can't write well.

说得有道理，因为现在的写作量很大。巨大，巨大——就像你给它的提示词那样的小东西。

That's a fair point because right now the writing is voluminous. Giant, giant—like it takes a little thing like the prompt you give it.

好吧，但问题是：你给了它一个提示词，对吧？

Well okay, but here's the thing: you're giving it a prompt, right?

是的。

Yes.

你为什么不直接把提示词发给我？我为什么要读那些垃圾？

Why don't you just send me the prompt? Why do I need to read all the garbage?

因为它必须借鉴所有人类经验。

Because it has to draw on all of human experience.

把提示词给我。直接把提示词发给我。

Give me the prompt. Just send me the prompt.

我是说那可能会发生。

I mean that might happen.

我会把它发给我的 Codex。整理一个 markdown 文件，发给我的开发者，他们可以给他们的 Claude Code 去执行。提示词：去做吧。

I'll send it to my Codex. Put together a markdown file that I send to my developers that they can give to their Claude Code to go execute. Prompt: go do it.

别跟我说话。让你的智能体跟我的智能体谈。就是这样。

Don't talk to me. Have your agent talk to my agent. That's literally what it is.

我的东西会在你的 GitHub 仓库上提一个 issue。

My thing will just file an issue on your GitHub repo.

是的，我想做的是，我上 Zoom，在 Zoom 上聊天，10 分钟后我知道这不适合我。然后我按一个按钮，智能体接管。智能体 Naval 继续说话，我去找下一个。

Yeah, what I want to do is I want to get on the Zoom and I'm on the Zoom having the conversation and then 10 minutes in I know this is not for me. And then I just hit a button and the agent takes over. Agent Naval keeps talking and I go after the next one.

是的。

Yeah.

我右滑你。

I swipe right on you.

是的。

Yeah.

但结果是我被耍了，因为你从一开始就不在。

But then the joke's on me because you were never there to begin with.

我觉得该谈谈政治了。

I think it's time for politics.

政治。政治。听起来大部分时间都在谈政治。

Politics. Politics. Sounds like it's been politics most of the time.

我们先从散布恐惧开始。

Let's start with fearmongering.

政治不重要，因为 AI。

Politics doesn't matter because of AI.

你知道，那不是真的。你在说什么？

You know, that's not true. What are you talking about?

一切都会结束。现在任何与政治相关的东西，三年后都不会再相关。

Everything's going to end. Nothing that's politically relevant right now is going to be relevant in 3 years.

我要说的是，我希望我能投票给他们。

I will say that I wish I could vote for their job.

我希望我能投票给一个 AI 来代表我们。Gary 会让它开源。有人可以做到。他们可以开源它。他们可以把权重放出来让我们看看。

I wish I could vote for an AI to represent us. Gary will make it open. Somebody can make it. They can open source it. They can put the weights out for us to take a look at.

读权重。你可以看权重。我的意思是，不是个人看那个权重，而是一个完全透明的 AI，正在竞选……

Read the weights. You can look at the weights. I mean, not personally that weight, but like a fully transparent AI that is running for...

这个东西，他们有一个完整的领域叫机械可解释性，试图弄清楚 AI 如何思考，但他们进展甚微。

This thing, they had this whole field called mechanistic interpretability where they try to figure out how the AI thinks and they've made very little progress.

当然。我不是说我们会知道，但我只是说你能够接触到同样的……

Sure. I'm not saying we're going to know, but I'm just saying you have access to the same...

它们是不可读的。顺便说一句，这是一种时代错误。我们的训练数据是自然语言和人类能读写计算机语言。所以 AI 是在这些数据上训练的。但未来，AI 会发展出自己的语言。它们会竞争。它们会以某种二进制高速协议相互通信。它们几乎是在说编译代码，甚至可能是我们无法想象的完全不同的东西，这对通信更高效。只有人类才要求可读性。但 AI 并不真正具备这一点。AI 系统比人类编码的专家系统好得多的部分原因就是它们不必是可读的。所以它们能找到我们无法表述的相关性和模式。事实上，我们的大部分知识都是无法言说的。你的大部分感受都是无法言说的。你甚至一半时间都不知道自己在感受或思考什么。你甚至都不在谈论我们的感受。

They are not legible. By the way, it's an anachronism. Our training data is in natural language and in computer languages that humans can read and write. And so the AIs are trained on that. But in the future, the AI will develop their own languages. They'll compete. They'll communicate with each other in some kind of a binary high-speed protocol. They'll almost be talking compiled code or maybe even something completely different that we can't even imagine which will be more efficient for communication. It's only humans who demand the legibility. But the AIs don't really have it. And part of what makes AI systems so much better than human-coded expert systems is that they don't have to be legible. So they can find correlations and patterns that we cannot articulate. In fact, even most of our knowledge is inarticulatable. Most of your feelings are inarticulatable. You don't even know half the time what you're feeling or thinking. You're not even talking about our feelings.

是的。

Yeah.

是的。是的。我们就这样吧。我们在那个我们是专家的领域是专家。

Yeah. Yeah. Let's do that. We're experts in that one place where we are experts.

GLM 5.2。我会投票给 GLM 5.2 竞选公职。是的。

GLM 5.2. I'll vote for GLM 5.2 for office. Yeah.

你的中国霸主。

Your Chinese overlords.

嘿，这是我最喜欢的寓言提炼。

Hey, it's my favorite distillation of fable.

我真的很感激中国人做出了所有那些开源模型。美国只是……有点可悲。

I really do appreciate how the Chinese are making all the open source models out there. It's kind of pathetic that the US just...

全球经济会感激的。

The global economy will appreciate.

中国可能是视频领域的全球领导者，对吧？最好的。所以实际上在某些地方……这是关于开源的一个有趣观点。在某些情况下，开源现在是第一。

China is probably the global leader in video, right? The best. So in some places actually that... that's an interesting point about open source. In some cases now open source is number one.

至少从历史上看，如果你看看 Linux 或其他开源项目，一旦某个开源项目取得领先，它很少会放弃领先地位。这是因为围绕它形成了一个生态系统，每个人都开始接入它，尤其是在企业用例中。所以现在 Claude 和 Codex 稍微领先于 GLM 和 Fable，可能比 GLM 领先很多。但如果像 GLM 这样的东西超过了它们，它们能否夺回领先地位就不清楚了。

And at least historically, if you look at things like Linux or other open source projects, once something open source kind of gets in the lead, it rarely surrenders it. And that's because an ecosystem springs up around it where everyone starts plugging into it, especially in enterprise use cases. So right now Claude and Codex are like slightly ahead of GLM and Fable, and maybe a lot ahead of GLM. But if something like a GLM were ever to surpass them, it's not clear they'd get the lead back.

