Anthropic CEO Dario Amodei: AI's Potential, OpenAI Rivalry, GenAI Business, Doomerism

And I want to make sure that we say what we believe and that we warn the world about the possible downsides. Even though, you know, no one can say what's going to happen. We're, you know, we're saying what we, you know, what we think might happen, what we think is likely to happen. You know, we back it up as best we can, although it's often, you know, extrapolations about the future where no one can be sure. But, you know, I think we see ourselves as kind of having the duty to, you know, to kind of warn the world about what's going to happen. And that's not to say, you know, I think there's an incredible number of like positive applications of AI, right? I've kind of continued to talk about that. I wrote this essay, Machines of Loving Grace. I feel, in fact, that I and Anthropic have often been able to do a better job of articulating the benefits of AI than some of the people who call themselves optimists or accelerationists. So I think we probably appreciate the benefits more than anyone. But for exactly the same reason, because we can have such a good world if we get everything right, I feel obligated to warn about the risks.

So all of this is coming from your timeline. Basically, it seems like you have a shorter timeline than most. And so you were feeling a sense of urgency to get out there because you think that this is imminent. Yes, I'm not sure. You know, I think it's very hard to predict, particularly on the societal side. So if you say, you know, when are people going to deploy AI or when are companies going to use, you know, X dollars of spend of AI or, you know, when will AI, you know, be used in these applications or when will it drive these medical cures? that's kind of harder to say. I think the underlying technology is more predictable, but still uncertain. Still no one knows. But I think on the underlying technology, I've started to become more confident. There isn't no uncertainty about it. I think the exponential that we're on could totally peter out. I think there's maybe, I don't know, 20% or 25% chance that sometime in the next two years, the models just stop getting better for reasons we don't understand, or maybe reasons we do understand, like data or compute availability.

And then everything I'm saying just seems totally silly. And everyone makes fun of me for all the warnings I've made. And I'm just totally fine with that, given the distribution that I see. And so I should say that this is part, our conversation is part of a profile I'm writing about you. I've spoken with more than two dozen people who've worked with you, who know you, who've competed with you, and I'm going to link that in the show notes. If anybody wants to read it, it's free to read. But one of the themes that has come through across everybody I've spoken with is that you have about the shortest timeline of any of the major lab leaders, and you just referenced it just now.

So why do you have such a short timeline, and why should we believe in yours? Yeah, it really depends what you mean by timeline. So one thing, and, you know, I've been consistent on this over the years, is, you know, there are these terms in the AI world like AGI and superintelligence. Like, you'll hear leaders of companies say, we've achieved AGI, we're moving on to superintelligence. Or, like, it's really exciting that someone stopped working on AGI and started working on superintelligence. So I think these terms are totally meaningless. I don't know what AGI is. I don't know what superintelligence is.

It sounds like a marketing term. It is marketing. Yeah, it sounds like, you know, something designed to activate people's dopamine. So you'll see in public, I never use those terms. And I, you know. I'm actually careful to criticize the use of those terms. But I think despite that, I am indeed one of the most bullish about AI capabilities improving very fast.

Dario AmodeiThe thing I think is real, that I've said over and over again, is the exponential. The idea that every few months we get an AI model

that is better than the AI model we got before, and that we get that by investing more compute in AI models, more data, data, more new types of training models. Initially, this was done by what's called pre-training, which is when you just feed a bunch of data from the internet into the model. Now we have a second stage that's reinforcement learning or test time compute or reasoning or whatever you want to call it. I think of it as a second stage that involves reinforcement learning. Now both of those things are scaling up together, as we've seen with our models and as we've seen with models from other companies. And I don't see anything blocking the further scaling of that. There's some stuff about how do we broaden the tasks. On the RL side of it, we've seen more progress on, say, math and code, where the models are getting pretty close to a high professional level and less on more subjective tasks. But I think that is very much a temporary obstacle. So when I look at it, I see this exponential and I say, look, people aren't very good at making sense of exponentials, right?

Like, you know, if something is doubling every six months, then, you know, two years before it happens, it looks like it's only one-sixteenth of the way there. And so we are sitting here in the middle of 2025, and the models are really starting to explode in terms of the economy, right? If you look at the capabilities of the model, they're starting to saturate all the benchmarks. If you look at revenue and you know anthropics revenue every year has grown 10x um uh every year we're kind of conservative and we say you know it can't grow 10x this time you know i i you know i never i never assume anything and actually always i'm very conservative in saying ah i think it's going to slow down on the business side but we went from zero to 100 million in 2023 we went from 100 million to a billion in 2024 and you know this year in this first half of the year we've gone from $1 billion to, you know, I think as of speaking today, it's well above four, it might be 4.5.

