Anthropic's leader in a major interview: AI is at the tail end of exponential growth, and in 2026, we will welcome the "genius nation in data centers," with revenue skyrocketing at a speed of 10 times

On the eve of an exponential explosion in AI technology, Dario Amodei, the head of Anthropic, made a shocking prediction that has reverberated throughout the industry: we are in the "twilight of exponential growth," and by as early as 2026, humanity will welcome a "nation of geniuses in data centers" composed of tens of thousands of top minds.

Recently, in a rare deep interview with Dwarkesh Patel, Anthropic's CEO Dario Amodei disclosed the company's astonishing revenue growth expectations and elaborated on the timeline for AGI (Artificial General Intelligence), the financial logic of computing power investments, and geopolitical risks. Amodei believes that AI technology is at a critical point of transition from quantitative to qualitative change, and the next 2-3 years will determine the trajectory of humanity for the next two hundred years.

(Dario Amodei, CEO of Anthropic, guest on Dwarkesh Patel's podcast)

AI is at the End of Exponential Growth

At the beginning of the interview, Dario Amodei pointed out that we are nearing the end of the AI exponential growth curve, while the world has yet to fully perceive this qualitative change.

From GPT-1 to today's professional-grade models, AI has transitioned from "a smart high school student" to "PhD level," even surpassing in fields like programming and mathematics. The underlying expansion laws have never failed, and the investment in computing power and data continues to yield clear returns.

The magic of exponential growth lies in the explosive end phase. Dario stated that Anthropic's annual revenue has seen a tenfold leap, Claude Code has significantly boosted engineer productivity, and the rapid breakthroughs in model context length and generalization capabilities all signal that "the end is near." This growth is not just a stacking of parameters but an upgrade of the essence of intelligence—from data fitting to autonomous generalization, AI is completing the last few critical pieces of the capability puzzle.

"A Nation of Geniuses in a Datacenter": Redefining 2026

Amodei introduced a highly impactful concept during the interview—"A Country of Geniuses in a Datacenter." He reviewed the technological evolution over the past three years and believes that AI models have evolved from "smart high school students" to "professionals."

He boldly predicts that by 2026 or 2027, the intelligence level, depth of knowledge, and logical reasoning ability exhibited by a single model will not only be equivalent to that of a Nobel Prize winner but will represent a collective of tens of thousands of top geniuses working in synergy.

Regarding the certainty of this timeline, Amodei expressed high confidence:

"I am 90% confident that this vision will be realized within 10 years; for it to happen in the next 1-2 years, I believe the possibility is 50/50." He pointed out that the only variable may come from geopolitical disasters (such as disruptions in the chip supply chain) or severe social unrest.

Revenue Surge: The "Terrifying" Curve from $100 Million to $10 Billion

In terms of financial data that the market is most concerned about, Amodei revealed Anthropic's astonishing growth curve. He disclosed that the company's revenue is experiencing a "bizarre 10x per year growth." Amodei candidly stated 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 billion to $10 billion. This exponential growth roughly aligns with my expectations, and even in the first month of this year, we added several billion dollars in revenue."

Amodei emphasized that despite the impact of the lagging speed of economic diffusion, the adoption of AI by enterprises requires a lengthy process of legal review and compliance checks, but the improvement in the technology's capabilities is driving this crazy growth curve.

Computing Power Gamble and Bankruptcy Risk: The CEO's Financial Balancing Act

Given such a certain technological outlook, why not borrow trillions of dollars now to stockpile chips? Amodei provided a highly pragmatic financial explanation: the expansion of computing power must be linked to revenue growth and forecasting accuracy, or it will face catastrophic risks.

"If I predict a trillion-level demand in 2027 and pre-purchase $1 trillion worth of computing power, but if the demand explosion is delayed by a year, or the growth rate slightly drops from 10x to 5x, no hedging measures can prevent the company from going bankrupt," Amodei explained, noting that this investment return based on the "law of logarithmic returns" requires precise calculations.

He pointed out that Anthropic's current strategy is "responsibly aggressive," meaning that the scale of computing power invested is sufficient to capture significant upside potential, but if the market explosion is delayed, the company can still survive due to its high gross margins and cash flow from its enterprise-level business.

He expects that Anthropic is likely to achieve profitability around 2028, at which point AI will become one of the most profitable industries in history.

The Endgame of Software Engineering: From Writing Code to Replacing Engineers

In terms of specific landing scenarios, Amodei views programming as the first fortress that AI will conquer. He divides the evolution of AI in the field of software engineering into three stages:

• Stage One: The model writes 90% of the lines of code (already achieved).
• Stage Two: The model handles 90% of end-to-end tasks, such as fixing bugs, configuring clusters, and writing documentation.
• Stage Three: The model possesses "work 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 responsibilities of a senior software engineer.

"This does not mean that engineers will be unemployed, but rather a tremendous explosion of productivity. The current models can do more than just complete code; they can directly take over high-difficulty tasks such as writing GPU kernels."

The full translation of the in-depth interview with Anthropic CEO Dario Amodei is as follows:

What exactly are we expanding?

Dwarkesh Patel (hereinafter referred to as Dwarkesh): We talked about this three years ago. In your view, what has been the biggest update in the past three years? What are the biggest differences in feeling between now and then?

Dario Amodei (hereinafter referred to as Dario): From a macro perspective, the exponential growth of underlying technology has largely met my expectations, although there is a year or two of discrepancy. I'm not sure if I predicted the specific development direction in coding. But when I look at this exponential curve, it roughly aligns with my expectations for model progress—from a smart high school student to a smart college student, and then starting to do PhD-level and professional work, even surpassing that level in the coding field. Although the cutting-edge progress is somewhat uneven, it generally meets expectations.

What is 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 just crazy—whether it's people inside or outside the field—talking about those outdated political hot topics while we are nearing 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 is scaling, and why is it effective?" Now I have a similar question, but it feels more complex.

At least from the public's perspective, three years ago there were well-known public trends, with computational amounts spanning multiple orders of magnitude, and 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 principles behind it are. Is this teaching the model skills? Is it teaching meta-learning? What are the current scaling assumptions?

Dario: Actually, my assumptions are the same as they were in 2017.

I think I talked about this last time, but I wrote a document called "The Big Compute Hypothesis." It wasn't specifically about scaling language models. I wrote it when GPT-1 had just come out; it was just one of many things.

At that time, there was robotics. People were trying to study reasoning as something independent of language models, as well as the kind of reinforcement learning scaling seen in AlphaGo and OpenAI's Dota. People remember DeepMind's StarCraft, AlphaStar.

This is a more general document. Rich Sutton published "The Bitter Lesson" years later. The hypothesis is basically the same.

It states that all the intelligence, all the technology, all the ideas of "we need new methods to do something" are not that important. Only a few things matter. I think I listed seven

First, how much raw computational power you have. Second, the quantity of data. Third, the quality and distribution of the data. It needs to be widely distributed. Fourth, how long you train. Fifth, you need a target function that can be scaled to the extreme. The pre-training target function is such a target function. Another is the reinforcement learning target function, which states that you have a goal, and you need to achieve that goal. Within this, there are objective rewards, like those seen in mathematics and coding, as well as more subjective rewards, like those seen in RLHF or higher-order versions.

Then the sixth and seventh items are about normalization or conditioning, just to achieve numerical stability, so that large computational blocks can flow in this laminar way without encountering issues.

This is the assumption, and it is one I still hold to this day. I haven't seen much that contradicts it.

The pre-training scaling law is an example we have observed. These laws have been ongoing. They are now widely reported, and we feel good about pre-training. It continues to yield benefits for us.

What has changed is that we now see the same situation emerging with reinforcement learning. We see a pre-training phase followed by a reinforcement learning phase based on that. For reinforcement learning, it is actually the same.

Even other companies have published content in some of their releases saying, "We trained models on math competitions—AIME or other competitions—the model's performance is logarithmically linearly related to the time we trained it." We have seen this, and not just in math competitions. This applies to a variety of reinforcement learning tasks.

We see that the scaling of reinforcement learning is the same as the scaling we see in pre-training.

Dwarkesh: You mentioned Rich Sutton and the "bitter lesson." I interviewed him last year, and he actually disagrees with large language models quite strongly. I don't know if this is his viewpoint, but to paraphrase his opposition: the core of human learning doesn't require all these billions of dollars of data and computation, nor these customized 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 embed these skills suggests that we are actually missing a core human learning algorithm. So we are scaling the wrong thing. This does raise a question. If we think there will be something akin to human-like instant learning capabilities, why are we doing all this reinforcement learning scaling?

Dario: I think this mixes together several things that should be thought of separately. There is indeed a real puzzle here, but it may not be important. In fact, I guess it may not be important.

Here is an interesting thing. Let me temporarily set reinforcement learning aside because I actually think saying that reinforcement learning is different from pre-training on this issue is misleading.

If we look at pre-training expansion, it's very interesting what Alec Radford did with GPT-1 in 2017. The models before GPT-1 were trained on datasets that did not represent a wide distribution of text. 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 portion of the text you could get.

At that time, it was about a billion words or so, so it represented a fairly narrow distribution of a small dataset of what you could see in the world. Its generalization ability was poor. If you did better on a certain fan fiction corpus, it wouldn't generalize well to other tasks. We had all these metrics. We had various measures of how it performed in predicting all other types of text.

Only when you train on all tasks on the internet—when you do general internet scraping from something like Common Crawl, or scrape links from Reddit (which is what we did for GPT-2)—do you start to gain generalization ability.