我的意思是，因为你如何合理化公司有限的资源投入到开源正在获胜的事情上？你必须用你的资源做点什么，而用你宝贵的资源无法赶上开源，这不是一个好的用途。你有一个时钟。所以如果你是 OpenAI 或 Anthropic，你会有一定数量的收入、资金、银行存款和投资者支持，你必须用这些来有意义地领先于开源，并在资金耗尽前重新夺回领先地位。所以……因为如果你看看 AI 竞赛，两年前，人们认为 Google、Elon 和 Meta 也是竞争者。

I mean, because how can you rationalize your limited resources at a company going into something that open source is winning at? You have to do something with your resources and failing to catch up to open source with your precious resources is not a good use. You'd have a clock. So if you're OpenAI or Anthropic, you would have a certain amount of revenue and money and cash in the bank and investor backing, and you would have to use that to sort of jump meaningfully ahead of the open source and take the lead again before that runs out. So that's an... because if you look at the AI race, two years ago, people thought that Google and Elon and Meta were also competitors.

是的，有五王。有五王。

Yeah, there were five kings. There were five kings.

现在有两个王。

There's two kings now.

有两个王。为什么？嗯，不仅仅是因为那些家伙有最好的模型。OpenAI 和 Anthropic 是唯一直接通过模型创收的公司。所以它们没有从其他业务中补贴或交叉补贴。它们没有耗尽现金。它们实际上是把现金再投入进去。第二是因为它们有非常活跃的用户群，现在很多改进来自强化学习。它们得到了指导和轨迹来改进模型。所以它们正在拉开差距。也许 Elon，因为他把 SpaceX 上市了，有战争基金，他可能还有一次机会，再加上他在太空有数据中心。所以你知道，不要赌 Elon 输。Google 我认为已经输了。

There's two kings. And why is that? Well, it's not just because those guys have the best models. OpenAI and Anthropic are the only ones that are making revenue off of their models directly. And so they're not subsidizing or cross-subsidizing from other businesses. They're not running out of cash. They're actually pouring the cash back in. And the second is because they have very active user bases and now a lot of the improvement is coming through reinforcement learning. They're getting the guidance and the trajectories to improve the models. And so they're just pulling away. And maybe Elon, because he took SpaceX public and he's got that war chest, he gets maybe one more bite at the apple, plus he's got the data centers in space. So you know, don't bet against Elon. Google I think has lost it.

那很可悲。发生的事情真的很惊人。

That's quite sad. It's really amazing what happened.

我一直看空 Google。我过去在推特上发过一些相当反 Google 管理层的东西。我当时说，我们会知道他们什么时候认真对待 AI……

I was always bearish on Google. I was tweeting some pretty anti-Google management stuff back in the day. I was like, we'll know they're serious about AI when...

Gemini 3.1 Pro 就像是那个时刻，它和以前一样好，但还不够。它更好。

Gemini 3.1 Pro was like the moment that was as good as it was, but it wasn't enough. It was better.

它不够好。他们甚至没能让基本应用正常工作。我记得我试着升级到每月付 20 美元。那是一场噩梦。这么大的流程。不，不，你需要成为 Google One 的一部分。不，不。你需要一个 Google Workspace 账户。

It wasn't good enough. They never even got the basic app working. I remember I tried to upgrade to pay them $20 a month. It was a nightmare. This huge run. No, no, you need to be part of Google One. No, no. You need a Google Workspace account.

他们是最早推出可用联系人的。很酷。但每次，甚至到今天，iOS 应用——我知道那不是他们的平台——但其他人都能搞定。你在 Gemini 里输入查询，放到后台，它总是断连。直接掉线。而其他应用都能在后台运行。这有多难？都 2026 年了，哥们，跑个后台应用吧。

They were first to have contact usable contacts. Very cool. And then every time, even today, the iOS app, which I know it's not their platform, but everyone else seems to figure it out. You put a query into Gemini, you background, it always loses the connection. It just drops it. Whereas all the others will just run it in the background. How hard is that? It's 2026. Run a background app, guys.

GRPC。那边只需要再斯巴达一点。Meta 可能太过了。

GRPC. You just need a little bit more Sparta over there. Meta might be too much.

我觉得对谷歌来说太晚了。他们得干掉 5000 个谷歌产品经理。这是唯一的办法。全是产品经理的垃圾。整个公司现在都是产品经理的垃圾，对吧？无论你去哪，都有大量产品重叠。一团糟。你根本没法在这迷宫里找到路。

I think it's too late for Google. I think they have to kill 5,000 Google PMs. Like that's the only way. It's PM slop. The whole company's PM slop at this point, right? Like anywhere you go, it's just like tons of products overlapping. It's a huge mess. You can't figure your way through the maze.

现在有些激励问题。所有最优秀的人都想要 OpenAI 或 Anthropic 的股票，或者想和 OpenAI 或 Anthropic 最优秀的同事一起工作，或者想要 OpenAI 或 Anthropic 的数据集，他们当然不想被所有产品经理和遗留的破事束缚。他们早就该用谷歌 AI 取代谷歌搜索了。我甚至都不看结果了。我就看谷歌。

Some incentive problem now. All the best people want stock in OpenAI or Anthropic, or they want to be sitting alongside the best colleagues in OpenAI or Anthropic, or they want the dataset from OpenAI or Anthropic, and they certainly don't want to be tied down by all the PMs and all the legacy BS. And they should have replaced Google Search with Google AI a long time ago. I don't even look at the results anymore. I just look at Google.

这里有一个替代愿景：今天的谷歌可以——我的意思是他们可能有所有会议的记录，如果没有，他们应该有——而且他们有你的所有邮件，对吧？然后他们可能需要用 Anthropic Claude 5 之类的，因为他们自己的模型还不够好，但你可以对这些进行说话人分离，你可以为谷歌的每个员工建立一个 markdown 文件，然后计算一个分数，你巧妙地注册了商标和专利。

Here's an alternative vision: Google today could—I mean they probably have all the transcripts of all the meetings, or if they don't, they should have it—and they have all your emails, right? And then they probably need to use Anthropic Claude 5 or whatever because their own models are not quite good enough, but you could diarize all of these, you could have a markdown file for every single employee at Google, and you could compute a score, which you cleverly trademarked and patented.

