Dario Amodei — “We are near the end of the exponential”

[00:00] So, we talked 3 years ago. I'm curious in your view, what has been the biggest update of the last 3 years? What has been the biggest difference between what it felt like last 3 years versus now? >> Yeah, I would say actually the underlying technology, like the exponential of the technology, has has gone, broadly speaking, I would say about about as I expected it to go. I mean, there's like plus or minus, you know, a couple. There's plus or minus a year or two here, there's plus or minus a year or two there. I don't know that I would predict the specific direction of code. Um, but but actually when I look
[00:32] at the exponential, it it is roughly what I expected in terms of the march of the models from like, you know, smart high school student to smart college student to like, you know, beginning to do PhD and professional stuff and in the case of code reaching beyond that. So, you know, the frontier is a little bit uneven. It's roughly what I expected. I will tell you though what the most surprising thing has been. The most surprising thing has been the lack of public recognition of how close we are to the end of the exponential. To me, it
[01:02] is absolutely wild that, you know, you have people, you know, within the bubble and outside the bubble, you know, but but you have people talking about these these, you know, just the same tired old hot button political issues and like, you know, or or or around us for like near the end of the exponential. I I I want to understand what what that exponential looks like right now because the first question I asked you when we recorded 3 years ago was, you know, what's up with scaling? How does it work? Um, and I have a similar
[01:32] question now, but I feel like it's a more complicated question because at least from the public's point of view, yes, 3 years ago there were these, you know, well-known public trends where across many orders of magnitude of compute you could see how the loss improves. And now we have RL scaling and there's no publicly known scaling law for it. It's not even clear what exactly the story is of is it supposed to be teaching the model skills? Is it supposed to be teaching meta learning? Um, what is the scaling hypothesis at this point? >> Yeah, so so I have actually the same hypothesis that I had even all the way
[02:04] back in 2017. So, in 2017, I think I talked about it last time, but I wrote a doc called the the big blob of compute hypothesis. And and and you know, it it wasn't about the scaling of language models in particular. When I when I wrote it, GPT-1 had had just come out, right? So, that was, you know, one among many things, right? There was back in those days there was robotics, people trying to work on reasoning as a separate thing from language models, there was scaling of the kind of RL that happened that you know, kind of happened in AlphaGo, and you know, that that
[02:34] happened at Dota at OpenAI, and you know, people remember StarCraft at DeepMind, you know, AlphaStar. Um, so uh it was written as a more general document. And and the specific thing I said was the following, that and you know, it's it's very, you know, Rich Sutton put out the bit bitter lesson a couple years later, um uh but you know, the the hypothesis is basically the same. So, so what it says is all the cleverness, all the techniques, all all the kind of we need a new method to to do something like
[03:05] that doesn't matter very much. There only a few things that matter, and I think I listed seven of them. One is like how much raw compute you have. The other is the quantity of data that you have. Then the third is kind of the quality and distribution of data, right? It needs to be a broad broad distribution of data. The fourth is I think how long you train for. Um, the fifth is you need an objective function that can scale to the moon. So, the pre-training objective function is one such objective function, right? An
[03:36] another objective function is, you know, the the kind of RL objective function that says like you have a goal, you're going to go out and reach the goal. Within that, of course, there's objective rewards like, you know, like you see in math and coding, and there's more subjective rewards like you see in RL from human feedback or kind of higher order higher order versions of that. And and then the sixth and seventh were things around kind of like normalization or conditioning. Like you know, just getting the numerical stability so that kind of the big blob of compute flows in
[04:07] this laminar way instead of instead of running into problems. So that was the hypothesis and it's a hypothesis I still hold. I I don't think I've seen very much that is not in line with that hypothesis. And so the pre-trained scaling laws were one example of what of of of of kind of what we see there. And indeed those have continued going. Like you know, you know, I think I think now it's been it's been widely reported like you know, we feel good about pre-training. Like pre-training is continuing to give us gains. What has
[04:39] changed is that now we're also seen the same thing for RL, right? So we're seeing a pre-training phase and then we're seeing like an RL phase on top of that. Um And with RL it's it's actually just the same like you know, even even other companies have have published um uh um like um you know, in some of their in some of their releases have published things that say, "Look, you know, we trained the model on math contests, you know, AIME or or the kind of other
[05:09] things and you know, how well the how well the model does is log linear and how long we've trained it. And we see that as well and it's not just math contest. It's a wide variety of RL tasks." And so we're seeing the same scaling in RL that we saw for pre-training. Um You mentioned Rich Richard Sutton and the bitter lesson. Yeah. I interviewed him uh last year and he is actually very non LLM pilled. And if I'm if I I don't know if this is disrespectful, but one way to paraphrase
[05:40] this objection is something like, "Look, something which possesses the true core of human learning will not require all these billions of dollars of data and compute and these bespoke environments to learn how to use Excel or how does an account you know how to how to use PowerPoint how to navigate a web browser and the fact that we have to build in these skills using these RL environments hints that we're actually lacking this core human learning algorithm
[06:10] and so we're scaling the wrong thing and so yeah that that is the reason question why are we doing all this RL scaling if we do think there's something that's going to be human-like in its ability to learn on the fly. Yeah, yeah. So I think I think this kind of puts together several things that should be kind of thought of thought of differently. I think there is a genuine puzzle here but it it may not matter. In fact I would guess it probably it probably doesn't matter. So let's take the RL out of it for a second cuz I actually think RL and it's a red herring to say that RL is any different
[06:41] from pre-training in this matter. So if we if we look at pre-training and scaling it it was very interesting back in you know 2017 when Alec Radford was doing GPT-1. If you look at the models before GPT-1 they were trained on these data sets that didn't represent a wide you know distribution of text right you had like you know these very standard you know kind of language modeling benchmarks and GPT-1 itself was trained
[07:11] on a bunch of I think it was fan fiction actually but you know it was it was like literary and it's like literary text which is a very small fraction of the text that you get and what we found with that you know and in those days it was like a billion words or something so small data sets and represented a pretty narrow distribution right like a narrow distribution of kind of what what you can see what you can see in the world and it didn't generalize well if you did better on you know the the you know I I I forgot what what some
[07:41] some kind of fan fiction corpus. Um, it wouldn't generalize that well to kind of the other task. You know, we had all these measures of like, you know, how well does it how well does a model do at predicting all of these other kinds of tasks? You really didn't see the generalization. It was only when you trained over all the tasks on the you know, the internet. When you when you kind of did a general internet scrape, right? From something like, you know, common crawl or scraping links on Reddit, which is what we did for GPT-2. It's only when you do that that you kind of started to get generalization. Um,
[08:13] and I think we're seeing the same thing on RL, that we're starting with first very simple RL tasks like training on math competitions, then we're kind of moving to, you know, kind of broader broader training that involves things like code as a task, and now we're moving to do kind of many many other tasks. And then I think we're going to increasingly get generalization. So, that that kind of takes out the RL versus the pre-training side of it. But, I think there is a puzzle here either way, which is that on pre-training, when
[08:43] we train the model on pre-training, you know, we we use like trillions of tokens, right? And and humans don't see trillions of words. So, there is an actual sample efficiency difference here. There there is actually something different that's that's happening here, which is that the models start from scratch, and you know, they have to get much more much more training. But, we also see that once they're trained, if we give them a long context length. The only thing a long context length is like inference, but if we give them like a
[09:14] context length of a million, they're very good at learning and adapting within that context length. And and so, I don't know the full answer to this, but but I think there's something going on that pre-training it's it's not like the process of humans learning. It's somewhere between the process of humans learning and the process of human evolution. It's like it's somewhere between like we get many of our priors from evolution. Our brain isn't just a blank slate, right? Whole books have been written about I think the language models, they're much more blank slates.
[09:45] They literally start as like random weights. Whereas the human brain starts with all these regions, it's connected to all these inputs and outputs. Um and and so maybe we should think of pre-training and for that matter RL as well as as being something that exists in the middle space between human evolution and you know, kind of human on on the spot learning. And as the in-context learning that the models do as as something between long-term human
[10:15] learning and short-term human learning. So, you know, there there's this hierarchy of like there's evolution, there's long-term learning, there's short-term learning, and there's just human reaction. And the LLM phases exist along this spectrum, but not necessarily exactly at the same points. That there's no analog to some of the human modes of learning. The LLMs are kind of falling between the points. Does that make sense? >> Um yes, although some things are still a bit confusing. For example, if the analogy is that this is like evolution,
[10:45] so it's fine that it's not that sample efficient, then like well, if we're going to get the kind of super sample efficient agent from in-context learning, why are we bothering to build in, you know, there's RL environment companies which are It seems like what they're doing is they're teaching it how to use this API, how to use Slack, how to use whatever. It's confusing to me why there's so much emphasis on that if the kind of agent that can just learn on the fly is emerging or is going to soon emerge or has already emerged. >> Yeah, yeah. So, I I mean, I can't speak for the emphasis of anyone else. I can I can only talk about how we how we think
[11:16] about it. I think the way we think about it is the goal is not to teach the model every possible skill within RL just as we don't do that within pre-training, right? Within pre-training, we're not trying to expose the model to, you know, every every possible you know, way that words could be put together, right? You know, we're it's it's rather that the model trains on a lot of things and then and then it reaches generalization across pre-training, right? That was that was the transition from GPT-1 to
[11:46] GPT-2 that I saw up close, which is like, you know, the the model reaches a point, you know, I I I I like had these moments where I was like, "Oh, yeah, you just give the model like you just give the model a list of numbers that's like, you know, um you know, this is the cost of the house, this is the square feet of the house." And the model completes the pattern and does linear regression. Like, not great, but it does it, but it's never seen that exact thing before. And and so, to you know, to to the extent that we are building these RL environments, the the
[12:18] goal is is very similar to what is be you know, to what was done 5 or 10 years ago with pre-training. With we're trying to get a whole we're trying to get a whole bunch of data not because we want to cover a specific document or a specific skill, but because we want to generalize. I mean, I I think the framework you're laying down obviously makes sense. Like, we're making progress towards AGI. I think the crux is something like nobody at this point disagrees that we're going to achieve AGI in the century. And the crux is you say we're
[12:49] hitting the end of the exponential um and somebody else looks at this and says, "Oh, yeah, we've we're making progress. We've been making progress since 2012 and then 2035 we'll have a human-like AGI." And so, I want to understand what it is that you're seeing which makes you think um yeah, obviously we're seeing the kinds of things that evolution did or that human within human lifetime learning is like in the these models. And why think that it's 1 year away and not 10 years away? >> I I I actually think of it as like two there's kind of two cases to be made
[13:19] here or like two two claims you could make, one of which is like stronger and the other of which is weaker. So, I think starting starting with the weaker claim, you know, when when I first saw the scaling back in like, you know, 2019, you know, I wasn't sure. You know, this was the the whole this was kind of a 50/50 thing, right? I thought I saw something that was, you know, and and and my claim was this is much more likely than anyone thinks it is. Like, this is wild. No one else would even consider this. Maybe there's a 50%
[13:49] chance this happens. Um on the basic hypothesis of you know, as you put it, within 10 years we'll get to, you know, you know, what I call kind of country of geniuses in a data center. I'm at like 90% on that. Um and it's hard to go much higher than 90% cuz the world is so unpredictable. Um maybe the irreducible uncertainty would be if we were at 95% where you get to things like, I don't know, may maybe multiple you know, multiple companies have, you know, kind of internal turmoil
[14:20] and nothing happens and then Taiwan gets invaded and like all the all the fabs get blown up by missiles and and, you know, and then now you were drinking the Astoria. Yeah. You know, just you could construct a a scenario where there's like a 5% chance that it it you know, or you know, you you can construct a 5% world where like things things get delayed for for for for for for for 10 years. That's maybe 5%. There's another 5% which is that I'm very confident on tasks that can be verified. So, I think I think with
[14:50] coding, I'm just except for that irreducible uncertainty, there's just there's I mean, I think we'll be there in one or two years. There's no way we will not be there there in 10 years in terms of being able to do it end-to-end coding. My one little bit, the one little bit of of fundamental uncertainty even on long time scales is this thing about tasks that aren't verifiable. Like, planning a mission to Mars, like, uh you know, doing some fundamental scientific discovery like like CRISPR, like, you know, writing a writing a novel.