I think we see the same thing in reinforcement learning. We started with simple reinforcement learning tasks, like training on math competitions, and then moved to broader training involving code and so on. Now we are moving to many other tasks.

I think we will increasingly gain generalization ability. So this somewhat eliminates the distinction between reinforcement learning and pre-training.

But either way, there is a puzzle, which is that in pre-training we used trillions of tokens. Humans do not see trillions of words. So there is indeed a sample efficiency difference here. There is indeed something different.

Models start from scratch and need more training. But we also see that once they are well trained, if we give them a million long contexts—the only thing that hinders long contexts is reasoning—they are 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 pre-training is not like the process of human learning, but it is somewhere between the process of human learning and the process of human evolution.

Much of our prior knowledge comes from evolution. Our brains are not just a blank slate. There have been entire books written about this. Language models are more like a blank slate. They really start from random weights, while the human brain has all these areas connected to all these inputs and outputs from the beginning.

Perhaps we should view pre-training—and reinforcement learning—as existing in an intermediate space between human evolution and human immediate learning. We should see the contextual learning that models undergo as something between human long-term learning and short-term learning.

So there is this hierarchy. There is evolution, long-term learning, short-term learning, and human immediate responses. The various stages of large language models exist on this spectrum, but not necessarily at exactly the same points.

There is no analogy corresponding to certain human learning patterns; large language models fall between these points. Does that make sense?

Dwarkesh: It's meaningful, although some things are still a bit confusing. For example, if the analogy is that this is like evolution, so low sample efficiency is acceptable, then if we want to obtain super sample-efficient agents from contextual learning, why bother building all these reinforcement learning environments?

Some companies seem to be working on teaching models how to use this API, how to use Slack, how to use other things. If such instantly learning agents are emerging or have already emerged, why is there so much emphasis on this? It confuses me.

Dario: I can't speak for others' focus. I can only talk about how we think.

The goal is not to teach the model every possible skill in reinforcement learning, just as we don't do that in pre-training. In pre-training, we are not trying to expose the model to every possible way words can combine.

Instead, the model is trained on a lot of things and then generalizes in pre-training. This is the transition I saw up close from GPT-1 to GPT-2. The model reached a point. I had moments where I thought, "Oh yes, you just give the model a column of numbers—this is the price of the house, this is the square footage of the house—and the model can complete the pattern and perform linear regression."

While it's not great, it did it, and it had never seen that exact thing before.

So in terms of building these reinforcement learning environments, the goal is very similar to what pre-training did five or ten years ago. We are trying to obtain a large amount of data, not because we want to cover specific documents or specific skills, but because we want to generalize.

I think the framework you proposed clearly makes sense. We are moving towards AGI. At this point, no one disagrees that we will achieve AGI this century. The key is that you say we are approaching the end of the exponential curve.

Others seeing this might say, "We have been making progress since 2012, and by 2035 we will have human-like agents."

Clearly, we see what evolution has done in these models, or what learning throughout a human life has done. I want to understand what you see that makes you think this will happen in a year rather than ten years.

Is Expansion an Excuse?

Dario: There are two ways to put this, one stronger and one weaker.

Starting with the weaker statement, when I first saw expansion in 2019, I was uncertain. It was a 50/50 thing. I thought I saw something. My statement is that it is more likely than anyone thinks. Maybe there’s a 50% chance it will happen.

Regarding what you said, that we will reach what I call the "genius nation in the data center" in ten years, I am 90% confident about that. It's hard to exceed 90% because the world is so unpredictable. Perhaps the irreducible uncertainty makes us reach 95%, and you will encounter situations like internal turmoil in multiple companies, Taiwan being invaded, all foundries being bombed, etc

Dwarkesh: Now you've cursed us, Dario.

Dario: You can build a 5% world where things are delayed by ten years.

There's another 5%, which is my confidence in verifiable tasks. For coding, aside from that irreducible uncertainty, I think we will reach our goals in a year or two. We can't possibly still be at a point where we haven't achieved end-to-end coding in ten years.

My little bit of fundamental uncertainty, even on a long time scale, is about those unverifiable tasks: planning Mars missions; making some foundational scientific discoveries like CRISPR; writing novels.

These tasks are hard to verify. I'm almost certain we have a reliable path to get there, but there's a little bit of uncertainty right there.

On a ten-year timeline, I'm 90% confident, which is about the most certain level you can achieve. I think it's crazy to say this won't happen by 2035. In some rational world, this would be considered an outlier view.

But the emphasis on verification implies a lack of belief in the generalization of these models. If you think about humans, we excel at both things that can yield verifiable rewards and those that cannot.

Dario: No, that's why I'm almost certain. We've seen quite a bit of generalization from verifiable things to unverifiable things. We've seen that.

But it seems you emphasize that this is a spectrum that will split, and we will see more progress in certain areas. That doesn't seem like a way for humans to get better.

Dario: The world we can't reach is one where we do all the verifiable things. Many of those will generalize, but we haven't fully gotten there. We haven't completely filled the other side of the box. It's not a binary thing.

Even if the generalization is weak, if you can only do verifiable areas, I'm not sure if you can automate software engineering in such a world. In a sense, you are a "software engineer," but part of the job of 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 engineers do involve design documents and other similar things. The models are already quite good at writing comments.

Again, the claim I'm making here is much weaker than I believe to distinguish the two things. We are almost there in software engineering.

Dwarkesh: By what standard? One standard is how many lines of code AI has written.

If you consider other productivity improvements in the history of software engineering, compilers wrote all the lines of software. There is a difference between how many lines were written and how much productivity improved. What does "we are almost there" mean?

Dario: The increase in productivity is not just about how many lines AI has written.

Dwarkesh: I actually agree with your point.

Dario: I made a series of predictions about code and software engineering. I think people repeatedly misunderstand them. Let me outline this spectrum.

About eight or nine months ago, I said AI models would write 90% of code lines within three to six months. This has happened, at least in some areas. It has happened at Anthropic and with many downstream users of our models.

But this is actually a very weak standard. People think I’m saying we don’t need 90% of software engineers. That’s far from the truth.

The spectrum is: 90% of code written by models, 100% of code written by models. There’s a significant difference in productivity.

90% of end-to-end software engineering tasks—including compiling, setting up clusters and environments, testing functionality, writing memos, etc.—are completed by models.

100% of today’s software engineering tasks are completed by models. Even if that happens, it doesn’t mean software engineers will be unemployed. They can do new, more advanced things; they can manage.

Then further along the spectrum, the demand for software engineers decreases by 90%, and I think that will happen, but it’s a spectrum.

I wrote about this in "The Adolescence of Technology," and I experienced this spectrum through agriculture.

Dwarkesh: I actually completely agree with your point. These are very different benchmarks from each other, but we are moving through them at an ultra-fast pace.

Part of your vision is that the jump from 90 to 100 will happen quickly and will bring huge productivity gains. But what I notice is that even in greenfield projects, when people start with Claude Code or something else, people report launching a lot of projects... Are we seeing a renaissance of software in the outside world, all these new features that wouldn’t exist otherwise? At least so far, it seems we haven’t seen that.

So it really makes me wonder. Even if I never need to intervene with Claude Code, the world is complex. Work is complex. On a closed-loop self-contained system, whether it’s just writing software or something else, how much broader benefit will we see from that?

Perhaps this should dilute our estimates of the "country of geniuses."

Dario: I simultaneously agree with your point, which is why these things won’t happen immediately, but at the same time, I think the effects will be very fast.

You can have these two extremes. One is that AI will not make progress. It’s slow. It will always diffuse into the economy.

Economic diffusion has become one of those buzzwords, a reason we won’t make progress with AI or that AI progress doesn’t matter.

The other axis is that we will achieve recursive self-improvement, the whole thing. Can’t you just draw an exponential line on the curve?

In the many nanoseconds after we achieve recursion, we will have a Dyson Sphere around the sun. I am completely being sarcastic about this point, but there are these two extremes.

But what we've seen from the beginning, at least if you look inside Anthropic, is this strange 10x revenue growth every year.

So in 2023, it's from zero to 100 million dollars. In 2024, it's from 100 million dollars to 1 billion dollars. In 2025, it's from 1 billion dollars to 9-10 billion dollars.

Dwarkesh: You should buy 1 billion dollars of your own products so that you can...

Dario: In the first month of this year, that exponential curve... you would think it would slow down, but we added several billion dollars in revenue again in January.

Obviously, that curve cannot continue forever. GDP is only so large. I even suspect it will curve this year, but it's a fast curve. It's a very fast curve. I bet even if it scales to the entire economy, it will maintain a fairly rapid pace.

So I think we should consider this middle world, where things are very fast, but not instantaneous; they take time because of economic diffusion, because of the need for closed loops.

Because it's cumbersome: "I have to do change management within my business... I set this up, but I have to change the security permissions of this to make it really work... I have this old software that checks the model before compiling and releasing, and 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 this kind of idea: there is a fast exponential curve, which is the capability of the model. Then there is another fast exponential curve, which is downstream, which is the model diffusing into the economy.

It's not instantaneous, it's not slow, it's much faster than any previous technology, but it has its limits.

When I look inside Anthropic, when I look at our customers: rapid adoption, but not infinitely fast.

Dwarkesh: Can I try a bold point?

Dario: Sure.

Dwarkesh: I feel like diffusion is the excuse people use. When the model can't do something, they say, "Oh, but it's a diffusion problem."

But you should compare it to humans. You would think the inherent advantages of AI would make the onboarding diffusion of new AI much easier than that of new humans.

AI can read your entire Slack and your drive in minutes. They can share all the knowledge owned by other copies of the same instance.