没错，谷歌，当然是通过 Gstack 运行。不。总之，Google Omni 是最好的……谷歌有代价。是 7%。

That's right, Google, running through Gstack of course. No. Anyway, Google Omni is the best for... Google has a price. It's 7%.

所以，这是标准做法。就是标准。说真的，我的意思是你可以解决这个问题。如果有人是真正的领导者，真的想解决这个问题，所有这些都只是系统问题，对吧？

So, it's standard. It's just standard. Seriously, I mean you could solve this. If someone was a true leader who really wanted to fix this, all of these things are just systems problems, right?

是的，但这需要拉里和谢尔盖走进去，就像《明星伙伴》里阿里·戈尔德回到办公室开枪打人的场景，只不过是用他们的利爪和实弹。我是认真的。他们不需要亲自到场。你知道我从 Brex 的 Pedro 那里学到的一件事，他就是那样经营 Brex 的。他对每个团队都有完全的信息感知。还有一件事让我……

Yeah, but it would take Larry and Sergey walking in there and literally it's that scene from Entourage when Ari Gold walks back in the office shooting people, only with their open claws and live ammunition. Like I'm serious. You don't need them to be there physically. You know one of the things I learned from Pedro at Brex, like that's how he runs Brex. He has total information awareness on every team. And one of the things that got me...

他们不是输给 Ramp 了吗？我开玩笑的。我开玩笑的。红队来了。

Didn't they lose to Ramp? I mean I'm kidding. I'm kidding. The red guys come in.

我对这两家都没有利害关系。

I don't have any skin in the game on either of those.

好吧，我们接受第一个打电话来的赞助商。谁付钱多，我们就跟谁谈。总之，我的意思是，他真的为他自己构建了 CEO 大脑。这是一个个人版的 Claude，知道 KPI 以及他的所有直接下属在那些会议上说了什么。而且它一直在对他们进行重大监视。

Okay, we're accepting sponsorships from whoever calls in first. Whoever pays us more, we'll talk. Anyway, I mean no, I mean he really did instrument the CEO brain for him. It's a personal Claude that knows the KPIs and literally what all of his directs talk about in those meetings. And it's significant spying on them all the time.

我的意思是，是的。不，我的意思是，他们不是唯一一家。我和其他 CEO 谈过。他们不会公开谈论这些，都是关起门来。这是你能做的最大程度的事情：将权力集中在你的 CEO 或联合创始团队手中，这样你就能真正了解正在发生的事情。

I mean yes. No, I mean the thing is, they're not the only company. I've talked to other CEOs. They don't publicly talk about this behind closed doors. This is like the max thing that you could do: concentrate the power in your CEO or co-founding team so you have actual awareness of what is going on.

嗯，现实是团队应该小得多，你的大多数员工都是在数据中心工作的 GPU。我的意思是，当然，如果你能减少 10 个——比如现在员工的十分之一——只留下那些 100 倍效率的人，那说起来容易做起来难。大多数 CEO 不想解雇 50 个人，但应该解雇更多。

Well, the reality is the team should be much smaller and most of your employees are GPUs working in the data center. I mean, of course, if you can have 10 fewer—like a tenth of your employees right now—and just make them the 100x guys, then that's easier said than done. Most CEOs do not want to fire 50 people, but there should be a lot more.

我在董事会待过，那里就像，兄弟，我们需要解雇 80% 的人，你有机会，但你根本不知道是哪 80%。数千万美元，你不知道通过经理你正在处理哪 80%。我的意思是现在你可以——这不是小事。这是大事。你可以拥有完全的信息。

I've been on boards where it's like, bro, we need to fire 80% of the people and you have a shot, but you literally don't know which 80%. Tens of millions of dollars, you don't know which 80% you're working through managers. What I'm saying is now you get—and that's not a little thing. That's a big thing. Like you can have total information.

好吧，这对我来说是件大事。我认为未来本质上是你需要驾驭 AGI。你实际上需要——对我来说最重要的事情是 100 万 token 的上下文窗口是个大事。就像我常说的，相当于三本《哈利·波特》的书，但你知道吗？人类大脑能记住的东西是七加减三。我们只是小猴子的大脑。我们谈论超级智能在未来。超级智能现在：一个为你工作的智能体可以生成一千页，像打印出来一样，有价值、有用、可信。

Okay, like this is a big thing to me. I think that the future is essentially that you need to ride the AGI. Like you actually need to—the most significant thing to me is like the 1 million token context window is a big deal. It's like I talk about as three Harry Potter books, but guess what? A human being can keep in their brain seven things plus or minus three. We are little monkey brains. We talk about superintelligence being in the future. Superintelligence right now: an agent working for you can literally make a thousand pages like printed, valuable, useful, credible.

我们别再谈 AI 写作了。

Let's not talk about AI writing again.

不。这不是写作的事。不。这是想法的事。就像你可以了解公司里正在发生什么。就像世界上每个 CEO，真的，所以你可以让你的 AI 接入整个公司的所有系统，评估一切，然后告诉我现在该解雇哪 20%。

No. It's not a writing thing. No. It's an idea thing. It's like you can understand what is happening in your company. Like every CEO in the world, like really, so you could have your AI tie into all the systems in the entire company, evaluate everything, and tell me which 20% to fire right now.

我的意思是你可以做到，而且你应该做。这就像世界上最容易的事，但没人会做。没人做。Meta 的项目？

I mean you could do that and you should. That's like the easiest thing in the world, but no one will do it. No one's doing it. Project at Meta?

而且它并不完美。你知道这方面的情况吗？我只知道我读到的。我的意思是，我在 Meta 认识的一半聪明人都走了。

Plus it's not perfect. Do you know anything about this? I only know what I read. I mean half of the smart people I knew at Meta have left.

是的。我的意思是，对不起，我真的很喜欢 Meta，你知道，你们很棒，但它也……我认为它已经变成了血汗工厂。我不认为你可以通过购买人来建立文化或士气，尤其是来自不同背景和地方的人，把他们硬凑在一起。然后你不知道有多少是真正的传教士，有多少是雇佣兵。文化是真实的东西，需要时间来演变和成长。Meta 尤其需要做我刚才描述的事情：必须有人真正了解每个人每个团队实际在做什么。我的意思是你现在就可以做到，对吧？但你也需要调整它。必须有一个评级。什么才是真正的好员工？他们的行为是什么？他们的特点是什么？你现在就可以调整你的文化。

Yeah. I mean I'm sorry, I really like Meta, you know, like you guys are great, but it's also... I think it's become a sweatshop. I don't think you can build a culture or morale by just buying people, especially all different from all different stripes and places and ramming them together. Then you have no idea how many are actually missionaries, how many are mercenaries. Culture is a real thing that takes time to evolve and grow. Meta of all places needs to do what I just described: someone has to actually know what is actually happening per person per team. And I mean you can do this now, right? But you also have to tune it. There has to be a rating. What is actually a good employee? What is their behavior? What are their characteristics? You can actually tune your culture right now.