[15:21] Hard to hard to verify those tasks. I am almost certain that we have a reliable path to get there, but like if there was a little bit uncertainty, it's there. So so so so So on the 10 years, I'm like, you know, 90% which is about as certain as you can be. Like I think it's I think it's crazy to say that this won't happen by by by 2035. Like in some sane world it would be outside the mainstream. But but the emphasis on verification hints to me as
[15:54] a lack of a lack of belief that these models will generalize. If you think about humans, yes. We are good at things that both of which we get verifiable reward and things which we don't. You're like you have to discard >> We we No no no this is this is why I'm almost sure. We already see substantial generalization from things that that verify to things that don't verify. We're already seeing that. >> Right. But but it seems like you were emphasizing this as a spectrum which will uh split apart which domains you see more progress in. And I'm like, but that doesn't seem like how humans get better. >> in which we don't make it or or or the
[16:25] world in which we don't get there is the world in which we do we do all the things that are that are verifiable and then they like, you know, many of them generalize, but but we kind of don't get fully there. We don't we don't we don't fully, you know, we don't fully color in this side of the box. It's it's it's not a it's not a binary thing. But it also seems to me even if even if in the world where generalization is weak when you only see it in verifiable domains, it's not clear to me in such a world you could automate software engineering because software like in some sense you are {quote} a software engineer. But you
[16:56] part of being a software engineer for you involves writing these like long memos about your grand vision about different things. And so >> think that's part of the job of SWE. That's part that's part of the job of the company. But I do think SWE involves like design documents and other things like that. Um which by the way that the models are not bad. They're already pretty good at writing comments. And so with with again I again I'm making like much weaker claims here than I believe to like, you know, to to to to to kind of set up a you know, to to distinguish between two things. Like, we're we're already almost there for software engineers. We are
[17:26] already almost there. By by what metric? There's one metric which is like how many lines of code are written by AI? And if you use if you consider other productivity improvements in the course of the history of it software engineering, compilers write all the lines of software. And but we there's a difference between how many lines are written and how big the productivity improvement is. Oh, yeah. So and then like We're almost there meaning like the how big is the productivity improvement and not just how many lines are written. >> Yeah, yeah. So so I actually I actually I actually agree with you on this. So I I've made this series of predictions on um code in software
[17:57] engineering. And and and I think people have repeatedly kind of misunderstood them. So so let me let me let me let me let me lay out the spectrum, right? Like I think it was like, you know, like, you know, eight or nine months ago or something I said, you know, the AI model will be writing 90 90% of the lines of code in like, you know, three to six months. Which which happened at least at some places, right? Happened happened at Anthropic, happened with many people downstream using our models. But but that's actually a very weak criterion, right? People thought I was saying like
[18:29] we won't need 90% of the software engineers. Those things are worlds apart, right? Like I would put the spectrum as 90% of code is written by the model, 100% of code is written by the model. And that's a big difference in productivity. 90% of the end-to-end SWE tasks, right? Including things like compiling, including things like setting up clusters and environments, testing features, writing memos. 90% of the SWE tasks are written by the models. 100% of today's SWE tasks are are are written by
[19:01] the models. And and even when when when that happens, it doesn't mean software engineers are out of a job. Like there's like new higher-level things they can do where they can they can manage. And then there's a further down the spectrum like you know, there's 90% less demand for sweets, which I think will happen, but like this is this this is a spectrum. And you know, I I wrote about it in in the adolescence of technology, where I went through this kind of spectrum with farming. Um And so, I I actually totally agree with you on that. It's just these are very different benchmarks from each other,
[19:32] but we're proceeding through them super fast. It seems like in part of your vision is like going from 90 to 100. Um first it's going to happen fast, and two, that somehow that leads to huge productivity improvements. Um whereas when I notice even in greenfield projects that people start with Claude code or something, people report starting a lot of projects, and I'm like, do we see in the world out there a renaissance of software, all these new features that wouldn't exist otherwise? And at least so far, it doesn't seem like we see that. And so, that does make me wonder, even if even if like I never
[20:03] had to intervene on Claude code, um there is this thing like there's this the world is complicated, jobs are complicated, and closing the loop on self-contained systems, whether I'm just writing software or something, how much sort of how much broader gains we would see just from that. And so, maybe that makes us this should dilute our estimation of the country of geniuses. >> well, I actually I I like I like simultaneously I simultaneously agree with you, agree that it's a reason why these things
[20:34] don't happen instantly, but at the same time, I think the the the effect is going to be very fast. So, like, I don't know, you could have these two poles, right? One is like, um you know, AI is like, you know, it's not going to make progress, it's slow, like it's going to take, you know, kind of forever to diffuse within the economy, right? Economic diffusion has become one of these buzzwords that's like a reason why we're not going to make AI progress, or why AI progress doesn't matter. And And you know, the other axis is like we'll get recursive self-improvement, you know, the whole thing, and you know,
[21:05] can't you just draw an exponential line on the on the curve? You know, it's it's we're going to have, you know, Dyson spheres around the sun and like, you know, you know, so many nanoseconds after you know, after after we get recurse. I mean, I'm completely caricaturing the view here, but like, you know, there there there are these two extremes, but what we've seen from from the beginning, you know, at least if you look within Anthropic, there's this bizarre 10x per year growth in revenue that we've seen. Right? So, you know, in 2023, it was like 0 to 100
[21:37] million. 2024, it was 100 million to a billion. 2025, it was a billion to like 9 or 10 billion. And then You guys should have just bought like a billion dollars worth of your own product so you could just like kind of clean 10 billion. >> [laughter] >> And and the first month of this year, like that that exponential is you would think it would slow down, but it would like, you know, we we added another few billion to like, you know, to to to we added another few billion to revenue in January. And and so, you know, obviously that curve can't go on forever, right?
[22:08] You know, the GDP is only so large. I don't you know, I I would even guess that it bends that it bends bends somewhat this year. But like, that is like a fast curve, right? That's like a that's like a really fast curve, and I would bet it stays pretty fast even as the scale goes to the entire economy. So like, I I think we should be thinking about this middle world where things are like extremely fast, but not instant, where they take time because of economic diffusion, because of the need to close
[22:38] the loop, because, you know, it's like this fiddly, oh man, I have to do change management within my enterprise, you know, I have to like, you know, you know, I I I like I set this up, but but, you know, I have to change the security permissions on this in order to make it actually work. Or, you know, I had this like old piece of software that, you know, that like, you know, checks the model before it's compiled and and and like released, and I have to rewrite it and yes the model can do that but I have to tell the model to do that and it has to it has to take time to do
[23:08] that and and and so I think everything we've seen so far is is compatible with the idea that there's one fast exponential that's the the capability of the model and then there's another fast exponential that's downstream of that which is the diffusion of the model into the economy. Not instant. Not slow. Much faster than any previous technology but it has its limits. And and and and this is what we you know when I when I look inside Anthropic,
[23:38] when I look at our customers, fast adoption but not infinitely fast. Um Can I try a hot take on you? Yeah. I feel like diffusion is cope that people use to say when it's like if the model isn't able to do something, they're like oh but the diffu- it's like a diffusion issue. But then you should use the comparison to humans. You would think that the inherent advantages that EIs have would make diffusion a much easier problem for new EIs getting on boarded than new humans getting on boarded. So an AI can read your entire Slack and your Drive in minutes. They
[24:08] can share all the knowledge that the other copy other copies of the same instance have. You don't have this adverse selection problem when you're hiring EIs cuz you can just hire copies of a vetted AI model. Um hiring a human is like so much more hassle and people hire humans all the time, right? We pay humans upwards of 50 trillion dollars in wages because they're useful uh even though it's like in principle it would be much easier to integrate EIs into the economy than it is to hire humans. So I think like the diffusion I feel like doesn't really explain >> I think diffusion is very real and and
[24:38] and and and doesn't have to you know doesn't exclusively have to do with limitation limitation limitations on the AI models. Like again, there are people who use diffusion to to you know as kind of a buzzword to say this isn't a big deal. I'm not talking about that. I'm not talking about you know AI will diffuse at the speed that previous I think AI will diffuse much faster than previous technologies have but but not infinitely fast. So I'll I'll just give an example of this, right? Like there's like Claude Code. Like Claude Code is
[25:08] extremely easy to set up. You know, if you're a developer, you can kind of just start using Claude Code. There is no reason why a developer at a large enterprise should not be adopting Claude Code as quickly as you know, individual developer or developer at a startup. And we do everything we can to promote it, right? We sell we sell Claude Code to enterprises and big enterprises like you know, big big financial companies, big pharmaceutical companies, all of them, they're adopting
[25:38] Claude Code much faster than enterprises typically adopt new technology, right? But but again, it like it it it it it it it takes time. Like any given feature or any given product like Claude Code or like Co-work will get adopted by the you know, the individual developers who are on Twitter all the time, by the like Series A startups many months faster than than you know, than they will get adopted by like you know,
[26:08] a like large enterprise that does food sales. There are a number of factors like you have to go through legal. You have to provision it for everyone. It has to you know, like it has to pass security and compliance. The leaders of the company who are further away from the AI revolution, you know, are are forward-looking but they have to say, "Oh, it makes sense for us to spend 50 million. This is what this Claude Code thing is. This is why it helps our company. This is why it makes us more productive." And then they have to explain to the people two levels below
[26:39] and they have to say, "Okay, we have 3,000 developers. Like here's how we're going to roll it out to our developers." And we have conversations like this every day. Like you know, we are doing everything we can to make Anthropic's revenue grow 20 or 30x a year instead of 10x a year. Um, you know, and and and again, you know, many enterprises are just saying, "This is so productive. Like, you know, we're going to take shortcuts in our usual procurement process, right?" They're moving much faster than, you know, when we tried to sell them just the ordinary API, which
[27:09] many of them use, but Claude code is a more compelling product. Um, but it's not an infinitely compelling product, and 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 three or five or 10x a year growth even when you're in the hundreds of billions of dollars, which is extremely hard to do and has never been done in history before, but not infinitely fast. >> I I I buy that it would be a slight slowdown. And maybe this is not your claim, but sometimes people talk about this like, "Oh, the capabilities are
[27:40] there, but because of the fusion, um, otherwise, like, we're basically at AGI and then >> I I I I don't believe we're basically at AGI. I think if you had the country of geniuses in a data center, if your company didn't adopt the country of geniuses in a data center, >> in a data center, we would know it. We would know everyone in this room would know it. Everyone in Washington would know it. Like, you know, people in rural rural parts of might not know it. But but but like, we would know it. I We don't have that
[28:11] now. That is very clear. As Dario was hinting at, to get generalization, you need to train across a wide variety of realistic tasks and environments. [music] For example, with a sales agent, the hardest part isn't teaching it to mash buttons in a specific database in Salesforce. It's training the agent's judgment across ambiguous situations. How do you sort through a database with thousands of leads to figure out which ones are hot? How do you actually reach out? What do you do when you get ghosted? When an AI lab wanted to train a sales agent, Labelbox brought in dozens of Fortune 500 sales people to
[28:41] build a bunch of different RL environments. They created thousands of scenarios where the sales agent had to engage with the potential customer, which was role-played by a second AI. Labelbox made sure that this customer AI had a few different personas, because when you cold call, you have no idea who's going to be on the other end. You need to be able to deal with a whole range of possibilities. Labelbox's sales experts monitored these conversations turn by turn, [music] tweaking the role-playing agent to ensure they did the kinds of things an actual customer would do. Labelbox could iterate faster than anybody else in the industry. This is super important
[29:12] because RL is an empirical science. [music] It's not a software problem. Labelbox has a bunch of tools for monitoring agent performance in real time. This lets their experts keep coming up with tasks, so that the model stays in the right distribution of difficulty and [music] gets the optimal reward signal during training. Labelbox can do this sort of thing in almost every domain. They've got head fund managers, radiologists, even airline pilots. So, whatever you're working on, Labelbox [music] can help. Learn more at labelbox.com/vorkash.
[29:42] Coming back to concrete predictions, because I think because there's so many different things to disambiguate, it can be easy to talk past each other when we're talking about capabilities. So, for example, when I interviewed you 3 years ago, I asked for a prediction about what we should expect 3 years from now. I think you were right in what you said. We should expect systems which, if you talk to them for the course of an hour, it's hard to tell them apart from a generally well-educated human. Yes. I think you were right about that. And I think spiritually I feel unsatisfied because my internal expectation was was
[30:12] that such a system could automate large parts of white-collar work. And so, it might be more productive to talk about the actual end capabilities you want such a system So, so I will I will I will basically tell you what what, you know, where where where I think we are. So, but let me let me ask you a very specific question so that we can figure out exactly what kinds of capabilities we should expect soon. So, maybe I'll ask about it in the context of a job I understand well, not because it's the most relevant job, but um just cuz I can evaluate the claims about it. Um take video editors, right? I video
[30:42] editors, and part of their job involves learning about our audience's preferences, learning about my preferences and taste and the different trade-offs we have, and how just over the course of many months building up this understanding of context. And so, the skill and ability they have 6 months into the job, a model that can pick up that skill on the job, on the fly. When should we expect such an AI system? Yeah, so I guess what you're talking about is like, you know, we've we're we're doing this interview for 3 hours, and then like, you know, someone's going to come in, someone's going to edit it, they're going to be like, oh, you know,
[31:13] you know, I don't know, Dario like, you know, scratched his head, and you know, we could we could edit that out, and you know, >> Magnify that. >> this like long there was this like long discussion that like is less interesting to people, and then then, you know, then there's other thing that's like more interesting to people, so, you know, let's let's let's kind of make this this edit. So, you know, I think the country of geniuses in a data center will be able to do that. The the way it will be able to do that is, you know, it will have general control of a computer screen, right? Like it, you know, and and and you'll be able to feed this in,
[31:43] and it'll be able to also use the computer screen to like go on the web, look at all your previous look at all your previous interviews, like look at what people are saying on Twitter in response to your interviews, like talk to you, ask you questions, talk to your staff, look at the history of kind of edits edits that you did, and from that like do the job. >> Yeah. Um so, I think that's dependent on several things. One, that's dependent and and and and I think this is one of the things that's actually blocking deployment, um getting to the point on computer use where the models are really
[32:13] masters at using the computer, right? And, you know, we've seen this climb in in benchmarks, and benchmarks are always, you know, imperfect measures, but like, you know, OS world is, you know, went from, you know, like 5%, you know, like uh I think when we first re-released, you know, uh uh computer use like a a year and a quarter ago, it was like maybe 15%, I don't remember exactly, but we've climbed from that to like 65 or 70% um and and you know, there may be harder measures as well, but but I think
[32:43] computer use has to pass a point of reliability. Can I just ask a follow-up on that before you move on to the next point? Um I often for years I've been trying to build different internal LLM tools for myself and I off often I have these text in text out tasks which should be dead center in the repertoire of these models and yet I still hire humans to do them just because it's if it's something like make identify what the best clips would be in this transcript and maybe they'll do like a seven out of 10 job at them but there's not this ongoing way I can engage with them to help them get
[33:13] better at the job the way I could with a human employee and so that missing ability even if you solve computer use would still block my ability to like offload an actual job to them. Again, there's there's this gets back to what we to kind of to kind of what what we were talking about before with learning on the job where it's it's very interesting. You know, I think I think with the coding agents like I don't think people would say that learning on the job is what is what is you know, preventing the coding agents from like you know, doing everything end to end like they keep they keep getting better.
