Dario Amodei predicted in a recent interview that by 2026-2027, a "Country of Geniuses in a Datacenter" composed of AI will emerge, with an intelligence density comparable to tens of thousands of Nobel Prize winners. Financially, he disclosed that the company is experiencing "terrifying" 10x annual growth, with projected revenue expected to hit the $10 billion mark in 2025. Amodei explained why he dares not throw trillions of dollars to stockpile chips in advance: if demand explodes a year later than expected, the massive cash flow pressure would directly lead to bankruptcy.
Author | Dong Jing
Editor | Hard AI
On the eve of AI technology's exponential explosion, Anthropic CEO Dario Amodei threw out a shocking prediction to the industry: we are in the "twilight of exponential growth," and as early as 2026, humanity will usher in a "Country of Geniuses in a Datacenter" composed of tens of thousands of top minds.
Recently, in an in-depth interview with Dwarkesh Patel, Dario Amodei, CEO of large model unicorn Anthropic, rarely disclosed the company's staggering revenue growth expectations and elaborated on the timeline for AGI (Artificial General Intelligence), the financial logic of computing power investment, and geopolitical risks. Amodei believes that AI technology is at a critical point from quantitative to qualitative change, and the next 2-3 years will determine the trajectory of humanity for the next two hundred years.
(Anthropic CEO Dario Amodei appears on Dwarkesh Patel podcast)
01. AI is at the End of Exponential Growth
Dario Amodei pointed out at the beginning of the interview that we are approaching the end of the AI exponential growth curve, and the world has not fully perceived this qualitative change.
From GPT-1 to today's professional-grade models, AI has completed the transition from "smart high school student" to "PhD level," even surpassing human levels in programming, mathematics, and other fields. The underlying scaling laws have never failed, and investments in computing power and data continue to deliver clear returns.
The magic of exponential growth lies in the explosion at the end. Dario stated that Anthropic's 10x leap in annual revenue, Claude Code doubling engineer productivity, and rapid breakthroughs in model context length and generalization capabilities all confirm the signal that "the endpoint is near." This growth is not just about stacking parameters, but an upgrade in the essence of intelligence—from data fitting to autonomous generalization, AI is filling in the last few key pieces of the capability puzzle.
02. "A Country of Geniuses in a Datacenter": Redefining 2026
Amodei proposed a highly impactful concept in the interview—"A Country of Geniuses in a Datacenter." He reviewed the technological evolution of the past three years, believing that AI models have evolved from "smart high school students" to "professionals."
He boldly predicted that by 2026 or 2027, the intelligence level, knowledge depth, and logical reasoning ability demonstrated by a single model will not only be equivalent to one Nobel Prize winner, but to the collective of tens of thousands of top geniuses working together.
Regarding the certainty of this timeline, Amodei expressed extremely high confidence:
"I have 90% confidence that this vision will be achieved within 10 years; and for this to happen in the next 1-2 years, I think it's a 50/50 possibility."
He noted that the only variables could be geopolitical disasters (such as chip supply chain disruption) or severe social turmoil.
03. Revenue Surge: The "Terrifying" Curve from $100 Million to $10 Billion
On the financial data that the market is most concerned about, Amodei disclosed Anthropic's jaw-dropping growth curve. He revealed that the company's revenue is experiencing "bizarre 10x per year growth." Amodei stated directly in the interview:
"In 2023, we grew from 0 to $100 million; in 2024, from $100 million to $1 billion; and in 2025, we expect to reach $9-10 billion. This exponential growth roughly matches my expectations, and even in the first month of this year, we added several billion dollars in revenue."
Amodei emphasized that despite the lagging impact of economic diffusion, enterprise adoption of AI requires lengthy processes like legal review and compliance checks, but the improvement in technology capability itself is driving this crazy growth curve.
04. Computing Power Gamble and Bankruptcy Risk: The CEO's Financial Balancing Act
Faced with such a certain technological prospect, why not borrow a trillion dollars now to stockpile chips? Amodei gave a highly realistic financial explanation: computing power expansion must be linked to revenue growth and prediction accuracy, otherwise it will face devastating risk.
"If I predict trillion-level demand in 2027 and thus purchase $1 trillion worth of computing power in advance, but demand explodes just one year later, or growth slightly drops from 10x to 5x, no hedging means can prevent the company from going bankrupt," Amodei explained. This investment return based on "logarithmic return laws" requires precise calculation.
He pointed out that Anthropic's current strategy is "responsibly aggressive," meaning the computing power invested is sufficient to capture huge upside, but if market explosion is delayed, the company can still survive with its high-margin enterprise business and cash flow.
He expects Anthropic to achieve profitability around 2028, when AI will become one of the most profitable industries in history.
05. The Endgame of Software Engineering: From Writing Code to Replacing Engineers
Regarding specific application scenarios, Amodei views programming as the first fortress AI will conquer. He divides AI's evolution in software engineering into three stages:
- Stage One: Models write 90% of code lines (already achieved).
- Stage Two: Models handle 90% of end-to-end tasks, such as fixing bugs, configuring clusters, writing documentation.
- Stage Three: Models have "workplace experience," able to understand the context of complex codebases and set technical directions.
Amodei predicts that within 1-3 years, AI will be able to perform all the duties of a senior software engineer.
"This doesn't mean engineers will lose their jobs, but rather a massive explosion in productivity. Current models can do more than just complete code—they can take over difficult tasks like GPU kernel writing."
Full translation of Anthropic CEO Dario Amodei's in-depth interview follows:
06. What Exactly Are We Scaling?
Dwarkesh Patel: We talked three years ago. In your view, what's the biggest update from the past three years? What's the biggest difference between now and then?
Dario Amodei: Broadly speaking, the exponential growth of underlying technology basically matches my expectations. Maybe off by a year or two. I'm not sure if I predicted the specific direction in coding. But when I look at this exponential curve, it roughly matches my expectations for model progress—from smart high school students to smart college students, to starting to do PhD and professional-level work, even surpassing that level in coding. Frontier progress has been somewhat uneven, but generally matches expectations.
What's most surprising is the public's lack of awareness of how close we are to the end of the exponential curve. To me, it's crazy—people both inside and outside the circle are talking about those tired old political hot button issues, and we're already near the end of the exponential curve.
Dwarkesh: I want to understand what this exponential curve looks like now. The first question I asked you three years ago was "What's up with scaling, why does it work?" Now I have similar questions, but it feels more complicated.
At least from the public's perspective, three years ago there were well-known public trends, across multiple orders of magnitude of computation, you could see how the loss function improved. Now we have reinforcement learning scaling, but there are no publicly known scaling laws. It's not even clear what the principle behind this is. Is this teaching models skills? Is it teaching meta-learning? What is the scaling hypothesis now?
Dario: Actually my hypothesis is the same as in 2017.
I think I talked about this last time, but I wrote a document called the "Big Compute Blob Hypothesis." It wasn't specifically about language model scaling. I wrote it when GPT-1 just came out, and that was just one of many things.
Back then there was robotics. People were trying to study reasoning as something separate from language models, and there was reinforcement learning scaling in AlphaGo and OpenAI's Dota. People remember DeepMind's StarCraft, AlphaStar.
It was a more general document. Rich Sutton published "The Bitter Lesson" a few years later. The hypothesis is basically the same.
It says that all the cleverness, all the techniques, all the "we need new methods to do something" ideas, these don't matter much. Only a few things matter. I think I listed seven.
One is how much raw computing power you have. Two is the quantity of data. Three is the quality and distribution of data. It needs to be a broad distribution. Four is how long you train. Five is you need an objective function that can scale to the extreme. The pretraining objective function is one such objective. Another is the reinforcement learning objective function, which says you have a goal and you need to achieve that goal. Within this, there are objective rewards, like what you see in math and coding, and more subjective rewards, like what you see in RLHF or higher-order versions.
Then the sixth and seventh items are about normalization or conditioning, just to get numerical stability so the big compute blob can flow in a laminar way instead of running into problems.
That's the hypothesis, and it's still the one I hold today. I haven't seen much that contradicts it.
The pretraining scaling laws are an example we've seen. These laws have continued. Now widely reported, we feel good about pretraining. It continues to give us returns.
What's changed is that now we're seeing the same thing happen with reinforcement learning. We see a pretraining phase, then a reinforcement learning phase on top of it. For reinforcement learning, it's actually the same.
Even other companies have published content in some of their releases saying, "We train models on math competitions—AIME or other competitions—model performance scales log-linearly with how long we train it." We're seeing this too, and not just on math competitions. It's all kinds of reinforcement learning tasks.
We're seeing the same scaling in reinforcement learning that we saw in pretraining.
Dwarkesh: You mentioned Rich Sutton and "The Bitter Lesson." I interviewed him last year, and he actually disagrees with large language models quite a bit. I don't know if this is his view, but to paraphrase his objection one way: truly having the core of human learning doesn't need all these billions of dollars of data and computation, and these custom environments, to learn how to use Excel, how to use PowerPoint, how to browse the web.
The fact that we have to use these reinforcement learning environments to build in these skills implies we're actually missing a core human learning algorithm. So we're scaling the wrong things. This does raise a question. If we think there will be something like human instant learning ability, why are we doing all this reinforcement learning scaling?
Dario: I think this conflates several things that should be thought about separately. There is a real puzzle here, but it might not matter. In fact, I suspect it might not matter.
Here's an interesting thing. Let me leave reinforcement learning aside for now, because I actually think saying reinforcement learning is any different from pretraining on this issue is misleading.
If we look at pretraining scaling, it was very interesting when Alec Radford did GPT-1 in 2017. The models before GPT-1 were trained on datasets that didn't represent broad text distributions. You had very standard language modeling benchmarks. GPT-1 itself was actually trained on a bunch of fan fiction. That was literary text, just a small fraction of the text you could get.
At that time it was probably a billion words or so, so it was a small dataset representing a rather narrow distribution of what you see in the world. It didn't generalize well. If you did better on some fan fiction corpus, it wouldn't generalize well to other tasks. We had all these metrics. We had various standards for measuring how well it performed in predicting all other types of text.