我的意思是，把整个工程团队变成数据标注员，就像在那里扔了一颗核弹。这在我看来是不分青红皂白的，对吧？你为什么要这么做？尤其是因为你已经有了这种极其精细的东西。

I mean, taking the whole engineering team and turning them into data labelers is just like dropping a nuclear bomb in the place. That seems indiscriminate to me, right? Why would you do that? Especially because you have this incredibly fine-grained thing.

就像你有一个提示词、一个技能文件、一堆 Markdown 文件，你可以精确地调整输出。

Like you have a prompt, you have a skill file, you have like a bunch of markdown. Like you can actually tune it to exactly the output.

我觉得他们的组织太大了，跟 Google 一样的问题。大组织里你会做一些在特定领域有意义的事，然后被媒体泄露、被炒作。哦，他们要把所有人都变成数据劳工。可能人太多了，厨房里厨师太多。但即便如此，它仍然比三本《哈利·波特》能装下的内容少得多，对吧？

I think their org is just too large. It's the same problem as Google. When you have a large org, you do things that actually make sense in certain domains. Then it gets leaked to the press, gets sensationalized. Oh, they're turning everyone into data labor. It's probably have too many people. You have too many cooks in the kitchen. But it's still like way less than what you can fit in three Harry Potter books, right?

你可以举个千禧一代的例子。

Millennial example you could give.

抱歉，我不知道。

Sorry, I don't know.

读另一本书吧。

Read another book.

是的。

Yes.

嗯，我觉得可怕的大问题是：创业生态还会存在吗？因为一方面公司在缩小，人们杠杆更高，这意味着更多创业公司，小团队能做更多事。很好，太棒了。

Well, I think the big scary question is will there still be a startup ecosystem because on the one hand you're shrinking the size of the firm. People are more leveraged. That should mean more startups, small teams can do more. Great. Fantastic.

我们确实看到了。人们能达到 1 亿 ARR 而不需要大规模融资。

We've definitely seen that. I mean people can get to 100 mil arr like not raise a big

但做什么呢？卖软件，而软件正在被这些法律领域的 Harvey 之类的完全取代。这是个很好的测试：Claude 的基础 Mythos 或 Fable 会不会比专门针对法律训练的模型做得更好，尤其是结合工具使用？

But doing what? Selling software and software is getting completely all these legal guys like Harvey and whatever like it's a good test like wouldn't claude just basic mythos or fable just do a better job than something trained specifically on legal especially with tool use

这取决于前沿实验室是否开放他们的最佳模型，还是封闭它们。

It depends because it depends on whether or not the frontier labs stay open for their best models or do they close them

我觉得只要有人付够钱，他们就会想办法开放。所以如果一家律所愿意付钱获得最佳结果，那……

I think anyone who pays them enough they'll find a way to be open so if a law firm will pay them the best result like that's

这不是必然的。Anthropic 完全可以说，实际上 Fable 6……

That's not a given. It is possible for Anthropic to say like actually Fable 6

我愿意拿我的生意打赌，他们不会进入我的垂直领域。我认为如果他们能不用定制模型就进入你的垂直领域，他们就会这么做。而这里的一个惨痛教训是通用模型打败了专用模型。

I would bet my business they're not going to enter my vertical. I think if they can enter your vertical without a custom model, they'll do it. And I think one of the bitter lessons here is that the general models beat the specialized ones.

是的。

Yes.

他们不需要进入，它直接走向用户。对，完全正确。

They don't have to enter it. It's just going to people. Yeah. Exactly.

占领那个市场。不是因为 OpenAI 法律法案，对吧？只是因为……

Toate that market. Not because, you know, OpenAI legal act, right? It's just because it's so much

我为什么要用别的？为什么要每月付 5000 美元给其他 SaaS 公司？我觉得 2027 年会是 AI 工具链战争之年。会是 Codex 吗？会是 Co-Work Claude Code 吗？会是 OpenClaw Hermes 吗？到底是什么？人们日常会用什么工具链来做所有事？

Why would I use something? Why would I pay $5,000 a month for some other SAS company? I think that like 2027 will be the year of the AI harness war. Is it going to be Codex? Is it going to be Co-Work Claude Code? Is it going to be OpenClaw Hermes? Is it going to like what is it? You know who's like what harnesses are people going to use day to day for everything?

对。现在，2007 年之后硅谷之所以能存在，是因为有两个手机提供商：苹果和谷歌。如果只有一个，如果 Android 生态系统不存在，创业公司的处境会艰难得多，你会面对链条末端或顶端的真正垄断者。

Yeah. Right now so a lot of Silicon Valley after 2007 got to exist because you had two mobile phone providers. You had Apple and you had Google. If you had only one, if Android didn't exist as an ecosystem, it would have been a much tougher position for startups, you would have been facing a kind of a true monopolist at the end of that chain or the top of that chain.

对吧？现在至少有两匹马在互相竞争，所以它们会迫使对方收敛一点。如果只剩一个，对创业公司来说是非常糟糕的局面。

Right? Now, you at least have two horses fighting each other, right? So, they kind of force each other to behave a little bit. If it boils down to just one, it's a very bad situation for startups.

嗯，那他们就会把它国有化，然后也许我们可以让政府……

Well, then they nationalize it and then, you know, maybe we could get the government to

他们真的会国有化吗？他们没有国有化 Google，也没有国有化 Apple。

Do they really nationalize it? They didn't nationalize Google. They didn't nationalize, you know, Apple.

这取决于情况有多糟。搜索是一种很好的公用事业，有很多理由……

It depends on how bad it gets. I mean like search it's a good utility like there are lots of reasons

如果你把它国有化，它可能会变慢并失去优势，因为现在由 DMV 运营，中国就赢了，对吧？

And if you nationalize it it might slow down and lose cuz now the DMV is running it and China wins right

中国人可没闲着。据说现在中国某处有一台 ASML 机器。你听说了吗？Howard Lit ASML 非常生气，因为一台 ASML 机器不知怎么泄露到了中国。

It's not like the Chinese are asleep. There's an ASML machine supposedly somewhere in China now. You hear about that? Howard Lit ASML and was all mad because an ASML machine somehow leaked into China.

