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

"At least from the public's point of view, three years ago there were well-known public trends across many orders of magnitude of compute where you could see how the loss improves. Now we have RL scali..."

"The hypothesis is basically the same. What it says is that all the cleverness, all the techniques, all the "we need a new method to do something", that doesn't matter very much. There are only a few t..."

"The pre-training scaling laws were one example of what we see there. Those have continued going. Now it's been widely reported, we feel good about pre-training. It’s continuing to give us gains. What ..."

"You mentioned Rich Sutton and "The Bitter Lesson". I interviewed him last year, and he's actually very non-LLM-pilled. I don’t know if this is his perspective, but one way to paraphrase his objection ..."

"In fact, I would guess it probably doesn't matter. There is an interesting thing. Let me take the RL out of it for a second, because I actually think it's a red herring to say that RL is any different..."

"If you did better on some fanfiction corpus, it wouldn't generalize that well to other tasks. We had all these measures. We had all these measures of how well it did at predicting all these other kind..."

"So that kind of takes out the RL vs. pre-training side of it. But there is a puzzle either way, which is that in pre-training we use trillions of tokens. Humans don't see trillions of words. So there ..."

"So I don’t know the full answer to this. I think there's something going on where pre-training is not like the process of humans learning, but it's somewhere between the process of humans learning and..."

"And we should think of the in-context learning that the models do as something between long-term human learning and short-term human learning. So there's this hierarchy. There’s evolution, there's lon..."

"For example, if the analogy is that this is like evolution so it's fine that it's not sample efficient, then if we're going to get super sample-efficient agent from in-context learning, why are we bot..."

"The goal is not to teach the model every possible skill within RL, just as we don't do that within pre-training. Within pre-training, we're not trying to expose the model to every possible way that wo..."

"The crux is you say we're hitting the end of the exponential. Somebody else looks at this and says, "We've been making progress since 2012, and by 2035 we'll have a human-like agent." Obviously we’re ..."

"On the basic hypothesis of, as you put it, within ten years we'll get to what I call a "country of geniuses in a data center", I'm at 90% on that. It's hard to go much higher than 90% because the worl..."

"My one little bit of fundamental uncertainty, even on long timescales, is about tasks that aren't verifiable: planning a mission to Mars; doing some fundamental scientific discovery like CRISPR; writi..."

"I think it's crazy to say that this won't happen by 2035. In some sane world, it would be outside the mainstream. But the emphasis on verification hints to me a lack of belief that these models are ge..."

"We already see substantial generalization from things that verify to things that don't. We're already seeing that. But it seems like you were emphasizing this as a spectrum which will split apart whic..."

"Many of them generalize, but we don't fully get there. We don’t fully color in the other side of the box. It's not a binary thing. Even if generalization is weak and you can only do verifiable domains..."

"But SWE does involve design documents and other things like that. The models are already pretty good at writing comments. Again, I’m making much weaker claims here than I believe, to distinguish betwe..."

"If you consider other productivity improvements in the history of software engineering, compilers write all the lines of software. There's a difference between how many lines are written and how big t..."

"I think people have repeatedly misunderstood them. Let me lay out the spectrum. About eight or nine months ago, I said the AI model will be writing 90% of the lines of code in three to six months. Tha..."

"The spectrum is: 90% of code is written by the model, 100% of code is written by the model. That's a big difference in productivity. 90% of the end-to-end SWE tasks — including things like compiling, ..."

"There are new higher-level things they can do, where they can manage. Then further down the spectrum, there's 90% less demand for SWEs, which I think will happen but this is a spectrum. I wrote about ..."

"Part of your vision is that going from 90 to 100 is going to happen fast, and that it leads to huge productivity improvements. But what I notice is that even in greenfield projects people start with C..."

"So that does make me wonder. Even if I never had to intervene with Claude Code, the world is complicated. Jobs are complicated. Closing the loop on self-contained systems, whether it’s just writing so..."

"You could have these two poles. One is that AI is not going to make progress. It's slow. It's going to take forever to diffuse within the economy. Economic diffusion has become one of these buzzwords ..."