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. We pay humans over 50 trillion dollars in wages because they are useful, even though in principle integrating AI into the economy should be much easier than hiring humans

Expansion cannot truly explain.

Dario: I believe diffusion is very real and not entirely related to the limitations of AI models.

Again, some people use diffusion as a buzzword to say it's no big deal. I'm not talking about that. I'm not saying that AI will diffuse at the speed of previous technologies.

I believe the diffusion speed of AI will be much faster than previous technologies, but not infinitely fast.

I'll give just one example. There’s Claude Code. Claude Code is very easy to set up. If you are a developer, you can start using Claude Code right away.

Developers at large enterprises have no reason not to adopt Claude Code as quickly as individual developers or startup developers.

We are doing everything we can to promote it. We are selling 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 technologies.

But again, it takes time. Any given feature or any given product, like Claude Code or Cowork, will be adopted by individual developers on Twitter months before it is adopted by large enterprises engaged in food sales.

There are just many factors. You have to go through legal reviews, you have to configure it for everyone. It has to pass security and compliance.

Company leaders are further away from the AI revolution; they have vision, but they have to say, "Oh, it makes sense for us to spend $50 million. This is what Claude Code is. This is why it helps our company. This is why it makes us more productive."

Then they have to explain it 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 are doing everything we can to grow Anthropic's revenue by 20 or 30 times a year, rather than 10 times.

Again, many enterprises are just saying, "This is so productive. We will shortcut our usual procurement process."

Their actions are much faster than when we try to sell them a regular API, which many enterprises are using. Claude Code is a more compelling product, but it is not an infinitely compelling product.

I believe even AGI or powerful AI or "the genius nation in the data center" will not be an infinitely compelling product. It will be a sufficiently compelling product that may achieve 3-5 times or 10 times growth per year, even at a scale of hundreds of billions, which is very hard to achieve and has never been done historically, but not infinitely fast.

Dwarkesh: I think there will be a slight slowdown. Maybe that’s not your claim, but sometimes people talk about it as if, "Oh, the capability is there, but because of diffusion... otherwise we are basically at AGI." "

Dario: I don't believe we are basically at AGI. I think if you have a "nation of geniuses in data centers"...

If we had a "nation of geniuses in data centers," we would know. If you have a "nation of geniuses in data centers," 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: Back to specific predictions... because there are so many different things to disambiguate, 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 three years later. You were right. You said, "We should expect systems that, if you talk to them for an hour, it's hard to distinguish them from well-educated humans." I think you were right.

I feel mentally unsatisfied because my internal expectation is that such systems could automate most white-collar jobs. So discussing the actual ultimate capabilities you want from such systems might be more productive.

I would 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 expect to consider soon. Maybe I will ask this question in a work context that I know well, not because it's the most relevant job, but simply because I can assess claims about it.

Take video editing, for example. I have video editors. Part of their job involves understanding our audience's preferences, understanding my preferences and tastes, and the different trade-offs we have. They build an understanding of the context over many months.

After six months of work, the skills and capabilities they have, when can we expect a model that can instantly master that skill?

Dario: I think what you're talking about is this three-hour interview we're doing. Someone will come in, and someone will edit it. They will say, "Oh, I don't know, Dario scratched his head, we can cut that out."

"Zoom in on that." "There's this long discussion that's not very interesting to people. There's something else that's more interesting to people, so let's do this edit."

I think a "nation of geniuses in data centers" will be able to do that. The way it will be able to do that is it will have general control over the computer screen. You will be able to input this.

It will also be able to browse the internet on the computer screen, look at all your previous interviews, see people's comments on your interviews on Twitter, talk to you, ask you questions, talk to your staff, look at your editing history, and complete the work from there.

I think it depends on a few things. I think one of the things that actually hinders deployment is reaching a level where the model is truly proficient at using the computer

We have seen this rise in benchmarks, which are never perfect measures. But I think when we first released computer usage a year and a quarter ago, OSWorld was around 15%.

I don't remember the exact number, but we have climbed from there to 65-70%. There may be harder metrics, but I believe computer usage must pass a reliability point.

Dwarkesh: Before you move on to the next point, can I follow up? For years, I have been trying to build different internal LLM tools for myself.

Typically, I have these text input and text output tasks, which should be the core capabilities of these models. However, I still hire humans to do them.

If it's something like "what is the best segment in this text," maybe the large language model has done seven-tenths of the work on that. But there isn't that continuous way for me to interact with them to help them do better at the job, like I can with human employees.

That missing capability, even if you solve computer usage, will still hinder my ability to outsource actual work to them.

Dario: This goes back to what we talked about earlier regarding learning on the job. This is very interesting. I don't think people would say that learning on the job is the reason coding agents can't complete everything end-to-end.

They have been 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 saying, "This GPU kernel, this chip, I used to write myself. I just let Claude do it." There is a huge increase in productivity.

When I look at Claude Code, the familiarity with the codebase or the model doesn't feel like someone who has worked at the company for a year; that's not high on the list of complaints I see.

I think what I'm saying is we are taking a different path.

Dwarkesh: Don't you think coding is like this because there is 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 do not have.

Dario: But when you say that, you imply 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 is written down or whether it is available—of a case where everything you need to know is obtained from the context window.

What we think of as learning—"I start this job, and I need six months to understand the codebase"—the model just does it 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 significant study where they had experienced developers try to close pull requests in repositories they were familiar with. Those developers reported an uplift. They reported feeling more productive using these models

But in fact, if you look at their output and the actual content that has been merged back, there is a 20% decline. The results from using these models indicate a decrease in productivity.

So I am trying to reconcile people's qualitative feelings about these models with the following points: 1) At a macro level, where is the renaissance of this software? And then 2) When people conduct these independent assessments, why are we not seeing the productivity gains we expect?

Dario: Internally at Anthropic, this is really unequivocal. We are under incredible commercial pressure, and we make it harder for ourselves because of all the safety work we do; I think we do more than other companies.

The pressure to survive economically while maintaining our values is incredible. We are striving to maintain this 10x revenue growth curve.

There is no time for nonsense. No time to feel productive when we are actually not.

These tools make us significantly more productive. Why do you think we are concerned about competitors using these tools? Because we believe we are ahead of the competition.

If this secretly reduced our productivity, we wouldn't be going through all this trouble. We can see the ultimate productivity every few months in the form of model releases.

There is no room for self-deception in this. These models make you more productive.

1. The feeling of productivity is qualitatively predicted by such research. But 2) If I only look at the final output, it is clear that you are making rapid progress.

But the idea should be that through recursive self-improvement, you create a better AI, which helps you build a better next AI, and so on and so forth.

Instead, what I see is—if I look at you, OpenAI, DeepMind—people are just moving positions on the podium every few months. Maybe you think that will stop because you won or something.

But if the last coding model had these huge productivity gains, why don't we see those who have the best coding models having this persistent advantage?

Dario: I think my model of the situation is that there is a gradually growing advantage. I would say that the total factor acceleration given by coding models now might be, I don’t know, 15-20%. That’s my view. Six months ago, it might have been 5%.

So it’s okay. 5% doesn’t count. It has just reached a point now where it is one of several factors, somewhat important. That will continue to accelerate.

I would say six months ago, there were several companies roughly at the same point because it wasn’t a significant factor, but I think it is starting to accelerate more and more.

I would also say that there are several companies writing models for code, and we are not entirely good at stopping some of those other companies from using our models internally.

So I think everything we see is consistent with this snowball model. Again, the theme I see in all of this is that it’s all a soft takeoff, a soft, smooth exponential curve, although the exponential curve is relatively steep

So we see this snowballing momentum, it’s like 10%, 20%, 25%, 40%. As you progress, according to Amdahl's Law, you have to clear away everything that prevents you from closing the loop.

But this is one of the biggest priorities internally at Anthropic.

Stepping back, when are we talking about getting this on-the-job learning?

It seems that the point you’re making about coding is that we actually don’t need on-the-job learning.

You can have huge productivity gains, you can bring potential trillions of dollars in revenue to AI companies without this fundamental human instant learning capability.

Maybe that’s not your claim, 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, and then over time they built context and understanding.”

It’s actually hard to define what we’re talking about here. But they got something, and now they are a powerful force, and they are so valuable to us.

If AI doesn’t develop this instant learning capability, I’m a bit skeptical that we would see the huge changes in the world without that capability.

Dario: I think there are two things here. The current state of technology.

Again, we have these two phases. We have the pre-training and reinforcement learning phases, where you throw a bunch of data and tasks into the model, and then they generalize.

So it’s like learning, but it’s more like learning from more data rather than learning over the lifetime of a human or a model.

So again, this sits between evolution and human learning. But once you’ve learned all these skills, you have them.

It’s like pre-training, like the model knows more; if I look at a pre-trained 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 broader than mine. So I think even just that could potentially get us to a point where the model is better at everything.

We also have, again, just by extending the types of existing setups, contextual learning. I would describe it as somewhat like human on-the-job learning, but a bit weaker, a bit more short-term.

You see contextual 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 how long it takes a model to read 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 to the "genius nation in the data center."

I’m not sure, but I think they would get you a large part of the way there. There may be gaps, but I certainly think that for now, it’s enough to generate trillions of dollars in revenue. That’s the first point

The second point is the idea of continuous learning, the idea of a single model learning while in operation. I think we are also researching this.

It is very likely that we will address this issue in the next year or two. Again, I think you have traveled most of the way without it.

A market worth trillions of dollars every year, perhaps all the national security implications and safety impacts I wrote about during the "puberty of technology" could occur without it.

But we, and I think others, are researching it. It is very likely that we will reach there in the next year or two.