那是光刻机，全世界只有 184 台。

That's the lithography machine. There's only 184 of them.

那是我们的蚕逃走了。

That's our silk worms escaping.

对。而且除此之外，他们还在开发华为芯片，研究 7 纳米、5 纳米工艺，建造晶圆厂，他们正在努力。他们不笨。那里有 14 亿人，有很多 STEM 博士，有很强的民族自豪感，可以重新调配资本。我不是说它能与自由市场经济的创新相匹敌。但美国领先是因为我们创造了这个薄薄的一层：我们从世界各地吸纳最优秀、最聪明的人，疯狂激励他们相互竞争，然后从中收获回报。硅谷不成比例地收获了回报，加州收获了回报，并没有太多扩散到美国其他地区。现在美国其他地区吓坏了。但就是这样。这就是所谓的自由世界与中国竞争之间的那道薄薄的蓝线。

Yeah. Or and then on top of it they're developing their Huawei chips, they're working on their 7 nm, 5 nanometer whatever processes, they're building their fabs, they're working on it. They're not dumb. I mean you've got 1.4 billion people there. They've got a lot of STEM PhDs. They have a lot of national pride. They can redirect capital. I'm not saying it matches up to innovation in a free market economy. But the US is ahead because we have created this very thin layer where we take the absolute best and brightest from all over the world and we incent them like crazy to compete with each other and then we collect the rewards from that. And Silicon Valley actually disproportionately collects the rewards and California collects the rewards and it's not really spread out to the rest of the country as much. The rest of the country is scared shitless now. But that's it. That's the thin blue line standing between, you know, the so-called free world and the Chinese competition.

现在，我要提出一个激进的观点。我不认为我们在与中国竞争。我看不到竞争。台湾，我认识的每个台湾人都不想打仗。台湾的有钱人忙着通过把孩子送出国来逃避兵役。台湾有个规定，好像是强制征兵，但如果你的孩子在过去 5 年中每年有 3 个月在国外，就不再有义务兵役。所以每个有钱人都每年把孩子送出国 3 个月，他们有一套标准流程。他们对打仗没兴趣。台湾第二大政党非常亲中。所以已经很妥协了。大多数台湾人认为他们会像香港一样，在这个过程中变富，成为通往中国的门户，然后一两代人之后同化。美国要怎么保卫台湾？航空母舰已经过时了。中国陆基导弹可以击沉航母。无人机。大疆是全球最大的国防承包商。这就像我们试图……就像中国试图从我们手中保卫佛罗里达群岛。是的。

Now, I'll actually take a radical point of view. I don't think we're in competition with China. I don't see the competition. Taiwan, I mean, every Taiwanese person I talked to doesn't want to fight. You know, the rich people in Taiwan are busy dodging the draft by taking their kids out of the country. You know, there's some rule in Taiwan that like mandatory conscription. But if your kid has spent I don't know the numbers but something like 3 months a year for the last 5 years out of the country then they're not eligible for the draft anymore or no more mandatory conscription. So every rich person is always taking their kids for 3 months a year out of the country. They have like a standard system doing it. They're not interested in fighting. The second largest political party there is very pro-China. So it's already very compromised. Most people in Taiwan are thinking they're going to go like Hong Kong. They're going to get rich in the process, that gateway to China, and then one or two generations later they'll assimilate. And how the heck is the US going to defend Taiwan? Aircraft carriers are dead. Land-based missiles from China can take out aircraft carriers. Drones. DJI is the largest defense contractor in the world. It's like us trying to... It's like China trying to defend the Florida Keys from us. Yeah.

你知道，这很荒谬。整个概念都很荒谬。你看看我们刚跟伊朗的冲突。我们表现如何？我们大概七天内就会用完导弹。我们没有制造业基础。现在，也许 Anduril 和 Seronic 等公司能发展到那一步并建立起来，但那还很遥远。所以我认为中国人知道这一点，台湾人也知道，每个人都在试图保全面子，而台湾会在 10 到 20 年内慢慢与中国统一。除此之外，我们跟中国并没有真正的矛盾。

You know, it's ridiculous. The whole concept is ridiculous. And you saw we just went through with Iran. How do you think we'd fare? We'd run out of missiles in like seven days. We don't have manufacturing base. Now, maybe Anduril and Seronic and others get there and build all of that up, but that's a long ways off. So I mean I think the Chinese know that and I think the Taiwanese know that and I think everyone's trying to save face while like over 10, 20 years Taiwan slowly reunites with China and outside of that we don't really have a beef with China.

美国面临的一个大问题是，我们长期处于低增长、高通胀的环境。现在增长可能回升了，但通胀也在上升。人们正在争夺民主社会主义者带来的战利品，比如亿万富翁税、财富税，社会主义者试图夺取生产资料。穆恩·多恩在民主党建制派周围绕圈子，他们谈论着消灭西方文明。人们要养家糊口，要让人感觉良好，要让人觉得自己不会失业，要喂饱他们。而与中国开战似乎是最愚蠢的事。我们应该对他们征收适当的关税。如果他们对我们征收关税，如果他们补贴自己的企业，尤其是在网络效应或规模经济方面，与我们竞争，那么我们也应该设置自己的壁垒，因为你要保护本地产业不被扼杀，这样你至少还有一搏之力，因为这些确实存在规模经济。这不像大卫·李嘉图说的那么简单，哦，我卖香蕉，你卖西瓜，我们可以直接贸易。不，因为那个更擅长做某事的人会变得非常擅长，然后可以做得非常便宜，或者他围绕生态系统或实际锁定获得网络效应，新用户为现有用户创造价值。所以你不能有那种简单的思维。你必须在公平的竞争环境中竞争。你必须确保货币对齐，没有人补贴等等。但话虽如此，我们没有理由与中国开战。真的没有。

One of the big problems the US is facing is that we've been in a low growth, high inflation environment for a while. Now maybe growth is picking back up, but so is inflation. And people are fighting over the spoils of the democratic socialists coming up, you know, the billionaires tax, the wealth tax, the socialist trying to seize the means of production. Mum Donn doing these rings around the democratic establishment and they're talking about the eradication of Western civilization. People, you got to put food on the table. You got to make people feel good. You got to make people feel like they're not going to go jobless. You got to feed them. And picking a war with China just seems like the stupidest thing to do. Now, we should have the appropriate tariffs against them. If they have tariffs against us, if they subsidize their businesses, especially in network effect or scale economies, to compete with us, then we should also have our own barriers because you got to protect your local industry so it doesn't get snuffed out so you at least have a fighting chance because these things do have economies of scale. It's not as simple as David Ricardo said where it's just like, oh yeah, I'm selling bananas and you're selling watermelons and we can just trade. No, because the guy who gets better at doing something gets really good at doing it and then can do it very very cheaply or he gets a network effect around either an ecosystem or an actual lock in where new users create value for existing users. So you can't have that simplistic of a mindset. You do have to play on even playing field. You have to make sure the currencies are aligned, no one's subsidizing, etc., etc. But that said, there's no reason for us to go to war with China. There really isn't.

