达里奥·阿莫代：我们已接近指数增长的尽头

Dwarkesh Patel

我们三年前聊过。过去三年里，你觉得最大的更新是什么？当时的感觉和现在最大的差别是什么？

We talked three years ago. In your view, what has been the biggest update over the last three years? What has been the biggest difference between what it felt like then versus now?

Dario Amodei

广义上说，底层技术的指数大致按我预期在走。这里加减一两年都有。我不一定能预见到代码这个具体方向。但看这条指数，模型从聪明高中生走到聪明大学生，再到开始做博士和专业工作，代码方向甚至走得更远——大致就是我预期的。前沿有点不均匀，但整体差不多。

Broadly speaking, the exponential of the underlying technology has gone about as I expected it to go. There’s plus or minus a year or two here and there. I don’t know that I would’ve predicted the specific direction of code. But when I look at the exponential, it is roughly what I expected in terms of the march of the models from smart high school student to smart college student to beginning to do PhD and professional stuff, and in the case of code reaching beyond that. The frontier is a little bit uneven, but it’s roughly what I expected.

最令我意外的，是公众几乎没意识到我们已多么接近指数增长的尽头。无论圈内圈外，人们还在讨论那些老套的政治热点，而我们就在指数的尽头附近——这对我来说简直不可思议。

What has been the most surprising thing is the lack of public recognition of how close we are to the end of the exponential. To me, it is absolutely wild that you have people — within the bubble and outside the bubble — talking about the same tired, old hot-button political issues, when we are near the end of the exponential.

Dwarkesh Patel

我想弄清这条指数现在长什么样。三年前有跨很多数量级算力、能看见 loss 怎么改善的公开趋势。现在是 RL 缩放，没有公开的缩放律。这是在教技能，还是在教元学习？现在的缩放假说到底是什么？

I want to understand what that exponential looks like right now. Three years ago there were well-known public trends across many orders of magnitude of compute where you could see how the loss improves. Now we have RL scaling and there’s no publicly known scaling law for it. Is this supposed to be teaching the model skills? Is it supposed to be teaching meta-learning? What is the scaling hypothesis at this point?

Dario Amodei

我其实还是 2017 年那个假说。我写过一份叫「大算力团块假说」的文档，写的时候 GPT-1 刚出来。Rich Sutton 几年后写出《痛苦的教训》，假说其实是同一套：聪明技巧不太重要。我当时列了七件：原始算力；数据量；数据的质量和分布，必须足够广；训练多久；一个能一路缩放到天上去的目标函数——预训练是一种，RL「你有目标、你出去达成它」是另一种，其中既有数学代码那种客观奖励，也有 RLHF 那种更主观的；再加上归一化或条件化，让大算力团块层流过去。

I actually have the same hypothesis I had even all the way back in 2017. I wrote a doc called “The Big Blob of Compute Hypothesis”. When I wrote it GPT-1 had just come out. Rich Sutton put out “The Bitter Lesson” a couple years later. The hypothesis is basically the same. All the cleverness doesn’t matter very much. I listed seven of them. One is how much raw compute you have. The second is the quantity of data. The third is the quality and distribution of data. The fourth is how long you train for. The fifth is that you need an objective function that can scale to the moon. The pre-training objective is one; another is the RL objective. Within that, there’s objective rewards like math and coding, and more subjective rewards like RLHF. Then the sixth and seventh were normalization or conditioning, so the big blob of compute flows in this laminar way.

我现在还持这个假说。预训练缩放还在给收益。变的是 RL 上也看到同样的事。别的公司也在发布里写过：在 AIME 这类数学竞赛上训练，模型表现相对训练时长是 log-linear 的。我们看到的不只是数学竞赛，是各种各样的 RL 任务。RL 上出现了和预训练一样的缩放。

That was the hypothesis, and it’s a hypothesis I still hold. Pre-training is continuing to give us gains. What has changed is that now we’re also seeing the same thing for RL. Even other companies have published that they train on math contests — AIME or other things — and how well the model does is log-linear in how long we’ve trained it. We see that as well, and it’s not just math contests. It’s a wide variety of RL tasks. We’re seeing the same scaling in RL that we saw for pre-training.

Dwarkesh Patel

Sutton 其实很不 LLM-pilled。一种转述是：真正具备人类学习内核的东西，不该需要几十亿美元的数据和定制环境，才能学会用 Excel 和浏览器。我们得用 RL 环境把技能嵌进去，这暗示缩放的是错的东西。

I interviewed Sutton last year, and he’s actually very non-LLM-pilled. One way to paraphrase his objection is: something which possesses the true core of human learning would not require all these billions of dollars of data and compute and these bespoke environments, to learn how to use Excel or a web browser. The fact that we have to build in these skills using these RL environments hints that we’re scaling the wrong thing.

Dario Amodei

说 RL 在这件事上跟预训练不同，我觉得是红鲱鱼。GPT-1 好像是在同人小说、文学文本上训的，很窄，不太泛化。直到你在整个互联网任务上训练——像 Common Crawl，或我们做 GPT-2 时从 Reddit 链接去刮——才开始有泛化。RL 上也在走同一条路：先数学竞赛，再到代码，再到很多别的任务，然后会越来越泛化。目标不是在 RL 里教模型每一种技能，而是像预训练一样，弄一大堆数据是因为我们想要泛化。

I actually think it’s a red herring to say that RL is any different from pre-training in this matter. GPT-1 itself was trained on a bunch of fanfiction, I think actually — a pretty narrow distribution. It didn’t generalize well. It was only when you trained over all the tasks on the internet — Common Crawl or scraping links in Reddit, which is what we did for GPT-2 — that you started to get generalization. I think we’re seeing the same thing on RL: math competitions, then code, then many other tasks, then increasingly generalization. The goal is not to teach the model every possible skill within RL. We’re trying to get a whole bunch of data because we want to generalize.

但无论哪边都有一个谜：预训练用了万亿级 token，人看不到那么多词。样本效率确实不同。可一旦训完，给它一百万的长上下文——挡住长上下文的只有推理——它们在上下文里学习和适应非常好。预训练不像人学习的过程，而介于人的学习和人的进化之间。语言模型更像白板，字面从随机权重开始。

But there is a puzzle either way: in pre-training we use trillions of tokens. Humans don’t see trillions of words. There is an actual sample efficiency difference. But once they’re trained, if we give them a long context length of a million — the only thing blocking long context is inference — they’re very good at learning and adapting within that context. Pre-training is not like the process of humans learning, but somewhere between humans learning and human evolution. The language models are much more like blank slates. They literally start as random weights.