"We've seen from the beginning, at least if you look within Anthropic, there's this bizarre 10x per year growth in revenue that we've seen. So in 2023, it was zero to $100 million. In 2024, it was $100..."

"The GDP is only so large. I would even guess that it bends somewhat this year, but that is a fast curve. That's a really fast curve. I would bet it stays pretty fast even as the scale goes to the enti..."

"Because it's fiddly: "I have to do change management within my enterprise… I set this up, but I have to change the security permissions on this in order to make it actually work… I had this old piece ..."

"So I think everything we've seen so far is compatible with the idea that there's one fast exponential that's the capability of the model. Then there's another fast exponential that's downstream of tha..."

"Can I try a hot take on you? Yeah. I feel like diffusion is cope that people say. When the model isn't able to do something, they're like, "oh, but it's a diffusion issue." But then you should use the..."

"An AI can read your entire Slack and your drive in minutes. They can share all the knowledge that the other copies of the same instance have. You don't have this adverse selection problem when you're ..."

"I think diffusion is very real and doesn't exclusively have to do with limitations on the AI models. Again, there are people who use diffusion as kind of a buzzword to say this isn't a big deal. I'm n..."

"I'll just give an example of this. There's Claude Code. Claude Code is extremely easy to set up. If you're a developer, you can just start using Claude Code. There is no reason why a developer at a la..."

"Big enterprises, big financial companies, big pharmaceutical companies, all of them are adopting Claude Code much faster than enterprises typically adopt new technology. But again, it takes time. Any ..."

"There are just a number of factors. You have to go through legal, you have to provision it for everyone. It has to pass security and compliance. The leaders of the company who are further away from th..."

"They have to say, "Okay, we have 3,000 developers. Here's how we're going to roll it out to our developers." We have conversations like this every day. We are doing everything we can to make Anthropic..."

"Again, many enterprises are just saying, "This is so productive. We're going to take shortcuts in our usual procurement process." They're moving much faster than when we tried to sell them just the or..."

"I don't think even AGI or powerful AI or "country of geniuses in a data center" will be an infinitely compelling product. It will be a compelling product enough maybe to get 3-5x, or 10x, a year of gr..."

"I buy that it would be a slight slowdown. Maybe this is not your claim, but sometimes people talk about this like, "Oh, the capabilities are there, but because of diffusion... otherwise we're basicall..."

"I think if you had the "country of geniuses in a data center"... If we had the "country of geniuses in a data center", we would know it. We would know it if you had the "country of geniuses in a data ..."

"Coming back to concrete prediction… Because there are so many different things to disambiguate, it can be easy to talk past each other when we're talking about capabilities. For example, when I interv..."

"I think you were right about that. I think spiritually I feel unsatisfied because my internal expectation was that such a system could automate large parts of white-collar work. So it might be more pr..."

"Take video editors. I have video editors. Part of their job involves learning about our audience's preferences, learning about my preferences and tastes, and the different trade-offs we have. They’re,..."

"They're going to be like, "Oh, I don't know, Dario scratched his head and we could edit that out." "Magnify that." "There was this long discussion that is less interesting to people. There's another t..."

"You'll be able to feed this in. It'll be able to also use the computer screen to go on the web, look at all your previous interviews, look at what people are saying on Twitter in response to your inte..."

"We've seen this climb in benchmarks, and benchmarks are always imperfect measures. But I think when we first released computer use a year and a quarter ago, OSWorld was at maybe 15%. I don't remember ..."

"Can I just follow up on that before you move on to the next point? For years, I've been trying to build different internal LLM tools for myself. Often I have these text-in, text-out tasks, which shoul..."

"But there's not this ongoing way I can engage with them to help them get better at the job the way I could with a human employee. That missing ability, even if you solve computer use, would still bloc..."

"This gets back to what we were talking about before with learning on the job. It's very interesting. I think with the coding agents, I don't think people would say that learning on the job is what is ..."

"When I see Claude Code, familiarity with the codebase or a feeling that the model hasn't worked at the company for a year, that's not high up on the list of complaints I see. I think what I'm saying i..."

"Coding made fast progress precisely because it has this unique advantage that other economic activity doesn't. But when you say that, what you're implying is that by reading the codebase into the cont..."