Only when you train on all tasks on the internet—when you do a general internet crawl from something like Common Crawl, or crawl links from Reddit (which is what we did for GPT-2)—do you start to get generalization.
I think we're seeing the same thing with reinforcement learning. We started with simple reinforcement learning tasks, like training on math competitions, then moved to broader training involving code. Now we're moving to many other tasks.
I think we'll get more and more generalization. So this somewhat eliminates the distinction between reinforcement learning and pretraining.
But either way there's a puzzle, which is that in pretraining we used trillions of tokens. Humans don't see trillions of words. So there is a sample efficiency difference here. There is something different.
Models start from scratch and need more training. But we also see that once they're trained, if we give them a million long context—the only thing stopping long context is inference—they're very good at learning and adapting in that context.
So I don't know the complete answer to this question. I think something is happening where pretraining is not like the human learning process, but it's somewhere between human learning and human evolution.
Many of our priors come from evolution. Our brains aren't just a blank slate. Whole books have been written about this. Language models are more like blank slates. They really start with random weights, while human brains start with all these regions connected to all these inputs and outputs.
Maybe we should think of pretraining—and reinforcement learning—as existing in an intermediate space between human evolution and human instant learning. We should think of the context learning that models do as somewhere between human long-term learning and short-term learning.
So there's this hierarchy. There's evolution, there's long-term learning, there's short-term learning, and there's human instant response. The various stages of large language models exist on this spectrum, but not necessarily at exactly the same points.
There's no analogue corresponding to some human learning modes, and large language models fall between these points. Does that make sense?
Dwarkesh: It makes sense, though some things are still a bit confusing. For example, if the analogy is that this is like evolution, so low sample efficiency is okay, then why do we bother building all these reinforcement learning environments if we're going to get super sample-efficient agents from context learning?
Some companies' work seems to be teaching models how to use this API, how to use Slack, how to use other things. If that kind of agent that can learn instantly is emerging or has already emerged, why is there so much emphasis on this? This confuses me.
Dario: I can't speak to other people's priorities. I can only talk about how we think about it.
The goal is not to teach the model every possible skill in reinforcement learning, just like we don't do that in pretraining. In pretraining, we're not trying to expose the model to every possible way words can be combined.
Instead, models train on many things and then achieve generalization in pretraining. This is what I saw up close from GPT-1 to GPT-2 transition. The model reached a point. I had moments where I thought, "Oh yeah, you just give the model a list of numbers—here's the house price, here's the square footage—and the model can complete the pattern and do linear regression."
Not very well, but it did it, and it had never seen that exact thing before.
So as far as we're building these reinforcement learning environments, the goal is very similar to what pretraining did five or ten years ago. We're trying to get lots of data, not because we want to cover specific documents or specific skills, but because we want generalization.
I think the framework you're proposing is clearly reasonable. We're moving toward AGI. At this point no one disagrees that we'll achieve AGI this century. The key is that you're saying we're approaching the end of the exponential curve.
Others look at this and say, "We've been making progress since 2012, and by 2035 we'll have human-like agents."
Clearly, we see in these models what evolution does, or what humans learn in a lifetime. I want to understand what you see that makes you think this is a year away rather than ten years away.
Is Scaling an Excuse?
Dario: There are two claims that can be made here, one stronger and one weaker.
Let's start with the weaker claim. When I first saw scaling in 2019, I wasn't sure. It was a 50/50 thing. I thought I saw something. My claim was that this is much more likely than anyone thinks. Maybe there's a 50% chance it will happen.
Regarding what you said, that within ten years we'll reach what I call "A Country of Geniuses in a Datacenter," I'm 90% confident about that. It's hard to go above 90% because the world is so unpredictable. Maybe irreducible uncertainty gets us to 95%, and you encounter internal turmoil at multiple companies, Taiwan being invaded, all the fabs being blown up by missiles.
Dwarkesh: Now you've put a curse on us, Dario.
Dario: You can construct a 5% world where things are delayed ten years.
There's another 5% where I'm very confident about verifiable tasks. For coding, except for that irreducible uncertainty, I think we'll reach the goal in a year or two. It's impossible that we won't have reached end-to-end coding capability in ten years.
My little bit of fundamental uncertainty, even on long time scales, is about unverifiable tasks: planning a Mars mission; making some basic scientific discoveries, like CRISPR; writing a novel.
These tasks are hard to verify. I'm almost certain we have a reliable path there, but if there's a little uncertainty, it's there.
On the ten-year timeline, I'm 90% confident, which is about as certain as you can get. I think saying this won't happen by 2035 is crazy. In some sane world, this would be considered an outside-the-mainstream view.
But the emphasis on verification implies I lack conviction that these models generalize. If you think about humans, we're good at both things that get verifiable rewards and things that don't.
Dario: No, that's why I'm almost certain. We've already seen considerable generalization from verifiable things to unverifiable things. We've seen this.
But it seems like you're emphasizing this as a spectrum that will split, and we'll see more progress in some areas than others. This doesn't seem like how humans get better.
Dario: The world where we can't get there is the world where we do all the verifiable things. Many of these will generalize, but we don't quite get all the way there. We don't completely fill the other side of the box. This isn't a binary thing.
Even if generalization is weak and you can only do verifiable domains, I'm not sure in such a world whether you can automate software engineering. In some sense, you're a "software engineer," but part of being a software engineer includes writing long memos about your grand vision.
Dwarkesh: I don't think that's part of a software engineer's job.
Dario: That's part of company work, not specifically for software engineers. But software engineering does involve design documents and other similar things. Models are already good at writing comments.
Again, the claim I'm making here is much weaker than what I actually believe, to distinguish two things. We're almost there in software engineering.
Dwarkesh: By what standard? One standard is how many lines of code AI writes.
If you consider other productivity improvements in software engineering history, compilers write all the software lines. There's a difference between how many lines are written and how much productivity improves. What does "we're almost there" mean?
Dario: How much productivity improves, not just how many lines AI writes.
Dwarkesh: I actually agree with you.
Dario: I've made a series of predictions about code and software engineering. I think people keep misunderstanding them. Let me list this spectrum.
About eight or nine months ago, I said AI models will write 90% of code lines in three to six months. This happened, at least in some places. It happened at Anthropic, it happened with many downstream people using our models.
But this is actually a very weak standard. People thought I meant we don't need 90% of software engineers. These things are far apart.
The spectrum is: 90% of code written by models, 100% of code written by models. There's a big difference in productivity.
90% of end-to-end software engineering tasks—including compiling, setting up clusters and environments, testing features, writing memos, etc.—done by models.
100% of today's software engineering tasks done by models. Even if this happens, it doesn't mean software engineers will lose their jobs. They can do new higher-level things, they can manage.
Then further along the spectrum, demand for software engineers drops 90%. I think this will happen, but it's a spectrum.
I wrote about this in "The Adolescence of Technology," using agriculture as an example of going through this spectrum.
Dwarkesh: I actually completely agree with you. These are very different benchmarks from each other, but we're passing through them at super-fast speed.
Part of your vision is that going from 90 to 100 will happen very quickly and will bring huge productivity gains. But what I notice is that even in greenfield projects where people start with Claude Code or something similar, people report starting a lot of projects... Are we seeing a software renaissance in the outside world, all these new features that otherwise wouldn't exist? At least so far, it seems we don't.
So this does make me wonder. Even if I never need to intervene with Claude Code, the world is complex. Work is complex. On closing the loop on self-contained systems, whether it's just writing software or something else, how much broader benefit will we see from this?
Maybe this should dilute our estimate of the "Country of Geniuses."
Dario: I simultaneously agree with you that this is why these things don't happen immediately, but at the same time, I think the effect will be very fast.
You can have these two extremes. One is that AI won't make progress. It's slow. It will diffuse into the economy forever.
Economic diffusion has become one of those buzzwords, becoming a reason why we won't make AI progress or AI progress doesn't matter.
The other axis is that we'll get recursive self-improvement, the whole thing. Can't you just draw an exponential line on the curve?
Within many nanoseconds of us getting recursion, we'll have Dyson spheres around the sun. I'm being completely sarcastic about this view here, but there are these two extremes.
But what we've seen from the beginning, at least if you look inside Anthropic, is this weird 10x annual revenue growth.
So in 2023, it was from zero to $100 million. In 2024, it was from $100 million to $1 billion. In 2025, it's from $1 billion to $9-10 billion.
Dwarkesh: You should buy $1 billion of your own product so you can...
Dario: In the first month of this year, that exponential curve... you'd think it would slow down, but we added several billion dollars of revenue in January.
Obviously that curve can't continue forever. GDP is only so big. I even guess it will bend this year, but that's a fast curve. That's a very fast curve. I bet even as it scales to the whole economy, it will remain quite fast.
So I think we should consider this intermediate world where things are very fast but not instant, they take time because of economic diffusion, because loops need to be closed.
Because it's cumbersome: "I have to do change management in my enterprise... I set this up, but I have to change the security permissions for this to actually work... I have this old software that checks models before compiling and releasing, I have to rewrite it. Yes, the model can do this, but I have to tell the model to do it. It has to take time to do it."
So I think everything we've seen so far is compatible with the idea that there's a fast exponential curve, which is model capability. Then there's another fast exponential curve downstream, which is model diffusion into the economy.
Not instant, not slow, much faster than any previous technology, but it has its limits.
When I look inside Anthropic, when I look at our customers: fast adoption, but not infinitely fast.
Dwarkesh: Can I try a bold view?
Dario: Sure.
Dwarkesh: I feel like diffusion is an excuse people use. When models can't do something, they say, "Oh, but it's a diffusion problem."
But you should compare with humans. You'd think AI's inherent advantages would make onboarding new AI much easier than onboarding new humans.
AI can read your entire Slack and your drive in minutes. They can share all the knowledge that other copies of the same instance have.
When you hire AI, you don't have this adverse selection problem, so you can just hire copies of vetted AI models.