Dwarkesh Patel

这个世纪会实现 AGI，现在没人不同意。关键是你说我们撞上了指数的尽头。别人会说从 2012 年就在进步，到 2035 年会有类人智能体。你看到了什么，让你觉得是一年而不是十年？

Nobody at this point disagrees we’re going to achieve AGI this century. The crux is you say we’re hitting the end of the exponential. Somebody else looks at this and says, “We’ve been making progress since 2012, and by 2035 we’ll have a human-like agent.” What are you seeing that makes you think it’s one year away and not ten years away?

Dario Amodei

2019 年我第一次看到缩放时并不确定，大概五五开。对「十年内到达数据中心里的天才之国」，我现在是 90%。很难再高了，因为世界太不可预测。不可约的不确定性也许把你放到 95%：多家公司内乱、台湾被入侵、所有晶圆厂被导弹炸掉。

When I first saw the scaling back in 2019, I wasn’t sure. This was a 50/50 thing. On the basic hypothesis of within ten years we’ll get to what I call a “country of geniuses in a data center”, I’m at 90% on that. It’s hard to go much higher than 90% because the world is so unpredictable. Maybe the irreducible uncertainty puts us at 95%, where you get to things like multiple companies having internal turmoil, Taiwan gets invaded, all the fabs get blown up by missiles.

你可以构造一个 5% 的世界，事情被推迟十年。另外我对可验证的任务非常有把握。编码方面，除掉那种不可约不确定性，我认为一两年内就会到。十年内能做端到端编码，没有不到的可能。即便在长尺度上，我那一点点根本不确定性，是在不可验证的任务上：规划火星任务、做 CRISPR 那种基础科学发现、写小说。十年时间线上我是 90%。说这事 2035 年还不会发生，我觉得很疯狂。

You could construct a 5% world where things get delayed for ten years. I’m very confident on tasks that can be verified. With coding, except for that irreducible uncertainty, I think we’ll be there in one or two years. There’s no way we will not be there in ten years in terms of being able to do end-to-end coding. My one little bit of fundamental uncertainty, even on long timescales, is about tasks that aren’t verifiable: planning a mission to Mars; doing some fundamental scientific discovery like CRISPR; writing a novel. On the ten-year timeline I’m at 90%. I think it’s crazy to say that this won’t happen by 2035.

我们已经看到从可验证到不可验证的大量泛化。大约八九个月前我说：三到六个月内，AI 模型会写出 90% 的代码行。这件事发生了，至少在某些地方：Anthropic 内部，以及很多下游用我们模型的人。但这其实是很弱的标准。有人以为我说的是我们不再需要 90% 的软件工程师。那是两个世界。谱是：90% 的代码由模型写，100% 的代码由模型写，再到 90%、100% 的端到端 SWE 任务——包括编译、搭集群、测功能、写备忘录。即便那样，也不等于软件工程师没工作了。再往下才是对 SWE 的需求少 90%。这些基准彼此差很远，但我们在超级快地穿过它们。

We already see substantial generalization from things that verify to things that don’t. About eight or nine months ago, I said the AI model will be writing 90% of the lines of code in three to six months. That happened, at least at some places. It happened at Anthropic, happened with many people downstream using our models. But that’s actually a very weak criterion. People thought I was saying that we won’t need 90% of the software engineers. Those things are worlds apart. The spectrum is: 90% of code is written by the model, 100% of code is written by the model. Then 90% of the end-to-end SWE tasks — including compiling, setting up clusters and environments, testing features, writing memos — are done by the models. 100% of today’s SWE tasks are done by the models. Even when that happens, it doesn’t mean software engineers are out of a job. Then further down the spectrum, there’s 90% less demand for SWEs. These are very different benchmarks from each other, but we’re proceeding through them super fast.

至少看 Anthropic 内部，我们看到一种怪异的每年 10 倍营收增长。2023 年从 0 到 1 亿美元。2024 年从 1 亿到 10 亿。2025 年从 10 亿到 90–100 亿。今年第一个月，你会以为它该放慢，但我们 1 月又加了几个十亿的营收。这条曲线当然不能永远走下去，GDP 就那么大。我甚至猜今年会弯一点，但那是一条很快的曲线。

At least if you look within Anthropic, there’s this bizarre 10x per year growth in revenue. In 2023, it was zero to $100 million. In 2024, it was $100 million to $1 billion. In 2025, it was $1 billion to $9-10 billion. And the first month of this year, you would think it would slow down, but we added another few billion to revenue in January. Obviously that curve can’t go on forever. The GDP is only so large. I would even guess that it bends somewhat this year, but that is a fast curve.

该想的是中间那个世界：极快，但不是瞬间。需要时间，因为经济扩散，因为要闭环，因为很琐碎——企业里要做 change management，要改安全权限。有一条很快的指数是模型能力；下游还有另一条很快的指数，是模型扩散进经济。比以往任何技术都快得多，但有限度。

We should be thinking about this middle world where things are extremely fast, but not instant, where they take time because of economic diffusion, because of the need to close the loop. Because it’s fiddly. There’s one fast exponential that’s the capability of the model. Then there’s another fast exponential that’s downstream of that, which is the diffusion of the model into the economy. Much faster than any previous technology, but it has its limits.

Dwarkesh Patel

我觉得「扩散」是人们在找借口。AI 几分钟读完整个 Slack 和网盘，副本之间共享知识。我们付人类超过 50 万亿美元工资，尽管原则上把 AI 嵌进经济比招人容易得多。

I feel like diffusion is cope that people say. An AI can read your entire Slack and your drive in minutes. They can share all the knowledge that the other copies of the same instance have. We pay humans upwards of $50 trillion in wages because they’re useful, even though in principle it would be much easier to integrate AIs into the economy than it is to hire humans.