And so if you think about it, you know, suppose that exponential continued for two years. I'm not saying it will. But suppose it continued for two years, you know, you're like well into the hundred billions. I'm not saying that'll happen. I'm saying the situation is that when you're on an exponential, you can really get fooled by it. Two years away from when the exponential goes totally crazy, it, you know, it looks like it's just starting to be a thing. And so that's the fundamental dynamic. You know, we saw that with the internet in the 90s, right, where it was like, you know, were getting fast. And over a few years, it became possible to basically build a digital global communications network on top of all this when it wasn't possible just a few years ago.

And almost no one except for a few people really saw the implications of that and how fast it would happen. And so that's where I'm coming from. That's what I think. Now, I don't know. If a bunch of satellites crashed, maybe the internet would have taken longer. If there was an economic crash maybe would have taken a little longer so we can't be sure of the exact timelines but i think people are getting fooled by the exponential and and not realizing how fast it it might be how fast i think it probably will although i'm not sure but so many folks in the ai industry are talking about diminishing returns from scaling now that really doesn't fit with the vision you just laid out are they wrong yeah uh i i from what we've seen i can only speak in terms of the models at anthropic um but what i've seen in terms of the models at anthropic if we look at you know let's take coding coding is one area where you know i think anthropic models have advanced very quickly adoption has been very quick we're not just a coding company we're planning to expand to many areas but if you if you look at if you look at coding um you know every you know we released 3.5 sonnet a model we call 3.5 sonnet v2 um uh which you know let's call it 3.6 sonnet now um

3.7 sonnet uh and then 4.0 sonnet and 4.0 opus and you know that series of four or five models each one got substantially better at coding than than the last if you want to look at benchmarks you can look at you know sweet bench growing from uh you know i think 18 months ago is that like three percent or something um growing all the way to you know 72 to 80 depending on how you how you measure it and and the real usage has grown exponentially as well where we're heading more and more towards autonomously, you can just use these models. I think the actual of code that's written at Anthropic is, you know, at this point, written by or at least with the involvement of one, you know, one of the Claude models.

And various other companies have said, you know, have said, you know, similar statements to similar statements to that. So we see the progress as being very fast and the exponential is continuing and we don't see any diminishing returns. But there are some liabilities, it seems like, with large language models. For instance, continual learning. We had Dwarkech on a couple weeks ago. Here's how he put it, and he wrote about it in his Substack. The lack of continual learning is a huge, huge problem. The LLM baseline at many tasks might be higher than an average human, but you're stuck with the abilities you get out of the box.

So you just make the model, and that's it. It doesn't learn. That seems like a glaring liability. What do you think about that? So first of all, I would say even if we never solved continual learning, Um, even if we never solve continual learning and memory, um, I think that the potential for the LLMs to do, you know, incredibly, uh, incredibly well to, you know, affect things at the scale of the economy will be very high, right? If I think of the field I used to be in, biology and medicine, like, you know, let's say I had a very smart Nobel prize winner and, you know, I, I, I said, okay, you know, you're, you know, you've, you've, you've discovered all these things.

You have this incredibly smart mind, but, you know, you can't, you know, you can't, like, read new textbooks or absorb any new information. I mean, that would be difficult, but, like, still, if you had, like, 10 million of those, like, they're still going to make a lot of biology breakthroughs. Like, they're going to be limited. They're going to be able to do some things humans can't, and there are some things humans can do that they can't. But even that, even if we impose that as a ceiling, like, man, that's pretty damned impressive and transformative. And even if I said you never solved that, like, I think, you know, I think people are underestimating the impact.

But look, context windows are getting longer, and models actually do learn during the context window, right? So as I, you know, talk to the model during the context window, I have a conversation. It absorbs information. The underlying weight of the model may not change. But, you know, just like I'm talking to you here, and we're having a conversation, and I listen to the things you say, and I think, and I respond to them, the models are able to do that. And from a machine learning perspective, from an AI perspective, there's no reason we can't make the- context length 100 million words today, right, which is roughly what a human hears in their lifetime. There's no reason that we can't do that. It's really inference support. And so again, even that fills in many of the gaps, not all the gaps, but it fills in many of the gaps.

And then there are a number of things like learning and memory that do allow us to update the weights. So, you know, there are a number of things around, you know, types of reinforcement learning, learning training, you know, you know, we used to many years ago talk about inner loops and outer loops, right? The inner loop is like, I have some episode and I learned some things in that episode, and I'm trying to optimize for the lifetime of that episode. And kind of the outer loop is the agents learning over episodes. And so I think maybe that inner loop, outer loop structure as a way to learn the continual learning.