Hiring humans is much more troublesome. People have been hiring humans all along. We pay humans over $50 trillion in wages because they're useful, even though in principle integrating AI into the economy should be much easier than hiring humans.
Scaling doesn't really explain this.
Dario: I think diffusion is very real and not entirely about AI model limitations.
Again, some people use diffusion as a buzzword to say this isn't a big deal. I'm not talking about that. I'm not talking about AI diffusing at the speed of previous technologies.
I think AI will diffuse much faster than previous technologies, but not infinitely fast.
Let me give just one example. There's Claude Code. Claude Code is very easy to set up. If you're a developer, you can just start using Claude Code.
There's no reason why developers at large enterprises shouldn't adopt Claude Code as quickly as individual developers or developers at startups.
We do everything possible to promote it. We sell Claude Code to enterprises.
Large enterprises, large financial companies, large pharmaceutical companies, all of these are adopting Claude Code, much faster than enterprises typically adopt new technology.
But again, it takes time. Any given feature or any given product, like Claude Code or Cowork, will be adopted by individual developers who are always on Twitter, by Series A startups, months earlier than by large enterprises doing food sales.
There are just many factors. You have to go through legal review, you have to configure it for everyone. It has to pass security and compliance.
Company leaders are further from the AI revolution, they have vision, but they have to say, "Oh, it makes sense for us to spend $50 million. Here's what this Claude Code thing is. Here's why it helps our company. Here's why it makes us more productive."
Then they have to explain to people two levels down. They have to say, "Okay, we have 3,000 developers. How are we going to roll this out to our developers."
We have these conversations every day. We're doing everything we can to make Anthropic's revenue grow 20 or 30 times a year instead of 10 times.
Again, many enterprises just say, "This is so productive. We're going to cut corners in our usual procurement process."
They're acting much faster than when we try to sell them ordinary APIs. Many enterprises are using them. Claude Code is a more compelling product, but it's not an infinitely compelling product.
I think even AGI or powerful AI or "A Country of Geniuses in a Datacenter" won't be an infinitely compelling product. It will be a sufficiently compelling product, maybe capable of 3-5x or 10x annual growth, even at hundreds of billions of dollars scale, which is very hard to do and has never been done in history, but not infinitely fast.
Dwarkesh: I think it will be a slight slowdown. Maybe this isn't your claim, but sometimes people talk about this like, "Oh, the capabilities are there, but because of diffusion... otherwise we'd basically be at AGI."
Dario: I don't believe we're basically at AGI. I think if you have "A Country of Geniuses in a Datacenter"...
If we had "A Country of Geniuses in a Datacenter," we would know. If you had "A Country of Geniuses in a Datacenter," we would know. Everyone in this room would know. Everyone in Washington would know. People in rural areas might not know, but we would know.
We don't have that right now. That's very clear.
Is Continuous Learning Necessary?
Dwarkesh: Going back to specific predictions... because there are so many different things that need to be disambiguated, it's easy to misunderstand each other when we talk about capabilities.
For example, when I interviewed you three years ago, I asked you for a prediction about what we should expect in three years. You were right. You said, "We should expect systems where if you talk to them for an hour, it's hard to distinguish them from well-educated humans." I think you were right.
I was spiritually unsatisfied because my internal expectation was that such systems could automate most white-collar work. So it might be more productive to talk about the actual end capabilities you want from such systems.
I'll basically tell you where I think we are.
Let me ask a very specific question so we can accurately figure out what kind of capabilities we should be thinking about soon. Maybe I'll ask this in the context of a job I know well, not because it's the most relevant job, but just because I can evaluate claims about it.
Take video editing. I have video editors. Part of their job involves understanding our audience's preferences, understanding my preferences and taste, and the various tradeoffs we have. They build understanding of context over many months.
When should we expect AI systems that can instantly master the skills and capabilities they have after working for six months?
Dario: I think what you're saying is we're doing this three-hour interview. Someone will come in, someone will edit it. They'll say, "Oh, I don't know, Dario scratched his head, we can cut that."
"Zoom in on that." "There's this long discussion, not very interesting to people. There's something else that's more interesting to people, so let's make this edit."
I think "A Country of Geniuses in a Datacenter" will be able to do this. The way it will be able to do this is that it will have general control over the computer screen. You'll be able to input this.
It will also be able to use the computer screen to go online, look at all your previous interviews, look at people's comments on Twitter about your interviews, talk to you, ask you questions, talk to your staff, look at the edit history you've made, and work from that.
I think this depends on a few things. I think this is actually one of the things blocking deployment: reaching the point where models are truly proficient at using computers.
We've seen this climb in benchmarks, and benchmarks are always imperfect measures. But I think when we first released computer use about a year and a quarter ago, OSWorld was around 15%.
I don't remember the exact number, but we've climbed from there to 65-70%. There may be harder measures, but I think computer use has to pass a reliability point.
Dwarkesh: Before you continue to the next point, can I follow up? For years, I've been trying to build different internal LLM tools for myself.
Usually I have these text-in, text-out tasks that should be the core competency of these models. Yet I still hire humans to do them.
If it's something like "identify what the best segment in this text is," maybe LLMs do seven out of ten jobs on it. But there's no consistent way for me to interact with them, help them get better at their work, like I can with human employees.
That missing capability, even if you solve computer use, will still prevent me from outsourcing actual work to them.
Dario: This goes back to what we were talking about before with learning on the job. This is very interesting. I think for coding agents, I don't think people would say that learning on the job is what's stopping coding agents from doing everything end-to-end.
They keep getting better. We have engineers at Anthropic who don't write any code. When I look at productivity, going back to your earlier question, we have people who say, "This GPU kernel, this chip, I used to write myself. I just let Claude do it." There's a huge productivity increase.
When I look at Claude Code, familiarity with the codebase or the model not having a feeling of working at the company for a year, that's not high on the complaint list I see.
I think what I'm saying is we're going down a different path.
Dwarkesh: You don't think coding is like that because there's an external memory scaffold instantiated in the codebase? I don't know how many other jobs have that.
Coding is making rapid progress precisely because it has this unique advantage that other economic activities don't have.
Dario: But when you say that, you're implying that by reading the codebase into context, I have everything a human needs to learn on the job.
So that would be an example—whether it's written down, whether it's available—a case where everything you need to know is obtained from the context window.
What we think of as learning—"I started this job, I need six months to understand the codebase"—the model just does in context.
I really don't know how to think about this, because someone qualitatively reported what you said. I believe you saw last year, there was a major study where they had experienced developers try to close pull requests in repositories they were familiar with. Those developers reported improvement. They reported feeling more productive using these models.
But in fact, if you look at their output and what actually got merged back, there was a 20% decrease. The result of them using these models was lower productivity.
So I'm trying to reconcile people's qualitative feelings about these models with: 1) At the macro level, where is this software renaissance? And then 2) When people do these independent evaluations, why don't we see the productivity gains we expect?
Dario: Inside Anthropic, this is really unambiguous. We're under incredible commercial pressure, and we make it harder for ourselves because of all the safety work we do, which I think we do more than other companies.
The pressure to survive economically while maintaining our values is incredible. We're struggling to maintain this 10x revenue curve growth.
There's no time for nonsense. There's no time to feel we're productive when we're actually not.
These tools make us much more productive. Why do you think we worry about competitors using these tools? Because we think we're ahead of competitors.
If this secretly lowered our productivity, we wouldn't go through all this trouble. We can see final productivity every few months, in the form of model releases.
There's no room for self-deception here. These models make you more productive.
1) People feeling productive is qualitatively predicted by such research. But 2) If I just look at final output, clearly you're making rapid progress.
But the idea should be that through recursive self-improvement, you build a better AI, the AI helps you build a better next AI, and so on.
Instead, what I see is—if I look at you, OpenAI, DeepMind—people just move positions on the podium every few months. Maybe you think that will stop because you win or something.
But if in fact the previous coding model had these huge productivity gains, why don't we see this persistent advantage for the people who have the best coding models.
Dario: I think my model of the situation is that there's a gradually growing advantage. I'd say right now the total factor acceleration from coding models is maybe, I don't know, 15-20%. That's my view. Six months ago, it might have been 5%.
So that's okay. 5% doesn't count. It's just now reaching a point where it's one of several factors that's somewhat important. That will continue to accelerate.
I'd say six months ago, there were several companies roughly at the same point because this wasn't a significant factor, but I think it's starting to accelerate more and more.
I'd also say there are multiple companies writing models for code, and we're not entirely good at stopping some of these other companies from using our models internally.
So I think everything we see is consistent with this snowball model. Again, my theme in all of this is that all of this is a soft takeoff, soft, smooth exponential curves, even though the exponential curves are relatively steep.
So we see this snowball gathering momentum, it's like 10%, 20%, 25%, 40%. As you go, Amdahl's Law, you have to clear out everything that's stopping you from closing the loop.
But this is one of the highest priorities inside Anthropic.
Stepping back, before we talked about when we'll get this on-the-job learning?
It seems like the point you're making about coding is that we actually don't need on-the-job learning.
You can have huge productivity improvements, you can potentially bring trillions of dollars in revenue to AI companies, without this basic human instant learning ability.
Maybe that's not your claim, and you should clarify. But in most areas of economic activity, people say, "I hired someone, they weren't that useful in the first few months, then over time, they built context, understanding."
It's actually hard to define what we're talking about here. But they got something, and now they're a powerful force, they're so valuable to us.
If AI doesn't develop this instant learning ability, I'm a bit skeptical that we'll see huge changes in the world without that ability.
Dario: I think there are two things here. There's the current state of technology.
Again, we have these two phases. We have pretraining and reinforcement learning phases, where you throw a bunch of data and tasks into the model, and then they generalize.
So this is like learning, but it's like learning from more data, rather than learning within one human or one model's lifetime.
So again, this sits between evolution and human learning. But once you learn all these skills, you have them.
Like pretraining, like the model knows more, if I look at a pretrained model, it knows more about the history of Japanese samurai than I do. It knows more about baseball than I do. It knows more about low-pass filters and electronics, all these things.