[33:45] We have engineers at Anthropic who like don't write any code and when I look at the productivity to your to your previous question, you know, we have folks who say this this GPU kernel this chip I used to write it myself. I just have Claude do it and so there's this there's this enormous improvement in productivity and I don't know like when I see Claude code like familiarity with the code base or like it you know, or or a feeling that the model hasn't worked at the company for for a year. That's
[34:15] not high up on the list of complaints I see and so I think what I'm saying is we're we're like we're kind of taking a different path. >> don't you think with coding that's because there is an external scaffold of memory which exist instantiated in the code base which I don't know how many other jobs have coding made fast progress precisely because it has its unique uh advantage that other economic activity doesn't. >> But but when you say that, what you're what you're implying is that by reading the code base into the context, I have
[34:45] everything that the human needed to learn on the job. 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, right? And that And that what we think of as learning, like, "Oh man, I started this job. It's going to take me 6 months to understand the code base." The model just did it in the context. Yeah, I honestly don't know how to think about this because there there are people who qualitatively report what you're saying. Um there was
[35:15] a meter study, I'm sure you saw last year, where they had experienced developers try to close pull request in repositories that they were familiar with. And those developers reported an uplift. They They reported that they felt more productive with the use of these models. But in fact, if you look at their output and how much was actually merged back in, there's a 20% downlift. They were less productive as a result of using these models. And so I'm trying to square the qualitative feeling that people feel with these models versus one in a macro level, where are all these Where is this like renaissance of
[35:46] software? And then, two, when people do these independent evaluations, why are we not seeing the Yeah, so productive benefits that you would expect? >> Within Anthropic, this is just really unambiguous, right? We're under an incredible amount of commercial pressure and make it even hard harder for ourselves cuz we have all this safety stuff we do that I think we do more than than than other companies. So, like, the the the pressure to survive economically while also keeping our values is is just incredible, right? We're trying to keep this 10x revenue curve going. There's
[36:18] like there is zero time for There is zero time for feeling like we're productive when we're not. Like, these tools make us a lot more productive. Like why why do you think we're concerned about competitors using the tools? Because we think we're ahead of the competitors and like we don't we don't want to we we wouldn't be going through all this trouble if this was secretly reducing reducing our productivity. Like we see the end productivity every few months in
[36:50] the form of model launches. Like there's no kidding yourself about this. Like the models make you more productive. Um one, that is people feeling like they're productive is qualitatively predicted by studies like this. But two, if I just look at the end output, obviously you guys are making fast progress. But the fact, you know, the the idea was supposed to be with recursive self-improvement is that you make a better AI, the AI helps you build a better next AI, etc. etc. And what I see instead, if I look at the U, Open AI, DeepMind, is that people are just
[37:20] shifting around the podium every few months. And maybe you think that stops cuz you you won or whatever. But um but why 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? >> no no no. I I I mean I mean I mean I think it's all like my my model of situation is there's there's an advantage that's gradually growing. Like I would say right now the coding models give maybe I don't know a a like 15
[37:53] maybe 20% total factor speedup. Like that's my view. Um and 6 months ago it was maybe 5% and so and so it didn't matter. Like 5% doesn't register. It's now just getting to the point where it's like one of several factors that that kind of matters. And and that's going to that's going to keep speeding up. And so I think 6 months ago like, you know, but there were several there were several companies that were at roughly the same point because uh, know, this this wasn't uh, this wasn't a notable factor, but I think it's starting to
[38:24] speed up more and more. I, you know, I I would I would also say there are multiple companies that, you know, write models that are used for code, and, you know, we're not perfectly good at, you know, preventing some of these other companies from from from using from from from kind of using our models internally. Um, so, uh, you know, I think I think everything we're kind kind of everything we're seeing is consistent with this kind of, um, this kind of snowball model where you where, you know, there's no hard Again, my my my my my theme in all of this is like
[38:56] all of this is soft takeoff, like soft smooth exponentials, although the exponentials are relatively steep. And so, and so, we're seeing this snowball gather momentum where it's like 10%, 20%, 25%, you know, 40% And as you go, yeah, Amdahl's law, you have to get all the like things that are preventing you from from closing the loop out of the way, but like this is one of the biggest priorities with an Anthropic. Um, this a a stepping back, I think before in the stack we were talking about, um, well, when do we get this
[39:28] on-the-job learning? And it seems like the coding the point you were making in the coding thing is we actually don't need on-the-job learning. Uh, that you can have tremendous productivity improvements, you can have potentially trillions of dollars of revenue for AI companies without this basic human ability Maybe that's not your claim, you should clarify. Um, but without this basic human ability to learn on the job. But I just look at like in 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 built up the context understanding." It's
[39:58] actually hard to define what we're talking about here. But they they got something, and then now now they're they're powerhouse, and they're so valuable to us. And if AI doesn't develop this ability to learn on the fly, I'm not I'm a bit skeptical that we're going to see huge changes to the world without that ability. >> I think I I two things here, right? There's the state of the technology right now, um, which is again, we have these two stages. We have the pre-training and RL stage where you throw you throw a bunch of data and tasks into the models and
[40:28] then they generalize. So it's like learning, but it's like learning from more data and and not, you know, not learning over kind of one human or one model's lifetime. So again, this is situated between evolution and and and and human learning. But once you learn all those skills, you have them. And and just like with pre-training, just how the models know more, you know, if if I look at a pre-trained model, you know, it knows more about the history of samurai in Japan than I do. It knows more about baseball than I do. It knows,
[40:58] you know, it knows more about, you know, low-pass filters in electronics than you know, all all all of these things. It's knowledge is way broader than mine. So I think I think even even just that, um, you know, may get us to the point where the models are better at you know, kind of better at everything. And then we also have, again, just with scaling the kind of existing setup, we have the in-context learning, which I would describe as kind of like human on the job learning, but like a little
[41:28] weaker and a little short-term. Like you look at in-context learning, the you give the model a bunch of examples, it does get it. There's real learning that happens in context and like a million tokens is a lot. That's that's, you know, that can be days of human learning, right? You know, if you think about the model, you know, you know, kind of re- reading reading a million words, you know, it it you know, takes me how long would it take me to read a million? I mean, you know, like days or weeks at least. Um, so you have these two things and and I think these two these two things within
[41:58] the existing paradigm may just be enough to get you the country's genius in the data center. I don't know for sure, but I think they're going to get you a large fraction of it. There may be gaps, but I I certainly think, just as things are, this I believe is enough to generate trillions of dollars of revenue. That's one. That's all one. Two is this idea of continual learning, this idea of a single model learning on the job. Um I think we're working on that, too, and I think there's a good chance that in the
[42:28] next year or two we also make we also solve that. Um I I again, I I I I I you know, I think you get most of the way there without it. I think the trillions of dollars of of, you know, the the I think the trillions of dollars a year market, maybe all of the national security implications and the safety implications that I wrote about in adolescence of technology can happen without it, but I I I also think we, and I imagine others, are working on it. And I think there's a good chance that that,
[43:00] you know, that we get there within the next year or two. There are a bunch of ideas. I won't go into all of them in detail, but um you know, one is just make the context longer. There's There's nothing preventing longer context from working. You just have to train at longer context and then learn to to serve them at inference. And both of those are engineering problems that we are working on and that I would assume others are working on as well. Yeah, so this context length increase It seemed like there was a period from 2020 to 2023 where from GPT-3 to GPT-4 Turbo there was an increase from like 2,000 context lengths to 128K.
[43:31] Do you feel like for the next for the two-ish years since then we've been in the same-ish ballpark? Yeah. And when model context lengths get much longer than that people report qualitative degradation in the ability of the model to consider that full context. Um so I'm curious what you're internally seeing that makes you think like oh, 10 million context, 100 million context to get human-like 6 months learning, billion billion context. >> isn't a research problem. This is a This is an engineering and inference problem, right? If you want to serve long context, you have to like store your
[44:01] entire KV cache, you have to, you know, um uh you know, it's it's it's it's difficult to store all the memory in the GPUs, to juggle the memory around. I don't even know the detail, you know, at this point this is at a level of detail that that that that I'm no longer able to follow, although you know, I I knew it in the GPT-3 era of like, you know, these are the weights, these are the activations you have to store. Um, but you know, you know, these days the whole thing has flipped cuz we have MoE models and and and kind of all of that. But, um,
[44:31] uh, and and this degradation you're talking about, like again, without getting too specific, like a question I would ask is like there's two things. There's the context length you train at, and there's a context length that you serve at. If you train at a small context length and then try to serve at a long context length, like maybe you get these degradations. It's better than nothing, you might still offer it, but you get these degradations. And maybe it's harder to train at a long context length. Yeah, so, you know, there's there's a lot. I I I want to at the same time ask about like maybe some rabbit holes of like, well, wouldn't you expect
[45:02] that if you have to train on longer context length that would mean that um, you're able to get sort of like less samples in for the same amount of computer. But, before maybe maybe it's not worth diving deep on that. I I want to get an answer to the bigger picture question, which is like, okay, so um, I don't feel a preference for a human editor that's been working for me for 6 months versus an AI that's been working with me for 6 months. What year do you predict that that will be the case? I my I mean, you know,
[45:33] my guess for that is you know, there's there's there's a lot of problems that are basically like, we can do this when we have the country of geniuses in a data center. Um, and so, you know, my my my my my picture for that is, you know, again, if you if you if you if you know, if you made me guess, it's like 1 to 2 years, maybe 1 to 3 years. It's really hard to tell. I have a I have a strong view 99 95% that like all this will happen in 10 years. Like, that's I think that's just a super safe bet. And then, I have a hunch, this is more like a 50/50 thing, that it's going to be more
[46:04] like one to two, maybe more like one to three. So, one to three years. The country of genius says um and in the slightly less economically valuable task of editing videos. >> [laughter] >> I I I I I it seems pretty economically valuable, let me tell you. It's just there are a lot of use cases like that, right? There are a lot of similar ones. So, you're predicting that within one to three years. Um and in generally, Anthropic has predicted that by late '26, early '27, we will have AI systems that are quote um have the ability to navigate interfaces available to humans doing digital work today, intellectual
[46:34] capabilities matching or exceeding that of Nobel Prize winners, and the ability to interface with the physical world. And then you gave an interview two months ago with DealBook, where you were emphasizing your um your company's more responsible compute scaling as compared to your competitors. And I'm trying to square these two views, where if you really believe that we're going to have a country of geniuses, you 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 is actually can do
[47:04] everything a Nobel Prize winner can do is like trillions of dollars. And so, I'm trying to square this conservatism uh which seems rational if you have more moderate timelines, with your stated views about AI progress. >> Yeah, so so it actually all fits together. And and we go back to this fast, but not infinitely fast diffusion. So, like, let's say that we're making progress at this rate. Um you know, the the the technology is making progress this fast. Again, I have, you know, very high conviction that like it's going,
[47:34] you know, the the the you know, we're we're we're going to get there within within a few years. I have a hunch that we're going to get there within a year or two. So, a a little uncertainty on the technical side, but like, you know, pretty pretty strong confidence that it won't be off by much. What I'm less certain about is, again, the economic diffusion side. Like, I really do believe that we could have models that are a country of geniuses a hundred country of geniuses in a data center in one to two years. One question
[48:04] is, how many years after that do the trillions in you know do do the do the trillions in revenue start rolling in. Um I don't think it's guaranteed that it's going to be immediate. Um you know I think it could be um one year, it could be two years, I could even stretch it to five years, although I'm like I'm skeptical of that. And so we have this uncertainty which is even if the technology goes as fast as I
[48:35] suspect that it will, we we don't know exactly how fast it's going to drive revenue. We 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, you know, in Machines of Loving Grace, I said, "Look, I think we might get this powerful AI of this country of geniuses in the data center." That description you gave comes from the Machines of Loving Grace. I said, "We'll get that 2026, maybe 2027." Again, that is that is my hunch. Wouldn't be surprised if I'm off by a year or two,
[49:05] but like that is my hunch. Let's say that happens. That's the starting gun. How long does it take to cure all the diseases, right? That's that's one of the ways that like drives a huge amount of of of of economic value, right? Like you cure you cure every disease, you know, there's a question of how much of that goes to the pharmaceutical company, to the AI company, but there's an enormous consumer surplus because everyone you know every I I assume we can get access for everyone, which I care about greatly. We you know we we cure all of these diseases. How long does it take? You have to do the biological discovery, you have to you
[49:36] know go you have to you know manufacture the new drug, you have to you know go through the regulatory process. I mean we saw this with like vaccines and COVID, right? Like it there's just this we we got the vaccine out to everyone, but it took a a and a half, right? And and so 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 everyone. How long from when that AI first exists in the lab to when diseases have actually been cured for
[50:07] everyone, right? In in you know, we've had a polio vaccine for 50 years. We're still trying to eradicate it in the most remote corners of Africa. And you know, the Gates Foundation is trying as hard as they can. Others are trying as hard as they can, but you know, that's difficult. I again, I you know, I don't expect most of the economic diffusion to be as difficult as that, right? That's like the most difficult case. But but there's a there's a real dilemma here. And and where I've settled on it is it will be it will be a it will be faster than
[50:37] anything we've seen in the world, but it it still has its limits. And and so then when we go to buying data centers, you know, you again again, the curve I'm looking at is okay, we you know, we've had a 10x a year increase every year. So beginning of this year, we're looking at 10 billion in in in annual in you know, rate of annualized revenue at the beginning of the year. We have to decide how much compute to buy. Um and
[51:07] you know, it takes a year or two to actually build out the data centers, to reserve the data centers. So basically I'm saying like in 2027, how much compute do I get? Well, I could assume um uh that uh the revenue will continue growing 10x a year. So it'll be one 100 billion at the end of 2026 and 1 trillion at the end of 2027. And so I could buy a trillion dollars. Actually
[51:38] it would be like 5 trillion dollars of compute cuz it would be a trillion dollar a year for for 5 years, right? I could buy a trillion dollars of compute that starts at the end of 2027. And if my if my revenue is not a trillion dollars, if it's even 800 billion, there's no force on Earth. There's there's no hedge on Earth that could stop me from going bankrupt if I if I buy that much compute. And and so, even though a part of my brain wonders if it's going to keep growing 10x, I can't buy a trillion dollars a year of
[52:09] compute in in in in in in in in in in in 2027 if I'm just off by a year in that rate of growth or if the the growth rate is 5x a year instead of 10x a year, then then, you know, that you go bankrupt. Um and and and and so, you end up in a world where, you know, you're supporting hundreds of billions, not trillions, and you accept you accept some risk that there's so much demand that you can't support the revenue, and you accept
[52:39] still some risk that, you know, you got it wrong and it's still slow. And so, when I talked about behaving responsibly, what I meant actually was not the absolute amount. That that actually was not um you know, I think it is true we're spending somewhat less than some of the other players. It's actually the other things like have we been thoughtful about it or are we yoloing and saying, "Oh, we're going to do a hundred billion dollars here or a hundred billion dollars there?" I kind of get the impression that, you know, some of the other companies have not
[53:09] written down the spreadsheet, that they don't really understand the risks they're taking. They're just kind of doing stuff cuz it sounds cool. Um uh and and we thought carefully about it, right? We're an enterprise business. Therefore, you know, we can rely more on revenue. It's less fickle than consumer. We have better margins, which is the buffer between buying too much and buying too little. And so, I think we bought an amount that allows us to capture pretty strong upside worlds. It won't capture the full 10x a year. Um and
[53:40] things would have to go pretty badly for us to be for us to be in financial trouble. So, I think we thought carefully and we've made that balance and and that's what I mean when I say that we're being responsible. Okay, so it seems like um it's possible that we're we actually just have different definitions of a country of geniuses in a data center. Because when I think of like actual human geniuses and actual country of human geniuses in a data center, I'm like I would happily buy 5 trillion dollars worth of compute to run uh actual country of human geniuses in a data center. So, let's say JP Morgan or
[54:10] Moderna or whatever doesn't want to use them. Also, I've got a country of geniuses. Like they'll they'll start their own company. And if like they they can't start their own company and they're bottlenecked by clinical trials, it is worth stating with clinical trials like most clinical trials fail because the drug doesn't work. There's no efficacy, right? >> exactly that point in in Machines of Love and Grace. I say the clinical trials are going to go much faster than we're used to. But not not instant not infinitely fast. And then suppose it takes a year to uh for the clinical trials to work out so that you're getting revenue from that and can make more drugs. Okay, well, you've got a country of geniuses and you're in the AI
[54:41] lab and you have you could use uh many more AI researchers. Um and you also think that there's these like self-reinforcing gains from you know, smart people working on AI tech. So, like okay, you can have the >> That's right. You can have the data center working on like AI progress. >> more gains from buying like substantially more gains from buying a trillion dollars a year of compute versus 300 billion dollars a year of compute? >> buying a trillion, yes, there is. >> Well, no, there's some gain, but then
[55:11] but again, there's this chance that they go bankrupt before you know, before again, if you're off by only a year, you destroy yourselves. That's the that's the balance. We're buying a lot. We're buying a hell of a lot. Like we're not we're we're you know, we're buying an amount that's comparable to that that you know, the the the the the the biggest players in the game are buying. Um but but if you're asking me why do why haven't we signed, you know, 10 10 trillion of compute starting in starting in mid
[55:42] 2027, first of all, it can't be produced. There isn't that much in the world. Um but but second, um what if the country of geniuses comes, but it comes in mid 2028 instead of mid 2027? You go bankrupt. So, if your projection is 1 to 3 years, it seems like you should have 1 10 trillion dollars of compute by um 2029. 2020 and maybe 2025. I mean, like >> I mean, you know, you you you you you >> But like are you interested in like it seems like even in your the longest version of the timelines you stated, the
[56:12] compute you are ramping up to build doesn't seem What what what what what makes you think that? Well, you you as you said, you would want the 10 trillion like human wages, let's say, are um on the order of 50 trillion a year. >> If if you look at so so I won't I won't talk about Anthropic in particular, but if you talk about the industry, like um the amount of compute the industry had you know, the the the the amount of compute the industry's building this year is probably in the you know, I don't know, very low tens of you may you
[56:42] know, call it 10 15 gigawatts. Next year, I you know, it it goes up by roughly 3x a year, so like next year's 30 or 40 gigawatts, and um 2028 might be 100, 2029 might be like 3 300 gigawatts, and like each gigawatt costs like um maybe 10 I mean, I'm doing the math in my head, but each gigawatt it costs maybe 10 billion dollars, you know, of order 10 to 15 billion dollars a year. So, you know, you you kind of you you
[57:12] you know, you put that all together, and you're getting about about what you described. You're getting multiple trillions a year by 2028 or 2029. So, you're you're you're getting exactly that. You're getting you're getting exactly what you predict. Um that's for the industry. >> That that's for the industry. That's right. >> So, suppose Anthropic's compute keeps 3x a year, and then by like 27 you have uh or 27 28 you have 10 gigawatts. And like multiply that by as you say 10 billion. So then it's like 100 billion a year. But then you're saying the TAM by 2028, 2029.