There are many ideas. I won't go into detail about all of them, but one is to make the context longer. There is nothing stopping longer context from working.

You just need to train on longer contexts and then learn to serve them during reasoning. Both of these are engineering problems we are researching, and I assume others are researching them too.

Dwarkesh: This increase in context length seems to have a period from 2020 to 2023, from GPT-3 to GPT-4 Turbo, with context length increasing from around 2000 to 128K.

I feel like for about two years since then, we have been within the same range. When the context length becomes much longer than that, people report a qualitative decline in the model's ability to consider the entire context.

So I am curious about what you have seen internally that makes you think, "10 million contexts, 100 million contexts, to achieve six months of human learning and context building."

Dario: This is not a research question. This is an engineering and reasoning question. If you want to provide long contexts, you have to store your entire KV cache.

Storing all memory in the GPU and processing memory is very difficult. I don't even know the details. At this point, this is a level of detail I can no longer keep up with, although I knew it during the GPT-3 era. "These are the weights, these are the activations you have to store..."

But now the whole thing has flipped because we have MoE models and all of this.

Regarding the degradation you mentioned, there are two things, don't be too specific. There is the context length you trained on and the context length you serve.

If you train on a small context length and then try to serve on a long context length, perhaps you will experience these degradations. It is better than nothing, and you might still provide it, but you will experience these degradations.

Perhaps training on long context lengths is more difficult. I think, at the same time, asking some possible rabbit holes.

If you have to train on longer context lengths, doesn't that mean that for the same amount of computation, the samples you can obtain will be fewer?

Maybe it's not worth delving into. I want to get answers to the bigger picture questions.

I don't feel a preference for the human editor who worked with me for six months and the AI that worked with me for six months; which year do you predict that will be?

Dario: My guess is that there are a lot of questions. Basically, when we have "a nation of geniuses in data centers," we can achieve this.

My view on this, if you make me guess, is one to two years, maybe one to three years. It's really hard to say. I have a strong opinion—99%, 95%—that all of this will happen within 10 years. I think that's just a super safe bet.

I have a hunch—this is more like a 50/50 thing—that it will be more like one to two years, maybe more like one to three years.

So one to three years. A nation of geniuses, along with slightly less economically valuable video editing tasks.

Dwarkesh: Sounds very economically valuable, let me tell you.

Dario: There are just a lot of use cases like that. There are many similar ones. So you predict within one to three years.

If AGI is coming, why not buy more computing?

Dwarkesh: Then, in general, Anthropic predicts that by the end of 2026 or early 2027, we will have AI systems "capable of interfacing with the digital work of today's humans, matching or exceeding the intellectual capabilities of Nobel Prize winners, and interacting with the physical world."

You emphasized in your DealBook interview two months ago that your company is more responsibly scaling computing compared to competitors. I'm trying to reconcile these two viewpoints.

If you really believe we will have a nation of geniuses, you would want the largest data centers possible. There’s no reason to slow down.

The TAM of a Nobel Prize winner, capable of doing everything a Nobel Prize winner can do, is in the trillions of dollars.

So I'm trying to reconcile this conservatism, if you have a more moderate timeline, which seems rational, with your statements about progress.

Dario: It actually all fits together. We come back to this rapid but not infinitely fast diffusion.

Assuming we make progress at this speed. Technology is advancing at such a fast pace. I am very confident we will get there in a few years.

I have a hunch that we will get there in one to two years. So there’s a bit of uncertainty on the technical side, but I’m very confident it won’t be too far off.

What I’m less certain about is, once again, the economic diffusion aspect. I truly believe we can have a model of a nation of geniuses in data centers within one to two years.

One question is: how many years after that will the 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, although I’m skeptical about that.

So we have this uncertainty. Even if the technological progress is as fast as I suspect, we don’t exactly know how quickly it will drive revenue

We know it's coming, but depending on how you purchase these data centers, a difference of a few years could be catastrophic.

It's like I wrote in "The Loving Machine." I said I think we might get this powerful AI, this "genius nation in the data center." The description you provided comes from "The Loving Machine."

I said we would get it in 2026, maybe 2027. Again, that's my intuition. If I'm off by a year or two, I wouldn't be surprised, but that's my intuition.

Assuming it happens. That's the starting gun. How long does it take to cure all diseases? That's one of the ways to drive massive economic value.

You cure every disease. One question is how much goes to pharmaceutical companies or AI companies, but there's a huge consumer surplus because—assuming we can provide access for everyone, which I care deeply about—we cure all these diseases.

How long does it take? You have to make biological discoveries, you have to manufacture new drugs, you have to go through the regulatory process. We've seen this with vaccines and COVID.

We rolled out the vaccine to everyone, but it took a year and a half.

My question is: from the time AI first exists in the lab to the time diseases are actually cured for everyone, how long does it take to have a method that cures everything for everyone—AI is theoretically capable of inventing?

We have had the polio vaccine for 50 years. We are still trying to eradicate it in the most remote corners of Africa. The Gates Foundation is doing everything it can to try. Others are also trying their best. But it's difficult.

Again, I don't expect most economic diffusion to be that difficult. That's the hardest case. But there's a real dilemma here.

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 purchase data centers, again, the curve I'm looking at is: we have 10x growth every year.

At the beginning of this year, what we saw was an annualized revenue of $10 billion. We have to decide how much computing to purchase.

Actually building data centers and booking data centers takes a year or two.

Basically, I'm saying, "By 2027, how much computing do I get?"

I can assume revenue will continue to grow 10x each year, so by the end of 2026 it will be $100 billion, and by the end of 2027 it will be $1 trillion.

It will actually be $5 trillion in computing because it will be $1 trillion a year for five years. I can purchase $1 trillion in computing starting at the end of 2027.

If my revenue is not $1 trillion, even if it's $800 billion, there is no force in the world, no hedge that can stop me from going bankrupt if I purchase that much computing.

Even though part of my brain wonders if it will continue to grow 10x, I can't purchase $1 trillion in computing every year in 2027

If I were just a year off that growth rate, or if the growth rate were 5 times a year instead of 10 times, then you would go bankrupt.

So you end up in a world where you support hundreds of billions instead of trillions. You accept some risks, that there is so much demand that you can't support the revenue, you accept some risks that you got it wrong, and it’s still slow.

When I talk about responsible behavior, I actually don't mean absolute numbers. I do think our spending is indeed a bit less than some other players.

It's actually about other things, like whether we are thoughtful or whether we are YOLOing and saying, "We're going to do $100 billion here or $100 billion there"?

My impression is that some other companies haven't written down spreadsheets, and they don't really understand the risks they are taking. They just do things because it sounds cool.

We have thought it through carefully. We are an enterprise business. Therefore, we can rely more on revenue. It’s not as volatile 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 10 times every year.

Things have to get quite bad for us to get into financial trouble. So we have thought it through carefully, and we have made that balance. That’s what I mean by saying we are responsible.

Dwarkesh: So it seems we might just have different definitions of "the genius nation in the data center." Because when I think of actual human genius, a real human genius nation in the data center, I would be happy to buy $5 trillion worth of computing to run a real human genius nation in the data center.

Suppose JP Morgan or Moderna or whatever doesn't want to use them. I have a genius nation. They would start their own companies.

If they can't start their own companies, they are bottlenecked by clinical trials... It’s worth noting that for clinical trials, most clinical trials fail because the drugs don't work. There is no efficacy.

Dario: I precisely made this point in "The Loving Machine," where I said clinical trials will be much faster than we are used to, but not infinitely fast.

Dwarkesh: Okay, then suppose clinical trials take a year to succeed, so you can generate revenue from it and make more drugs.

Alright, then you have a genius nation, and you are an AI lab. You can use more AI researchers.

You also think that smart people working on AI technology have these self-reinforcing benefits. You can have the data center work on AI advancements.

Is there substantially more benefit from purchasing $1 trillion worth of computing a year versus $300 billion worth of computing a year?

Dario: If your competitors are buying $1 trillion, yes, there is

Well, no, there are some gains, but also, there’s this opportunity; they went bankrupt before. Again, if you’re only a year off, you’ll ruin yourself. That’s the balance.

We are buying a lot. We are buying quite a bit. The amount we are purchasing is comparable to what the biggest players in the game are buying.

But if you ask me, "Why haven’t we signed for a $10 trillion computation starting in mid-2027?"... First of all, it can’t be produced. There isn’t that much in the world.

But secondly, what if the genius country comes, but it arrives in mid-2028 instead of mid-2027? You’re bankrupt.

Dwarkesh: So if your prediction is one to three years, it seems you should want to have $10 trillion in computing power by 2029 at the latest? Even in the longest version of the timeline you stated, the computation you’re expanding seems inconsistent.

Dario: What makes you think that?

Human wages, let’s say, are about $50 trillion a year—

Dwarkesh: So I won’t specifically talk about Anthropic, but if you’re talking about this industry, the amount of computation being built this year in the industry might be, let’s say, 10-15 terawatts.

It grows about three times a year. So next year it’s 30-40 terawatts. By 2028 it could be 100 terawatts. By 2029 it could be like 300 terawatts.

I’m doing the math in my head, but each terawatt might cost $10 billion, about $10-15 billion a year.

You add all that up, and you get roughly what you described. You get exactly that. By 2028 or 2029, you’re getting multiple trillions a year.

Dwarkesh: That’s for the industry.

Dario: That’s for the industry, that’s right.

Dwarkesh: Assuming Anthropic’s computation continues to grow threefold each year, then by 2027-28, you have 10 terawatts. Multiply by, as you said, $10 billion. So that’s like $100 billion a year.