Its knowledge is much more extensive than mine. So I think even just that might get us to a point where models are better at everything.
We also have, again, just by expanding existing types of setups, context learning. I'd describe it as somewhat like human on-the-job learning, but slightly weaker, slightly shorter-term.
You look at context learning, if you give the model a bunch of examples, it really can understand. Real learning does happen in context.
A million tokens is a lot. That could be several days of human learning. If you think about a model reading a million words, how long would it take me to read a million? At least several days or weeks.
So you have these two things. I think these two things in the existing paradigm might be enough to get you "A Country of Geniuses in a Datacenter."
I'm not sure, but I think they'll get you a large part of the way there. There may be gaps, but I certainly think for now, this is enough to generate trillions of dollars in revenue. That's the first point.
Second point, this idea of continuous learning, the idea of a single model learning on the job. I think we're working on this too.
It's quite possible that in the next year or two, we'll solve this too. Again, I think you get most of the way there without it.
Trillions of dollars annually in market, maybe all the national security implications and safety implications I wrote about in "The Adolescence of Technology" could happen without it.
But we, and I assume others, are working on it. It's quite possible we'll get there in the next year or two.
There are many ideas. I won't go into detail on all of them, but one is to make context longer. There's nothing stopping longer context from working.
You just have to train on longer context, then learn to serve them at inference time. Both of these are engineering problems we're working on, and I assume others are working on them too.
Dwarkesh: This context length increase, it seems like there was a period from 2020 to 2023, from GPT-3 to GPT-4 Turbo, going from 2000 context length to 128K.
I feel like for about two years since then, we've been in the same range. When context length gets much longer than that, people report qualitative decline in the model's ability to consider full context.
So I'm curious what you've seen internally that makes you think, "10 million context, 100 million context, to get six months of human learning and context building."
Dario: This isn't a research problem. This is an engineering and inference problem. If you want to serve long context, you have to store your entire KV cache.
Storing all memory in GPU, processing memory is difficult. I don't even know the details. At this point, this is a level of detail I can't keep up with anymore, even though I knew it in the GPT-3 era. "These are the weights, these are the activations you have to store..."
But now the whole thing is flipped because we have MoE models and all that.
Regarding this degradation you're talking about, there are two things, not to get too specific. There's the context length you train on and the context length you serve.
If you train on small context length and then try to serve on long context length, maybe you get these degradations. It's better than nothing, you might still serve it, but you get these degradations.
Maybe training on long context length is harder. I think, at the same time, ask about possible rabbit holes.
If you have to train on longer context length, doesn't that mean for the same amount of compute, you get fewer samples?
Maybe it's not worth going into in depth. I want to get answers to bigger picture questions.
I don't feel preference for a human editor who's worked for me for six months versus an AI that's worked with me for six months, what year do you predict that will be?
Dario: My guess on this is there are many issues, basically when we have "A Country of Geniuses in a Datacenter," we can do this.
My view on this, if you let me guess, is one to two years, maybe one to three years. Really hard to say. I have a strong view—99%, 95%—that all of this will happen in 10 years. I think that's just a super safe bet.
I have a hunch—this is more of a 50/50 thing—that it's going to be more like one to two years, maybe more like one to three years.
So one to three years. Country of Geniuses, and slightly less economically valuable video editing tasks.
Dwarkesh: Sounds very economically valuable, let me tell you.
Dario: There are just many use cases like that. There are many similar ones. So you're predicting within one to three years.
If AGI is Coming, Why Not Buy More Compute?
Dwarkesh: Then, generally speaking, Anthropic predicts that by the end of 2026 or early 2027, we'll have AI systems "with the ability to browse the interfaces humans use today for digital work, matching or exceeding the intellectual capacity of Nobel Prize winners, and the ability to interact with the physical world."
You emphasized in a DealBook interview two months ago that your company is scaling computing more responsibly compared to competitors. I'm trying to reconcile these two views.
If you really believe we'll have a Country of Geniuses, you want the largest possible data center. There's no reason to slow down.
The TAM for one Nobel Prize winner who can actually do everything a Nobel Prize winner can do is trillions of dollars.
So I'm trying to reconcile this conservatism, which seems rational if you have more moderate timelines, with your stated views on progress.
Dario: It actually all fits together. We go back to this fast but not infinitely fast diffusion.
Suppose we make progress at this rate. Technology makes progress this fast. I'm very confident we'll get there in a few years.
I have a hunch that we'll get there in one or two years. So there's a little uncertainty on the technology side, but very confident it won't be off by much.
What I'm less sure about, again, is the economic diffusion side. I really believe we could have models of a Country of Geniuses in a Datacenter in one or two years.
One question is: How many years after that will trillions in revenue start rolling in? I don't think it's guaranteed to be immediate.
It could be one year, it could be two years, I could even stretch it to five years, though I'm skeptical of that.
So we have this uncertainty. Even if technology progresses as fast as I suspect, we don't know exactly how fast it will drive revenue.
We know it's coming, but depending on how you buy these data centers, if you're off by a few years, that could be devastating.
It's like what I wrote in "Machines of Loving Grace." I said I think we might get this powerful AI, this "Country of Geniuses in a Datacenter." That description you gave comes from "Machines of Loving Grace."
I said we'll get it in 2026, maybe 2027. Again, that's my hunch. If I'm off by a year or two, I wouldn't be surprised, but that's my hunch.
Let's say that happens. That's the starting gun. How long does it take to cure all diseases? That's one way to push massive economic value.
You cure every disease. There's a question about how much goes to pharmaceutical companies or AI companies, but there's huge consumer surplus because—assuming we can provide access for everyone, which I care a lot about—we cure all these diseases.
How long does it take? You have to do biological discovery, you have to make new drugs, you have to go through regulatory processes. We saw this with vaccines and COVID.
We rolled out vaccines to everyone, but it took a year and a half.
My question is: From when AI first exists in the lab to when diseases are actually cured for everyone, how long does it take for everyone to get the cure for everything—AI is a genius that could theoretically invent it?
We've had the polio vaccine for 50 years. We're still trying to eradicate it in the most remote corners of Africa. The Gates Foundation is doing everything it can to try. Others are doing everything they can to try. But it's difficult.
Again, I don't expect most economic diffusion to be that difficult. That's the hardest case. But here's a real dilemma.
My view on this is that it will be faster than anything we've seen in the world, but it still has its limits.
So when we go to buy data centers, again, the curve I see is: we have 10x growth every year.
At the beginning of this year, we're looking at $10 billion in annualized revenue. We have to decide how much compute to buy.
Actually building data centers, booking data centers takes a year or two.
Basically I'm saying, "In 2027, how much compute do I get?"
I could assume revenue will continue growing 10x annually, so by the end of 2026 it will be $100 billion, and by the end of 2027 it will be $1 trillion.
It would actually be $5 trillion in compute because it would be $1 trillion a year for five years. I could buy $1 trillion in compute starting from the end of 2027.
If my revenue isn't $1 trillion, even if it's $800 billion, there's no force in the world, no hedge that can stop me from going bankrupt if I buy that much compute.
Even though part of my brain wonders if it will continue growing 10x, I can't buy $1 trillion a year in compute in 2027.
If I'm off by just one year on that growth rate, or if the growth rate is 5x annually instead of 10x, then you go bankrupt.
So you end up in a world where you support hundreds of billions, not trillions. You accept some risk that there's so much demand you can't support revenue, and you accept some risk that you got it wrong and it's still slow.
When I talk about responsible behavior, I don't actually mean absolute amounts. I think it's true that we spend a bit less than some other players.
It's actually other things, like whether we're thoughtful, or whether we're YOLOing and saying, "We're going to do $100 billion here or $100 billion there"?
My impression is some other companies haven't written down spreadsheets, they don't really understand the risks they're taking. They just do things because they sound cool.
We've thought about it carefully. We're an enterprise business. Therefore, we can rely more on revenue. It's not as fickle as consumers. We have better margins, which is a buffer between buying too much and buying too little.
I think the amount we buy allows us to capture a fairly strong upside world. It won't capture the full 10x annually.
Things would have to get pretty bad for us to get into financial trouble. So we've thought about it carefully, and we made that balance. That's what I mean when I say we're responsible.
Dwarkesh: So it seems we might actually just have different definitions of "Country of Geniuses in a Datacenter." Because when I think about actual human geniuses, an actual Country of Geniuses in a data center, I'd be happy to buy $5 trillion worth of compute to run an actual Country of human Geniuses in a data center.
Let's say JPMorgan or Moderna or whoever doesn't want to use them. I have a Country of Geniuses. They'll start their own companies.
If they can't start their own companies, they're bottlenecked by clinical trials... It's worth noting that for clinical trials, most clinical trials fail because the drug doesn't work. No efficacy.
Dario: I made exactly this point in "Machines of Loving Grace," I said clinical trials will be much faster than we're used to, but not infinitely fast.
Dwarkesh: Okay, so let's say clinical trials take a year to succeed so you can get revenue from it and make more drugs.
Okay, then, you have a Country of Geniuses, you're an AI lab. You can use more AI researchers.
You also think that smart people working on AI technology have these self-reinforcing returns. You can have data centers working on AI progress.
Is there substantially more gain from buying $1 trillion in compute annually versus $300 billion annually?
Dario: If your competitor buys $1 trillion, yes, there is.
Well, no, there are some gains, but again, there's this chance that they go bankrupt first. Again, if you're off by just one year, you destroy yourself. That's the balance.
We're buying a lot. We're buying quite a lot. We're buying amounts comparable to what the biggest players in the game are buying.
But if you ask me, "Why haven't we signed up for $10 trillion in compute starting mid-2027?"... First, it can't be produced. There's not that much in the world.
But second, what if the Country of Geniuses comes, but it comes in mid-2028 instead of mid-2027? You go bankrupt.