[57:43] >> I don't want to give exact numbers for Anthropic, but but these numbers are too small. These numbers are too small. Okay, interesting. I'm really proud that the puzzles I've worked on with Jane Street have resulted in them hiring a bunch of people for my audience. Well, they're still hiring and they just sent me another puzzle. For this one, they spent about 20,000 GPU hours training backdoors into three different language models. Each one has a hidden prompt that elicits completely different behavior. You just have to find the trigger. This is particularly cool because finding backdoors is actually an open question in frontier AI research. Anthropic actually released a
[58:15] couple of papers about sleeper agents and they showed that you can build a simple classifier on the residual stream to detect when a backdoor is about to fire. [music] But they already knew what the triggers were because they built them. Here, you don't and it's not feasible to check the activations for all possible trigger [music] phrases. Unlike the other puzzles they've made for this podcast, Jane Street isn't even sure this one is solvable. But they've set aside $50,000 for the best [music] attempts and write-ups. The puzzle's live at janestreet.com/twarcash and they're accepting submissions until
[58:45] April 1st. All right, back to Dario. You've told investors that you plan to be profitable starting in 2028. And this is the year where we're like potentially getting the country of geniuses a data center. And we either this is like going to now unlock all this progress and medicine and health and etc. etc. and new technologies. It wouldn't this be a particular the exactly the time where you'd like want to reinvest in the business and build bigger countries so they can make more discoveries?
[59:15] >> I mean profit profitability is this kind of like weird thing in this field. I like like I don't think I I don't think in this field profitability is actually a measure of uh you know kind of spending down versus investing in the business like let's let's just let's just take a model of this. I actually think profitability happens when you underestimated the amount of demand you were going to get and loss happens when you overestimated the amount of demand you were going to
[59:45] get because you're buying the data centers ahead of time. So think about it this way. Ideally you would like and again these are stylized facts. These numbers are not exact for anything. I'm just trying to make a toy model here. Let's say half of your compute is for training and half of your compute is for inference and you know the inference has some gross margin that's like more than 50% and so what that means is that if you were in steady state you build a data center if you knew exactly exactly exactly the demand you were getting you
[60:15] would you know you would you would you you would you would get a certain amount of revenue say I don't know let's say you pay a hundred billion dollars a year for compute and on 50 billion dollars a year you support 150 billion dollars on of of of of of of of revenue and the other 50 billion the other 50 billion are used for training. So basically you're profitable you make you make you make 50 billion dollars of profit. Those are the economics of the industry today or or sorry not today but like that's where we're where we're
[60:45] projecting forward in a year or two. The only thing that makes that not the case is if you get less demand than 50 billion then you have more than 50% of your your data center for research and you're not profitable. So you you know you train stronger models but you're like not profitable. If you get more demand than you thought then your research gets squeezed but you know you're you're you're kind of able to support more inference and you're more profitable. So it's
[61:15] maybe I'm not explaining it well but but the thing I'm trying to say is you decide the amount of compute first and then you have some target desire of of inference versus versus training, but that gets determined by demand. It doesn't get determined by you. >> what I'm hearing is the reason you're predicting profit is that you are systematically underestimat- under investing in compute, right? Because if you actually like >> no. I'm saying I'm saying it's hard to predict. So so these things about 2028 and when it will happen, that's our that's our attempt to do the best we can
[61:45] with investors. All of this stuff is really uncertain because of the cone of uncertainty. Like we could be profitable in 2026 if the if the revenue grows fast enough. And then and then um you know, if we if we overestimate or underestimate the next year, that could swing wildly. Like I I I What I'm trying to get at is you have a model in your head of like the the business invests, invests, invests, invests, gets scale, and and and and kind of then becomes profitable. There's a single point at which things turn around. I don't think
[62:16] the economics of this industry work that way. I see. So if I'm understanding correctly, you're saying because of the discrepancy between the amount of compute we should have gotten and the amount of compute we got, we we were like sort of forced to make profit, but that that doesn't mean we're going to continue making profit. We're going to like reinvest the money because well, now AI's made so much progress and we want the bigger country of geniuses. And so then back into uh revenue's high, but losses are also high. >> If we If we If we predict If every year we predict exactly what the demand is
[62:47] going to be, we'll be profitable every year. Because grow because spending spending 50% of your compute on on um 50% of your compute on research, roughly, um plus a gross margin that's higher than 50% and and correct demand prediction leads to profit. That's the profit That's That's the profitable business model that I think is kind of like there, but like obs- obscured by these like building ahead and prediction errors. >> I I guess you're treating the 50% as a uh as a sort of like
[63:18] you know, just like a given constant. Whereas you In fact, if you if the AI progress is fast and you can increase the progress by scaling up more, you just have more than 50% and not make progress. >> what I'll say. You might want to scale up it more. You might want to scale it up more. But but but, you know, remember the law returns to scale, right? If if 70% would get you a very little bit of a smaller model through a factor of of 1.4x, right? Like that extra 20 billion dollars is is is is, you know, that each each dollar there is worth much less to you because
[63:48] of because of because the law linear setup. And so, you might find that it's better to invest that that that that it's better to invest that 20 billion dollars in, you know, in in serving inference or in hiring engineers who are who are who are kind of better who are who are kind of better who are kind of better at what they're doing. So, the the reason I said 50% That's not that's not exactly our target. It's not exactly going to be 50%. It'll probably vary vary over time. What what I'm saying is the the the the the like law linear return what it leads to is you spend of
[64:19] order one fraction of the business, right? Like not 5% not 95% and then it then it then you know, then then then you get diminishing returns because of the because the law law law law law law >> like convincing Dario to like believe in AI progress or something. But like you Okay, you you don't invest in research because it has diminishing returns, but you invest in the other things you mentioned. Again, again, we're talking about diminishing returns. After you're spending 50 billion a year. Right. Like This is a point I'm I'm I'm sure you'd make, but like diminishing returns on a genius is could
[64:51] be quite high. And more generally, like what is profit in a market economy? Profit is basically saying the other companies in the market can like do more things with this money that I can't. >> I then thought I'm just trying to like cuz I I you know, I don't want to give information about Anthropic is why I'm giving these stylized numbers. But like let's just derive the equilibrium of the industry, right? I think the So so so why doesn't everyone spend 100% of their you know, 100% of their compute on
[65:21] training and not serve any customers, right? It's because if they didn't get any revenue, they couldn't raise money, they couldn't do compute deals, they couldn't buy more compute the next year. So there's going to be an equilibrium where every every company spends less than 100% on on on on training and certainly less than 100% on inference. It should be clear why you don't just serve the current models and and you know, and and and and and and and and never train another model because then you don't have any demand because you'll because you'll fall behind. So there's some equilibrium. It's it's not going to be 10%, it's not going to be 90%. Let's
[65:53] just say as a stylized fact it's 50%. That's what I'm getting at. And and and I think we're going to be in a position where that equilibrium of how much you spend on training is less than the gross margins that that you're that that that that you're able to get on compute. And so the the the the underlying economics are profitable. The problem is you have this this hellish demand prediction problem when you're when you're buying the next year of compute and you might guess under and be very profitable but have no compute for research or you might guess over and
[66:26] you know, you're you're you're you you are not profitable and you have all the compute compute for research in the world. Does it does it does that make sense? It's just a dynamic model of the industry. >> Maybe stepping back I'm like I I I'm not saying I I think the country of geniuses going to come in two years and therefore you should buy this compute. To me what you're saying the end conclusion you're arriving at makes a lot of sense, but that's because like oh, it seems like country of geniuses is hard and there's
[66:56] a long way to go. And so the stepping back the thing I'm trying to get at is more like it seems like your worldview is compatible with somebody who says Uh, like 10 years away from a world in which like we're generating trillions of dollars worth. >> that's just not my view. That is that is not my view. Like I I so so so I'll like I'll like make another prediction. It is hard for me to see that that there won't be trillions of dollars in revenue before 2030. Um, like uh, I can I can construct a plausible world. It takes
[67:26] maybe 3 years. So that would that would you know, that would be the end of what I think it's plausible like in 2028 we get the the real country of geniuses in the data center. You know, the revenue's been been go you know, the revenue's been going into the maybe is a is in the low hundreds of billions by by by by 2028. And and and then the country of geniuses accelerates it to trillions, you know, and and we're basically we're basically on the slow end of diffusion. It takes 2 years to get to the trillions. That that that would that that that would be the world where it
[67:56] takes until That would be the world where it takes until 2030. I I I suspect even composing the technical exponential and diffusion exponential will get there before 2030. So you laid out a model where Anthropic makes profit because it seems like fundamentally we're in a compute-constrained world. And so it's like eventually we keep growing compute. >> No, I think I think the way the profit comes is again and and you know, let's let's just abstract the whole industry here. Like we have a you know, let's just imagine we're we're we're in like an economics textbook. We have a small
[68:27] number of firms. Each can invest a limited amount in you know, or or or like each can invest some fraction fraction in R&D. They have some marginal cost to serve. The margins on that the profit margin the gross profit margins on that marginal cost are like very high because because because inference is efficient. There's some competition, but the models are also differentiated. There's some there's some um, you know, companies will compete to push their research budgets up, but like because there's a small number of players, you
[68:58] know, we have the I what is it called the the Cournot equilibrium? I think is what the what the uh small number of firm equi- equilibrium is. It The point is it doesn't equilibrate to perfect competition with with with with with with with zero margins. If there's like three firms, if there's three firms in the economy, all are kind of independently behaving behaving rationally, it doesn't equilibrate to zero. Um Help me understand that cuz right now we do have three leading firms and they're not making profit. Um and so what what
[69:29] what uh Yeah, what what what is changing? Yeah, so the the again, the gross margins right now are very positive. What's happen- What what what's happening is a combination of two things. One is we're still in the exponential scale up phase of compute. >> Yeah. Um So what basically what that means is we're training like a model gets trained. It costs, you know, let's say a model got trained that costs uh a billion dollars last year. Um and then uh this year it produced uh $4 billion
[70:02] of revenue and cost $1 billion to to uh to to to inference from. Um so you know, again, I'm using stylized number here, but you know, that would be 75% you know, gross gross gross margins and you know, this this 25% tax. So that model as a whole makes $2 billion. Um but at the same time, we're spending $10 billion to train the next model because there's an exponential scale up. And so the company loses money. Each model makes money, but the company loses
[70:32] money. The equilibrium I'm talking about is an equilibrium where we have the country of geniuses, we have the country of geniuses in the data center, but that that um model training scale up has equilibrated more. Maybe maybe it's still it's still going up. We're still trying to predict the demand, but it's more it's more um leveled out. Um I'm giving you a couple of things there. So um let's start with the current world. Um, in the current world, you're right that as you said before, if you treat each
[71:02] individual model as a company, it's profitable. Yeah, of course. A big part of the production function of being a frontier lab is training the next model, right? So, if you if you didn't do that, then you'd make profit for 2 months, and then you wouldn't have margins cuz you wouldn't have the best model, and then so, yeah, you you can make profit for 2 months on the current system. >> point that reaches the biggest scale that it can reach. And then and then in equilibrium, we have algorithmic but we're spending roughly the same amount to train the next model as as as we as we spent to train the current model. Um,