But then you’re saying by 2028 the TAM is $200 billion.

Dario: Again, I don’t want to give the exact numbers for Anthropic, but those numbers are too small.

Dwarkesh: Okay, interesting.

How will AI labs really profit?

Dwarkesh: You tell investors you plan to be profitable starting in 2028. That year we might gain the genius country as a data center. This will now unlock all advancements in medicine, health, and new technologies. Isn’t this exactly when you want to reinvest in the business and build a bigger “country” so they can make more discoveries?

Dario: In this field, profitability is a bit of a strange thing. I don't think profitability is actually a measure of consumption versus investment in this business. Let's take this as an example. I actually believe that when you underestimate the demand you will get, you will make a profit, and when you overestimate the demand you will get, you will incur a loss because you are buying data centers in advance.

Think of it this way. Again, these are formulaic facts. These numbers are not accurate. I'm just trying to build a toy model here. Assume you use half of your computing power for training and half for inference. Inference has some gross margins exceeding 50%. So this means that if you are in a steady state, you have built a data center, and if you know exactly the demand you will get, you will generate a certain amount of revenue.

Assume you pay $100 billion a year for computing power. On $50 billion a year, you support $150 billion in revenue. The other $50 billion is for training. Basically, you are profitable, making a profit of $50 billion. That is the economics of this industry today, or not today, but what we predict will happen in a year or two.

The only situation that makes this not hold is if the demand you get is less than $50 billion. Then you have more than 50% of the data center for research, and you are not profitable. So you trained a stronger model, but you are not profitable. If the demand you get is more than you imagined, then research will be squeezed, but you can support more inference, and you will be more profitable. Maybe I didn't explain it well, but what I'm trying to say is that you first decide the amount of computing power. Then you have some goals for inference and training, but that is determined by demand. It is not determined by you.

Dwarkesh: What I'm hearing is that the reason you predict profitability is that you are systematically under-investing in computing power?

Dario: No, no, no. I'm saying it's hard to predict. Regarding 2028 and when things will happen, that is what we are trying to do for investors. All of these things are very uncertain due to the uncertainty cone. If revenue grows fast enough, we could be profitable in 2026. If we overestimate or underestimate the next year, it could fluctuate dramatically.

Dwarkesh: What I'm trying to figure out is that you have a business model in mind where you invest, invest, invest, gain scale, and then become profitable. There is a single point in time when things will improve.

Dario: I don't think the economics of this industry works that way.

Dwarkesh: I see. So if I understand correctly, you are saying that due to the difference between the amount of computing power we should have and the amount we actually have, we are somewhat forced to be profitable. But that doesn't mean we will continue to be profitable. We will reinvest the money because AI has made such great progress now, and we want a bigger genius kingdom. So back to high revenue, but also high losses

Dario: If we could accurately predict demand every year, we would be profitable every year. Because spending 50% of computing power on research, roughly speaking, combined with a gross margin above 50% and accurate demand forecasting would lead to profitability. That’s the profitable business model that I think exists but is somewhat obscured by these premature constructions and forecasting errors.

Dwarkesh: I think you treat 50% as a fixed constant, whereas in fact, if AI progresses rapidly, you can scale up to increase progress, and you should have more than 50% and not be profitable.

Dario: But what I’m saying is, you might want to scale it up more. Remember the law of logarithmic returns. If 70% only allows you to get a slightly smaller model by a factor of 1.4... that extra $20 billion, every dollar there is worth much less to you because it’s a logarithmic linear setup. So you might find it better to invest that $20 billion in service inference or hiring engineers who are better at what they do.

So the reason I say 50%... that’s not entirely our goal. It won’t necessarily be 50%. It may change over time. What I’m saying is logarithmic linear returns, which leads you to spend a small portion of the business. For example, not 5%, not 95%. Then you get diminishing returns.

Dwarkesh: I find it strange that I’m trying to convince Dario of the progress of AI or something else. Okay, you don’t invest in research because it has diminishing returns, but you invest in the other things you mentioned.

Dario: I think the macro-level profits—again, I’m talking about diminishing returns, but after you spend $50 billion every year. I believe you would raise this point, but the diminishing returns on genius can be quite high.

More generally, what are profits in a market economy? Profits essentially say that other companies in the market can do more with that money than I can. Set Anthropic aside. I don’t want to provide information about Anthropic. That’s why I give these stylized numbers. But let’s derive the equilibrium of the industry. Why doesn’t everyone spend 100% of their computing power on training without serving any customers? It’s because if they don’t generate any revenue, they can’t raise funds, can’t trade computing power, and can’t buy more computing power the next year.

So there will be an equilibrium where each company spends less than 100% on training, and of course, spends less than 100% on inference as well. It’s obvious why you wouldn’t just serve the current model and never train another model, because then you wouldn’t have any demand, because you would 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 will be in a position where the equilibrium you spend on training is less than the gross margin you can achieve on computing power

So the underlying economics are profitable. The problem is that when you purchase computing power for the next year, you have this hellish demand forecasting problem where you might underestimate and be very profitable but have no computing power for research. Or you might overestimate, not be profitable, and have all the computing power 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 "Genius Nation" will arrive in two years, so you should buy this computing power. To me, your ultimate conclusion makes a lot of sense. But it seems that the "Genius Nation" is difficult and has a long way to go. So stepping back, I'm trying to figure out that your worldview seems compatible with those who say, "We are 10 years away from a world that generates trillions of dollars in value."

Dario: That is not my point at all. So I would make another prediction. I find it hard to see that there won't be trillions of dollars in revenue before 2030. I can construct a reasonable world. Maybe it takes three years. That would be my reasonable endpoint. Like in 2028, we get the real "Genius Nation in the data center." Revenue entering the low hundreds of billions by 2028, and then the Genius Nation accelerates it to trillions. We are basically at the slow end of diffusion. It takes two years to reach trillions. That would be the world until 2030. I doubt even with the comprehensive technology index and diffusion index, we will get there before 2030.

Dwarkesh: So you propose a model where Anthropic is profitable because it seems fundamentally we are in a computing power-constrained world. So ultimately we continue to grow computing power—

Dario: I think the source of profit is... again, let's abstract the entire industry. Let's imagine we are in an economics textbook. We have a few companies. Each can invest a limited amount. Each can allocate part of the investment to R&D. They have certain marginal service costs. The gross profit margin of that marginal cost is very high because inference is efficient. There is some competition, but the model is also differentiated. Companies will compete to push their research budgets higher.

But because there are only a few players, we have... what do you call it? Cournot equilibrium, I think, which is the equilibrium of a few companies. The point is it won't balance to zero profit in perfect competition. If there are three companies in the economy, and they are all acting rationally to some extent independently, it won't balance to zero.

Dwarkesh: Help me understand because we do have three leading companies now, and they are not profitable. So what is changing?

Dario: Again, the gross margins are very positive right now. What is happening is a combination of two things. One is that we are still in the exponential expansion phase of computing power. A model is being trained. Suppose training a model last year cost $1 billion. Then this year it generates $4 billion in revenue, with inference costs of $1 billion. Again, I'm using stylized numbers here, but that would be a 75% gross margin and 25% tax So that model overall made $2 billion.

But at the same time, we spent $10 billion to train the next model, because there is exponential scaling. So the company is losing money. Each model is profitable, but the company is losing money. What I mean by equilibrium is an equilibrium where we have a "genius nation in the data center," but the scaling of that model training has become more balanced. Maybe it's still on the rise. We are still trying to predict demand, but it has become more stable.

Dwarkesh: I'm confused about a few things there. Let's start with the current world. In the current world, you are right, as you said before, if you treat each individual model as a company, it is profitable. But of course, a large part of the production function of frontier labs is training the next model, right?

Dario: Yes, that's correct.

Dwarkesh: If you don't do that, then you will be profitable for two months, and then you won't have profits because you won't have the best model. But at some point, this reaches its maximum scale. Then in equilibrium, we have algorithm improvements, but we spend roughly the same amount to train the next model as we do to train the current model. At some point, you run out of money in the economy. The fallacy of fixed labor supply... the economy will grow, right? That's one of your predictions. We will have data centers in space.

Dario: Yes, but that's another example of the topic I'm talking about. I believe that with AI, economic growth will be faster than ever before. Now computing power is growing threefold each year. I don't believe the economy will grow by 300% each year. I mentioned in "The Kind Machine" that I think we might achieve 10-20% economic growth per year, but we won't achieve 300% economic growth. So I think ultimately, if computing power becomes a large part of economic output, it will be limited by that.

Dwarkesh: So let's assume a model where computing power has an upper limit. A world where frontier labs make money is a world where they continue to make rapid progress. Because fundamentally, your profits are limited by how good the alternatives are. So you are able to make money because you have the frontier model. If you don't have the frontier model, you won't make money. So this model requires that there is never a steady state. You are constantly making more algorithmic progress.

Dario: I don't think that's true. I mean, I feel like we're in an economics class. Do you know Tyler Cowen's famous saying? We will never stop talking about economics.

Dwarkesh: We will 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 enter an industry with only a few players has a very high barrier to entry. Cloud is like that. I think cloud is a good example of this. There are three, maybe four players in the cloud space. I think AI is the same, three, maybe four. The reason is that it’s too expensive. Running a cloud company requires so much expertise and so much capital. You have to put in all that capital. Besides putting in all that capital, you also have to have all these other things that require a lot of skills to achieve.