Dwarkesh: So if your prediction is one to three years, it seems like you should want to have $10 trillion in compute by 2029 at the latest? Even in the longest version of your stated timeline, the compute you're scaling up construction for doesn't seem consistent.
Dario: What makes you think that?
Human wages, let's say, are about $50 trillion annually—
Dwarkesh: So I won't talk specifically about Anthropic, but if you talk about the industry, the amount of compute the industry is building this year is probably, let's call it, 10-15 gigawatts.
It's growing about 3x annually. So next year is 30-40 gigawatts. 2028 could be 100 gigawatts. 2029 could be like 300 gigawatts.
I'm doing the math in my head, but cost per gigawatt might be $10 billion, about $10-15 billion annually.
You add all this up, and you get about what you described. By 2028 or 2029, you get over a trillion annually.
Dwarkesh: That's for the industry.
Dario: That's for the industry, exactly.
Dwarkesh: Let's say Anthropic's compute continues growing 3x annually, and then by 2027-28, you have 10 gigawatts. Multiply by, as you said, $10 billion. So that's like $100 billion annually.
But then you're saying by 2028 TAM is $200 billion.
Dario: Again, I don't want to give exact Anthropic numbers, but these numbers are too small.
Dwarkesh: Okay, interesting.
How Will AI Labs Actually Make Money?
Dwarkesh: You told investors you plan to be profitable starting in 2028. This is the year we might get a Country of Geniuses in a Datacenter. This will now unlock all these advances in medicine, health, and new technology. Isn't this exactly when you'd want to reinvest in the business and build a bigger "country" so they can make more discoveries?
Dario: In this field, profitability is a somewhat strange thing. I don't think in this field, profitability is actually a measure of burn versus investing in the business. Let's take an example. I actually think profitability happens when you underestimate the amount of demand you're going to get, and losses happen when you overestimate the amount of demand you're going to get, because you're buying data centers in advance.
Think about it this way. Again, these are stylized facts. These numbers aren't exact. I'm just trying to build a toy model here. Let's say half your compute is for training, and half your compute is for inference. Inference has more than 50% gross margin. So that means if you're in a steady state, you build a data center, if you know exactly the demand you're getting, you get a certain amount of revenue.
Let's say you pay $100 billion annually for compute. On $50 billion annually, you support $150 billion in revenue. The other $50 billion is for training. Basically you're profitable, you made $50 billion in profit. That's the economics of this industry today, or not today, but what we predict for a year or two from now.
The only thing that breaks this is if you get less than $50 billion in demand. Then you have more than 50% of your data center for research, and you're not profitable. So you trained stronger models, but you're not profitable. If you get more demand than you imagined, then research gets squeezed, but you can support more inference, and you're more profitable. Maybe I'm not explaining it well, but what I'm trying to say is, you first decide the amount of compute. Then you have some target aspiration for inference vs training, but that's determined by demand. It's not determined by you.
Dwarkesh: What I'm hearing is that you predict profitability because you're systematically underinvesting in compute?
Dario: No, no, no. I'm saying it's hard to predict. Regarding 2028 and when things happen, that's us trying our best for investors. All this stuff is very uncertain because of the uncertainty cone. If revenue grows fast enough, we could be profitable in 2026. If we overestimate or underestimate the next year, that could swing wildly.
Dwarkesh: What I'm trying to figure out is, do you have a business model in your head, invest, invest, invest, get scale and then become profitable. There's a single point in time when things turn around.
Dario: I don't think the economics of this industry work that way.
Dwarkesh: I see. So if I understand correctly, you're saying that because of the difference between the compute we should get and the compute we actually get, we're somewhat forced to be profitable. But that doesn't mean we'll continue to be profitable. We'll reinvest the money because now AI has made so much progress, we want a bigger Country of Geniuses. So revenue goes back up, but so do losses.
Dario: If we accurately predict every year what demand is, we'd be profitable every year. Because spending 50% of compute on research, roughly, plus greater than 50% gross margin and correct demand prediction leads to profitability. That's the profitable business model I think exists somewhat but is obscured by all this pre-building and prediction error.
Dwarkesh: I think you're treating 50% as a given constant, when in fact, if AI progresses fast, you could increase progress by scaling up, and you should have more than 50% and not be profitable.
Dario: But that's what I'm saying. You might want to scale it more. Remember logarithmic scale returns. If 70% only gets you a slightly smaller model through a 1.4x factor... that extra $20 billion, every dollar there is worth much less to you because of the log-linear setup. So you might find it better to invest that $20 billion in serving inference or hiring engineers who are better at what they do.
So the reason I say 50%... that's not exactly our target. It won't exactly be 50%. It might change over time. What I'm saying is log-linear returns, which lead you to spend a fraction of the business. Like not 5%, not 95%. Then you get diminishing returns.
Dwarkesh: I feel weird that I'm convincing Dario about AI progress or something. Okay, you don't invest in research because it has diminishing returns, but you invest in other things you mentioned.
Dario: I think macro-level profit—again, I'm talking about diminishing returns, but after you've spent $50 billion annually. This I believe you'll bring up, but diminishing returns on geniuses could be quite high.
More generally, what is profit in a market economy? Profit basically says that other companies in the market can do more with this money than I can. Put Anthropic aside. I don't want to provide information about Anthropic. That's why I'm giving these stylized numbers. But let's derive the industry equilibrium. Why doesn't everyone spend 100% of their compute on training and serve no customers? Because if they don't get any revenue, they can't raise money, can't trade compute, can't buy more compute next year.
So there will be an equilibrium where every company spends less than 100% on training, and of course less than 100% on inference. It's obvious why you don't just serve current models and never train another model, because then you have no demand because you fall behind. So there's some equilibrium. It won't be 10%, it won't be 90%. Let's say as a stylized fact, it's 50%. That's what I'm trying to express. I think we'll be in a position where the equilibrium is you spend less on training than the gross margin you're able to get on compute.
So the underlying economics are profitable. The question is when you buy next year's compute, you have this hellish demand prediction problem, where you might guess low and be very profitable but have no compute for research. Or you might guess high, you're not profitable and have all the compute in the world for research. Does that make sense? Just as a dynamic model of the industry?
Dwarkesh: Maybe stepping back, I'm not saying I think the "Country of Geniuses" is coming in two years, so you should buy this compute. To me, the final conclusion you came to makes a lot of sense. But it seems that's because the "Country of Geniuses" is hard, and there's still a long way to go. So stepping back, what I'm trying to figure out is, your worldview seems compatible with people who say "we're 10 years away from a world that generates trillions of dollars in value."
Dario: That's not my view at all. So I'll make another prediction. I find it hard to see there won't be trillions of dollars in revenue before 2030. I can construct a plausible world. Maybe it takes three years. That would be what I think is a reasonable endpoint. Like in 2028, we get the real "Country of Geniuses in a Datacenter." Revenue by 2028 enters the low hundreds of billions, then the Country of Geniuses accelerates it to trillions. We're basically on the slow end of diffusion. It takes two years to reach trillions. That would be the world until 2030. I suspect even combining the technology exponential and diffusion exponential, we'll get there before 2030.
Dwarkesh: So you're proposing a model where Anthropic is profitable because it looks like fundamentally we're in a compute-constrained world. So eventually we continue to grow compute—
Dario: I think the source of profit is... again, let's abstract out the whole industry. Let's imagine we're in an economics textbook. We have a few companies. Each can invest a limited amount. Each can invest a fraction in R&D. They have certain marginal serving costs. The gross margin over that marginal cost is very high because inference is efficient. There's some competition, but models are also differentiated. Companies compete to push up their research budgets.
But because there are only a few players, we have... what's that called? Cournot equilibrium, I think, which is the equilibrium for a few companies. The point is it won't balance to zero profit perfect competition. If there are three companies in the economy, and they all act rationally to some extent independently, it won't balance to zero.
Dwarkesh: Help me understand, because right now we do have three leading companies, and they're not profitable. So what changes?
Dario: Again, margins are very positive right now. What's happening is a combination of two things. One is that we're still in the exponential expansion phase of compute. A model gets trained. Let's say training a model last year cost $1 billion. Then this year it generated $4 billion in revenue, with inference costs of $1 billion. Again, I'm using stylized numbers here, but that would be 75% gross margin and this 25% tax. So that model overall made $2 billion.
But at the same time, we spent $10 billion training the next model because of exponential scaling. So the company loses money. Every model makes money, but the company loses money. What I'm calling equilibrium is an equilibrium where we have a "Country of Geniuses in a Datacenter," but the scaling of that model training has become much more balanced. Maybe it's still going up. We're still trying to predict demand, but it's much steadier.
Dwarkesh: I'm confused about a few things there. Let's start from the current world. In the current world, you're right, as you said before, if you treat each individual model as a company, it's profitable. But of course, a big part of the production function of being a frontier lab is training the next model, right?
Dario: Yes, that's right.
Dwarkesh: If you don't do that, then you're profitable for two months, then you have no profit because you won't have the best models. But at some point, this reaches the maximum scale it can reach. Then in equilibrium, we have algorithm improvements, but we spend roughly the same amount to train the next model as we spent to train the current model. At some point, you run out of money in the economy. Fixed lump of labor fallacy... the economy will grow, right? That's one of your predictions. We'll have data centers in space.
Dario: Yes, but this is another example of the theme I've been talking about. I think with AI, economic growth will be faster than ever. Now compute is growing 3x annually. I don't believe the economy will grow 300% annually. I said in "Machines of Loving Grace" that I think we might get 10-20% annual economic growth, but we won't get 300% annual economic growth. So I think eventually, if compute becomes a large part of economic output, it will be constrained by this.
Dwarkesh: So let's assume a model where compute remains capped. The world where frontier labs make money is a world where they continue to make rapid progress. Because fundamentally, your profit is limited by how good the substitutes are. So you're able to make money because you have the frontier models. If you don't have frontier models, you won't make money. So this model requires there's never a steady state. You're constantly making more algorithm progress.
Dario: I don't think that's true. I mean, I feel like we're in an economics class. You know the Tyler Cowen quote? We'll never stop talking about economics.