Dario AmodeiOne thing we've learned in AI is whenever it feels like there's some fundamental obstacle, like two years ago, we thought there was this fundamental obstacle around reasoning.

Turned out just to be RL. You just train with RL and you let the model write some stuff down. You know, you let the model write things down to try and figure out objective math problems without being too specific. You know, I think, and we already have maybe some evidence to suggest that this is another of those problems that is not as difficult as it seems, that will fall to scale plus a slightly different way of thinking about things. Do you think your obsession with scale might blind you to some of the new techniques like Demesis Abbas says, you know, to get to AGI, or you might call it super powerful AGI, whatever, human level intelligence is what we're all talking about. We might need a couple new techniques for that to happen.

If you're so focused on scale, can you get there? We're developing new techniques every day. Okay. Claude is very good at code, and we don't really talk externally that much about why Claude is so good at code. Why is it so good at code? Like I said, we don't talk externally about it. I have to ask. So every new version of Claude that we make has improvements to the architecture, improvements to the data that we put into it improvements to the methods that we use to train it so we're developing new techniques all the time um new techniques are a part of every you know model that we build and you know that's why we you know i've said these things about like you know we're trying to optimize for like talent density as much as possible. Like, you need that talent density in order to invent the new techniques. You know, there's one thing that's been hanging over this conversation, which is that maybe Anthropic is the company with the right idea, but the wrong resources. Because you look at what's happening with XAI and inside Meta, where Elon's built his massive cluster. Mark zuckerberg is building this five gigawatt data center and they are putting so much resources towards scaling up um is it possible i mean anthropic obviously you have raised billions

of dollars but these are trillion dollar companies yeah so we've raised i think at this point a little short of 20 billion dollars it's not bad so that's that's not that's not nothing and i would also say if you look at the size of the data centers that we're building with for example, Amazon, I don't think our data center scaling is substantially smaller than that of any of the other companies in the space. You know, in many cases, these things are limited by energy, they're limited by capitalization. You know, when people talk about, you know, these large amounts of money, they're talking about it over several years, right? And when you hear some of these announcements. Sometimes they're not funded yet. We've seen the size of the data centers that folks are building. And we're actually pretty confident that we will be within a rough range of the size of the data centers that they built. You talked about talent density. What do you think about what Mark Zuckerberg is doing on the talent density front? I mean, combining that with these massive data centers, it seems like he's going to be able to compete.

Yeah, so this is actually very interesting because, you know, one thing we noticed is that relative to other companies, you know, I think a lot fewer people from Anthropic have been caught by these. And it's not for lack of trying. I've talked to plenty of people who got these offers at Anthropic and who just turned them down, who wouldn't even talk to Mark Zuckerberg, who said, no, I'm staying at Anthropic.

Dario AmodeiAnd our general response to this was, you know, I posted something to the whole company Slack where I said, look, you know, we are not willing to compromise our, you know, our compensation principles, our principles of fairness to respond individually to these offers.

the way things work. philanthropic is there's a series of levels. When candidate comes in, they get assigned a level and we don't negotiate that level because we think it's unfair. We want to have a systematic way. If Mark Zuckerberg throws a dart at a dartboard and hits your name, that doesn't mean that you should be paid 10 times more than the guy next to you who's just as skilled, who's just as talented. um uh and and and my view of the situation is that you know the only way you can really be hurt by this is if you allow it to destroy the culture of your company by panicking by treating people unfairly in an attempt to to defend the company um and i think actually this was a unifying moment for the company where um you know we we didn't give in we refused to compromise our principles because we have the confidence that people are anthropic because they truly believe in the mission.

And, you know, I think that gets to kind of how I see this. I think that what they are doing is trying to buy something that cannot be bought, and that is alignment with the mission. And, you know, I think there are selection effects here. Like, you know, are they getting the people who are most enthusiastic, who are most mission-aligned, who are most excited. But they have talent and GPUs. You're not underestimating them? I will see how it plays out. I am pretty bearish on what they're trying to do. So let's talk a little bit about your business because a lot of people have been wondering, is the business of Generative AI a real thing?

And I'm also curious. I have questions all the time. You talked about how much money you've raised, close to $20 billion. You've raised $3 billion from Google, $8 billion from Amazon. $3.5 billion from a new round led by Lightspeed, who I've spoken with. What is your pitch? Because you are not part of like a big tech company. You're out there on your own. Do you just bring the scaling laws and say, can I have some money? So my view of this has always been that talent is the most important thing. So, you know, if you go back three years ago, you know, we were in a position where we had raised mere hundreds of hundreds of millions.