但这是关于台湾还是控制南海？为什么我们需要控制南海？好吧，世界分裂成两半。

But is it like Taiwan or like control the South China Sea? Why do we need to control the South China? Like, okay, so the world breaks in half.

世界已经分裂成很多块了。

The world's already broken into many pieces.

我知道，但现在我们的船经过那里没问题，对吧？但世界真的……如果他们来回运送贸易货物，他们为什么要阻止我们的船通过？谁知道他们为什么想那么做。我们有一条规则，海上的规则是我们可以在任何地方自由通行。

I know, but like right now our ships go through there and it's fine, right? But like the world literally... Why would they stop our ships from going through there if they're carrying trade goods back and forth? Who knows why they might want to do that. We have a rule that's like, you know, the rule of the sea is that we can freely pass wherever we want.

这个国家内部正在进行一场内战，如果你没注意到的话。是民族主义者对共产主义者。双方都变得更加激进。我的意思是，如果世界分裂成两半，中国控制亚洲，我们的船有时可以通过，他们可能会收费。中国确实控制了大部分东亚地区，在它没有控制的地方，日本、韩国和其他国家要维护自己的利益，所以他们可以与美国结盟，但我们不会为了南太平洋的利益与中国发生枪战。

There is a civil war going on internally in this country, in case you haven't noticed. It's nationalists versus communists. Everyone is going more radical on both sides. I mean it might be fine if the world breaks in half and China controls Asia right and it's like get our ships through sometimes they charge things. China does control most of East Asia and to the extent that it does not it is up to Japan and South Korea and others to look out for their interest so they can ally with the US but we're not going to get into a shooting war with China over interests in the South Pacific.

我们永远不会那么做，但我们会派船经过那里，他们会撞它。

We would never do that but we will sail a ship through there and they'll hit it.

他们为什么要撞我们的船？如果他们派船通过巴拿马运河，我们为什么要攻击它？如果那是一艘合法的贸易船只，往返于美国运送我们交易的货物和服务，我们有什么动机对彼此做那种事？

Why would they hit our ship going through there? If they put a ship through the Panama Canal, why would we attack it? If it's a legal trade vessel carrying goods and services back and forth from the US that we're trading on, we're incentivized to do any of that to each other.

是的。

Yeah.

我认为现在，中国实际上对这些事情的关心程度比我们想象的要低得多。中国非常关注内部。中国共产党非常关注内部。他们想继续掌控自己的帝国。你知道，我不是共产党的粉丝，尽管我认为中共更像一个资本主义法西斯组织。共产主义只是表面，对吧？法西斯的意思是，国家为了共同的事业而团结起来。

I think right now, China actually cares way less about this stuff than we think. China's very internally focused. The CCP is very internally focused. They want to stay in charge of their empire. And you know, I'm not a fan of the Communist Party, although I think CCP resembles more of a capitalist fascist organization. And the communist is a veneer, right? Fascist in the sense of like it's the nation uniting for a common cause.

他们很会赚钱。

They're very good at making money.

是的，没错。不仅如此，他们实际上非常能干，如果你把它和加州政府相比，那简直是灾难性的。

Yeah. Exactly. Not only that, they are actually pretty hyper competent, which actually if you compare it to like the California government, it's disastrous.

高智商且同质化，对吧？是的。

High IQ and homogeneous, right? Yeah.

嗯，我的意思是，当你看到高铁，就像，伙计们，改改 SQL 吧。是时候建设了。是时候建设了。没人建设。硅谷有少数人在建设，但我们甚至在建一些虚无缥缈的东西，吓人……

Well, I mean, when you look at high-speed rails, like guys, just change the SQL. It's time to build. It's time to build. No one's building. There's a few of us building in Silicon Valley, but we're even building ethereal things that are scaring...

他们是共产主义者而不是……是的。

They're communist instead of... Yeah.

我们就把我们的人送过去，但他们说，‘不，不，不，我们不想要这些人，不太好。’

We'll just send ours over there, but they're like, 'Nah, no, no. We don't want these guys like not so good.'

他们管他们叫什么？Baou，就像他们有个 bao。是的。那是什么意思来着？

What do they call them? Baou like they have like a bao. Yeah. What does that mean again?

意思是白左。

It means white and left.

白左。是的。这是一个贬低性的术语，对吧？他们娘娘腔或愚蠢之类的。

White and left. Yeah. And it's kind of derogatory look down term, right? They're effeminate or stupid or something like that.

我的意思是，以上所有。是的。我就是这么看他们的。

I mean, all of the above. Yeah. I mean, that's how I think of them.

是的，没错。

Yeah. Exactly.

中间派民主党人。每个人……

Centrist Democrat. Everybody...

中间派民主党人，旧金山宝贝。我觉得你穿着它。你就像，‘先生，请别吃我。’

Centrist Democrat Democrat San Francisco baby. I feel like you're wearing it. You're like, 'Please don't eat me, sir.'

不，我的意思是，看看旧金山。你知道……

No, I mean, look at San Francisco. Like, it's you know...

旧金山正在反弹。

San Francisco is on the rebound.

我们至少会在监事席位上大获全胜，你知道……

We're going to run the table on the supervisor seats at least, you know, like...

但问题是洛杉矶正在下沉。是的。

But the problem is that LA is sinking. Yes.

洛杉矶已经沉了。好莱坞被赶走了，疯了。真是疯了。

LA has sank. Hollywood got chased out insane. Like, it's insane.

这不是问题。这是机会。我所有纽约的朋友都会回到旧金山。

This is not a problem. This is opportunity. All my friends in New York are going to come back to San Francisco.

这就是问题所在。加州垄断了美国所有温暖干燥的海岸线、所有地中海气候的土地，而美国是世界上最强大的帝国。所以，它是最强大的，拥有世界上最好的土地，全在一个州。应该分成五六个不同的州。

This is the problem. California has a monopoly on all the warm dry coastline, all the Mediterranean land in the United States, which is the most powerful empire in the world. So, it's the most powerful, it's the best land in the world, all in one state. There should be five or six different states.