Dwarkesh: We'll never stop talking about economics.
Dario: So no, I don't think this field will become a monopoly. All my lawyers don't want me to say the word "monopoly." I don't think this field will become a monopoly. You do have some industries with only a few players. Not one, but a few players. Usually, you get monopolies like Facebook or Meta—I always call them Facebook—through these network effects.
The way you get industries with only a few players is very high entry costs. Cloud is like that. I think cloud is a good example of this. There are three, maybe four players in cloud. I think AI is the same, three, maybe four. The reason is it's so expensive. Running a cloud company requires so much expertise and so much capital. You have to put in all this capital. Besides putting in all this capital, you also have to have all these other things that require a lot of skill to achieve.
So if you go to someone and say, "I want to disrupt this industry, here's $100 billion." You'd be like, "Okay, I put in $100 billion, and I'm also betting you can do all these things other people have been doing." The result just lowers profit. Your entry impact is margin compression. So, we've always had this equilibrium in the economy with only a few players. Profits aren't astronomical. Margins aren't astronomical, but they're not zero. This is what we see with cloud. Cloud is very homogeneous.
Models are more differentiated than cloud. Everyone knows Claude is good at different things than GPT is good at, than Gemini is good at. It's not just that Claude is good at programming, and GPT is good at math and reasoning. It's more subtle than that. Models are good at different types of programming. Models have different styles. I think these things are actually quite different from each other, so I expect to see more differentiation than cloud.
Now, there's actually a counter-argument. That counter-argument is, if the process of producing models, if AI models themselves can do that, then that might diffuse throughout the economy. But that's not an argument for making AI models universally commoditized. That's kind of an argument for commoditizing the entire economy at once.
I don't know what happens in that world, where basically anyone can do anything, anyone can build anything, with no moats at all. I don't know, maybe we want that world. Maybe that's the ultimate state here. Maybe when AI models can do everything, if we've solved all the safety and security issues, that's one of the mechanisms where the economy flattens itself again. But that's a bit far after "A Country of Geniuses in a Datacenter."
Dwarkesh: Maybe a more refined way to express that underlying view is: 1) It seems AI research particularly depends on raw intelligence, which will be particularly abundant in an AGI world. 2) If you just look at today's world, it seems few technologies diffuse as quickly as AI algorithm progress. So this does suggest the industry is structurally diffusive.
Dario: I think coding is progressing fast, but I think AI research is a superset of coding, where some aspects aren't progressing that fast. But I do think, again, once we nail coding, once we have AI models progressing fast, then this will accelerate the ability of AI models to do everything else.
So while coding is progressing fast now, I think once AI models are building the next AI models and building everything else, the whole economy will develop at the same speed. But I'm a bit worried geographically. I'm a bit worried that just being close to AI, having heard of AI, might be a differentiator. So when I say 10-20% growth rates, what I'm worried about is Silicon Valley and the parts of the world socially connected to Silicon Valley might grow at 50%, while elsewhere isn't much faster than now. I think that would be a pretty bad world. So one thing I think about a lot is how to prevent this.
Dwarkesh: Do you think once we have this Country of Geniuses in a Datacenter, robotics will be solved very soon after? Because it seems like a big problem with robotics is that humans can learn how to teleoperate current hardware, but current AI models can't, at least not in a super efficient way. So if we have this ability to learn like humans, shouldn't that solve robotics immediately too?
Dario: I don't think it depends on learning like humans. It could happen in different ways. Again, we could have models train on many different video games, which is like robot control, or many different simulated robot environments, or just train them to control computer screens, and then they learn to generalize. So it will happen... it doesn't necessarily depend on human-like learning.
Human-like learning is one way it could happen. If the model is like, "Oh, I pick up a robot, I don't know how to use it, I learn," that could happen because we discovered continuous learning. That could also happen because we train models on many environments and then generalize, or it could happen because the model learns that in context length. Actually it doesn't matter which way.
If we go back to our discussion from an hour ago, that kind of thing could happen through several different paths. But I do think, for whatever reason, when models have these skills, robotics will be completely transformed—whether in the design of robots, because models will be much better at this than humans, or the ability to control robots. So we'll be better at building physical hardware, building physical robots, and we'll also get better at controlling them.
Now, does this mean the robotics industry will also generate trillions of dollars in revenue? My answer is yes, but with the same extremely fast but not infinitely fast diffusion. So will robotics be completely transformed? Yes, maybe add another year or two. That's how I think about these things.
Dwarkesh: That makes sense. There's a general skepticism about extremely fast progress. Here's my view. It sounds like you're going to solve continuous learning somehow in a few years. But like a few years ago people weren't talking about continuous learning, and then we realized, "Oh, why aren't these models as useful as they could be, even though they obviously pass the Turing test and are experts in many different fields? Maybe it's this thing." And then we solve this thing, and we realize, actually, there's another thing human intelligence can do that these models can't, and it's fundamental to human labor. So why not think there will be more things like this, where we discover more fragments of human intelligence?
Dario: Well, to be clear, I think continuous learning, as I said before, might not be an obstacle at all. I think we might get there just through pretraining generalization and RL generalization. I think there might not be such a thing at all. In fact, I'd point to the history of ML, where people raise something that looks like an obstacle, and then it dissolves in the "big compute blob."
People say, "How does your model track nouns and verbs?" "They can understand syntactically, but not semantically? It's just statistical correlation." "You can understand a paragraph, you can't understand a word. There's reasoning, you can't do reasoning." But suddenly, it turns out you can do code and math very well. So I think actually there's a stronger history that shows some of these things look like big problems, and then they kind of dissolve. Some of them are real. The need for data is real, maybe continuous learning is a real thing.
But again, I'd ground us in things like code. I think we might reach a point in a year or two where models can do SWE end-to-end. That's a complete task. That's a complete domain of human activity, and we just say models can do it now.
Dwarkesh: When you say end-to-end, do you mean setting technical direction, understanding the context of the problem, etc.?
Dario: Yes. I mean all of that.
Dwarkesh: Interesting. I feel like that's AGI-complete, which is maybe internally consistent. But it's not like saying 90% of code or 100% of code.
Dario: No, I gave this spectrum: 90% of code, 100% of code, 90% of end-to-end SWE, 100% of end-to-end SWE. New tasks get created for SWE. Eventually those get done too. There's a long spectrum, but we're traversing this spectrum very fast.
Dwarkesh: I do find it interesting that I've watched several podcasts you've done, and the host will be like, "But Dwarkesh wrote about continuous learning." It always makes me laugh out loud, because you've been an AI researcher for 10 years. I'm sure there's some feeling of, "Okay, so a podcast host wrote an article, and I get asked about this every interview."
Dario: The reality is we're all trying to figure this out together. There are aspects where I can see things others can't. These days that's probably more about seeing a bunch of things inside Anthropic and having to make a bunch of decisions, rather than me having some great research insight others don't have. I'm running a 2500-person company. Actually having specific research insights is much harder for me than it was 10 years ago or even two or three years ago.
Dwarkesh: As we move toward a world that can directly replace remote workers, is the API pricing model still the most reasonable? If not, what's the right way to price or serve AGI?
Dario: I think there will be many different business models being tried simultaneously here. I actually do think the API model is more durable than many imagine. One way I think about this is, if technology is progressing fast, if it's exponentially progressing, this means there's always a surface area of new use cases developed in the past three months. Any product surface you build always faces the risk of becoming irrelevant.
Any given product surface might make sense for a certain range of model capabilities. Chatbots have already encountered limits where making them smarter doesn't really help the average consumer that much. But I don't think that's a limit of AI models. I don't think that's evidence that models are good enough and them getting better is economically irrelevant. It's irrelevant for that particular product.
So I think the value of APIs is that APIs always provide an opportunity very close to the ground floor to build on the latest stuff. There will always be this frontier of new startups and new ideas that were impossible a few months ago and are now possible because models have progressed.
I actually predict it will coexist with other models, but we'll always have the API business model, because there's always a need for a thousand different people trying to experiment with models in different ways. 100 of them become startups, 10 of them become successful big startups. Two or three really end up being how people use a certain generation of models. So I basically think it will always exist.
At the same time, I'm sure there will be other models. Not every token output by a model is worth the same money. Think about when someone calls and says, "My Mac is broken," or whatever, and the model says, "Restart it." What's the value of those tokens. Someone hasn't heard it before, but the model has said it 10 million times. Maybe that's worth a dollar or a few cents or whatever. Whereas if the model goes to a pharmaceutical company and says, "Oh, you know, this molecule you're developing, you should take that aromatic ring off that end of the molecule and put it on this end. If you do that, amazing things will happen." Those tokens might be worth tens of millions of dollars.
So I think we'll definitely see business models that recognize this. At some point we'll see some form of "pay for results," or we might see labor-like compensation forms, that kind of hourly billing work. I don't know. I think because this is a new industry, many things will be tried. I don't know what will ultimately prove to be the right thing.
Dwarkesh: I agree with you that people have to try various things to figure out the best way to use this blob of intelligence. But I find Claude Code striking. I don't think in startup history there's been a single application as competitively intense as programming agents. Claude Code is the category leader here. This seems surprising to me. It didn't inherently have to be Anthropic that built it. I wonder if you have an explanation for why it had to be Anthropic, or how Anthropic ended up building an application that succeeded beyond just the underlying model.
Dario: Actually the way it happened was very simple, which is that we have our own coding models, and they're good at coding. Around early 2025, I said, "I think the time has come, if you're an AI company, you can achieve extraordinary acceleration of your own research by using these models." Of course, you need an interface, you need a tool to drive them. So I encouraged people internally. I didn't say this is something you have to use.
I just said people should try this. I think it was originally maybe called Claude CLI, and then the name eventually changed to Claude Code. Internally, it's something everyone uses, and it saw rapid internal adoption. I looked at it and said, "Maybe we should release this externally, right?" It saw such rapid adoption inside Anthropic. Coding is a lot of what we do. We have an audience of hundreds of people, representing in some ways at least an external audience.