"So that would be an example of—whether it's written or not, whether it's available or not—a case where everything you needed to know you got from the context window. What we think of as learning—"I st..."

"I honestly don't know how to think about this because there are people who qualitatively report what you're saying. I'm sure you saw last year, there was a major study where they had experienced devel..."

"So I'm trying to square the qualitative feeling that people feel with these models versus, 1) in a macro level, where is this renaissance of software? And then 2) when people do these independent eval..."

"The pressure to survive economically while also keeping our values is just incredible. We're trying to keep this 10x revenue curve going. There is zero time for bullshit. There is zero time for feelin..."

"Why do you think we’re concerned about competitors using the tools? Because we think we're ahead of the competitors. We wouldn't be going through all this trouble if this were secretly reducing our pr..."

"The models make you more productive. 1) People feeling like they're productive is qualitatively predicted by studies like this. But 2) if I just look at the end output, obviously you guys are making f..."

"But the idea was supposed to be that with recursive self-improvement, you make a better AI, the AI helps you build a better next AI, et cetera, et cetera. What I see instead—if I look at you, OpenAI, ..."

"But why are we not seeing the person with the best coding model have this lasting advantage if in fact there are these enormous productivity gains from the last coding model. I think my model of the s..."

"Six months ago, it was maybe 5%. So it didn't matter. 5% doesn't register. It's now just getting to the point where it's one of several factors that kind of matters. That's going to keep speeding up. ..."

"I would also say there are multiple companies that write models that are used for code and we're not perfectly good at preventing some of these other companies from using our models internally. So I t..."

"Again, my theme in all of this is all of this is soft takeoff, soft, smooth exponentials, although the exponentials are relatively steep. So we're seeing this snowball gather momentum where it's like ..."

"As you go, Amdahl's law, you have to get all the things that are preventing you from closing the loop out of the way. But this is one of the biggest priorities within Anthropic."

"Stepping back, before in the stack we were talking about when do we get this on-the-job learning? It seems like the point you were making on the coding thing is that we actually don't need on-the-job ..."

"Maybe that's not your claim, you should clarify. But in most domains of economic activity, people say, "I hired somebody, they weren't that useful for the first few months, and then over time they bui..."

"If AI doesn't develop this ability to learn on the fly, I'm a bit skeptical that we're going to see huge changes to the world without that ability. I think two things here. There's the state of the te..."

"So it's like learning, but it's like learning from more data and not learning over one human or one model's lifetime. So again, this is situated between evolution and human learning. But once you lear..."

"It knows more about baseball than I do. It knows more about low-pass filters and electronics, all of these things. Its knowledge is way broader than mine. So I think even just that may get us to the p..."

"I would describe it as kind of like human on-the-job learning, but a little weaker and a little short term. You look at in-context learning and if you give the model a bunch of examples it does get it..."

"If you think about the model reading a million words, how long would it take me to read a million? Days or weeks at least. So you have these two things. I think these two things within the existing pa..."

"There may be gaps, but I certainly think that just as things are, this is enough to generate trillions of dollars of revenue. That's one. Two, is this idea of continual learning, this idea of a single..."

"Again, I think you get most of the way there without it. The trillions of dollars a year market, maybe all of the national security implications and the safety implications that I wrote about in "Adol..."

"There are a bunch of ideas. I won't go into all of them in detail, but one is just to make the context longer. There's nothing preventing longer contexts from working. You just have to train at longer..."

"Both of those are engineering problems that we are working on and I would assume others are working on them as well. This context length increase, it seemed like there was a period from 2020 to 2023 w..."

"I feel like for the two-ish years since then, we've been in the same-ish ballpark. When context lengths get much longer than that, people report qualitative degradation in the ability of the model to ..."

"This isn't a research problem. This is an engineering and inference problem. If you want to serve long context, you have to store your entire KV cache. It's difficult to store all the memory in the GP..."

"At this point, this is at a level of detail that I'm no longer able to follow, although I knew it in the GPT-3 era. "These are the weights, these are the activations you have to store…" But these days..."

"If you train at a small context length and then try to serve at a long context length, maybe you get these degradations. It's better than nothing, you might still offer it, but you get these degradati..."