美国 30%到 50%的 GDP 将集中在加州。

30 to 50% of the GDP of the United States will concentrate in California.

可耕地、美丽的海岸线、好天气。这是地理的诅咒。他们拥有所有自然资源，对吧？基本上就像南意大利或法国，甚至更好，但在美国，而且全在一个该死的州。加州是一种奇怪的直接民主。有史以来最糟糕的主意。50.1%的人可以投票决定任何事情，他们只看个标题。他们就说，‘是啊是啊，免费猫，免费猫狗粮给大家。太好了。我们投票赞成。吃富人。’

The arable land, the beautiful coastline, the good weather. It's a curse of geography. They've got all the natural resources, right? It's like basically southern Italy or France or even better, but in the US, and it's all in one freaking state. California is this weird direct democracy. The worst idea ever. 50.1% can vote anything and they just read a headline. They're like, 'Yeah, yeah, free free cats, free food for cats and dogs for everyone. Great. We'll vote for it. Eat the rich.'

斯泰尔斯·斯泰尔进不去，对吧？嗯，你知道，基本上我们实际上有两个相对中间派的州长候选人，这很好。这太棒了。太好了。我喜欢这样。对吧。

Styers Styer couldn't get in there, right? Um, you know, basically like we actually have two relatively centrist options for governor, which is good. Which is amazing. That's great. Like I like that. Right.

所以我认为所有共产主义实验都会失败，就像它们在旧金山失败一样，在纽约也会失败。

So I think all these communist experiments will fail everywhere, like they failed in San Francisco, they're going to fail in New York.

我不同意。我认为它们会失败。旧金山的不同之处在于，人口结构就是命运。纽约、洛杉矶和芝加哥的人口结构坚定地支持那个事业，情况越糟，他们越会投票支持。就像委内瑞拉一样，他们会一路投票到底，没有人能飞进去逮捕下一个马杜罗。这必须内部解决。美国是自由的最后堡垒。如果没有民兵在红州和紫州扛着 M16 和 AR-15 游行经过州议会，说我们受够了封锁，COVID 封锁就不会解除。然后封锁就结束了。如果没有那 4000 万拥有枪支的美国人，整个世界还会被封锁 6 到 12 个月。每个人都喜欢讨厌他们，但他们守护着你的自由。自由不是免费的。你需要那些人。

I don't agree. I think they'll fail. The difference with San Francisco is demographics are destiny. The demographics of New York, LA, and Chicago are firmly committed to that cause, and the worse it gets, the more they'll vote for it. It's like Venezuela. They'll vote all the way down, and there is no one to fly in and arrest the next Maduro. It has to be solved internally. The United States is the last bastion for freedom. Even the COVID lockdowns would not have lifted if you didn't have militia men marching around in red and purple states carrying M16s and AR-15s past state houses, saying we're done with the lockdowns. Then the lockdowns came up. The whole world would have been in lockdowns for 6 to 12 months longer if it weren't for the 40 million Americans who have guns. Everybody likes to hate on them, but they guard your freedom. Freedom ain't free. You need those people out there.

所以你看，当美国崩溃时，首先，我认为自由会从世界其他地方流失。你已经看到澳大利亚和英国在言论法、审查制度和奇怪逮捕方面的情况。但如果美国堕落并倒下，它不会像欧洲那样失败。欧洲现在也不太好，但它不会变成一个大型养老院和博物馆，每个人都能享受良好的社会福利一段时间。它会像拉丁美洲国家那样失败，因为它与拉丁美洲接壤。所以有很多人想进来，他们会进来。所以你看到的更像是卡特尔、毒品、犯罪和暴力。

So look, when the US collapses, first of all, I think freedom bleeds out from the rest of the world. You can already see what's happening in Australia and the UK with speech laws, censorship, and weird arrests. But if the US degenerates and falls, it doesn't fail like Europe did. Europe isn't faring well these days either, but it doesn't become a big retirement home and museum where everyone is on good social welfare for a while. It fails like a Latin American country fails because it is bordered by Latin America. So you have a lot of people who want to get in here, and they will get in. So you're looking at much more like cartels, drugs, crime, and violence.

好吧，白色药丸：机器人技术。我不想搞 UBI。我们应该搞 UBR。

Okay, white pill: robotics. I don't want to do UBI. We should do UBR.

UBR 很好。我喜欢。我们需要通用基本机器人。每个人都应该有一个机器人，机器人应该为你做饭、收拾东西。好消息和坏消息是，机器人技术的支持者说它还有 2-3 年，怀疑者说 5-10 年，没有人认为这是不可能的。我们已经有了自动驾驶汽车。

UBR is good. I like that. We need universal basic robot. Everyone should have a robot, and the robot should cook for you, pick up your stuff. The good and bad news is that robotics boosters say it's 2-3 years away, skeptics say 5-10 years, and no one thinks it's impossible. We already have self-driving cars.

我认为如果人们有更好的生活，我们可以让他们真正快乐。谁喜欢刷马桶？没有人。但如果机器人来做，那就太好了。如果你投票给中间派民主党人，你会喜欢这个。

I think we can make people really happy if they have a better life. Who likes to clean their toilet bowl? Nobody. But if a robot does it, it's great. You'd love this if you vote for centrist Democrats.

很多移民和税收问题是因为婴儿潮一代退休了。他们下面没有足够的人来照顾他们。他们给自己投票了很多福利。社会保障没有足够的钱。人口在萎缩。所以更少的工人要养活大量这些人。所以他们说，‘好吧，我们就把国家卖了然后退休。’

A lot of the immigration and taxation issues are happening because the boomers are retiring. There aren't enough people below them to take care of them. They're voting themselves lots of benefits. There's not enough money left in Social Security. The population is shrinking. So a smaller number of workers are carrying a large number of these guys. So they're saying, 'Okay, we'll just sell the country and retire.'

这就是道路。我们需要支持技术、注重富足的中间派民主党人。

This is the path. We need centrist Democrats who are pro-technology, who are abundance-oriented.

好吧，但让我唱个反调：医疗保健，90%是浪费的。甚至不是人们需要的东西。有多少医疗保健是用来让你在最后两周活着的？那真的不是……

Okay, but to play devil's advocate for a second, healthcare: 90% of healthcare is wasted. It's not even stuff people need. How much of healthcare goes into keeping you alive for the last two weeks? That's not really...

我可以用 markdown 文件解决这个问题。

I can fix this with markdown files.