So it looks like we have product-market fit. Let's release this thing. Then we released it. I think just that we're developing the models ourselves, and we ourselves know best how we need to use the models, I think this is creating this feedback loop.
Dwarkesh: I see. So, like, a developer at Anthropic is like, "Ah, if only it were better at this X thing." And then you bake that into the next model you build.
Dario: That's one version, but there's also just ordinary product iteration. We have a bunch of programmers at Anthropic who use Claude Code every day, so we get rapid feedback. This was more important early on. Now, of course, there are millions of people using it, so we get a bunch of external feedback too. But being able to get this rapid internal feedback is great. I think that's why we released a coding model and didn't release a pharmaceutical company. My background is biology, but we don't have the resources needed to release a pharmaceutical company.
Will Regulations Destroy the AGI Dividend?
Dwarkesh: Now let me ask you about making AI go well. It seems like whatever vision we have for AI going well has to be compatible with two things: 1) The ability to build and run AI is diffusing extremely rapidly, and 2) The amount of AI, both the quantity we have and their intelligence, will also increase very rapidly. This means many people will be able to build large amounts of unaligned AI, or AI that's just trying to expand its footprint or has weird psychology like Sydney Bing but is now superhuman. In a world where we have massive amounts of diverse AI running around, some of which are unaligned, what's our vision for an equilibrium?
Dario: I think in "The Adolescence of Technology," I'm skeptical about power balance. But specifically I'm skeptical that you have three or four companies all building models derived from the same thing, and they'll check each other. Or that any number of them will check each other. We might live in a world where offense is dominant, where one person or one AI model is smart enough to do something that damages everything else.
In the short term, we have a limited number of players right now. So we can start with a limited number of players. We need to put safeguards in place. We need to make sure everyone does the right alignment work. We need to make sure everyone has bio classifiers. These are things we need to do immediately.
I agree this doesn't solve the long-term problem, especially if the ability of AI models to make other AI models surges, then the whole thing could become harder to solve. I think in the long run, we need some kind of governance architecture. We need some governance architecture that both preserves human freedom and allows us to govern large numbers of human systems, AI systems, hybrid human-AI companies or economic units.
So we'll need to think about: How do we protect the world from bioterrorism? How do we protect the world from mirror life? We'll probably need some kind of AI monitoring system to watch all these things. But we also need to build it in a way that preserves civil liberties and our constitutional rights. So I think like anything else, this is a new safety landscape with a new set of tools and a new set of vulnerabilities.
My worry is that if we had 100 years for this to happen very slowly, we'd get used to it. We've gotten used to explosives existing in society, or various new weapons existing, or cameras existing. We'd get used to it in 100 years, and we'd make governance mechanisms. We'd make mistakes. My worry is just that this is happening too fast. So maybe we need to think faster about how to make these governance mechanisms work.
Dwarkesh: It seems like in a world where offense is dominant, over the course of the next century—the idea is that AI is making the progress that would happen in the next century happen in some period of five to ten years—even if humans are the only players, we'd still need the same mechanisms, or the power balance is equally hard to handle.
Dario: I think we have AI's suggestions. But fundamentally, this doesn't seem like a completely different ball game. If checks and balances are going to work, they'll work for humans too. If they don't work, they won't work for AI either. So maybe this dooms human checks and balances to fail too. Again, I think there's some way to make this happen. Governments around the world might have to work together to make this happen. We might have to discuss with AI about establishing social structures to make these defenses possible. I don't know. I don't want to say this is too far away in time, but it's too far away in technical capability, and might happen in a very short time, so it's hard for us to predict it in advance.
Dwarkesh: Speaking of government getting involved, on December 26, the Tennessee legislature introduced a bill that says, "It's a crime if a person knowingly trains artificial intelligence to provide emotional support, including through open-ended dialogue with users." Of course, one of the things Claude tries to be is a thoughtful, knowledgeable friend.
Overall, it looks like we're going to have this patchwork of state laws. A lot of the benefits that ordinary people might experience from AI will be cut off, especially as we get into the kinds of things you discuss in "Machines of Loving Grace": biological freedom, mental health improvements, etc. It seems easy to imagine a world where these get knocked down like whack-a-mole by different laws, and bills like this don't seem to solve the actual existential threats you worry about.
I'm curious to understand, in the context of things like this, Anthropic's position opposing a federal pause on state AI laws.
Dario: There are many different things happening at once. I think that particular law is stupid. It was clearly made by legislators who probably don't know much about what AI models can and can't do. They're like, "AI models serve us, that sounds scary. I don't want that to happen." So we don't actually support that.
But that's not what's being voted on. What's being voted on is: We're going to ban all states from regulating AI for 10 years, without any obvious federal AI regulatory plan, which requires Congress to pass, which is a very high bar. So this idea that we're going to ban states from doing anything for 10 years... people say they have a federal plan, but there's no actual proposal on the table. No actual attempt. Considering the serious dangers I listed in "The Adolescence of Technology" about bioweapons and bioterrorism autonomous risks, and the timelines we've been talking about—10 years is an eternity—I think that's crazy.
So if that's the choice, if you force us to choose, then we'll choose not to pause. I think the benefits of that position outweigh the costs, but if that's the choice, that's not a perfect position. Now, as for what we should do, what I would support is, the federal government should step in, not saying "states you can't regulate," but "we're going to do this, and states you can't deviate from this standard." I think as long as the federal government says: "Here's our standard. This applies to everyone. States can't do something different." That kind of preemption I can accept. If done in this right way, I'd support it.
But if the current idea is that states are saying: "You can't do anything, and we (the federal government) won't do anything either," this seems very unreasonable to us. I think this approach won't stand the test of time, and actually with all the opposition you're seeing, it's already starting to look outdated.
As for what we want, the things we've discussed are starting with transparency standards to monitor some of these autonomy risks and bioterrorism risks. As risks become more serious, as we get more evidence, I think we can take more targeted aggressive measures, like stipulating: "Hey, AI bioterrorism is really a threat. Let's pass a law requiring people to use classifiers." I can even imagine... but it depends. It depends on how serious the threat ultimately is. We can't be sure right now.
We need to proceed in an intellectually honest way, where we state in advance that currently risks haven't manifested. But I can certainly imagine, given the current pace of developments, later this year we might say: "Hey, this AI bioterrorism thing is really serious. We should act. We should write it into federal standards. If the federal government doesn't act, we should write it into state standards." I can totally see this happening.
What I'm worried about is a world where: if you consider the speed of progress you anticipate, and then consider the legislative lifecycle... as you said, because of diffusion lag, benefits come more slowly, so I think this patchwork of state laws, on the current trajectory, will really be obstructive. I mean, if having an emotional chatbot friend is something that drives people crazy, then imagine the real benefits we want ordinary people to be able to experience from AI. From improvements in health and healthspan, to mental health improvements, etc.
Yet at the same time, it seems like you think danger is already on the horizon, but I don't see that many... compared to AI dangers, this seems particularly harmful to AI benefits. So that might be where I think the cost-benefit analysis doesn't quite work.
Dwarkesh: There are a few things here. People talk about there being thousands of these state laws. First, the vast, vast majority of laws won't pass. Theoretically the world works a certain way, but just because a law is passed doesn't mean it will actually be enforced. The people enforcing it might think: "My god, this is so stupid. This means shutting down everything ever built in Tennessee." A lot of times, laws are interpreted in ways that make them less dangerous or harmful than they appear.
Of course, on the other hand, if you pass a law to stop bad things from happening; you run into the same problem.
Dario: My basic view is, if we could decide what laws to pass and how to do things—of course we're just one small voice among many—I'd deregulate a lot around AI health benefits. I'm less worried about chatbot laws. Actually, I'm more worried about drug approval processes, which I think AI models will greatly accelerate our ability to discover drugs, and the approval pipeline will get clogged. The approval pipeline won't be ready to handle everything pouring in.
I think regulatory process reform should tilt more toward the fact that we're going to have a lot of things coming through where their efficacy and safety will actually be very clear and straightforward, which is a beautiful thing, and very effective. Maybe we don't need all this superstructure built around it, which was designed for an era when that drug barely worked and often had serious side effects.
At the same time, I think we should dramatically strengthen legislation on safety and security. Like I said, starting with transparency is what I think is the view that tries not to hinder industry development, trying to find the right balance. I'm worried about this. Some people criticize my article, saying: "That's too slow. If we do that, AI dangers will come too fast."
Well, basically, I think the past six months and the next few months will be about transparency. Then, if and when we're more certain about these risks—I think as early as later this year we might be certain—if these risks appear, then I think we need to act very quickly in the areas where we actually see risks.
I think this is the only way to do it, which is to be flexible. The legislative process now usually isn't flexible, but we need to emphasize the urgency of this to everyone involved. That's why I'm sending this urgent message. That's why I wrote "The Adolescence of Technology." I want policymakers, economists, national security professionals, and decision-makers to read it so they hopefully act faster than they otherwise would.
Dwarkesh: Is there anything you can do or advocate for to make people more certain that AI benefits can be better realized? I feel like you've already worked with legislatures saying: "Okay, we're going to prevent bioterrorism here. We're going to increase transparency, we're going to strengthen protections for whistleblowers."
But I think, by default, the actual benefits we expect seem very fragile in the face of various moral panics or political economy issues.
Dario: Actually for developed countries, I don't quite agree with this view. I think in developed countries, markets work quite well. When something has a lot of money-making potential, and it's obviously the best available alternative, the regulatory system is actually quite hard to stop it. We've seen this with AI itself.
If we're talking about drug and technology benefits, I'm less worried about these benefits being blocked in developed countries. I'm a bit worried they're progressing too slowly. As I said, I do think we should try to speed up FDA approval processes. I do think we should oppose these chatbot bills you described. Individually, I oppose them. I think they're stupid.
But I actually think the bigger worry is developing countries, where markets don't work well, and we often can't build on existing technology. I'm more worried about those people being left behind. I'm also worried that even if treatments are developed, maybe someone in rural Mississippi can't access it well either. That's a smaller version of our worry in developing countries.