[71:33] so, this equilibrium relies I mean, at some point at some at at some point you run out of money in the economy. Uh, fixed lump of labor file the economy is going to grow, right? That's one of your predictions. Well, yes. But this is this is space. >> But this is another example of the theme I was talking about, which is that the economy will grow much faster with AI than I think it ever has before, but it's not like right now the computer is growing 3x a year. Yeah. I don't believe the economy is going to grow 300% a year. Like I said this in Machines of
[72:04] Loving Grace. Like I think we we may get 10 or 20% for a year growth in the economy, but we're not going to get 300% growth in the economy. So, I think I think in the end, you know, if compute becomes the majority of what the economy produces, it's it's going to it's going to be capped by that. >> So, okay, now let's assume a model where compute stays capped. Yeah. The world where frontier labs are making money is one where they continue to make um, fast progress because fundamentally your margin is limited by how good the
[72:35] alternative is. And so, you are able to make money cuz you have a frontier model. Um, if you didn't have a frontier model, you wouldn't be making money. Well, you you I mean >> And and so, this this this this model requires there never to be a steady state. Like forever and ever, you keep making more algorithmic progress. >> think that's true. I mean, I feel I feel I like we're we're like we're talk we're we're you know we're I feel like this is an economics like you know this is this is like an economics class >> Tyler Cowen code? We never stop talking about economics. >> we never stop talking about economics. So, no, but but there there are there
[73:05] are worlds in which um you know there So, I I don't think this field's going to be a I don't think this field's going to be a monopoly. All my lawyers never want me to say the word monopoly. Um but I don't think this field's going to be a monopoly, but but you do get you get industries in which there are small number of players, not one but a small number of players. And ordinarily like the way you get monopolies like Facebook or or Meta or I always call them Facebook, but um is is these kind of net is these kind of
[73:35] these kind of network effects. Yeah. The way you get industries in which there are small number of players are very high costs of entry, right? Um so, you know, uh cloud is like this. I think cloud is a good example of this. You have three, maybe four players within cloud. I think I think that's the same for AI, three maybe four. Um and the reason is that it's it's so expensive, it requires so much expertise and so much capital to like run a cloud
[74:05] company, right? And so you have to put up all this capital and then in addition to putting up all this capital, you have to get all of this other stuff that like, you know, requires a lot of skill to you know to make it happen. And so it's like if you go to someone and you're like, I want to disrupt this industry, here's $100 billion or like, "Okay, I'm putting $100 billion and also betting that you can do all these other things that these people have been doing and so like >> Only you decrease the profit in the industry. >> And and and then and then the effect of your entering is is the profit margins go down. So, you know, we have equilibria like this all the time in the economy where we have a few we have a
[74:35] few players, profits are not astronomical, margins are not astronomical, but they're they're not zero, right? Um and and you know, I think I I that's what we see on cloud. Cloud is very undifferentiated. Models are more differentiated than cloud, right? Like everyone knows Claude is Claude Claude is good at different things than GPT is good at is than than Gemini is good at. And it's not just Claude's good at coding, GPT is good at, you know, math and reasoning, you know, um uh it it's
[75:05] more subtle than that. Like models are good at different types of coding. Models have different styles. Like I think I think these things are actually, you know, quite different from each other. And so expect more differentiation than you see in in um cloud. Now, there there actually is a uh counter there there there is one counterargument. Um and that counterargument is that if all of that, the process of producing models, um becomes uh if AI models can do that themselves, then that could spread
[75:36] throughout the economy. But that is not an argument for commoditizing AI models in general. That's kind of an argument for commoditizing the whole economy at once. Um I don't know what what quite happens in that world where basically anyone can do anything, anyone can build anything, and there's like no moat around anything at all. I mean, I don't know, maybe we want that world. Like like maybe that's the maybe that's the end state here. Like maybe maybe um you know, when what maybe when when when kind of AI models can do you know, when what when when AI models can do everything, if we've solved all the
[76:07] safety and security problems, like, you know, that's one of the one of the one of the mechanisms for for uh you know, um uh uh you know, just just kind of the economy flattening itself again. But but that's kind of like post like far post country of geniuses in a data center. Um I maybe uh a finer way to put that uh potential point is one it seems like AI research is especially loaded on raw intellectual power, which will be especially abundant in a world of AGI. And two, if you just look at the
[76:39] world today, there's very few technologies that seem to be diffusing as fast as um as AI algorithmic progress. And so, that does hint that this industry is sort of structurally diffusive. So, I think coding is going fast, but I think AI research is a super set of coding and there are aspects of it that are not going fast. Um uh but I but I do think again, once we get coding, once we get AI models going fast, then, you know, AI you know, that will speed up the ability of AI models to kind to kind of do
[77:09] everything else. So, I think while coding is going fast now, I think once the AI models are building the next AI models and building everything else, the kind of whole the whole economy will start to kind of go at the same pace. I am I am worried geographically, though. I'm a little worried that like just proximity to AI, having heard about AI, um uh that that that may be one differentiator. And so, when I said the like, you know, 10 or 20% growth rate, a worry I have is that the growth rate
[77:40] could be like 50% in Silicon Valley and, you know, parts of the world that are kind of socially connected to Silicon Valley and, you know, not that much faster than its current pace elsewhere. And I think that'd be a pretty messed up world. So, I one of the things I think about a lot is how to prevent that. Yeah. Do you think that once we have uh the center of genius in a data center, that robotics is sort of quickly solved afterwards because it seems like a big problem with robotics is that um a human can learn how to teleoperate current hardware, but current AI models
[78:12] can't, at least not not in a way that's super productive. And so, if we have this ability to learn like a human, should it solve robotics immediately as well? >> I don't think it's dependent on learning like a human. It could happen in different ways. Again, we could have trained the model on many different video games, which are like robotic controls or many different simulated robotics environments or just, you know, train them to control computer screens and they learn to generalize. So, it will happen. It's not necessarily dependent on human-like learning. Human-like learning
[78:42] is one way it could happen if the model's like, "Oh, I pick up a robot. I don't know how to use it. I learn." That that could happen because we discovered discovering continual learning. That could also happen because we train the model on a bunch of environments and then it generalized or it could happen because the model learns that in the context length. It It doesn't actually matter which way. If we go back to the discussion we had like like an hour ago, that type of thing can happen in that type of thing can happen in several different ways. Um but but I do think when for for whatever
[79:12] 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 also the the ability to kind of control robots. So, we'll get better at the physical building the physical hardware, building the physical robots and we'll also get better at controlling it. Now, you know, does that mean the robotics industry will also be generating trillions of dollars of revenue? My answer there is yes, but there will be the same extremely fast
[79:42] but not infinitely fast diffusion. So, will robotics be be revolutionized? Yeah, maybe tack on another year or two. Mhm. That's that's my That's the way I think about these things. >> Uh there's a general skepticism about extremely fast progress. Like here here's my view which is like it sounds like you are going to solve continual learning one way or another within the matter of years, but just as people weren't talking about continual learning a couple years ago and then we realized, "Oh, why aren't these models as useful as they could be right now even though they are clearly passing the Turing test
[80:12] and are experts in so many different domains? Maybe it's this thing." And then we solve this thing and we realize, "Actually, there's another another thing that human intelligence can do and that's a basis of human labor that these models can't do." And then so why not think there will be more things like this? Why I think that like we're we're, you know, we've like found the pieces of human intelligence. >> Well, well, to be clear, I mean, I think continual learning as I've said before might not be a barrier at all, right? Like like, you know, I think I think we maybe just get there by pre-training generalization and and and and and and
[80:42] and and RL generalization. Like I I think there just might not be um there there there basically might not be such a thing at all. In fact, I would point to the history in in ML of people coming up with things that are barriers that end up kind of dissolving within the big blob of compute, right? That, you know, people talk talked about, you know, you know, how do you have you know, how do how do your models keep track of nouns and verbs? And, you know, how do they, you know, they can understand semantic syntactically, but they can't understand semantically, you
[81:14] know, it's only statistical correlations. You can understand a paragraph, you can understand a word. There's reasoning, you can't do reasoning, but then suddenly it turns out you can do code and math very well at all. So, I I think there actually there's there's actually a stronger history of some of these things seeming like a big deal and then and then kind of and then kind of dissolving. Some of them are real. I mean, the need for data is real. May maybe continual continual learn continual learning is a real thing, but again, I would ground us
[81:44] in something like code. Like, I think we may get to the point in like a year or two where the models can just do sweet end end. Like, that's a whole task. That's a whole sphere of human activity that that we're just saying models can do it now. Um When you say end to end, do you mean um setting technical direction, understanding the context of the problem, etc.? Okay. >> Yes, I mean all of that. >> Interesting. I mean, it that that is I feel like EJ complete. Um maybe it's maybe it's internally consistent, but I
[82:15] um it's not like saying 90% of code or 100% of code. It's like >> No, no. I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I Eventually those get done as well. But that's a long spectrum there. But we're traversing the spectrum very quickly. Yeah. Um I do think it's funny that I I I've seen a couple of podcasts you've done where um the host will be like, "Oh, but we're catching up with the essay about the control learning thing." And it always makes me crack up cuz you're like, you know, you've you've
[82:46] been an AI researcher for like 10 years and uh I'm sure there's like some uh feeling of like, "Okay, so podcasts are writing an essay." And [laughter] like every interview I get asked about it. You know, the the truth of the the truth of the matter is that we're all trying to figure this out together. >> Yeah. Right? There there are some ways in which I'm able to see things that others aren't. These days that probably has more to do with like I can see a bunch of stuff within Anthropic and have to make a bunch of decisions than I have any great research insight that that that others don't, right? I you know,
[83:17] I'm running a 2,500 person company. Like it's it's actually pretty hard for me to concrete research insight, you know, my uh much harder than you know, than it than it would have been, you know, 10 years ago or or you know, or even two or three years ago. Um As we go towards a world of a full drop-in remote worker replacement, does a API pricing model still make the most sense? And if not, what is the correct way to price AGI or serve AGI? >> Yeah, I mean I think there's going to be a bunch of different business models
[83:48] here sort of all at once that are going to be that are going to be experimented with. Um I I I actually do think that the the API um model is is more durable than many people think. Um one way I think about it is if the technology is kind of advancing quickly, if it's advancing exponentially, what that means is there's there's always kind of like a surface area of of kind of new use cases that have been developed in in the last uh in the last
[84:18] 3 months. And any kind of product surface you put in place is always at risk of sort of becoming irrelevant, right? Any given product surface probably makes sense for our, you know, a range of capabilities of the model, right? The the chatbot is already running into limitations of, you know, making it smarter doesn't really help the average consumer that much. But I don't think that's a limitation of AI models. I don't think that's evidence that, you know, the models are are the models are good enough and they're
[84:48] they're, you know, them getting better doesn't matter to the economy. It doesn't matter to that particular product. Um and and so I think the value of the API is the API always offers an opportunity, you know, very close to the bare metal to build on what the latest thing is. Um and so there, you know, there's there's there's kind of always going to be this, you know, this this kind of front of new startups and new ideas that weren't possible a few months ago and are possible because the model is advancing.