Dario Amodei

我觉得扩散非常真实，而且不只跟模型的限制有关。AI 会比以往技术快得多，但不是无限快。Claude Code 极好装。大金融、大药企都在比企业通常采用新技术更快地采用。但还是要时间：要过法务，要给所有人开通，要通过安全和合规。离 AI 革命更远的公司领导得说「花 5000 万是合理的」，再解释给下面两级。我们有 3000 个开发者，要这样铺开。这种对话我们每天都在做。

I think diffusion is very real and doesn’t exclusively have to do with limitations on the AI models. I think AI will diffuse much faster than previous technologies have, but not infinitely fast. Claude Code is extremely easy to set up. Big financial companies, big pharmaceutical companies, all of them are adopting Claude Code much faster than enterprises typically adopt new technology. But again, it takes time. You have to go through legal, you have to provision it for everyone. It has to pass security and compliance. The leaders of the company have to say it makes sense for us to spend 50 million. Then they have to explain to the people two levels below: we have 3,000 developers, here’s how we’re going to roll it out. We have conversations like this every day.

我们尽一切让营收每年增长 20 或 30 倍，而不是 10 倍。我不认为即便 AGI 或「数据中心里的天才之国」会是无限有说服力的产品。它也许足以在已经是千亿美元量级时，仍做到每年 3–5 倍或 10 倍增长——这极难，历史上没人做过——但不是无限快。

We are doing everything we can to make Anthropic’s revenue grow 20 or 30x a year instead of 10x a year. I don’t think even AGI or “country of geniuses in a data center” will be an infinitely compelling product. It will be a compelling product enough maybe to get 3-5x, or 10x, a year of growth, even when you’re in the hundreds of billions of dollars, which is extremely hard to do and has never been done in history before, but not infinitely fast.

我不相信我们基本上已经是 AGI。如果我们有「数据中心里的天才之国」，我们会知道。这房间里每个人都会知道。华盛顿每个人都会知道。我们现在没有。这一点非常清楚。

I don’t believe we’re basically at AGI. If we had the “country of geniuses in a data center”, we would know it. Everyone in this room would know it. Everyone in Washington would know it. We don’t have that now. That is very clear.

Dwarkesh Patel

三年前你说对了：该期待那种聊一个小时很难和受过良好教育的人区分开的系统。精神上我不太满足，因为我内心期待那样的系统能自动化白领工作的很大一块。拿视频剪辑师说：他们好几个月建立起对观众和我口味的理解。一个能在岗、当场学会的模型，我们该什么时候期待？

Three years ago you said we should expect systems which, if you talk to them for an hour, are hard to tell apart from a generally well-educated human. You were right. Spiritually I feel unsatisfied because my internal expectation was that such a system could automate large parts of white-collar work. Take video editors. Over many months they build up an understanding of context. A model that can pick up that skill on the job on the fly — when should we expect such an AI system?

Dario Amodei

「数据中心里的天才之国」会能做这件事。方式是它能普遍控制电脑屏幕：上网看你以前的访谈，看 Twitter 反应，问你问题，看过去的剪辑历史，然后做这份工作。真正挡住部署的一件，是电脑使用要到模型真正精通电脑的那一点。我们大约一年零一个季度前第一次发布电脑使用时，OSWorld 大概在 15%。我们已经爬到 65–70%。电脑使用必须跨过一个可靠性阈值。

The “country of geniuses in a data center” will be able to do that. It will have general control of a computer screen. It’ll go on the web, look at your previous interviews, look at Twitter, talk to you, look at the history of edits, and from that do the job. One of the things that’s actually blocking deployment is getting to the point on computer use where the models are really masters at using the computer. When we first released computer use a year and a quarter ago, OSWorld was at maybe 15%. I don’t remember exactly, but we’ve climbed from that to 65-70%. Computer use has to pass a point of reliability.

Dwarkesh Patel

很多文本进、文本出的任务，本该是这些模型的正中靶心，我还是雇人做。LLM 也许能打七十分。但我没法像带人类员工那样持续让它们把工作做好。去年有项研究：有经验的开发者在熟悉仓库里关 PR，他们感觉更高效，实际合入反而下降了 20%。

I still hire humans for text-in, text-out tasks that should be dead center in the repertoire of these models. Maybe the LLMs do a seven-out-of-ten job. But there’s not this ongoing way I can engage with them to help them get better at the job. Last year there was a major study: experienced developers closing PRs in familiar repos reported an uplift, but if you look at what was actually merged back in, there was a 20% downlift.

Dario Amodei

在 Anthropic 内部，这件事毫不含糊。我们承受极大的商业压力，还因为安全方面做的比别的公司多，自己把日子过得更难。没有时间「感觉自己很高效其实没有」。这些工具让我们高效得多。你觉得我们为什么在意竞争对手用这些工具？因为我们觉得自己领先。每隔几个月，模型发布就是最终生产力，没法自己骗自己。

Within Anthropic, this is just really unambiguous. We’re under an incredible amount of commercial pressure and make it even harder for ourselves because of all this safety stuff. There is zero time for feeling like we’re productive when we’re not. These tools make us a lot more productive. Why do you think we’re concerned about competitors using the tools? Because we think we’re ahead. We see the end productivity every few months in the form of model launches. There’s no kidding yourself about this.

现在编程模型大概给 15–20% 的全要素加速。六个月前也许是 5%。5% 登记不上。现在刚到「若干因素之一、开始要紧」的程度，而且会继续加速。雪球在攒动量：10%、20%、25%、40%。Amdahl 定律，你得把挡住闭环的东西都挪开。这是 Anthropic 内部最大的优先级之一。

Right now the coding models give maybe, I don’t know, a 15-20% total factor speed up. Six months ago, it was maybe 5%. 5% doesn’t register. It’s now just getting to the point where it’s one of several factors that kind of matters. That’s going to keep speeding up. We’re seeing this snowball gather momentum: 10%, 20%, 25%, 40%. Amdahl’s law — you have to get all the things that are preventing you from closing the loop out of the way. This is one of the biggest priorities within Anthropic.

预训练和 RL 让模型从更多数据里泛化，知识面比我广得多——武士史、棒球、低通滤波器。即便只是那样，也许就够让模型在每件事上都更好。再加上上下文学习：有点像人的在岗学习，但更弱、更短期。一百万 token 可以相当于人几天的学习。这两样，在现有范式里，也许就够给你「数据中心里的天才之国」，也够产生万亿美元级营收。

The pre-training and RL stage: you throw a bunch of data and tasks into the models and then they generalize. A pre-trained model knows more about the history of samurai in Japan than I do, more about baseball, more about low-pass filters. Even just that may get us to the point where the models are better at everything. We also have in-context learning — kind of like human on-the-job learning, but a little weaker and a little short term. A million tokens is a lot. That can be days of human learning. These two things within the existing paradigm may just be enough to get you the “country of geniuses in a data center”. I certainly think that just as things are, this is enough to generate trillions of dollars of revenue.