So what we've been doing is working with philanthropists. We work with people who provide medicines and health interventions to developing countries, sub-Saharan Africa, India, Latin America, and other developing regions of the world. I think this is the thing that won't automatically happen without intervention.
Claude's Constitution
Dwarkesh: You recently announced that Claude will have a constitution that conforms to a set of values, not necessarily just for the end user. I can imagine a world where if it were for the end user, it would preserve the power balance we have in today's world, because everyone has their own AI defending them. The ratio of bad people to good people stays the same.
This seems to work pretty well for our world today. Why not do that, and instead have a specific set of values that AI should follow?
Dario: I'm not sure I'd draw the distinction that way. There might be two related distinctions here. I think you're talking about a mix of both. One is, should we give the model a set of instructions about "do this" vs "don't do this"? The other is, should we give the model a set of principles about how to act?
This is purely something we've observed in practice and empirically. By teaching models principles and letting them learn from principles, their behavior is more consistent, easier to cover edge cases, and the model is more likely to do what people want it to do. In other words, if you give it a list of rules—"don't tell people how to hotwire car ignitions, don't speak Korean"—it doesn't really understand these rules, and it's hard to generalize from these rules. This is just a "do and don't" list.
But if you give it principles—it has some hard guardrails, like "don't make bioweapons," but—overall you're trying to understand what it should aim to do, how it should aim to operate. So purely from a practical perspective, this has proven to be a more effective way to train models. That's the rules versus principles tradeoff.
Then there's the other thing you're talking about, which is the tradeoff between corrigibility and intrinsic motivation. To what extent should the model be like a "straitjacket" that just directly follows the instructions of whoever gives it instructions, versus to what extent should the model have a set of internal values and do things on its own?
On that I'd actually say, everything about models is closer to the direction where it should primarily do what people want it to do. It should primarily follow instructions. We're not trying to build something that runs off to rule the world. We're actually very biased toward the corrigible side.
Now, we do say there are some things models won't do. I think we've said in the constitution in various ways, under normal circumstances, if someone asks the model to do a task, it should do that task. That should be the default. But if you ask it to do something dangerous, or something that hurts people, then the model won't want to do that. So I actually see it as a primarily corrigible model that has some limits, but these limits are principle-based.
Then the fundamental question is, how are these principles determined? This isn't an Anthropic-specific problem. This will be a problem for any AI company.
Dwarkesh: But because you're the ones who actually wrote down the principles, I can ask you this. Usually, constitutions are written down, carved in stone, and there's a process for updating and changing it, etc. In this case, it seems like it's a document written by Anthropic people that can be changed at any time, guiding the behavior of a system that will become the basis of a lot of economic activity. How do you think about how these principles should be set?
Dario: I think there might be three sizes of loops here, three ways to iterate. One is we iterate inside Anthropic. We train a model, we're not happy with it, we modify the constitution. I think doing this is good. Releasing public updates to the constitution every so often is good, because people can comment on it.
The second layer of loop is different companies have different constitutions. I think this is useful. Anthropic releases a constitution, Gemini releases a constitution, other companies release a constitution. People can look at them and compare. Outside observers can criticize and say: "I like this provision in this constitution, and that provision in that constitution." This creates a soft incentive and feedback for all companies to take the best and improve.
Then I think there's a third loop, which is society outside of AI companies, and not just commentators without hard power. There we've done some experiments. A few years ago, we did an experiment with the Collective Intelligence Project, basically polling people on what should be in our AI constitution. At the time, we incorporated some of those changes.
So you could imagine doing something similar with our new approach to constitutions. It's a bit harder because when the constitution is a "do and don't" list, this approach was easier to take. At the principle level, it has to have a certain coherence. But you can still imagine getting views from a wide variety of people.
You could also imagine—this is a wild idea, but this whole interview is about wild ideas—representative government systems have inputs. I wouldn't do this today, because legislative processes are too slow. This is exactly why I think we should be careful about legislative processes and AI regulation.
But in principle there's no reason you couldn't say: "All AI models must have a constitution, it starts with these things, and then you can append other things later, but this special chapter must take precedence." I wouldn't do that. That's too rigid, sounds too prescriptive, like what I consider overly aggressive legislation. But it is something you could try to do.
Is there some less hard-nosed version? Maybe. I really like the second control loop.
Dwarkesh: Obviously, this isn't how actual government constitutions work or should work. There isn't this vague sense that the Supreme Court will feel how people feel—what the vibe is—and update the constitution accordingly. For actual government, there's a more formal, procedural process.
But you have a vision of competition between constitutions, which is actually very reminiscent of what some libertarian charter city people used to say, about archipelagos of different governments. There would be choice between who can run most effectively and where people are happiest. In some sense, you're recreating that archipelago utopian vision.
Dario: I think that vision has things to recommend it, and also things that will go wrong. It's an interesting, in some ways convincing vision, but there are things that will go wrong that you didn't expect. So I like the second loop too, but I think the whole thing has to be some mix of first, second, and third loops, depending on proportion. I think that has to be the answer.
Dwarkesh: When someone eventually writes the equivalent of "The Making of the Atomic Bomb" for this era, what's the thing that's hardest to collect in the historical record, that they're most likely to miss?
Dario: I think there are a few things. One is, at every moment of this exponential growth, the extent to which the outside world doesn't understand it. This is a bias that often exists in history. Anything that actually happens looks inevitable in retrospect.
When people look back, it will be hard for them to put themselves in the shoes of people who were actually betting on this thing that wasn't inevitable happening. We had arguments like the ones I made for scaling or continuous learning being solved. Some of us internally thought there was a high probability of this happening, but the outside world just didn't act on it.
I think its weirdness, and unfortunately its closedness... if we're a year or two away from it happening, the average person on the street has no idea. This is one of the things I'm trying to change through memos, through talking to policymakers. I don't know, but I think that's just a crazy thing.
Finally, I'd say—this probably applies to almost every crisis moment in history—how absolutely fast it happened, and how everything happened at the same time. Decisions you might think were carefully calculated were actually decisions you had to make, and then you had to make 30 other decisions the same day, because everything was happening so fast.
You don't even know which decisions will end up being critical. One of my worries—though this is also an insight into what's happening—is that some very critical decisions will be someone walking into my office saying: "Dario, you have two minutes. Should we do A or B on this?"
Someone gives me this random half-page memo asking: "Should we do A or B?" I say: "I don't know. I'm going to lunch. Do B." And that ends up being the most important thing ever.
Dwarkesh: Last question. Usually tech CEOs don't write 50-page memos every few months. It looks like you've successfully built a role for yourself, and built a company around you, that's compatible with this more intellectual type of CEO role. I want to understand how you structured this. How does that work? Do you just disappear for a few weeks and then tell your company, "Here's the memo. This is what we're doing"? There are also reports that you write a lot of this stuff internally.
Dario: For this particular one, I wrote it during winter break. It's hard for me to find time to actually write it. But I think about this in a broader way. I think this has to do with company culture. I might spend a third, maybe 40% of my time making sure Anthropic's culture is good.
As Anthropic gets bigger, it's getting harder to directly participate in training models, releasing models, building products. There are 2500 people. I have certain intuitions, but it's hard to participate in every detail. I try as much as possible, but one thing that's very leveraged is making sure Anthropic is a good place to work, people like working there, everyone sees themselves as team members, everyone is working together and not against each other.
We see with the growth of some other AI companies—not naming names—we've started to see disconnection and people fighting each other. I'd argue there was a lot of this even from the beginning, but it's gotten worse. I think we've done very well in keeping the company united, even if not perfectly, making everyone feel the mission, we're sincere about the mission, everyone believes that others there are working for the right reasons. We're a team, people aren't trying to get promoted at others' expense or stabbing each other in the back, and again, I think this happens a lot at some other places.
How do you do this? It's many things. It's me, it's Daniela who runs the company day-to-day, it's the co-founders, it's the other people we hire, it's the environment we try to create. But I think an important part of culture is, other leaders too, but especially me, have to articulate what the company is about, why it's doing what it's doing, what its strategy is, what its values are, what its mission is, what it stands for.
When you reach 2500 people, you can't do this person by person. You have to write, or you have to speak to the whole company. That's why I speak for an hour in front of the whole company every two weeks. I wouldn't say I write articles internally. I do two things. First, I write this thing called DVQ, Dario Vision Quest.
I didn't name it that. That's the name it got, which is one of the names I want to push back against, because it sounds like I'm going to consume peyote or something. But the name stuck. So I show up every two weeks in front of the company. I have a three- or four-page document, and I just talk about three or four different topics, about what's happening internally, the models we're producing, products, the external industry, the whole world's relationship with AI, and geopolitics, etc. Just a mix of these.
I speak very honestly, and I say: "Here's what I'm thinking, here's what Anthropic leadership is thinking," and then I answer questions. This direct connection has a lot of value, which is hard to achieve when you pass things down a chain of six levels. A large portion of the company comes, either in person or virtually. This really means you can communicate a lot.
The other thing I do is I have a channel in Slack where I just write a bunch of stuff and comment frequently. Usually it's in response to things I see in the company or questions people ask. We do internal surveys, there are things people care about, so I write them down. I'm very honest about these things. I just say it very directly.
The key is to build a reputation for telling the company the truth, being matter-of-fact, admitting problems, avoiding corporate-speak, that defensive communication that's usually necessary in public, because the world is big and full of people who maliciously interpret things. But if you have a company of people you trust, and we try to hire people we trust, then you can really be completely unfiltered.
I think this is a huge advantage for the company. It makes it a better workplace, it makes people more than the sum of their parts, and increases the likelihood that we accomplish our mission, because everyone is aligned on the mission, and everyone is debating and discussing how best to accomplish it.
Dwarkesh: Well, as an external version of "Dario Vision Quest," we have this interview. This interview is kind of like that. This was fun, Dario. Thank you for the interview.
Dario: Thank you, Dwarkesh.
Thank you for reading!