[85:18] And and so I I actually I I I kind of actually predict that we are it's going to exist alongside other models, but we're always going to have the API business model because there's there's always going to be a need for a thousand different people to try experimenting with the model in different way and a hundred of them become startups and ten of them become big successful startups and, you know, two or three really end up being the the way that people use the model of a of a given generation. So I basically think
[85:49] it's always going to exist. At the same time I'm sure there's going to be other models as well. Like not every token >> [snorts] >> that's output by the model is worth the same amount. Think about you know, how how how what is the value of the tokens that are like, you know, that the model outputs when someone, you know call, you know, someone, you know, calls them up and says my Mac isn't working or something, you know, the model's like restart it, right? Um and like, you know, someone hasn't heard that before, but like, you know, the model said that
[86:20] 10 million times, right? Um, you know, that's that maybe that's worth like a dollar or a few cents or something. Um, whereas if uh the model, you know, the model goes to, you know, one of the one of the pharmaceutical companies and it says, "Oh, you know, this molecule you're developing, you should take the aromatic ring from that end of the molecule and put it on that end of the molecule. Um, and and you know, if you do that, wonderful things will happen." Um, uh uh like like those tokens could could be [laughter] worth, you know, tens of millions of dollars, right? Um,
[86:51] uh so so I think we're definitely going to see business models that that recognize that, you know, at some point we're going to see, you know, pay for results or you you know, in some in some form or we may see forms of compensation that are like labor. Um, uh you know, that that kind of work by the hour. Um, I I I you know, I don't know. I think I think I think because it's a new industry, a lot of things are going to be tried and I, you know, I don't know what will turn out to be the right thing. Um, what I find uh
[87:23] I I take your point that people will have to try things to figure out what is the best way to use this blob of intelligence, but what I find striking is Claude Code. So, I don't think in the history of startups there has been a single application that has been as hotly competed in as coding agents. And um and and Claude Code is a category leader here. And that seems surprising to me. Like, it doesn't seem intrinsically like Anthropic had to build this. And I wonder if you have an accounting of why it had to be Anthropic
[87:55] or why how Anthropic ended up building an application in addition to the model underlying it. >> Yeah, so it actually happened in a pretty simple way, which is we had our own um you know, we had our coding models, which were good at coding. And and you know, around the beginning of 2025, I said, "I I think the time has come where you can have non-trivial acceleration of your own research. Um, if you're an AI company by using these models. And of course, you know, we you need an interface. You need a harness to use them. And so I encouraged
[88:26] people internally, you know, I didn't say this is one thing that, you know, that you have to use. I just said people should experiment with this. And then, you know, this thing I you know, I think it might have been originally called Claude CLI and then then and then eventually got changed to Claude Code internally, um, was the thing that kind of everyone was using it and was seeing fast internal adoption. And I looked at it and I said, "Well, probably we should launch this externally, right? Um, you know, it's it's seen such fast adoption within Anthropic like, you know, like, you
[88:57] know, coding is a lot of what we do and and so, you know, we have a we have a audience of many many hundreds of people that's in some ways at least representative of the external audience. So, it looks like we already have product market fit. Let's launch this thing." Um, and and then we launched it and and and I think, you know, just just the fact that we ourselves are kind of developing the model and we ourselves know what we most need to use the model. I think it's it's kind of creating this feedback loop. I see. In the sense that you let's say a developer at Anthropic
[89:27] is like, "Ah, it it'd be better if it was better at this X thing." And then you bake that into the next model that you build. >> That that's that's one version of it. But but then there's just the ordinary product iteration of like, you know, we have a bunch of we have a bunch of coders within Anthropic like, we um, you know, they they like use Claude Code every day and so we get fast feedback. That was more important in the early days. Now, of course, there are millions of people using it. Um, and so we get a bunch of external feedback as well, but it's you know, it's just great to be able to
[89:57] get, you know, kind of kind of uh um, fast fast internal feedback. You know, I think this is the reason why we launched a coding model and you know, didn't launch a pharmaceutical company, right? It you know, you know, my background's in in my background's in in like biology, but like we don't have any of the resources that are needed to launch a pharmaceutical company. So, there's been a ton of hype around Open Claw, and I want to check it out for myself. I've got a day coming up this weekend, and I don't have anything planned yet. So, I gave Open Claw a Mercury debit card. I set a couple
[90:27] hundred dollar limit, and I said, "Surprise me." Okay, so here's the Mercury money it's on, and besides having access to my Mercury, it's totally quarantined. Naturally, I feel quite comfortable giving it access to a debit card because Mercury makes it super easy to set up guardrails. I was able to customize permissions, cap the spend, and restrict the category of purchases. I wanted to make sure the debit card worked, so I asked Open Claw to just make a test transaction, and decided to donate a couple bucks to Wikipedia. Besides that, I have no idea what's going to happen. I will report back on the next episode about how it goes. In the meantime, if you want a personal banking solution that can
[90:57] accommodate all the different ways that people use their money, even experimental ones like this one, visit mercury.com/personal. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column NA, members FDIC. You know, she thinks we're getting coffee and walking around the neighborhood. >> [laughter] >> Um let me ask you about and now um making AI go well. Um it seems like whatever vision we have about how AI goes well has to be
[91:28] compatible with two things. One is the ability to build and run AIs is diffusing extremely rapidly. And two is that the population of AIs, the amount we have in their intelligence, will also increase very rapidly. And that means that lots of people will be able to build huge populations of misaligned AIs or uh AIs which are just like companies which are trying to increase their uh footprint or have weird psychics like Sydney Bing, but now they're superhuman. What is a vision for a world in which
[92:00] we have an equilibrium that is compatible with lots of different AIs, some of which are misaligned, running around. >> Yeah, yeah. So, I think, you know, in the adolescence of technology, I was kind of you know, skeptical of like the balance of power. But, I would I think I was particularly skeptical of or the thing I was specifically skeptical of is you have like three or four of these companies like kind of all building models that are kind of dry, you know, sort of sort of um uh uh like derived from the
[92:30] like derived from the same thing and uh you know, that that these would check each other. Or or even that kind of, you know, any number of them would would would uh would would check each other. Like, we might live in a offense-dominant world where, you know, like one person or one AI model is like smart enough to do something that like causes damage for everything else. Um I think in the I mean, in the short run, we have a limited number of players now. So, we can start by within the limited number of players, we uh you know, we
[93:00] kind of, you know, we we need to put in place the, you know, the safeguards. We need to make sure everyone does the right alignment work. We need to make sure everyone has bio classifiers. Like, you know, those are those are kind of the immediate things we need to do. I agree that, you know, 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, you know, the the whole thing can kind of um you know, it can become harder to solve. I you know, I think I think in the long run, we need some architecture of
[93:30] governance, right? Some architecture of governance that preserves human freedom, but but kind of also allows us to like, you know, govern the the very large number of kind of um you know, uh uh human systems, AI systems, hybrid hybrid human human um you know, hybrid hybrid human AI like, you know, companies or or like or like or like economic units. So, you know, we're going to we're going to need to think about like, you know, how do we
[94:01] how do we protect the world against, you know, bioterrorism? How do we protect the world against like, you know, against like against like mirror life? Like, you know, probably probably we're going to need to, you know, need some kind of like AI monitoring system that like monitor, you know, kind of monitors for for all of these things, but then we need to build this in a way that like, you know, preserves civil liberties and like our constitutional rights. So, I think just just as is as is anything else, like it's it's like a new security landscape with a new set of, you know,
[94:34] a new set of tools and a new set of vulnerabilities. And I think my worry is if we had 100 years for this to happen all very slowly, we'd get used to it. You know, like we've gotten used to like, you know, the presence of, you know, the presence of explosives in society or like the, you know, the presence of various, um, you know, like new weapons or the, you know, the the presence of video cameras. Um, we would get used to it over over over over 100 years and we'd develop governance mechanisms. We'd make our mistakes. My
[95:04] My worry is just that this happening all so fast. And so, I think maybe we need to do our thinking faster about how to make these governance mechanisms work. >> Yeah. It seems like in an offense-dominant world, over the course of the next century, so the idea is that AI is making the progress that would happen over the next century happen in some period of 5 to 10 years. But we would still need the same mechanisms or balance of power would be similarly intractable even if humans were the only game in town. Um, and so, I guess we have the advice of
[95:35] AI. We If fundamentally it doesn't seem like a totally different ball game here. If checks and balances were going to work, they would work with humans as well. If they aren't going to work, they won't work with AIs as well. Um, and so, maybe this is just doomed as human checks and balances as well, but >> Yeah, again again I think there's some way to I think there's some way to make this happen. Like it you know, it it it it it it just it just you know, the governments of the world may have to work together to make it happen. Like, you know, we may have to you may have to talk to AIs about kind of you know,
[96:06] building societal structures in such a way that like these these defenses are possible. I I I don't know. I mean this is so this is you know, I I don't want to say so far ahead in time, but like so far ahead in technological ability that may happen over a short period of time that it's hard for us to anticipate it in advance. Um speaking of governments getting involved, on December 26th, the Tennessee legislature introduced a bill which uh said quote, um it would be an offense for a person to knowingly train artificial intelligence to provide emotional support including through
[96:36] open-ended conversations with a user. And of course, one of the things that Claude attempts to do is be uh a thoughtful um uh thoughtful friend, thoughtful knowledgeable friend. And in general, it seems like we're going to have this patchwork of state laws. A lot of the benefits that normal people could experience as a result of AI are going to be curtailed, especially when we get into the kinds of things you discuss in Machines of Loving Grace, biological freedom, mental health improvements, etc. etc. It seems easy to imagine worlds in which these get whac-a-moled away by different laws. Um whereas
[97:09] bills like this don't seem to address the actual existential threats that you're concerned about. So, I'm curious about to understand in the context of things like this, your Entropic's position against the federal moratorium on state AI laws. Yes. So, I don't know. There's there's many different things going on at at at once, right? I think I think that that I think that particular law is is dumb. Like, you know, I think it was it was clearly made by legislators who just probably had little idea what AI models could do and not do. They're like, AI models serving as that
[97:39] that just sounds scary. Like, I don't want I don't want that to happen. So, you know, we're we're we're not we're not in favor of that, right? But but but that you know, that that wasn't the thing that was being voted on. The thing that was being voted on is we're going to ban all state regulation of AI for 10 years with no apparent plan to to do any federal regulation of AI, which would take Congress to pass, which is a very high bar. Um so, you know, the idea that we'd ban states from doing anything for 10 years
[98:09] and people said they had a plan for federal government, but you know, there was no actual there was no proposal on the table. There was no actual attempt. Um given the serious dangers that I lay out in adolescence of technology around things like the you know, kind of biological weapons and bioterrorism, autonomy risk and the timelines we've been talking about, like 10 years is an eternity. Like that's that's a that's a I I think that's a crazy thing to do. So, if if that's the choice, if that's what you force us to choose, then then we're
[98:41] going to we're going to choose not to have that moratorium. And you know, I I think the the the benefits of that position exceed the costs, but it's it's not a perfect position if that's the choice. Now, I think the thing that we should do, the thing that I would support is the federal government should step in, not saying states you can't regulate, but here's what we're going to do and and states you can't differ from this, right? Like I think preemption is fine in the sense of saying that federal government says, here is our standard,
[99:11] this applies to everyone, states can't do something different. That would be something I would support if it would be done in the way. What um But, but this idea of states you can't do anything and we're not doing anything either, that that struck that struck us as, you know, very much not making sense and I think will not age well, was already starting to not age well with with all the um backlash that that you've seen. Now, in terms of in terms of what we would want, I mean, you know, the things we've talked about are are starting with
[99:41] transparency standards. um uh you know, in order to monitor some of these autonomy risks and bioterrorism risks. As the risks become more serious, um as we as we get more evidence for them, then I think we could be more aggressive in some targeted ways and and say, "Hey, AI bioterrorism is really a threat. Let's Let's pass a law that kind of forces people to have classifiers." And I could even imagine it it depends it depends how serious a threat it ends up being. We don't know for sure. Then
[100:11] we need to pursue this in an intellectually honest way where we say ahead of time, the risk has not emerged yet. But I could certainly imagine with the pace that things are going that, you know, I could imagine a world where later this year we say, "Hey, this this AI bioterrorism stuff is really serious. We should do something about it. We should put it in the federal We should, you know, put it in a federal standard. And if the federal government won't act, we should put it in a state state standard." I could totally see that. I I'm concerned about a world where if you just consider the the pace of
[100:41] progress you're expecting, the life cycle of of legislation, you know, the the benefits are, as you say, because of diffusion lag, the benefits are slow enough that I really do think this patchwork of on the current trajectory, this patchwork of state laws would prohibit I mean, having an emotional chatbot friend is something that freaks people out, then just imagine the kinds of actual benefits from AI we want normal people to be able to experience from improvements in health and health span and improvements in mental health and so forth. Whereas at the same time,
[101:11] uh it seems like you think the dangers are already on the horizon. And I just don't see that much um it seems like it would be especially injurious to the benefits of AI uh as compared to the the dangers of AI. And so that that that's maybe the where the cost-benefit makes less sense to me. So so so there's a few things here, right? I mean, people talk about there being thousands of these state laws. First of all, the vast vast majority of them do not pass. Um and you know, the the the the you know, the world works a certain way in theory, but like just
[101:41] because a law's been passed doesn't mean it's really enforced, right? The people the people, you know, implementing it may be like, "Oh my god, this is stupid. It would mean shutting off like, you know, everything that's ever been built and everything that's ever been built in Tennessee." So, you know, very often laws are interpreted in like, you know, a way that makes them that that that makes them not as dangerous or not as harmful on on the same side. Of course, you have to worry if you're passing a law to stop a bad thing, you have this you have this problem as well. Um, look, my my look, I mean, my basic view
[102:13] is, you know, if if if, you know, we could decide, you know, what laws were passed and how things were done, which, you know, we're only one small input input into that, you know, I would deregulate a lot of the stuff around the health benefits of AI. Um, I think you know, I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I greatly accelerate
[102:43] um the rate at which we discover drugs and just the the pipeline will get jammed up. Like the pipeline will not be prepared to like process all all of the stuff that's going through it. So, um you know, I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I regulatory process to bias more towards we have a lot of things coming where the safety and the efficacy is actually going to be really crisp and clear. Like I mean, a beautiful thing. Really really crisp and clear and like really really effective, but, you know, and and and maybe we don't need all this all this um
[103:15] like all this super structure around it that was designed around an era of drugs that barely work and often have serious side effects. But at the same time, I think we should be ramping up quite significantly the uh, know, this this kind of safety and security legislation. And, you know, like I've said, um, you know, starting with transparency is is my view of trying not to hamper the industry, right? Trying to find the right balance. I'm worried about it.