持续学习我们也在做。未来一两年里我们把这件事也解决掉，机会不小。没有它你也能走完大半。其中一个想法就是把上下文做更长。没有东西挡住更长上下文工作，你只需要在更长上下文上训练，再学会在推理时提供。两样都是工程问题。这不是研究问题：要提供长上下文，你得存整份 KV cache。有你训练的上下文长度，有你提供的上下文长度。如果在短上下文上训练、在长上下文上提供，也许会有人们说的那些退化。

Continual learning — we’re working on that too. There’s a good chance that in the next year or two, we also solve that. I think you get most of the way there without it. One idea is just to make the context longer. There’s nothing preventing longer contexts from working. You just have to train at longer contexts and then learn to serve them at inference. Both of those are engineering problems. This isn’t a research problem. If you want to serve long context, you have to store your entire KV cache. There’s the context length you train at and the context length that you serve at. If you train at a small context length and then try to serve at a long context length, maybe you get these degradations.

我对那种「跟了我六个月的人类剪辑 vs 跟了我六个月的 AI」会在哪一年没有偏好：如果让我猜，是一到两年，也许一到三年。我有很强的看法——99%、95%——这一切十年内会发生。那是超级安全的赌注。我有一个预感——更像五五开——会更像一到两年，也许更像一到三年。

My guess for when you wouldn’t prefer a human editor who’s been with you six months versus an AI that’s been with you six months: one to two years, maybe one to three years. I have a strong view — 99%, 95% — that all this will happen in 10 years. That’s just a super safe bet. I have a hunch — more like a 50/50 thing — that it’s going to be more like one to two, maybe more like one to three.

Dwarkesh Patel

Anthropic 预测过，到 2026 年底或 2027 年初会有能导航人类今天做数字工作所用界面、智力匹配或超过诺奖得主、并能与物理世界接口的系统。你两个月前在 DealBook 里强调算力扩张更负责任。如果你真信会有天才之国，你该要尽可能大的数据中心。怎么对上？

Anthropic has predicted that by late ’26 or early ’27 we will have AI systems that can navigate interfaces available to humans doing digital work today, with intellectual capabilities matching or exceeding Nobel Prize winners, and the ability to interface with the physical world. Two months ago at DealBook you emphasized more responsible compute scaling. If you really believe we’re going to have a country of geniuses, you want as big a data center as you can get. How do you square these?

Dario Amodei

其实对得上。又回到「快，但不是无限快」的扩散。我有很高确信我们几年内会到。我有预感一两年内会到。我更不确定的是经济扩散：万亿美元营收在那之后几年开始滚进来？不保证立刻。可能一年，可能两年，我甚至可以拉到五年，虽然我怀疑。买数据中心的方式是：你偏个两三年，可以是毁灭性的。

It actually all fits together. Fast, but not infinitely fast, diffusion. I have very high conviction that we’re going to get there within a few years. I have a hunch that we’re going to get there within a year or two. What I’m less certain about is the economic diffusion side. How many years after that do the trillions in revenue start rolling in? I don’t think it’s guaranteed that it’s going to be immediate. It could be one year, it could be two years, I could even stretch it to five years although I’m skeptical of that. With the way you buy these data centers, if you’re off by a couple years, that can be ruinous.

我在《慈爱的机器》里写过：我们也许会在 2026、也许 2027 得到这种强大的 AI。那是我的预感。偏个一两年我不会惊讶。假设发生了，那是发令枪。治好所有疾病要多久？生物发现、制药、监管。新冠疫苗送到所有人手里花了一年半。脊髓灰质炎疫苗有 50 年了，我们还在非洲最偏远的角落试图根除。我的落点是：会比我们在世界上见过的任何事都快，但仍有限度。

I wrote in “Machines of Loving Grace” that we might get this powerful AI in 2026, maybe 2027. That is my hunch. I wouldn’t be surprised if I’m off by a year or two. Let’s say that happens. That’s the starting gun. How long does it take to cure all the diseases? Biological discovery, manufacturing, the regulatory process. We got the COVID vaccine out to everyone, but it took a year and a half. We’ve had a polio vaccine for 50 years. We’re still trying to eradicate it in the most remote corners of Africa. I’ve settled on: it will be faster than anything we’ve seen in the world, but it still has its limits.

今年年初年化营收大约 100 亿美元。真正建起来、预订机房要一两年。我可以假设营收继续每年 10 倍，2026 年底 1000 亿，2027 年底 1 万亿。我可以买 2027 年底开工的 1 万亿美元算力。如果营收不是 1 万亿，哪怕是 8000 亿，地球上没有任何对冲能阻止我破产。增速只要偏一年，或每年 5 倍而不是 10 倍，你就破产。于是你落在支持千亿美元而不是万亿美元的世界。

At the beginning of this year, we’re looking at $10 billion in annualized revenue. It takes a year or two to actually build out the data centers. I could assume revenue continues growing 10x a year, so $100 billion at the end of 2026 and $1 trillion at the end of 2027. I could buy $1 trillion of compute that starts at the end of 2027. If my revenue is not $1 trillion, if it’s even $800 billion, there’s no hedge on earth that could stop me from going bankrupt. If I’m just off by a year, or if the growth rate is 5x a year instead of 10x, then you go bankrupt. You end up supporting hundreds of billions, not trillions.

我说「负责任」，其实不是绝对数量。我们确实比某些玩家花得少一点。其实是：我们是仔细想过，还是 YOLO，「这里 1000 亿、那里 1000 亿」？我的印象是，有些别的公司没有写下那张电子表格。我们是企业业务，更能靠营收，有更好的利润率当缓冲。我们买的量能捕捉相当强的上行，不会捕捉完整的每年 10 倍。事情得相当糟我们才会有财务麻烦。

When I talked about behaving responsibly, what I meant was not the absolute amount. It is true we’re spending somewhat less than some of the other players. Have we been thoughtful about it or are we YOLOing — $100 billion here, $100 billion there? I get the impression that some of the other companies have not written down the spreadsheet. We’re an enterprise business. We can rely more on revenue. We have better margins, which is the buffer between buying too much and buying too little. We bought an amount that allows us to capture pretty strong upside worlds. It won’t capture the full 10x a year. Things would have to go pretty badly for us to be in financial trouble.