So if you go to someone and say, “I want to disrupt this industry, it’s $100 billion.” You’d be like, “Okay, I’ll put in $100 billion, and I’m also betting that you can do all these other things that everyone else has been doing.” The result just lowers the profits. The impact of your entry is a decrease in profit margins. So, we’ve always had this equilibrium in the economy with only a few players. Profits are not astronomical. Profit margins are not astronomical, but they are not zero. That’s what we see in the cloud. The cloud is very homogeneous.

Models are more differentiated than the cloud. Everyone knows that what Claude is good at is different from what GPT is good at, which is different from what Gemini is good at. It’s not just that Claude is good at programming, GPT is good at math and reasoning. It’s more subtle than that. Models excel at different types of programming. Models have different styles. I think these things are actually very different from each other, so I expect to see more differentiation than in the cloud.

Now, there is actually a counterargument. That counterargument is that if the process of producing models, if AI models themselves can do that, then that might spread throughout the economy. But that’s not the argument for making AI models universally commoditized. That’s a bit like arguing for commoditizing the entire economy all at once.

I don’t know what would happen in that world, where basically anyone can do anything, anyone can build anything, and there are 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 solve all the safety and security issues, that’s one of the mechanisms for the economy to flatten itself again. But that’s a bit far beyond the “genius nation in the data center.”

Dwarkesh: Perhaps a more nuanced way to express that potential point is: 1) It seems that AI research particularly relies on raw intelligence, which will be particularly abundant in an AGI world. 2) If you look at today’s world, it seems that few technologies spread as quickly as AI algorithms do. So this does imply that the industry is structurally diffusive.

Dario: I think programming is progressing quickly, but I think AI research is a superset of programming, where some aspects haven’t progressed as quickly. But I do think, to emphasize again, once we nail programming, once we get AI models to progress quickly, then that will accelerate the ability of AI models to do everything else

So even though programming is progressing rapidly now, I believe that once AI models are building the next AI models and everything else, the entire economy will develop at the same speed. However, I am a bit concerned geographically. I worry that just being close to AI, having heard of AI, might be a distinguishing factor. So when I mention a growth rate of 10-20%, I am concerned that the growth rate in Silicon Valley and parts of the world socially connected to Silicon Valley might be 50%, while elsewhere it may not be much faster than it is now. I think that would be a rather bleak world. So one thing I often think about is how to prevent that from happening.

Dwarkesh: Do you think that once we have this genius nation in the data center, robotics will be solved quickly afterward? Because it seems that a big problem with robotics is that humans can learn to remotely operate the current hardware, but the current AI models cannot, at least not in a super-efficient way. So if we have this human-like learning ability, shouldn't we also solve robotics immediately?

Dario: I don't think it relies on human-like learning. It can happen in different ways. Again, we can train models on many different video games, which is like robot control, or many different simulated robotic environments, or just train them to control computer screens, and then they learn to generalize. So it will happen... it doesn't necessarily rely on humanoid learning.

Humanoid 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 continual learning. That could also happen because we trained the model in many environments and then generalized, or it could happen because the model learned that within the context length. It doesn't really matter which way it is.

If we go back to our discussion an hour ago, that kind of thing can happen in several different ways. But I do believe that for whatever reason, when the model has those skills, robotics will be fundamentally transformed—whether it's the design of robots, because the model will be much better than humans in that regard, or the ability to control robots. So we will be better at building physical hardware, building physical robots, and we will also get better at controlling them.

Now, does that mean the robotics industry will also generate trillions of dollars in revenue? My answer is yes, but there will be the same extremely rapid but not infinitely fast diffusion. So will robotics be fundamentally transformed? Yes, maybe in another year or two. That's how I think about these things.

Dwarkesh: That makes sense. There is a general skepticism about extremely rapid progress. That's my view. It sounds like you will solve continual learning in some way in a few years. But just a few years ago, people weren't talking about continual learning, and then we realized, "Oh, why aren't these models as useful as they could be, even though they clearly passed the Turing test and are experts in many different fields? Maybe it's this thing." Then we solved that thing, and we realized that, in fact, there are other things that human intelligence can do that these models cannot do, which is the basis of human labor So why don't we think there will be more things like this, as we discover more fragments of human intelligence?

Dario: Well, to be clear, I think continuous learning, as I mentioned before, may not be a barrier at all. I think we might just get there through pre-training generalization and RL generalization. I think there may not be such a thing at all. In fact, I would point to the history of ML, where people have raised things that seemed like barriers, and they ultimately dissolved in the "compute big chunk."

People talk about, "How does your model track nouns and verbs?" "It can understand syntactically but not semantically? That's just statistical correlation." "You can understand a passage, but 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 there is actually a stronger historical precedent that some of these things seem like big problems, and then they kind of dissolve. Some of them are real. The demand for data is real, and maybe continuous learning is a real thing.

But again, I would 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 whole domain of human activity, and we're just saying the model can do it now.

Dwarkesh: When you say end-to-end, do you mean setting the technical direction, understanding the context of the problem, and so on?

Dario: Yes. I mean all of that.

Dwarkesh: Interesting. I feel like that's AGI complete, and that might be internally consistent. But it's not like saying 90% of the code or 100% of the code.

Dario: No, I gave this spectrum: 90% of the code, 100% of the code, 90% of end-to-end SWE, 100% of end-to-end SWE. New tasks are created for SWE. Ultimately, those will also be completed. There's a long spectrum there, but we're moving through this spectrum very quickly.

Dwarkesh: I do find it interesting; I've watched a few podcasts you've done where the host would say, "But Dwarkesh wrote an article about continuous learning." That always makes me laugh because you've been an AI researcher for 10 years. I'm sure there's some sense of, "Okay, so a podcast host wrote an article, and every time I get interviewed, I'm asked this question."

Dario: The fact is we're all trying to figure this out together. There are aspects where I can see things that others can't. These days, it may be more about seeing a bunch of things internally at Anthropic and having to make a bunch of decisions rather than I have great research insights that others don't. I'm running a company of 2,500 people. In fact, for me, having specific research insights is much rarer than it was 10 years ago or even two or three years ago

Dwarkesh: As we move towards a world where remote workers can be completely replaced, is the API pricing model still the most reasonable? If not, what is the correct way to price AGI or service AGI?

Dario: I think there will be many different business models being experimented with simultaneously. I actually believe the API model is more durable than many people imagine. One way I think about this is that if technology is advancing rapidly, if it is exponential progress, that means there is always a surface area of new use cases developed in the past three months. Any product surface you build is always at risk of becoming irrelevant.

Any given product surface may make sense for a certain range of capabilities of the model. Chatbots have encountered limitations, and making them smarter doesn't really help the average consumer that much. But I don't think that's a limitation of the AI model. I don't think that's evidence that the model is good enough and that it becoming better is irrelevant to the economy. That is irrelevant to that specific product.

So I think the value of the API is that it always provides a very close opportunity to build on the latest things. There will always be this frontier of new startups and new ideas that were impossible a few months ago but are now possible because of model advancements.

I actually predict it will coexist with other models, but we will always have the API business model because there will always be a thousand different people trying to experiment with the model in different ways. Among those, 100 will become startups, and 10 will become successful large startups. Two or three will truly end up being the way people use a certain generation of models. So I basically think it will always exist.

At the same time, I am sure there will be other models. Not every token output by the model is worth the same amount of money. Think about when someone calls and says, "My Mac is broken," or something, and the model says, "Restart it." What is the value of those tokens? Someone has never heard that before, but the model has said it 10 million times. Maybe that is worth a dollar or a few cents or something. And if the model goes to a pharmaceutical company and says, "Oh, you know, this molecule you are developing, you should take that aromatic ring from one end of the molecule and put it on the other end. If you do that, amazing things will happen." Those tokens could be worth tens of millions of dollars.

So I think we will definitely see business models recognizing this. At some point, we will see some form of "pay for results," or we might see something like compensation for labor, that kind of hourly billing work. I don't know. I think because this is a new industry, a lot of things will be tried. I don't know what will ultimately prove to be the right thing.

Dwarkesh: I agree with your point that people have to try various things to figure out the best way to use this bundle of intelligence. But I find Claude Code quite remarkable. I don't think there has been a single application in the history of startups that is as competitive as programming agents. Claude Code is the category leader here This seems quite surprising to me. It doesn't seem essential for it to be Anthropic to build this. I wonder if you have an explanation for why it must be Anthropic, or how Anthropic ultimately built a successful application in addition to the underlying model.

Dario: The way it actually happened is quite simple, which is that we have our own programming models that are good at programming. Around early 2025, I said, "I think the time has come, if you are an AI company, you can achieve extraordinary acceleration of your research by using these models." Of course, you need an interface, you need a tool to navigate them. So I encouraged the internal team. I didn't say this is something you must use.

I just said people should try this. I think it was initially called Claude CLI, and later the name was changed to Claude Code. Internally, it was something everyone was using, and we saw rapid internal adoption. I looked at it and said, "Maybe we should release this externally, right?" It saw such rapid adoption within Anthropic. Programming is a lot of what we do. We have an audience of hundreds of people, which at least represents the external audience in some ways.

So it seems we already had product market fit. Let's release this thing. Then we released it. I think just the fact that we were developing the models ourselves, and we knew how we needed to use the models the most, I think that was creating this feedback loop.

Dwarkesh: I see. So, for example, an Anthropic developer might say, "Ah, it would be better if it were better at this X thing." And then you bake that into the next model you build.

Dario: That's one version of it, but there's also the 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 in the early days. Now, of course, there are millions of people using it, so we also get a lot of external feedback. But being able to get that rapid internal feedback was just fantastic. I think that's why we released a programming model and not a pharmaceutical company. My background is in biology, but we didn't have the resources needed to release a pharmaceutical company.