[103:45] Some people criticize my essay for saying, "That's too slow. The dangers of AI will come too soon if we do that." Well, basically, I kind of think like the last 6 months and maybe the next few months are going to be about transparency and then if these if these risks emerge when we're more certain of them, which I think we might be as soon as as later this year, then I think we need to act very fast in the areas that we've actually seen the risk. Like, I think the only way to do this is to be nimble. Now, the legislative process is normally
[104:15] not nimble, but we we need to emphasize to everyone involved the urgency of this. That's why I'm sending this message of urgency, right? That's why I wrote Adolescence of Technology. I wanted policy makers to read it. I wanted economists to read it. I want national security professionals to read it. You know, I want decision makers to read it so that they have some hope of acting faster than they would have otherwise. Is there anything you can do or advocate that would make it's more certain that the benefits
[104:47] of AI are um are better instantiated where I feel like you have worked with legislatures to be like, okay, we're going to prevent bioterrorism here. We're going to increase transparency. We're going to increase whistleblower protection. And I just think by default, the actual like the things we're looking forward to here it just seems very easy. They seem very fragile to uh different kinds of moral panics or political economy problems. >> Yeah, I don't actually So, so I don't actually agree that much in the developed world. I feel like you know, in the developed world like markets
[105:17] function pretty well and when there's when there's like a lot of money to be made on something and it's clearly the best available alternative, it's actually hard for the regulatory system to stop it. You know, we're we're seeing that in AI itself, right? I you know, like I think I've been trying to fight for is export controls on chips to China, right? And like that's in the national security interests of the US. Like, you know, that's like square within the you know, the the policy beliefs of you know, every almost
[105:48] everyone in Congress of both parties. But and you know, I think the case is very clear. The counterarguments against it are I'll politely call them fishy. Um, uh and yet it doesn't happen and we sell the chips because there's there's so much money. There's so much money riding on it. Um, and you know, the the that money wants to be made and and in that case, in my opinion, that's a bad thing. Um, and but but it also it also applies when when it's a good thing. And and so I I don't think that if we're talking
[106:19] about drugs and benefits of the technology, I I I I am not as worried about those benefits being hampered in the developed world. I am a little worried about them going too slow. And I as I said, I do think we should work to speed the approval process in the FDA. I do think we should fight against these chatbot bills that you're describing, right? Described individually, I'm against them. I think they're stupid. Um, but I actually think the bigger
[106:49] worry is the developing world. Um, where we don't have functioning markets, where um, you know, we often can't build on the technology that that we've had. I worry more that those folks will get left behind. And I worry that even if the cures are developed, you know, maybe there's someone in rural Mississippi who who doesn't get it as well, right? That's a That's a That's a kind of smaller version of the thing the concern we have in the in the developing world. And so the things we've been doing are, you know, you know, we work with, you know, we work with, you know, philanthropists, right? You know, we
[107:19] work with folks um who you know, who you know, deliver, you know, medicine and health interventions to, you know, to to developing world, to sub-Saharan Africa, you know, India, Latin America, you know, you know, other other developing parts of the world. That's the thing I think that won't happen on its own. Mhm. You mentioned export controls. Yeah. Why can't US and China both have a country of geniuses on a data center? >> Why can't, you know, why won't it happen
[107:49] or why shouldn't No, like why why shouldn't it happen? Why shouldn't it happen? Um you know, I think I think if this does happen, um you know, then then we kind of have a Well, we could have a few situa- If we have like an offense-dominant situation, we could have a situation like nuclear weapons, but like more dangerous, right? Where it's like um you know, kind of kind of either side could could easily destroy everything. Um we could also have a world where it's kind of it's unstable. Like nuclear equilibrium is stable, right? Because it's you know,
[108:20] it's like deterrents. But let's say there were uncertainty about like if the two AIs fought, which AI would win. Um that could create instability, right? You often have conflict when the two sides have a different assessment of their likelihood of winning, right? If one side is like, "Oh, yeah, there's a 90% chance I'll win." And the other side's like, "There's a 90% chance I'll win." Then then then a fight is much more likely. Um they can't both be right, but they can both think that. But this is like a fully general argument against the diffusion of AI technology, which it may which is that's an
[108:51] implication of this world. >> eventually. The other concern I have is that people the governments will oppress their own people with AI. And and and so, um you know, I'm I'm just I'm worried about some world where you have a country that's already a you know, kind of a uh you know, uh you know, there's there's a government that kind of kind of already um you know, is is kind of kind of building a you know, a tech a high-tech authoritarian state. Um, and to be clear, this is about the government.
[109:21] This is not about the people. Like, people We need to find a way for people everywhere to benefit. Um, my worry here is about governments. Um, so, yeah, my, you know, my my worry is that 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. Um, now, will will governments eventually get powerful AI and and, you know, there's risk of authoritarianism? Yes, will governments eventually get powerful AI and there's risk of, um, uh, you know, of of kind of bad bad bad
[109:52] equilibria? Yes, I think both things. But, the initial conditions matter, right? You know, at some point we're need 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 is a would be a better setup, although it requires more international cooperation than we currently seem to want to make. Um, but, you know, I don't I don't think a coalition of democracies or or certainly one country should just say, "These are the rules of the road." There's going to be some negotiation,
[110:22] right? The world is going to have to grapple with this. And what I would like is that the the the you know, the democratic nations of the world, those with, you know, who are close whose governments have represent closer to pro-human values are are holding the stronger hand then, have have more leverage when the rules of the road are set. And and so, I'm I'm very concerned about that initial condition. I um, I was re-listening to an interview from 3 years ago, and one of the ways it aged poorly is that I kept asking questions
[110:53] assuming there was going to be some key fulcrum moment in 2 or 3 years from now, when in fact, being that far out, it just seems like progress continues, AI improves, AI is more diffuse, and people use it for more things. It seems like you're imagining a world in the future where the countries get together, and here's Here's rules of the road, and here's the leverage we have, here's the leverage you have. When it seems like on current trajectory, everybody will have more AI. Some of that AI will be used by authoritarian countries, some of that within the authoritarian countries will be by private actors versus state
[111:23] actors. It's not clear who will benefit more. It's always unpredictable to tell tell in advance. You know, it seems like the internet privileged authoritarian countries more than you would have expected. And maybe the AI will be the opposite way around. So, I I I want to better understand what you're imagining here. >> Yeah, yeah. So, so just to be precise about it, I think the exponential of the underlying technology will continue as it has before, right? The models get smarter and smarter, even when they get to country of geniuses in the data center, you know,
[111:53] I think you can continue to make the model smarter. There's a question of like getting diminishing returns on their value in the world, right? How much does it matter after you've already solved human biology or, you know, you know, at some point you can do harder math, you can do more abstruse math problems, but nothing after that matters. But putting that aside, I do think the the exponential will continue, but there will be certain distinguished points on the exponential and companies, individuals, countries will
[112:24] reach those points at different times. And and so, you know, there's there's you know, could there be some you know, I you know, I talk about is a nuclear deterrent still in adolescence of technology is a nuclear deterrent still stable in the world of of of AI. I don't know, but that's that's an example of like one thing we've taken for granted that like the technology could reach such a level that it's no longer like you know, we can no longer be certain of it at least. You know, think of think of others, you know, there there there you know, there there kind of points where if you if you
[112:55] reach a certain point, you maybe you have offensive cyber dominance and like every every computer system is transparent to you after that. I unless the other side has has a kind of equivalent defense. So, I don't know what the critical moment is or if there's a single critical moment, but I think there will be either a critical moment, a small number of critical moments, or some critical window where it's like AI is AI confers some large advantage from the perspective of
[113:26] national security and one country or coalition has reached it before others. That that you know that that that you know I'm not advocating that they're just like okay we're in charge now or that's not that's not how that's not how I think about it. You know that there's always the the other side is catching up. There's extreme actions you're not willing to take and and and it's not right to take you know to take complete um to take complete control anyway, but but at at the point that that happens I think people are going to understand
[113:56] that the world has changed and there there's going to be some negotiation implicit or implicit about what what is the what is the post AI world order look like and and I think my interest is in you know making that negotiation be one in which you know classical liberal democracy has you know has a strong hand. Well well I don't understand what that better means cuz you say in the essay quote autocracy is
[114:28] simply not a form of government that people can accept in the post powerfully AI age and that sounds like you're saying the CCP as an institution cannot exist after we get AGI um and that seems like a like a very strong demand and it seems to imply a world where the leading lab or the leading country will be able to and by that language should get to determine how the world is governed or what kinds of governments are allowed and not allowed. Yeah, so
[115:00] when I when I I believe that paragraph was I think I said something like you could take it even further and say X. So I wasn't I wasn't necessarily endorsing that that that I wasn't necessarily endorsing that view. I you know, I was saying like here's a first, you know, here here's a weaker thing that I believe but you know, I think I you know, I think I said, you know, we have to worry a lot about authoritarians and you know, we should try and you know, kind of kind of check them and limit their power like you could take this kind of further much more
[115:30] interventionist view that says like authoritarian countries with AI are these you know, the the the you know, the these kind of self-fulfilling cycles that that you can't that are very hard to displace and so you just need to get rid of them from from the beginning. That that has exactly all the problems you say which is you know, you know, if you were to make a commitment to overthrowing every authoritarian country. I mean, they then they would take a bunch of actions now that like Right. you know, that that that could could lead to instability. So that that may or you know, that that that just that just may not be possible.
[116:02] But the point I was making that I do endorse is that it is it is quite possible that you know, today you know, the view or at least my view or the view in most of the Western world is is democracy is a better form of government than authoritarianism. But it's not like if a country's authoritarian we don't react the way we reacted if they committed a genocide or something, right? And and I'm I guess what I'm saying is I'm a little worried that in the age of AGI authoritarianism will
[116:32] have a different meaning. It will be a graver thing. Um and and we have to decide one way or another how to how how how how how how how to deal with that. And the interventionist view is one possible view. I was exploring such views. Um you know, it may end up being the right view. It it may end up being too extreme to be the right view. But I do have hope. And one piece of hope I have is there there is we have seen that as new technologies are invented
[117:02] forms of government become obsolete. I I mentioned this in adolescence of technology where I said, you know, like feudalism was basically, you know, like a form of government, right? And and then when when we invented industrialization feudalism was no longer sustainable, no longer made sense. Why is that hope? Why couldn't that imply that democracy is no longer going to be a competitive system? >> It could right. It could go It could go either way, right? But but I actually so
[117:32] I these problems with authoritarianism, right? That the problems of authoritarianism get deeper. I just I wonder if that's an indicator of other problems that authoritarianism will have, right? In other words, people become because authoritarianism becomes worse, people are more afraid of authoritarianism. They work harder to stop it. It's it's more of a like you have to think in terms of total equilibrium, right? Um I just wonder if
[118:04] it will motivate new ways of thinking about with the with with the new technology how to preserve and protect freedom. And and even more optimistically will it lead to a collective reckoning and you know, a kind of a more emphatic realization of how important some of the things we take as individual rights are, right? A more emphatic realization that we just we really can't give these away. There's there we've seen there's no other way to
[118:34] live that actually works. Um I I I am actually I am actually hopeful that I I guess one way to say it it sounds too idealistic, but I actually believe it could be the case is is that is that dictatorships become morally obsolete. They become morally unworkable forms of government. Um and that and that and that the the the the crisis that that creates is is is sufficient to force us to find another way. Um I I think there is genuinely a tough
[119:05] question here, which I'm not sure how you resolve. Uh for and we've had to come out one way or another on it through history, right? So, with China in the '70s and '80s, we decided, "Even though it's an authoritarian system, we will engage with it." And I think in retrospect that was the right call because it a state authoritarian system, but a billion plus people are much wealthier and better off than they would have otherwise been. Um and it's not clear that it would have stopped being an authoritarian country otherwise. You can just look at North Korea uh as an example of that, right? And I don't know if that takes that much that much intelligence to remain an authoritarian
[119:36] country that continues to coalesce its own power. And you can just imagine a North Korea with an AI that's much worse than everybody else's, but still enough to keep power. And I I and and and and so in general, it seems like should we just have this attitude of the benefits of AI will, in the form of all these empowerments of humanity and health and so forth, will be big. And in historically we have decided it's good to spread the benefits of technology widely, even with even to people whose governments are authoritarian. And I think I guess it is a tough question how
[120:06] to think about it with AI, but um historically we have said yes, this is a positive sum world, and it's still worth diffusing the technology. Yeah, so so there are a number of choices we have. I you know, I think framing this as a kind of government-to-government decision and you know, in in national security terms, that's like one lens, but there are a lot of other lenses. Like you could imagine a world where, you know, we produce all these cures to diseases and like the you know, the the the cures to diseases are fine to sell to authoritarian countries, the data
[120:36] centers just aren't, right? The chips in the data centers just aren't. Um and and that the AI industry itself. Um you know, like like another possibility is and and I think folks should think about this like, you know, could there be developments we can make either that naturally happen as a result of AI or that we could make happen by building technology on AI. Could we create an equilibrium where where it becomes infeasible for authoritarian countries to deny their people kind of private use
[121:07] of the benefits of the technology. Um, you know, are there are there are there are there equilibria where we can kind of give everyone in an authoritarian country their own AI model that kind of you know, you know, like defends themselves from surveillance and there isn't a way for the authoritarian country to like crack crack down on this while while retaining power. I don't know. That that sounds to me like if that went far enough it would be it would be a reason why authoritarian countries would disintegrate from the inside. Um, but but maybe there's a middle world where like there there's an
[121:37] equilibrium where if they want to hold on to power the authoritarians can't deny kind of individualized access access to the technology. But I actually do have a hope for the for the for the for the more radical version which is, you know, is it possible that the technology might inherently have properties or that by building on it in certain ways we could create properties that that that that have this kind of dissolving effect on authoritarian structures. Now, we we hoped originally, right? If we think about back to the
[122:07] beginning of the Obama administration, we thought originally that that, you know, social media and and the internet would have that property and turns out not to. But but I don't know. What what if we could what if we could try again with with the knowledge of how many things could go wrong and that this is a different technology. I don't know that it would work but it's worth a try. Yeah. I I think it's just I it's very unpredictable. Like there's first principles reasons why authoritarians might develop AI. >> very unpredictable. I don't think I mean we got it we we just got to we kind of we got to recognize the problem and then
[122:38] we got to come up with 10 things we can try and we got to try those and then assess whether they're working or which ones are working if any and and then try new ones if the old ones aren't working. >> what I'm getting at today as you say we will not sell data centers or sorry uh chips and then the ability to make chips in China. And so in some sense you are denying there would be some benefits to That's right. >> the Chinese economy, Chinese people, etc. because we're doing that. And then there'd also be benefits to the American economy because it's a positive sum world. We could trade. They could have their country doing data centers doing
[123:08] one thing. We could have ours doing another. And already we you're saying it's not worth that positive sum uh stipend to empower this country that What I would say is that you know we are we are about to be in a world where growth and economic value will come very easily. If right if we're able to build these powerful AI models, growth and economic value will come very easily. What will not come easily is distribution of benefits, distribution of wealth, political freedom,
[123:39] um you know, these are the things that are going to be hard to achieve. And so when I think about policy, I think I think that the technology and the market will deliver all the fundamental benefits, you know, almost almost faster than we can take them. Um uh and and that these questions about about distribution and political freedom and rights are are are the ones that that will actually matter and that policy should focus on. Okay, so speaking of distribution as you're mentioning, we have developing countries and um
[124:11] in many cases catch-up growth has been weaker than we would have hoped for. But when catch-up growth does happen, it's fundamentally because they have underutilized labor. You know, we can bring the capital and know-how from developed countries to these countries and then they can grow quite rapidly. >> Yes. Obviously in a world where labor is no longer the constraining factor, this mechanism no longer works. And so is the hope basically to rely on philanthropy from the people who immediately get wealthy from AI or from the countries that get wealthy from create AI. What is what is the hope for? >> I mean philanthropy should obviously play some role as it has the
[124:42] you know as it has as it has in the past, but I think growth is always growth is always better and stronger if we can make it endogenous. So, you know, what are the relevant industries in like in like in like in like an AI driven world? Look, there's lots of stuff, you know, like there's you know, I said I said we shouldn't build data centers in China, but there's no reason we shouldn't build data centers in Africa, right? In fact, I think it'd be great to build data centers in Africa. You know, as long as they're not owned by China we should build we should build data centers in Africa. I think that's a that's that's a I think that's a great
[125:12] thing to do. You know, we should also build you know, there's no reason we can't build you know, a pharmaceutical industry that's like AI driven. Like, you know, the the if if AI is accelerating accelerating drug discovery, then you know, there will be a bunch of biotech startups. Like, let's make sure some of those happen in the developing world and certainly during the transition and we can talk about the point where humans have no role, but but humans will have still have some role in starting up these companies and supervising supervising the AI models. So, let's make sure some of those humans are
[125:42] humans in the developing world so that fast growth can happen there as well. Mhm. You guys recently announced Quad is going to have a constitution that's aligned to a set of values and not necessarily just to the end user. And there's a world I could imagine where if it is aligned to the end user, it preserves the balance of power we have in the world today because everybody gets to have their own AI that's advocating for them. And so the ratio of bad actors and good actors stays constant. It seems to work out for our world today. Um why is it better not to do that but to have a specific set of values that the
[126:12] AI should carry forward? Uh yeah, so I'm not sure I'd quite draw the distinction in that way. There there may be two relevant distinctions here which are I think you're talking about a mix of the two. Like one is should we give the model a set of instructions about do this and do versus don't do this. And the other you know, versus should we give the model a set of principles for, you know, for kind of how to act. Um and and and there it's it's you know, it's it it it you know, it's it's just it's
[126:44] it's kind of purely a practical and empirical thing that we've observed that by teaching the model principles, getting it to learn from principles, its behavior is more consistent, it's easier to cover edge cases, and the model is more likely to do what people want it to do. Another words, if you you know, if you're like, you know, don't tell people how to hot-wire a car, don't speak in Korean, don't you know, you know, just you know, if you give it a a list of rules it doesn't really understand the rules and it's kind of hard to generalize from them.