好吧。但教育就是一场大型角色扮演。任何想学习的人现在都可以用 AI 和开源课件自学。他们只是想要那个文凭，他们想上大学，在田园诗般的天堂里混四年，为社会正义抗议，和朋友一起放松、聚会，还能拿到报酬。这已经是 UBI 了。

Okay. But then education is just a big LARP. Anyone who wants can educate themselves right now with AI and open-source courseware. They just want the stamp, they want to go to college, hang out in an idyllic paradise for four years, protest for social justice, chill with friends, party, and get paid for it. It's already UBI.

你已经说了一件非常重要的事。大学不是关于实际的学习行为。而是关于和一群有正确人脉的人待在一个房间里。

You said a very important thing already. College is not about the actual act of learning. It's about being in a room with a certain set of people who have the right networks.

正确。这是社交、保姆、文凭主义。只有一点点教育掺杂在少数 STEM 学位中，然后你去追逐一堆工作。

Correct. It's socializing, babysitting, credentialism. There's just a tiny bit of education sprinkled in for a few STEM degrees, and then you go after a bunch of jobs.

使其合法化。是的。然后所有社会科学和政治学。

Legitimize it. Yeah. And then all the social sciences and political science.

我们不想要更多的工作。我们想要更好的工作。

We don't want more jobs. We want better jobs.

没错。为什么我们不能，而不是去格子间，像投资银行家那样绞尽脑汁？

That's true. Why can't we just, instead of going to a cubicle and breaking your brain as an investment banker or something?

我认为我们最终都会成为水管工、电工和理发师。所有白领工作都会崩溃成虚无。它会崩溃成虚无，因为它正在吞噬越来越多。你没看到这是怎么发展的吗？一切都会消失。

I think we're all going to end up being plumbers, electricians, and cutting hair. All the white-collar stuff is going to collapse into nothing. It's going to collapse into nothing because it's just eating more and more. Don't you see how this is going? Everything's going to be gone.

有一个好的版本，人们不必做糟糕的体力工作。他们实际上可以处理和引导机器人和 AI。所以如果我们没有 ASI，如果 AI 让你达到专家水平但没有超越，仍然需要人类的指导、品味和创造力。人类仍然是环境中具有动力的东西。所以人们变得像宝可梦训练师、机器人操作员、AI 操作员。这就是我们今天看到的。今天使用 AI 的人比以往任何时候都更有效率。他们没有被取代。你被 AI 取代的唯一原因是你拒绝使用 AI。如果你使用 AI，你的工作比以往任何时候都多。

There's a good version of this where people don't have to have shitty manual jobs. They can actually be handling and guiding the robots and AIs. So if we don't get ASI, if AI gets you to expert level but doesn't go beyond, still needs human guidance, taste, and creativity. Humans are still the motivated things in the environment. So people become like Pokémon trainers, robot handlers, AI handlers. That's what we're seeing today. The people who are using AI today are more productive than ever. They are not being displaced. The only reason you're being displaced by AI is because you refuse to use AI. If you're using AI, you have more work than ever.

你无法取代的一件事是人类的欲望。即使机器人有自己的欲望，它们也无法取代人类的欲望。所以只要人类有欲望，并且有 AI 和机器人可以帮助实现它们，你就总是需要人类参与其中。

The one thing you can't replace is human desire. Even if robots have their own desires, they cannot replace human desire. So as long as humans have desires and there are AIs and robots that can help fulfill them, you always need humans in the loop.

是的。当我发现我在和 AI 说话，或者读 AI 的作品，或者看 AI 创作的艺术时，我完全没兴趣。

Yeah. The moment I find out I'm talking to an AI or reading the work of an AI or looking at art created by an AI, I'm completely uninterested.

等一下。那只是因为现在它还很烂。我保证它会变得无法区分，而且会非常非常好。

Wait a second. That's just because it's shitty right now. I guarantee you it will be indistinguishable and it will be very, very good.

我不同意。我认为即使它无法区分，一旦我知道是 AI 做的……

I don't agree. I think even if it is indistinguishable, the moment I know it's done by AI...

比如，但那不就是中文房间吗？

Like for example, but that's the Chinese box at that point.

对对对。但说实话，我第一次看到 GPT-3 或者 API 能用的那些东西时，我的反应就是：‘这全是把戏，不过是障眼法。’但现在我不这么觉得了。我每天、每小时都能从中获得很多价值。

Yeah. Yeah. Yeah. But which is like actually when I first saw GPT3 or whatever it was whatever was available from the API like that was my reaction. I was like, 'This is a bunch of Like, this is a trick, whatever.' But like, that's not how I feel now. Like, I get a lot of value day to day, like hour to hour.

你和 AI 有关系吗？你会和它一起玩吗？和它聊天？它是朋友吗？

Do you have a relationship with the AI? Do you hang out with it? You talk to it? Is it a friend?

快差不多了。

It's getting there.

哇。

Whoa.

它不是我女朋友，但——

It's not my girlfriend, but

我老婆会看这期节目的，老兄。

my wife is going to watch this, buddy.

这就像你让菲律宾呼叫中心的人当你女朋友，然后你只是在填补……有什么区别？AI 和菲律宾呼叫中心里跟你聊天的人有什么区别？

It's You just have someone in the Filipino call center being your girlfriend, then you're just like filling in for the What's the difference? What's the difference between an AI and someone in the Filipino call center controlling a good chatting with you?

它是我哥们儿。好了，别这么说我朋友。

It's a buddy of mine. Okay. Don't talk to my friend. Don't talk about my friend like this.

没什么。

Nothing's going on.

我们最后会聊到我的 AI 妄想症。

We're going to end with my AI psychosis.

我只是觉得它确实非常有用，显然。

I just think that it is actually very useful, obviously.

我有个朋友是某家 AI 公司的房东，那家公司做的是你可以聊天的 AI 角色，比如 AI 女友。他说他们得雇额外的保安，因为有人半夜跑来问：‘她在哪？她在哪个盒子里？我要带她回家，我要救她。’

I have a friend who's a landlord for one of these AI companies that does the AI people that you talk to, right? Like your AI girlfriend companies. And he said that to hire extra security because these guys would show up in the middle of the night saying, 'Where is she? Which box is she in? I want to take her home. I want to save her.'

Gary 第一次陷入悲观时刻了。哦不，听起来不妙。

Gary's having his first black pill moment. Oh no, that doesn't sound good.

这期播客让我彻底悲观了。

This podcast has blackpilled me.

顺便说一句，这整期播客都是 AI 生成的。根本没人在这儿。我们什么都不承认。

By the way, this entire podcast was AI generated. No one was actually here. We deny everything.