Will regulations destroy the benefits of AGI?

Dwarkesh: Now let me ask you about steering AI in a positive direction. It seems that whatever vision we have for AI to develop positively must be compatible with two things: 1) the capability to build and run AI is spreading extremely rapidly, and 2) the number of AIs we have and their intelligence will also increase very quickly. This means many people will be able to build a large number of misaligned AIs, or those that are just trying to increase their footprint or have strange psychologies like Sydney Bing but are now superhuman AIs In a world where we have a large number of different AIs (some of which are misaligned) running everywhere, what is our balanced vision?

Dario: I am skeptical about the balance of power in "The Adolescence of Technology." But I am particularly skeptical that you have three or four companies building from the same model of things, and they will balance each other out. Even any number of them will balance each other out. We may live in a world dominated by offense, where one person or one AI model is smart enough to do something that harms all other things.

In the short term, the number of players we have now is limited. So we can start with a limited number of players. We need to implement safeguards. We need to ensure that everyone is doing the right alignment work. We need to ensure that everyone has biological classifiers. These are the things we need to do immediately.

I agree that this does not solve the long-term problem, especially if the ability of AI models to create other AI models surges, then the whole thing may become harder to solve. I think in the long run, we need some kind of governance framework. We need some kind of governance framework that preserves human freedom while allowing us to govern a large number of human systems, AI systems, hybrid human-AI companies, or economic units.

So we will need to think: How do we protect the world from bioterrorism? How do we protect the world from mirror life? Perhaps we will need some kind of AI monitoring system to oversee 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 security landscape with a new set of tools and a new set of vulnerabilities.

My concern is that if we have 100 years for all this to happen very slowly, we will get used to it. We have already gotten used to explosives existing in society, or various new weapons, or cameras. We will get used to it in 100 years, and we will establish governance mechanisms. We will make mistakes. My concern is simply that all this is happening too quickly. So perhaps we need to think faster about how to make these governance mechanisms work.

Dwarkesh: It seems that in a world dominated by offense, during the course of the next century—this idea that AI is making the progress that will happen in the next century occur within a span of five to ten years—even if humans are the only players, we still need the same mechanisms, or the balance of power is equally difficult to manage.

Dario: I think we have the suggestion of AI. But fundamentally, this does not seem to be a completely different ball game. If checks and balances are to work, they will also work for humans. If they do not work, they will not work for AI either. So perhaps this destined that human checks and balances will also fail. Again, I emphasize that I think there is a way to make this happen. Governments around the world may have to work together to make this happen. We may have to discuss with AI the establishment of social structures that make these defenses possible. I don’t know. I don’t want to say this is too far off in time, but it is too far off in technical capability, and it may happen in a very short time, making it difficult for us to predict it in advance

Dwarkesh: Speaking of government intervention, on December 26, the Tennessee legislature proposed a bill stating, "If a person intentionally trains artificial intelligence to provide emotional support, including through open dialogue with users, it constitutes a crime." Of course, one of the things Claude is trying to do is to be a thoughtful and knowledgeable friend.

Overall, it seems we will have this patchwork of state laws. Many benefits that ordinary people might experience due to AI will be curtailed, especially as we enter the kinds of issues you discuss in "The Kind Machine": biological freedom, mental health improvements, etc. It seems easy to imagine a world where these are knocked down like whack-a-mole by different laws, and bills like this seem to not address the actual existential threats you are concerned about.

I am curious to understand, in the context of things like this, Anthropic's stance against a federal pause on state AI laws.

Dario: There are many different things happening at the same time. I think that specific law is foolish. It is clearly crafted by legislators who know very little about what AI models can and cannot do. They are like, "AI models serve us, that sounds scary. I don't want that to happen." So we actually do not support that.

But that is not what is being voted on. What is being voted on is: we will ban all state regulation of AI for 10 years, without any apparent federal AI regulatory plan, which requires Congress to pass, and that is a very high bar. So the idea that we are going to ban states from doing anything for 10 years... People say they have a federal government plan, but there is no actual proposal on the table. No actual attempts. Considering the serious dangers I listed in "The Adolescence of Technology," such as autonomous risks of biological weapons and bioterrorism, and the timeline we have been discussing—10 years is an eternity—I think that is a crazy thing.

So if that is the choice, if you force us to choose, then we will choose not to pause. I think the benefits of that stance outweigh the costs, but if that is the choice, it is not a perfect stance. Now, in terms of what we should do, what I would support is that the federal government should intervene, not to say "states, you cannot regulate," but to say "we are going to do this, and states cannot deviate from this standard." I think as long as the federal government says, "This is our standard. It applies to everyone. States cannot do anything else." I would accept that kind of preemption. If it is done in the right way, I would support it.

But if the current idea is for states to say, "You can’t do anything, and we (the federal government) won’t do anything either," that seems very unreasonable to us. I think this approach will not stand the test of time, and in fact, with all the opposing voices you see, it has begun to seem outdated.

As for what we want, what we have discussed is starting with transparency standards to monitor some of those autonomy risks and bioterrorism risks. As the risks become more severe, and as we gather more evidence, I think we can take more targeted radical measures, such as stipulating, "Hey, AI bioterrorism is really a threat." "Let's enforce a law that requires people to use classifiers." I can even imagine... but it depends on the situation. It depends on how serious the ultimate threat is. We can't be sure yet.

We need to advance this matter in a rational and honest way, and we should clarify in advance that the current risks have not yet manifested. But I can certainly imagine that at the rate things are developing, later this year we might say, "Hey, this artificial intelligence bioterrorism thing is really serious. We should take action. We should write it into federal standards. If the federal government doesn't act, we should write it into state standards." I can completely see that happening.

What worries me is a world where, if you consider the expected pace of progress and then consider the legislative lifecycle... as you said, due to diffusion lags, the benefits come slowly, to the point where I think these patchwork state laws, at the current trajectory, will really be obstructive. I mean, if having an emotional chatbot friend drives people crazy, then imagine the real benefits from AI that we want ordinary people to experience. From improvements in health and healthy lifespan to mental health improvements, and so on.

However, at the same time, it seems you believe the dangers are already on the horizon, but I don't see that many... Compared to the dangers of AI, this seems particularly harmful to the benefits of AI. So this might be where I think the cost-benefit analysis is not very reasonable.

Dwarkesh Patel: There are a few things here. People talk about thousands of such state laws. First, the vast majority, the vast majority of laws will not pass. Theoretically, the world operates in a certain way, but just because a law is passed doesn't mean it will actually be enforced. The people enforcing it might think, "Wow, this is stupid. This means shutting down everything built in Tennessee." Many times, the way laws are interpreted makes them not seem as dangerous or harmful as they appear.

Of course, on the other hand, if you pass a law to prevent bad things from happening; you will encounter the same problem.

Dario Amodei: My basic point is that if we can decide what laws to pass and how to do things—of course, we are just a small part of the opinion—I would lift a lot of the regulations around the health benefits of AI. I'm less concerned about chatbot laws. In fact, I'm more worried about drug approval processes, as I believe AI models will greatly accelerate our drug discovery speed, while the approval pipeline will be clogged. The approval pipeline will not be ready to handle all the influx.

I think regulatory reform should lean more towards the fact that we will have a lot of things coming out whose effectiveness and safety will actually be very clear, which is a wonderful thing and very effective. Perhaps we don't need to build all this superstructure around it, which was originally designed for drugs that were almost ineffective and often had serious side effects

At the same time, I believe we should significantly strengthen legislation regarding safety and security. As I have said, starting with transparency is my perspective on trying not to hinder industry development and trying to find the right balance. I am very concerned about this. Some people criticize my article, saying, "That's too slow. If we do that, the dangers of artificial intelligence will come too quickly."

Well, basically, I think the past six months and the coming months will be about transparency. Then, if we become more certain about these risks— I think we may be able to determine this as early as later this year—if these risks materialize, then I believe we need to act very quickly in areas where we actually see the risks.

I think the only way to do this is to be flexible. The current legislative process is usually not flexible, but we need to emphasize the urgency of this matter to all relevant parties. That is why I want to send out this urgent message. That is why I wrote the article "Adolescence of Technology." I want policymakers, economists, national security professionals, and decision-makers to read it so that they hopefully can take action faster than they otherwise would.

Dwarkesh Patel: Is there anything you can do or advocate for that would make people more certain that the benefits of artificial intelligence can be better realized? I feel like you've already worked with legislators saying, "Okay, we want to prevent bioterrorism here. We want to increase transparency, and we want 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 economic issues.

Dario Amodei: Actually, I don't quite agree with that view for developed countries. I feel that in developed countries, the market operates quite well. When something has a large profit margin and is clearly the best available alternative, the regulatory system actually finds it hard to stop it. We see this with artificial intelligence itself.

If we talk about the benefits of drugs and technology, I am not as worried about those benefits being hindered in developed countries. I am a bit concerned that they are progressing too slowly. As I said, I do believe we should work to accelerate the FDA approval process. I do believe we should oppose the chatbot legislation you described. I oppose them on their own. I think they are foolish.

But I actually think the bigger concern is developing countries, where there are no well-functioning markets, and we often cannot build on existing technological foundations. I am more worried that those people will be left behind. I also worry that even if treatments are developed, perhaps someone in rural Mississippi may not be able to access them well. That is a smaller version of what we worry about in developing countries.