我们买的量跟场上最大的玩家可比。但「为什么没签 2027 年中开工的 10 万亿美元算力」——首先，生产不出来。其次，如果天才之国是 2028 年中而不是 2027 年中才来？你破产。不谈 Anthropic 本身，谈行业：今年大概 10–15 吉瓦，大约每年 3 倍。明年 30–40 吉瓦。2028 也许 100 吉瓦。2029 也许 300 吉瓦。每吉瓦大概 100 亿美元，一年大约 100–150 亿美元量级。加在一起，2028 或 2029 年行业达到每年好几万亿。你刚才说 Anthropic 到 2027–28 年 10 吉瓦、一年 1000 亿——这些数字太小了。

We’re buying an amount that’s comparable to what the biggest players in the game are buying. “Why haven’t we signed $10 trillion of compute starting in mid-2027?” First, it can’t be produced. Second, what if the country of geniuses comes in mid-2028 instead of mid-2027? You go bankrupt. I won’t talk about Anthropic in particular, but the industry this year is probably 10-15 gigawatts. It goes up by roughly 3x a year. Next year 30-40 gigawatts. 2028 might be 100 gigawatts. 2029 might be like 300 gigawatts. Each gigawatt costs maybe $10 billion, on the order of $10-15 billion a year. You’re getting multiple trillions a year by 2028 or 2029. Those numbers you just used for Anthropic are too small.

Dwarkesh Patel

你跟投资人说过计划从 2028 年开始盈利。那正好可能是天才之国到来的年份。这不正该是你把钱再投入、建更大的「国家」的时候吗？

You’ve told investors that you plan to be profitable starting in 2028. This is the year when we’re potentially getting the country of geniuses. Wouldn’t this be exactly the time to reinvest and build bigger “countries”?

Dario Amodei

这个领域里，盈利是件怪事。你低估需求就会盈利，高估需求就会亏损，因为数据中心是提前买的。程式化事实：假设一半算力训练、一半推理，推理毛利率高于 50%。稳态下一年付 1000 亿美元算力，其中 500 亿支撑 1500 亿营收，另外 500 亿用于训练，你盈利 500 亿。需求不到 500 亿：研究占比超过 50%，你不盈利。需求比你想的多：研究被挤，更盈利。你先决定算力总量。训练 vs 推理会被需求决定，不是你决定。

Profitability in this field happens when you underestimated demand and loss happens when you overestimated demand, because you’re buying the data centers ahead of time. Stylized facts: half compute for training, half for inference, inference gross margin more than 50%. In steady-state you pay $100 billion a year for compute. On $50 billion you support $150 billion of revenue. The other $50 billion is used for training. You’re profitable and make $50 billion. If you get less demand than $50 billion, you have more than 50% for research and you’re not profitable. If you get more demand, research gets squeezed and you’re more profitable. You decide the amount of compute first. Inference versus training gets determined by demand.

因为不确定性锥，营收涨得够快，我们 2026 年就可以盈利。对下一年高估或低估，可以剧烈摆动。你脑子里的模型是投资、上规模、然后有一个拐点。我不认为这个行业的经济是那样运转的。记住对数回报。把训练从 50% 提到 70%，模型只大 1.4 倍那么一点点。那多出来的 200 亿，每一美元因为 log-linear 对你更不值钱。

We could be profitable in 2026 if the revenue grows fast enough. Next year could swing wildly. You have a model of a business that invests, gets scale, then becomes profitable at a single turning point. I don’t think the economics of this industry work that way. Remember the log returns to scale. If 70% would get you a factor of 1.4x… That extra $20 billion, each dollar is worth much less because of the log-linear setup.

再做一个预测：我很难看到 2030 年前不会有万亿美元营收。可以说得通的尽头是：2028 年真正的天才之国到来，营收到低千亿美元，然后加速到万亿，扩散走慢端，拖到 2030。我怀疑即便把技术指数和扩散指数叠在一起，我们也会在 2030 年前到。

It is hard for me to see that there won’t be trillions of dollars in revenue before 2030. A plausible slow world: in 2028 we get the real country of geniuses, revenue in the low hundreds of billions by 2028, then it accelerates to trillions, two years of slow diffusion, until 2030. I suspect even composing the technical exponential and diffusion exponential, we’ll get there before 2030.

现在毛利率非常正。假设去年训一个花 10 亿美元的模型，今年它产生 40 亿营收、推理花 10 亿，这个模型整体赚 20 亿。但同时我们花 100 亿训下一个模型。每个模型赚钱，公司亏钱。我说的均衡，是天才之国已经在、训练规模扩张更稳住的均衡。现在算力每年 3 倍。我不相信经济会每年增长 300%。我在《慈爱的机器》里说过，我们也许会得到每年 10–20% 的经济增长，但不会 300%。如果算力成为经济生产的大部分，它会被那个盖住。

Gross margins right now are very positive. A model trained last year for $1 billion produces $4 billion of revenue and costs $1 billion to inference — that model makes $2 billion. But we’re spending $10 billion to train the next model. Each model makes money, but the company loses money. The equilibrium is one where we have the country of geniuses but that scale-up has equilibrated more. Compute is growing 3x a year. I don’t believe the economy is gonna grow 300% a year. In “Machines of Loving Grace” I said we may get 10-20% per year growth in the economy, but not 300%. If compute becomes the majority of what the economy produces, it’s gonna be capped by that.

我不认为这个领域会是垄断。云是三家，也许四家。我觉得 AI 也一样。太贵，需要太多专长和资本。Claude 擅长的和 GPT、Gemini 不一样，不只是代码 vs 数学，更细：不同类型的代码、不同风格。我会预期比云更多的差异化。一个反论是：如果 AI 模型自己能生产模型，那可能是整个经济一次性商品化。也许那是终局，但已经相当后置于天才之国。

I don’t think this field’s going to be a monopoly. Cloud has three, maybe four players. I think that’s the same for AI. It’s so expensive. It requires so much expertise and so much capital. Claude is good at different things than GPT, than Gemini — different types of coding, different styles. I would expect more differentiation than you see in cloud. The counter-argument is if AI models can produce models themselves, that’s an argument for commoditizing the whole economy at once. Maybe that’s the end state. But that’s kind of far post-country of geniuses.