"So I think the “country of geniuses in a data center” will be more than capable of supporting such tasks as video editing by one to two years from now."

"My guess for that is there's a lot of problems where basically we can do this when we have the "country of geniuses in a data center". My picture for that, if you made me guess, is one to two years, m..."

"I have a strong view—99%, 95%—that all this will happen in 10 years. I think I have a hunch—this is more like a 50/50 thing—that it's going to be more like one to two, maybe more like one to three."

"So one to three years. Country of geniuses, and the slightly less economically valuable task of editing videos. It seems pretty economically valuable, let me tell you. It's just there are a lot of use..."

"So you're predicting that within one to three years. And then, generally, Anthropic has predicted that by late '26 or early '27 we will have AI systems that "have the ability to navigate interfaces av..."

"You gave an interview two months ago with DealBook where you were emphasizing your company's more responsible compute scaling as compared to your competitors. I'm trying to square these two views."

"If you really believe that we're going to have a country of geniuses, you want as big a data center as you can get. There's no reason to slow down. The TAM of a Nobel Prize winner, that can actually d..."

"So I'm trying to square this conservatism, which seems rational if you have more moderate timelines, with your stated views about progress. It actually all fits together. We go back to this fast, but ..."

"Let's say that we're making progress at this rate. The technology is making progress this fast. I have very high conviction that we’re going to get there within a few years. I have a hunch that we’re ..."

"I really do believe that we could have models that are a country of geniuses in the data center in one to two years. One question is: How many years after that do the trillions in revenue start rollin..."

"It could be one year, it could be two years, I could even stretch it to five years although I'm skeptical of that. So we have this uncertainty. Even if the technology goes as fast as I suspect that it..."

"We know it's coming, but with the way you buy these data centers, if you're off by a couple years, that can be ruinous. It is just like how I wrote in "Machines of Loving Grace". I said I think we mig..."

"I said we'll get that in 2026, maybe 2027. Again, that is my hunch. I wouldn't be surprised if I'm off by a year or two, but that is my hunch. Let's say that happens. That's the starting gun."

"How long does it take to cure all the diseases? That's one of the ways that drives a huge amount of economic value. You cure every disease. There's a question of how much of that goes to the pharmaceu..."

"How long does it take? You have to do the biological discovery, you have to manufacture the new drug, you have to go through the regulatory process. We saw this with vaccines and COVID. We got the vac..."

"My question is: How long does it take to get the cure for everything—which AI is the genius that can in theory invent—out to everyone? We've had a polio vaccine for 50 years. We're still trying to era..."

"Others are trying as hard as they can. But that's difficult. Again, I don't expect most of the economic diffusion to be as difficult as that. That's the most difficult case. But there's a real dilemma..."

"looking at $10 billion in annualized revenue. We have to decide how much compute to buy. It takes a year or two to actually build out the data centers, to reserve the data center. Basically I'm saying..."

"So it seems like it's possible that we actually just have different definitions of the "country of a genius in a data center". Because when I think of actual human geniuses, an actual country of hum..."

"Human wages, let's say, are on the order of $50 trillion a year— So I won't talk about Anthropic in particular, but if you talk about the industry, the amount of compute the industry is building this..."

"You've told investors that you plan to be profitable starting in 2028. This is the year when we're potentially getting the country of geniuses as a data center. This is now going to unlock all this pr..."

"What I'm trying to get at is that you have a model in your head of a business that invests, invests, invests, gets scale and then becomes profitable. There's a single point at which things turn arou..."

"That's the profitable business model that I think is kind of there, but obscured by these building ahead and prediction errors. I guess you're treating the 50% as a sort of given constant, whereas i..."

"Why doesn't everyone spend 100% of their compute on training and not serve any customers? It's because if they didn't get any revenue, they couldn't raise money, they couldn't do compute deals, they ..."

"Maybe stepping back, I'm not saying I think the "country of geniuses" is going to come in two years and therefore you should buy this compute. To me, the end conclusion you're arriving at makes a lo..."

"I think the way the profit comes is… Again, let's just abstract the whole industry here. Let's just imagine we're in an economics textbook. We have a small number of firms. Each can invest a limited a..."