[127:14] Um you know, if if it's just kind of a like you know, list of do do's and don'ts, whereas if you give it principles and then it you know, it has some hard guardrails like don't make biological weapons, but overall you're trying to understand what it should be aiming to do, how it should be aiming to operate. So, just from a practical perspective, that turns out to be just a more effective way to train the model. That's one piece of it. So, that you know, that's the kind of rules versus principles trade-off. Then there's another thing you're talking about, which is kind of like the corrigibility versus um
[127:46] like, you know, I would say kind of intrinsic intrinsic motivation trade-off, which is like how much should the model be a kind of I don't know, like a a skin suit or something where you know, you you know, you know, you're just kind of you know, it it just kind of directly follows the instructions that are given to it by whoever is giving it those instructions. Um versus how much should the model have an inherent set of values and and go off and do things on its own. Um and and and and and there I I would actually say
[128:17] everything about the model is actually closer to the direction of of like, you know, it should mostly do what people want. It should mostly follow the We're not trying to build something that like you know, goes off and runs the world on its own. We're actually pretty far on the corrigible side. Now now what we do say is there are certain things that the model won't do, right? That it's like, you know, that that that I think we say it in various ways in the constitution that under normal circumstances if someone asked the model to do a task it should do that task. That that should be the default. Um but if you've asked it
[128:49] to [snorts] do something dangerous or if you've, you know, if you've um asked it to um you know, uh uh to kind of harm someone else, um then the model is unwilling to do that. So I I actually think of it as like a mostly a mostly corrigible model that has some limits, but those limits are based on principles. Yeah, I mean then the fundamental question is how are those principles determined? And this is not a special question for Anthropic, this would be a question for any AI company, but um
[129:19] uh because you have been the ones to actually write down the principles, I get to ask you this question. Uh normally a constitution is like you write it down, it's set in stone, and there's a process of updating it, and changing it and so forth. In this case it seems like a document that people in Anthropic write that can be changed at any time that guides the behavior of systems that are going to be the basis of a lot of economic activity. What is the how do you think about how those principles should be set? Yes.
[129:49] Um so I think there's there's two there's maybe three three kind of sizes of loop here, right? Three three ways to iterate. One is you can iterate we iterate within Anthropic, we train the model, we're not happy with it, and we kind of change the constitution. And I think that's good to do. Um and you know, putting out publicly, you know, making updates to the constitution every once in a while saying here's a new constitution. Right. I think that's good to do cuz people can comment on it. The second level of loop is different companies will have different constitutions. Um and you
[130:19] know, I think it's useful for like Anthropic puts out a constitution and you know you know the Gemini model puts out a constitution and you know other companies put out a constitution and then they can kind of look at them, compare, outside observers can critique and say this this I like this one this thing from this constitution and this thing for that constitution and and then kind of that that creates some kind of you know soft incentive and feedback for all the companies to like take the best of each elements and improve. Then I think there's a third loop which is you
[130:50] know society beyond the AI companies and beyond just those who kind of you know who who comment on the constitutions without hard power and and there you know we've done some experiments like you know a couple years ago we did an experiment with I think it was called the collective intelligence project to like um you know to to basically poll people and ask them what should be in our AI constitution. Um and and you know I think at the time we incorporated some of those changes and so you could imagine with the new approach we've taken to the constitution doing
[131:21] something like that. It's a little harder because it's like that was actually an easier approach to take when the constitution was like a list of do's and don'ts. Um at the level of principles it has to have a certain amount of coherence. Um but but you could you could still imagine getting views from a wide variety of people and I think you could also imagine and this is like a crazy idea but hey you know this whole interview is about crazy ideas right? So Um uh you know you could even imagine systems of of kind of representative government having having input right? Like you know I wouldn't I
[131:52] wouldn't do this today because the legislative process is so slow like this is exactly why I think we should be careful about the legislative process in AI regulation but there's no reason you couldn't in principle say like you know all AI you know all AI models have to have a constitution that starts with like these things and then like you can append you can append other things after it but like there has to be this special section that like takes precedence. I wouldn't do that That's too rigid. That that sounds um you know, that that that that sounds kind of overly prescriptive in a way
[132:23] that I think overly aggressive legislation is. But like that is a thing you could that you know, like like that is a that is a thing you could try to do. Is is there some much less heavy-handed version of that? Maybe. I really like control loop two where obviously this is not how constitutions of actual governments do or should work where there's not this vague sense in which the Supreme Court will feel out what how people are feeling and what are the vibes and then update the update the constitution accordingly. So there's with actual governments there's a more procedural
[132:53] process. Yeah, exactly. But you actually have a vision of competition between constitutions which is actually very reminiscent of how some libertarian charter cities people used to talk about an archipelago of different kinds of governments could look like and then there'd be selection among them of who could operate the most effectively in which place people would be the happiest. And in a sense you're actually yeah, there's this vision >> I'm kind of recreating that. >> Yeah, yeah. Like the this like utopia of archipelagoes, you know. >> again you know, I think I think that
[133:23] vision has has you know, it things to recommend it and things that things that things that will kind of kind of go wrong with it. You know, I think I think it's a I think it's an interesting and some ways compelling vision but also things will go wrong with it that you hadn't that you hadn't imagined. So you know, I I I like loop two as well. But I I I feel like the whole thing has got to be some some mix of loops one, two, and three and it's a it's a matter of the proportions, right? I I think that's got to be the the answer. Um when somebody eventually writes the
[133:54] equivalent of the making of the atomic bomb for this era, what is the thing that will be hardest to glean from the historical record that they're most likely to miss? 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 This is a bias that's often present in history where anything that actually happened looks inevitable in retrospect. And And so, you know, I I think when people when people look back, it will be hard
[134:25] for them to put themselves in the place of people who were actually making a bet on this thing to happen that wasn't inevitable, that we had these arguments, like the arguments that you know that I make for scaling or that continual learning will be solved. Um uh uh you know, that that you know, some of us internally in our heads put a high probability on this happening, but but it's like there's there's a world outside us that's not that's not acting on that's not kind of not acting on that
[134:55] at all. Um uh and and and I think I think the the weirdness of it um I I I I think unfortunately like the insularity of it, like, you know, if if we're 1 year or 2 years away from it happening, like the average person on the street has no idea. And that's one of the things I'm trying to change, like with the memos, with talking to policy makers, but like I don't know. I think I I I think that's just a That's just like a crazy That's just like a crazy thing. Yeah. Um Finally, I would say and and this
[135:27] probably applies to almost all historical moments of crisis. Um how absolutely fast it was happening, how everything was happening all at once. And so, decisions that you might think you know, were kind of carefully calculated, well, actually, you have to make that decision and then you have to make 30 other decisions on the on the same day because it's all happening so fast. And And you don't even know which decisions are going to turn out to be consequential. So, you know, one of my one of my I guess worries, although it's also an insight into into
[135:57] you know, in into kind of what's happening is that, you know, some very critical decision will be will be some decision that, you know, someone just comes into my office and is like, "Dario, you have 2 minutes. Like, you know, should we should we do you know, should we do thing thing A or thing B on this like, you know, someone gives me this random, you know, half-page half-page memo and is like, should we should we do A or B?" And I'm like, "I don't know, I have to eat lunch. Let's do B." And and then, you know, that ends up being the most consequential thing ever. Mhm. So, final question. Uh it
[136:28] seems like you have there's not tech CEOs are usually writing 50-page memos every few months. And it seems like you have managed to build a role for yourself and a company around you which is compatible with this more intellectual type role of CEO. And I want to understand how you construct that and how like how does that work to be you just go away for a couple of weeks and then you tell your company this is the memo, like here's what we're doing. It's also reported you write a bunch of these internally. Yeah, so I mean, for
[136:59] this particular one, you know, I wrote it over winter break. Um so, there was the time, you know, and I was having a hard time finding the time to actually find it to actually write it. But, I actually think about this in a broader way. I actually think it relates to the culture of the company. So, I probably spend a third, maybe 40% of my time making sure the culture of Anthropic is good. As Anthropic has gotten larger, it's it's gotten harder to just, you know, get involved in like, you know, directly involved in like the training of the models, the launch of the models, the building of the products. Like, it's
[137:30] 2,500 people. It's like, you know, there's just, you know, I have certain instincts, but like there's only, you know, it's very difficult to get in to get to get involved in every single detail, you know. I like I I try as much as possible, but one thing that's very leveraged is making sure Anthropic is a good place to work, people like working there, everyone thinks of themselves as team members, everyone works together instead of against each other. And you know, we've seen as some of the other AI companies have grown without naming any names, you know, we're starting to see
[138:00] decoherence and people fighting each other. And you know, I would argue there was even a lot of that from the beginning but but you know, that it's it's gotten worse but I I think we've done an extraordinarily good job even if not perfect of holding the company together, making everyone feel the mission that we're sincere about the mission and that you know, everyone has faith that everyone else there is working for the right reason that we're a team that people aren't trying to get ahead of each other's expense or backstab each other which again can happen a lot at some of
[138:30] the other places um and and how do you make that the case? I mean it's a lot of things, you know, it's me, it's it's it's Daniela who you know, 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 in the culture is I some and just you know, the the you know, the other leaders as well but especially me have to articulate what the company is about, why it's doing what it's doing, what its strategy is,
[139:00] what its values are, what its mission is and what it stands for. And um you know, when you get to 2,500 people you can't do that person by person. You have to write or you have to speak to the whole company. This is why I get up in front of the whole company every 2 weeks and speak for an hour. It's actually I mean, I wouldn't say I write essays internally. I do two things. One, I write this thing called a DVQ. Dario Vision Quest. Um uh uh I wasn't the one who named it that. That's the name it it it received and it's one of these names that I kind of I tried to fight it cuz
[139:31] it made it sound like I was like going off and smoking peyote or something. Um uh but but the name just stuck. Um so I get up in front of the company every 2 weeks. I have like a three or four page document and I just kind of talk through like three or four different topics about what's going on internally, the you know, the the models we're producing, the products, the outside industry, the world as a whole as it relates to AI and geopolitically in general, you know, just some mix of that. And I just go through very very honestly. I just go through and I just I
[140:02] just say, you know, this is this is what I'm thinking and this is what Anthropic leadership is thinking. And then I answer questions. And and that direct connection I think has a lot of value that is hard to achieve when you're passing things down the chain, you know, six six levels deep. Um and you know, large fraction of the company comes comes to attend either either in person or um either in person or virtually. And it it you know, it really means that you can communicate a lot. And then the other thing I do is I just, you know, I
[140:32] have a channel in Slack where I just write a bunch of things and comment a lot. Um and often that's in response to, you know, just things I'm seeing at the company or questions people ask or like you know, we do internal surveys and there are things people are concerned about and so I'll write them up. And I'm like I'm, you know, I'm I'm I'm just I'm very honest about these things, you know, I just I just say them very directly. And the point is to get a reputation of telling the company the truth about what's happening. To call things what they are, to acknowledge
[141:03] problems, to avoid the sort of corpo speak, the kind of defensive communication that often is necessary in public because, you know, the world is very large and full of people who are, you know, interpreting things in bad faith. Um but, you know, if you have a company of people who you trust and we try to hire people that we trust, then then you know, you can you can you can you know, you can you can really just be entirely unfiltered. Um and uh you know, I think I think that's an enormous strength of
[141:33] the company. It makes it a better place to work. It makes people more, you know, more than the sum of their parts. It increases likelihood that we accomplish the mission cuz everyone is on the same page about the mission. And everyone is debating and discussing how to best to accomplish the mission. Mhm. Well, in lieu of an external Dario vision quest, we have this interview. This this this interview is a little like that. Uh this is Inquiring Minds with Dwarkesh. Thanks for doing it. Yeah, thank you, Dwarkesh. Hey everybody. I hope you enjoyed that episode. If you did, the most helpful thing you can do is just share it with other people who you think might enjoy it. It's also
[142:03] helpful if you leave a rating or a comment on whatever platform you're listening on. If you're interested in sponsoring the podcast, you can reach out [music] at dwarkesh.com/advertise. Otherwise, I'll see you on the next one.