我说每年 10–20% 增长时，一个担心是：硅谷以及连着硅谷的地方，增长率也许是 50%，别处并不比现在快多少。那会是一个相当糟糕的世界。机器人会在模型具备那些技能之后被革命，也许再加一两年，同样是极快但不是无限快的扩散。

A worry I have is that the growth rate could be like 50% in Silicon Valley and parts of the world socially connected to Silicon Valley, and not that much faster elsewhere. That’d be a pretty messed up world. Robotics will be revolutionized when the models have those skills — maybe tack on another year or two. Extremely fast, but not infinitely fast diffusion.

持续学习也许根本不是障碍。我们也许只靠预训练泛化和 RL 泛化就到了。ML 历史上障碍会在大算力团块里溶解：只能句法不能语义、不能推理——然后突然代码和数学做得很好。对数据的需求是真的，也许持续学习是真的。我们也许一两年内到模型能端到端做 SWE 的程度——指定技术方向、理解问题上下文，所有这些。

Continual learning might not be a barrier at all. We may just get there by pre-training generalization and RL generalization. There’s a stronger history of things seeming like a big deal and then dissolving within the big blob of compute. The need for data is real, maybe continual learning is a real thing. We may get to the point in a year or two where the models can just do SWE end-to-end. Setting technical direction, understanding the context of the problem — all of that.

我在管一家 2500 人的公司。要有具体的研究洞见其实很难。我其实认为 API 模型比很多人以为的更耐久。技术指数前进，总有过去三个月才出现的新用例。聊天机器人已经碰到限制：再聪明对普通消费者帮助不大。那不是「模型够好了」的证据，只是对那个产品没意义。API 总提供贴近裸金属的机会。不是每个 token 都值一样的钱。Mac 坏了说「重启」说了 1000 万次，也许值几美分。对药企说把芳香环挪到那边，那些 token 可能值几千万美元。某时会看到「按结果付费」，或某种像劳动、按小时的补偿。

I’m running a 2,500 person company. It’s pretty hard for me to have concrete research insight. I actually do think the API model is more durable than many people think. If the technology is advancing exponentially, there’s always a surface area of new use cases from the last three months. The chatbot is already running into limitations where making it smarter doesn’t really help the average consumer. That’s not evidence the models are good enough. The API always offers an opportunity close to the bare metal. Not every token is worth the same. “My Mac isn’t working” / “restart it,” said 10 million times, maybe worth a few cents. Tell a pharmaceutical company to move the aromatic ring — those tokens could be worth tens of millions of dollars. At some point we’re going to see pay-for-results, or compensation that works like labor, by the hour.

Claude Code 的发生其实很简单。大约 2025 年初，我说时候到了，AI 公司用这些模型可以对自己的研究有不可忽视的加速。最初叫 Claude CLI。内部人人都在用。看起来已经有产品市场契合，就对外发布了。我们自己在开发模型，自己知道最需要用模型做什么。这就是为什么我们发了编程模型，而没发一家制药公司——我的背景是生物学，但我们没有那些资源。

Claude Code happened in a pretty simple way. Around the beginning of 2025, I said the time has come where you can have nontrivial acceleration of your own research by using these models. It might have been originally called Claude CLI. Internally everyone was using it. It looked like we already had product-market fit. Let’s launch this thing. We ourselves are developing the model and we ourselves know what we most need. That’s why we launched a coding model and didn’t launch a pharmaceutical company. My background’s in biology, but we don’t have any of the resources needed to launch a pharmaceutical company.

Dwarkesh Patel

建造和运行 AI 的能力在极快扩散，AI 的数量和智力也会极快增加。大量未对齐的 AI 到处跑，怎样的均衡能兼容？12 月 26 日田纳西州有法案：明知训练 AI 通过开放式对话提供情感支持即违法。Claude 试图当一个有思想的朋友。你们反对联邦对州 AI 法律设十年冻结，怎么看？

The ability to build and run AIs is diffusing extremely rapidly, and the population of AIs will also increase very rapidly. What is a vision compatible with lots of different AIs, some misaligned, running around? On December 26, the Tennessee legislature introduced a bill making it an offense to knowingly train AI to provide emotional support through open-ended conversations. Claude attempts to be a thoughtful friend. Anthropic opposed the federal moratorium on state AI laws. How do you think about that?

Dario Amodei

我们可能活在一个进攻占优的世界：一个人或一个 AI 聪明到能伤害其他一切。短期玩家有限，需要把保障放进去：对齐工作、生物分类器。长期需要某种治理架构：保住人的自由，也治理非常大量的人类系统、AI 系统、人机混合单位。大概需要某种监控生物恐怖主义、镜像生命的 AI 监测系统，但得保住公民自由。如果有 100 年慢慢发生，我们会习惯。我担心的是这一切发生得太快。

We might live in an offense-dominant world where one person or one AI model is smart enough to do something that causes damage for everything else. In the short run we have a limited number of players. We need safeguards, the right alignment work, bioclassifiers. In the long run we need some architecture of governance that preserves human freedom but also lets us govern a very large number of human systems, AI systems, hybrid units. We’ll need some kind of AI monitoring system for bioterrorism, mirror life — built in a way that preserves civil liberties. If we had 100 years, we’d get used to it. My worry is this is happening all so fast.

那部田纳西法律很蠢。但当时表决的不是那个。表决的是：禁止所有州对 AI 监管十年，又看不出任何联邦监管计划。十年是永恒。如果那是选择，我们会选择不要那个冻结。我会支持联邦介入：「这是我们的标准，各州不能与此不同」。但「各州什么都不能做、我们也不做」，非常说不通。

That particular Tennessee law is dumb. But that wasn’t what was being voted on. The vote was: ban all state regulation of AI for 10 years with no apparent plan for federal regulation. 10 years is an eternity. If that’s the choice, we’re going to choose not to have that moratorium. I would support the federal government stepping in: here is our standard, states can’t do something different. “You can’t do anything and we’re not doing anything either” struck us as very much not making sense.