"Then this year it produced $4 billion of revenue and cost $1 billion to inference from. Again, I'm using stylized numbers here, but that would be 75% gross margins and this 25% tax. So that model as ..."

"A fixed lump of labor fallacy… The economy is going to grow, right? That's one of your predictions. We're going to have the data centers in space. Yes, but this is another example of the theme I was ..."

"So no, I don't think this field's going to be a monopoly. All my lawyers never want me to say the word "monopoly". But I don't think this field's going to be a monopoly."

"But I do think when for whatever reason the models have those skills, then robotics will be revolutionized—both the design of robots, because the models will be much better than humans at that, and a..."

"No, I gave this spectrum: 90% of code, 100% of code, 90% of end-to-end SWE, 100% of end-to-end SWE."

"I actually do think that the API model is more durable than many people think."

"At some point we're going to see "pay for results" in some form, or we may see forms of compensation that are like labor, that kind of work by the hour."

"That means that lots of people will be able to build huge populations of misaligned AIs, or AIs which are just companies which are trying to increase their footprint or have weird psyches like Sydney ..."

"I agree that that doesn't solve the problem in the long run, particularly if the ability of AI models to make other AI models proliferates, then the whole thing can become harder to solve."

"We've gotten used to the presence of explosives in society or the presence of various new weapons or the presence of video cameras."

"But I don't know. I don’t want to say this is so far ahead in time, but it’s so far ahead in technological ability that may happen over a short period of time, that it's hard for us to anticipate it i..."

"I don't worry as much about the chatbot laws. I actually worry more about the drug approval process, where I think AI models are going to greatly accelerate the rate at which we discover drugs, and th..."

"At the same time, I think we should be ramping up quite significantly the safety and security legislation. Like I've said, starting with transparency is my view of trying not to hamper the industry, t..."

"I feel like you have worked with legislatures to say, "Okay, we're going to prevent bioterrorism here. We're going to increase transparency, we're going to increase whistleblower protection." But I th..."

"We're seeing that in AI itself. A thing I've been trying to fight for is export controls on chips to China. That's in the national security interest of the US. That's squarely within the policy belief..."

"So if we're talking about drugs and benefits of the technology, I am not as worried about those benefits being hampered in the developed world. I am a little worried about them going too slow. As I sa..."

"Why shouldn't it happen or why shouldn't it happen? Why shouldn't it happen. If this does happen, we could have a few situations. If we have an offense-dominant situation, we could have a situation li..."

"My worry here is about governments. My worry is if the world gets carved up into two pieces, one of those two pieces could be authoritarian or totalitarian in a way that's very difficult to displace. ..."

"At some point, we're going to need to set up the rules of the road. I'm not saying that one country, either the United States or a coalition of democracies—which I think would be a better setup, altho..."

"...Today, the view, my view, in most of the Western world is that democracy is a better form of government than authoritarianism. But if a country’s authoritarian, we don’t react the way we’d react if..."

"We have to decide one way or another how to deal with that. The interventionist view is one possible view. I was exploring such views. It may end up being the right view, or it may end up being too ex..."

"I am actually hopeful that—it sounds too idealistic, but I believe it could be the case—dictatorships become morally obsolete. They become morally unworkable forms of government and the crisis that th..."

"Speaking of distribution, as you were mentioning, we have developing countries. In many cases, catch-up growth has been weaker than we would have hoped for. But when catch-up growth does happen, it's ..."

"In fact, I think it'd be great to build data centers in Africa. As long as they're not owned by China, we should build data centers in Africa. I think that's a great thing to do. There's no reason we ..."

"One is, should we give the model a set of instructions about "do this" versus "don't do this"? The other is, should we give the model a set of principles for how to act? It's kind of purely a practica..."

"I think there are maybe three sizes of loop here, three ways to iterate. One is we iterate within Anthropic. We train the model, we're not happy with it, and we change the constitution. I think that's..."

"I think a few things. One is, at every moment of this exponential, the extent to which the world outside it didn't understand it. This is a bias that's often present in history. Anything that actually..."

"It's a lot of things. It's me, it's Daniela, who runs the company day to day, it's the co-founders, it's the other people we hire, it's the environment we try to create. But I think an important thing..."