So what we have been doing is working with philanthropists. We are collaborating with those who provide drugs and health interventions to developing countries, sub-Saharan Africa, India, Latin America, and other developing regions of the world. I think that is something that will not happen automatically without intervention

Summary of Claude

Dwarkesh Patel: You recently announced that Claude will have a constitution aligned with a set of values, rather than just for the end user. I can imagine a world where, if it is for the end user, it retains the power balance we have in today's world, as everyone has their own AI to advocate for them. The ratio of bad to good people remains unchanged.

This seems to work well in our world today. Why not do it that way, instead of having a specific set of values that AI should follow?

Dario Amodei: I'm not sure I would draw that distinction. There may be two relevant distinctions here. I think you are talking about a mix of the two. One is whether we should give the model a set of instructions about "do this" versus "don't do this." The other is whether we should give the model a set of principles about how to act.

This is purely something we have observed as a practical and empirical matter. By teaching the model principles and allowing it to learn from those principles, its behavior becomes more consistent, it is 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 a car, don't speak Korean"—it doesn't really understand those rules and has a hard time generalizing from them. It's just a checklist of "do's and don'ts."

However, if you give it principles—there are some hard boundaries, like "don't create biological weapons," but—overall you are trying to understand what it should aim to do and how it should operate. So from a practical standpoint, this has proven to be a more effective way to train models. This is the trade-off between rules and principles.

Then there is another thing you are talking about, which is the trade-off between corrigibility and intrinsic motivation. To what extent should the model be like a "straightjacket," simply following the instructions of those who give it commands, versus to what extent should the model have a set of intrinsic values and do things on its own?

In that regard, I would actually say that everything about the model is closer to the direction that it should primarily do what people want it to do. It should primarily follow instructions. We are not trying to build something that runs off to take over the world. We are actually very much inclined towards the corrigible side.

Now, we do say there are some things the model will not do. I think we have stated in the constitution in various ways that, under normal circumstances, if someone asks the model to perform a task, it should do that task. That should be the default. But if you ask it to do something dangerous or to harm others, then the model would be unwilling to do so. So I actually see it as a primarily corrigible model that has some limitations, but those limitations are based on principles.

Then the fundamental question is, how are those principles determined? This is not a problem unique to Anthropic. It will be a question for any AI company

Dwarkesh Patel: But because you are the ones who truly write down the principles, I can ask you this question. Typically, constitutions are written down, carved in stone, and have a process for updating and changing them, etc. In this case, it seems to be a document written by people at Anthropic that can be changed at any time, guiding the behavior of a system that will become the foundation of many economic activities. How do you think these principles should be set?

Dario Amodei: I think there might be three sizes of loops, three ways of iteration here. One is the iteration we do internally at Anthropic. We train the model, we are not satisfied with it, so we modify the constitution. I think that’s a good thing. It’s good to release public updates to the constitution every once in a while because people can comment on it.

The second layer of the loop is that different companies have different constitutions. I think that’s useful. Anthropic releases one constitution, Gemini releases one, and other companies release one. People can look at them and compare. External observers can criticize, saying, “I like this clause in this constitution and that clause in that constitution.” This creates a kind of soft incentive and feedback for all companies to take the best parts and improve upon them.

Then I think there’s a third loop, which is society outside of AI companies, and not just those commentators without hard power. There we did some experiments. A few years ago, we did an experiment with the Collective Intelligence Project, basically a poll asking people what should be in our AI constitution. At that time, we incorporated some of those changes.

So you can imagine doing something similar with the new approach we take to the constitution. It’s a bit tricky because when the constitution is a checklist of “dos and don’ts,” it’s easier to take that approach. At the level of principles, it has to have some coherence. But you can still imagine gathering perspectives from a wide variety of people.

You can also imagine—this is a crazy idea, but this whole interview is about crazy ideas—a representative government system having input. I wouldn’t do that today because the legislative process is too slow. That’s exactly why I think we should be cautious about the legislative process and AI regulation.

But in principle, there’s no reason you can’t say, “All AI models must have a constitution that starts with these things, and then you can append other things later, but this particular section must take precedence.” I wouldn’t do that. That’s too rigid and sounds overly prescriptive, like I think overly radical legislation does. But it is indeed something you could try to do.

Is there a less rigid version? Maybe. I really like the second control loop.

Dwarkesh Patel: Obviously, this is not how an actual government constitution operates or should operate. There’s no sense that the Supreme Court would feel the sentiments of the people—what the atmosphere is—and update the constitution accordingly For actual governments, there is a more formal, procedural process.

But you have a vision of competition between constitutions, which actually reminds one of what some libertarian charter city advocates used to say about what different governments in an archipelago would look like. There would be choices about who can operate most effectively and where people are happiest. In a sense, you are recreating that archipelago utopia vision.

Dario Amodei: I think that vision has its merits, but also places where it could go wrong. It’s an interesting and, in some ways, compelling vision, but there are things that will go wrong that you didn’t anticipate. So while I also like the second loop, I think the whole thing has to be some mix of the first, second, and third loops, depending on the proportions. I think that has to be the answer.

Dwarkesh Patel: When someone eventually writes a book equivalent to The Making of the Atomic Bomb for this era, what are the hardest things to collect in the historical record that they are most likely to miss?

Dario Amodei: I think there are a few things. One is, to what extent the outside world does not understand it at every moment of this exponential growth. This is a bias that often exists in history. Anything that actually happens seems inevitable in hindsight.

When people look back, they will find it hard to put themselves in the position of those who were actually betting on this non-inevitable thing happening. We’ve had debates like the ones I made about scaling or that continuous learning would be solved. Some of us internally believed the probability of this happening was high, but the outside world took no action on it at all.

I think it’s bizarre, and unfortunately its insularity… If we are a year or two away from it happening, the average person on the street has no idea. That’s one of the things I’m trying to change through memos and conversations with policymakers. I don’t know, but I think that’s just a crazy thing.

Finally, I want to say—this probably applies to almost all historical crisis moments—how absolutely fast it happens, and how everything happens simultaneously. You might think it’s a carefully calculated decision, but in reality, you have to make that decision, and then you have to make 30 other decisions on the same day because everything is happening too quickly.

You don’t even know which decisions will ultimately become critical. One of my concerns—though this is also an insight into what’s happening—is that some very key decisions will be someone walking into my office and saying, “Dario, you have two minutes. Should we do A or B?”

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. Let’s do B.” As it turns out, that became one of the most important things in history

Dwarkesh Patel: The last question. Typically, no tech CEO writes a 50-page memo every few months. It seems you have successfully established a role for yourself and built a company around it, which is compatible with this more intellectual type of CEO role. I want to understand how you constructed this. How does it work? Do you just leave for a few weeks and then tell your company, “Here’s the memo. This is what we’re going to do”? There have also been reports that you write a lot of these internally.

Dario Amodei: For this particular piece, I wrote it during the winter break. I find it hard to find time to really write it. But I think about it in a broader way. I think it relates to company culture. I probably spend a third, maybe 40% of my time ensuring that Anthropic has a good culture.

As Anthropic grows larger, it becomes increasingly difficult to be directly involved in model training, model releases, and product building. There are 2,500 people. I have certain intuitions, but it’s hard to be involved in every detail. I try to participate as much as I can, but one thing that is very leveraged is ensuring that Anthropic is a good place to work, where people enjoy working, everyone sees themselves as team members, and everyone works together rather than against each other.

We’ve seen with the growth of some other AI companies—without naming names—we’ve started to see disconnection and people fighting against each other. I would argue that there was a lot of this from the beginning, but it has gotten worse. I think we do a very good job of keeping the company united, even if not perfectly, making everyone feel the mission, that we are sincere about the mission, and that everyone believes that others there are working for the right reasons. We are a team, and people are not trying to get promoted at the expense of others or stab each other in the back, which I think often happens in some other places.

How do you do that? It’s a lot of things. It’s me, it’s Daniela who is responsible for the company’s daily operations, 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 aspect of culture is that other leaders also play a role, but especially I must clarify what the company is about, why it is doing what it is doing, what its strategy is, what its values are, what its mission is, and what it stands for.

When you reach 2,500 people, you can’t do this one by one. 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 internal articles. I do two things. First, I write something called DVQ, Dario Vision Quest.

I didn’t name it that. That’s the name it got, and it’s one of the names I wanted to oppose because it sounds like I’m going to go smoke peyote or something. But that’s what it’s called. So I show up in front of the company every two weeks I have a document of three to four pages where I discuss three to four different topics, including what is happening internally, the models we are producing, products, the external industry, the relationship between the whole world and artificial intelligence, as well as geopolitics, and so on. It's a mix of these.

I speak very honestly, saying, "This is what I'm thinking, this is what the leadership at Anthropic 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 the chain six layers. A large part of the company comes to participate, either in person or virtually. This really means you can communicate a lot.

Another thing I do is have a channel on Slack where I write a bunch of things and comment frequently. Usually, that responds to things I see in the company or questions people ask. We conduct internal surveys, and there are some things people care about, so I write them down. I am very honest about these matters. I speak very directly.

The key is to build a reputation for speaking the truth in the company, being pragmatic, acknowledging problems, and avoiding that kind of corporate jargon, that defensive communication that is often necessary in public settings because the world is large and filled with people who maliciously misinterpret things. But if you have a company of people you trust, and we also try to hire people we trust, then you can really speak without any filters.

I think this is a huge advantage for the company. It makes it a better workplace, it makes people more than just the sum of their parts, and increases our chances of accomplishing our mission because everyone is aligned on the mission, and everyone is debating and discussing how to best achieve that mission.

Dwarkesh Patel: Well, as an external version of "Dario's vision exploration," we have this interview. This interview is somewhat like that. It's interesting, Dario. Thank you for the interview.

Dario Amodei: Thank you, Dwarkesh