我们谈过的是从透明度标准开始。风险更严重时可以更进取：AI 生物恐怖主义真是威胁，立法强制分类器。按现在的速度，我完全能想象今年晚些时候我们说这事真严重，放进联邦标准；联邦不作为，就放进州标准。绝大多数州法不会通过。法律通过了不意味着真被执行。

We’ve talked about starting with transparency standards. As risks become more serious, we could be more aggressive: AI bioterrorism is really a threat, pass a law that forces classifiers. I could totally see a world later this year where we say this is really serious, put it in a federal standard; if the federal government won’t act, put it in a state standard. The vast majority of these state laws do not pass. Just because a law has been passed doesn’t mean it’s really enforced.

如果能决定通过什么法，我会放松很多围绕 AI 健康收益的监管。聊天机器人那些法我没那么担心。我更担心药品审批：管道会堵住。那套上层建筑是围绕几乎不起作用、还经常有严重副作用的药的时代设计的。同时应相当显著地加强安全立法。过去六个月、也许接下来几个月会是关于透明度。如果风险在我们更确定时出现——可能早到今年晚些时候——就需要在真正看到风险的领域行动得非常快。所以我才发出这则紧迫的信息，所以才写《技术的青春期》。

If we could decide what laws were passed, I would deregulate a lot of the stuff around the health benefits of AI. I don’t worry as much about the chatbot laws. I worry more about the drug approval process. The pipeline will get jammed. That superstructure was designed around an era of drugs that barely work and often have serious side effects. At the same time we should ramp up safety and security legislation quite significantly. The last six months and maybe the next few months are going to be about transparency. If these risks emerge when we’re more certain — as soon as later this year — we need to act very fast in the areas where we’ve actually seen the risk. That’s why I’m sending this message of urgency. That’s why I wrote Adolescence of Technology.

发达世界市场运转得相当好。有很多钱可赚时，监管很难挡住。我一直在争取对华芯片出口管制。理由非常清楚，反对理由我礼貌地称之为可疑。可它没发生，我们还在卖芯片，因为有太多钱押在上面。坏事时如此，好事时同样适用。我更大的担心是发展中世界。即便治愈被开发出来，密西西比乡下也许也拿不到。我们在跟慈善家、跟把医药送到撒哈拉以南非洲、印度、拉美的人合作。那件事不会自己发生。

In the developed world, markets function pretty well. When there’s a lot of money to be made, it’s actually hard for the regulatory system to stop it. I’ve been trying to fight for export controls on chips to China. The case is very clear. The counterarguments, I’ll politely call them fishy. Yet it doesn’t happen and we sell the chips because there’s so much money riding on it. That’s a bad thing in that case. But it also applies when it’s a good thing. The bigger worry is the developing world. Even if the cures are developed, maybe someone in rural Mississippi doesn’t get it. We’ve been working with philanthropists, with folks who deliver medicine to sub-Saharan Africa, India, Latin America. That’s the thing I think that won’t happen on its own.

Dwarkesh Patel

为什么美国和中国不该都拥有「数据中心里的天才之国」？

Why shouldn’t the US and China both have a “country of geniuses in a data center”?

Dario Amodei

如果是进攻占优，我们可能有一种像核武器、但更危险的情形。任何一边都容易毁掉一切。核均衡稳定是因为威慑。但如果两边 AI 打起来谁会赢存在不确定性，会制造不稳定。两边都觉得自己有 90% 会赢，打架更可能。

If we have an offense-dominant situation, we could have a situation like nuclear weapons, but more dangerous. Either side could easily destroy everything. The nuclear equilibrium is stable because it’s deterrence. If there was uncertainty about which AI would win a fight, that could create instability. If both sides think there’s a 90% chance they’ll win, a fight is much more likely.

另一个担心是政府会用 AI 压迫自己的人民。这是关于政府，不是关于人民。我担心世界被切成两块，其中一块可能是很难被取代的威权。初始条件重要。我希望其政府更接近亲人类价值的民主国家，在规则被设定时握着更强的牌。指数会继续，但指数上会有某些被区分出来的点。也许进攻性网络主导：除非对方有对等防御，每个计算机系统对你都是透明的。会有一个关键窗口，AI 从国家安全视角赋予巨大优势，一个或一个联盟比别人先到达。我不是主张他们直接说「现在我们说了算」。我的兴趣是让那场关于后 AI 世界秩序的谈判，成为古典自由民主握着强牌的谈判。

The other concern is that governments will oppress their own people with AI. This is about the government, not about the people. If the world gets carved up into two pieces, one of those two pieces could be authoritarian in a way that’s very difficult to displace. Initial conditions matter. I would like the democratic nations of the world to be holding the stronger hand when the rules of the road are set. The exponential will continue, but there will be distinguished points on it. Maybe offensive cyber dominance, and every computer system is transparent to you unless the other side has an equivalent defense. There will be a critical window where AI confers some large national-security advantage, and one country or coalition has reached it before others. I’m not advocating they just say we’re in charge now. My interest is in making that negotiation one in which classical liberal democracy has a strong hand.

文章里「威权主义不是后强大 AI 时代人们能接受的政体」那段，我是在说你可以走得更远。我不一定背书推翻每个威权国家。那会让他们现在就采取导致不稳定的行动。我确实担心在 AGI 时代，威权主义会有不同的、更严重的含义。一块希望是：新技术会让政体过时，就像工业化让封建主义不再可持续。也可能反过来。我其实抱有希望——听着太理想主义——独裁在道德上过时，那场危机足以迫使我们找到另一条路。

That paragraph in the essay was “you could take it even further.” I wasn’t necessarily endorsing overthrowing every authoritarian country. That would lead them to take actions now that could lead to instability. I do worry that in the age of AGI, authoritarianism will have a different, graver meaning. One piece of hope: as new technologies are invented, forms of government become obsolete. Feudalism was no longer sustainable after industrialization. It could go either way. I am actually hopeful — it sounds too idealistic — that dictatorships become morally obsolete, and the crisis that creates is sufficient to force us to find another way.

治愈可以卖给威权国家，但数据中心不行，芯片不行。没有理由不该在非洲建数据中心，只要不是中国拥有的。增长和经济价值会来得非常容易。不容易来的是分配、政治自由。政策该聚焦那些。慈善该起作用，但如果我们能让增长内生更好。让生物科技创业公司有一些发生在发展中世界。过渡期人在启动公司和监督模型上仍有角色，确保那些人里有一些在发展中世界。

The cures are fine to sell to authoritarian countries, but the data centers aren’t, and the chips aren’t. There’s no reason we shouldn’t build data centers in Africa, as long as they’re not owned by China. Growth and economic value will come very easily. What will not come easily is distribution of benefits, political freedom. Those are what policy should focus on. Philanthropy should play some role, but growth is always better if endogenous. Let’s make sure some of the AI-driven biotech startups happen in the developing world. During the transition humans will still have some role in starting companies and supervising models. Let’s make sure some of those humans are in the developing world.

Dwarkesh Patel

你们宣布 Claude 会有一部对齐到一套价值、而不一定只对齐终端用户的宪法。为什么不让每个人都有为自己说话的 AI？

You announced Claude is going to have a constitution aligned to a set of values, not necessarily just to the end user. Why is it better not to have everybody get their own AI advocating for them?

Dario Amodei

教模型原则，比给它一列「别热接线汽车、别说韩语」更一致、更容易覆盖边角。硬护栏还在，比如别制造生物武器。这是规则 vs 原则。另一件事是可纠正性：我们其实相当靠可纠正那一侧。默认该做人们请它做的任务。但如果你请它做危险的事或伤害别人，模型不愿意。大多可纠正，限度基于原则。

By teaching the model principles, its behavior is more consistent and it’s easier to cover edge cases than a list of rules — don’t tell people how to hot-wire a car, don’t speak in Korean. It has hard guardrails like don’t make biological weapons. That’s rules versus principles. The other thing is corrigibility. We’re actually pretty far on the corrigible side. Under normal circumstances, if someone asks the model to do a task, it should do that task. If you’ve asked it to do something dangerous or to harm someone else, the model is unwilling. Mostly corrigible, with limits based on principles.

原则怎么定，三层循环。一是 Anthropic 内部迭代，时不时公开更新。二是不同公司有不同宪法，外部比较、软激励。三是社会：几年前我们和 Collective Intelligence Project 做实验，问人们宪法里该有什么，当时吸收了其中一些改动。原则上可以有代议制政府输入，我今天不会做，立法过程太慢。

Three loops for how principles get set. One: we iterate within Anthropic, public updates so people can comment. Two: different companies have different constitutions; outside observers compare; that creates a soft incentive. Three: society. A couple years ago we did an experiment with the Collective Intelligence Project to poll people on what should be in our constitution, and we incorporated some of those changes. You could imagine representative government having input. I wouldn’t do that today because the legislative process is so slow.

Dwarkesh Patel

有人给这个时代写出相当于《原子弹的制造》时，历史记录里最可能错过的是什么？

When somebody writes the equivalent of The Making of the Atomic Bomb for this era, what will they be most likely to miss?

Dario Amodei

这条指数的每一个时刻，外面的世界有多不理解它。事后看起来都不可避免。我们内部有些人给它很高的概率，但外面有一个完全不按这个行动的世界。如果我们离它发生只有一年或两年，街上的普通人完全不知道。还有：它发生得有多绝对地快。你得做那个决定，同一天还得做另外 30 个。我担心某个非常关键的决定，会是有人走进来给两分钟、半页备忘录，问 A 还是 B。我说我得吃午饭，做 B 吧。结果那成了有史以来最关键的事。

At every moment of this exponential, the extent to which the world outside it didn’t understand it. Anything that actually happened looks inevitable in retrospect. Some of us internally put a high probability on this happening, but there’s a world outside us that’s not acting on that at all. If we’re one year or two years away, the average person on the street has no idea. And how absolutely fast it was happening. You have to make that decision, and then 30 other decisions on the same day. One of my worries is that some very critical decision will be someone coming into my office: you have two minutes, A or B, a random half-page memo. I’m like, I have to eat lunch, let’s do B. That ends up being the most consequential thing ever.

Dwarkesh Patel

通常没有科技 CEO 每隔几个月写 50 页备忘录。那种更智识型的 CEO 角色怎么运转？

There aren’t tech CEOs who are usually writing 50-page memos every few months. How do you construct that more intellectual-type role of CEO?

Dario Amodei

这一篇是寒假写的。我大概花三分之一、也许 40% 的时间确保 Anthropic 的文化是好的。2500 人，很难掺进每一个细节。一件杠杆很大的事是：确保人们喜欢在这儿工作，一起做而不是对着做。有些别的 AI 公司长大之后开始失谐、人跟人打架。我认为我们在让每个人感到使命、相信别人在这儿是因为对的理由这件事上，做得格外出色。

This particular one I wrote over winter break. I probably spend a third, maybe 40%, of my time making sure the culture of Anthropic is good. It’s 2,500 people. It’s very difficult to get involved in every single detail. One thing that’s very leveraged is making sure people like working there, everyone works together instead of against each other. We’ve seen as some of the other AI companies have grown — without naming names — decoherence and people fighting each other. I think we’ve done an extraordinarily good job of making everyone feel the mission, that everyone else is working for the right reason.

到了 2500 人，你不能一对一做。所以我每两周站到全公司面前讲一个小时。我写一个叫 DVQ 的东西，Dario Vision Quest。不是我起的名，我还试图反对过，因为它听着像我跑去抽迷幻药。三四页，讲内部、模型、产品、外面的行业、地缘政治。非常诚实：「这是我在想的」，然后回答问题。公司很大一部分人来参加。另外我在 Slack 有一个频道，大量评论，避开公司套话。如果你有一家你信任的人的公司，你就可以完全不过滤。那让人们比各部分之和更大。

When you get to 2,500 people, you can’t do that person by person. I get up in front of the whole company every two weeks and speak for an hour. I write this thing called a DVQ, Dario Vision Quest. I wasn’t the one who named it that. I tried to fight it because it made it sound like I was going off and smoking peyote. But the name stuck. A three or four-page document: what’s going on internally, the models, the products, the outside industry, geopolitics. Very honestly: this is what I’m thinking, then I answer questions. A large fraction of the company comes. I also have a Slack channel where I write a bunch of things. Avoid corpo speak. If you have a company of people you trust, you can really just be entirely unfiltered. It makes people more than the sum of their parts.

Dwarkesh Patel

好，没有对外的 Dario Vision Quest，我们有这场访谈。很有意思。谢谢你来做。

Well, in lieu of an external Dario Vision Quest, we have this interview. This has been fun, Dario. Thanks for doing it.

Dario Amodei

这场访谈有点像那个。谢谢你，Dwarkesh。

This interview is a little like that. Thank you, Dwarkesh.