OpenAI had already raised, you know, $13 billion from Microsoft. And, of course, you know, the large hyper-cap tech companies were sitting on $100 billion, $200 billion. dollars. And basically, the pitch we made then is we know how to make these models better than others do, right? There may be a curve. There may be a curve of scaling laws. But look, if we are in a position where we can do for $100 million what others can do for $1 billion, and we could do for $10 billion what they can do for $100 billion, then it's 10 times more capital efficient to invest in Anthropic than it is to invest in these other companies. Would you rather be in a position where you can do anything for 10 times cheaper or where you start with a large pile of money.

If you can do things 10 times cheaper, the money is a temporary defect that you can remedy. If you have this intrinsic ability to build things for the same price much better than anyone else or as good as anyone else for a much lower price, investors aren't idiots, or at least they aren't always idiots. It depends which one you go to. I'm not going to name any names. But, you know, they basically understand the concept of capital efficiency. And so we've been in a position, you know, three years ago where, you know, these differences were like 1,000x. And now you're saying, you know, with $20 billion, can you compete with $100 billion?

And my answer is basically yes, because of the talent density. You know, I've said this before, but, you know, the Anthropic is actually the fastest growing software company in history at the scale that it's at. So we grew from zero to 100 million in 2023, 100 million to billion in 2024. And this year we've grown from 1 billion to, I think I said this before, 4.5. So that 10x a year, I mean, you know, every year I like suspect that we'll grow at that scale. And every year I'm almost afraid to say it publicly because I'm like, no, it couldn't possibly happen again. So, you know, I think the growth at that scale like kind of speaks for itself in terms of our ability to compete with the big players.

OK, so CNBC says 60 to 75 percent of Anthropic sales are through the API. That was according to internal documents. Is that still accurate? I won't give exact numbers, but the majority does come through the API, although we also do have a flourishing apps business. And, you know, I think more recently, the, you know, max tier, which power users use, as well as cloud code, which coders use. So, you know, I think we have a thriving and fast-growing apps business, but yes, the majority comes through the API. So you're making the most pure bet on this technology. Like, you know, OpenAI might be betting on ChatGPT and Google might be betting on the fact that no matter where the technology goes, it can, you know, integrate into Gmail and Calendar.

So why have you made this bet on this, the pure bet on the tech itself? Yeah, I mean, I would say I wouldn't quite put it that way. I think we've, I would describe it more as we've bet on business use cases of the model, more so than we've bet on the API per se. And it's just that the first business use cases of the model come through the API. So, you know, as you mentioned, OpenAI is very focused on the consumer side. Google is very focused on kind of the existing products that Google has.

Dario AmodeiOur view is that, if anything, the enterprise use of AI is going to be greater even than the consumer use of AI, right?

I should say the business use because it's enterprise, it's startups, it's developers and it's kind of, you know, power users using the model for productivity. I also think that being a company that's focused on the business use cases actually gives us better incentives to make the models better. A thought experiment that I think is worth running is, you know, suppose I have this model and it's, you know, it's as good as an undergrad at biochemistry. And then I improve it and it's as good as a PhD student at biochemistry. If I go to a consumer, right, if I give them the chat bot and I say, great news, I've improved the model from, you know, undergrad to graduate level in biochemistry, you know, maybe, I don't know, 1% of consumers care about that at all, right?

99% are just going to be like, I don't understand it either way. But now suppose I go to Pfizer and I say, you know, I've improved this from undergrad at biochemistry to grad at biochemistry. Like this is going to be the biggest deal in the world, right? They might pay 10 times more for something like that. It might have 10 times more value to them. And so the general aim of making the models solve the problems of the world to make them smarter and smarter, but also able to bring many of the positive applications, right? The things I wrote about in like Machines of Loving Grace of like solving the problems of biomedicine, solving the problems of geopolitics, solving the problems of, you know, of economic development, you know, as well as more prosaic things like finance or legal or productivity or insurance, I think it gives a better incentive to kind of develop the models as far as, as, as, as.

as far as possible. And I think in many ways, it's like it may even be a more positive business. So I would say we're making a bet on the business use of AI because it's most aligned with kind of the exponential. Okay, then briefly, how did you decide to go with the coding use case? Yes. So, you know, originally, as happens with most things, we're trying to optimize for making the model better at a bunch of stuff. And, you know, coding particularly stood out in terms of how valuable it was. You know, I've worked with thousands of engineers, and there was a point about a year, a year and a half ago, where one of the best I'd ever worked with said, you know, every previous coding model has been useless to me, and this one finally was able to do something I wasn't able to do. And then after we released it, it started getting quick adoption.

This was around the time that, you know, a lot of the coding companies like Cursor, Windsurf, GitHub augment code started exploding in popularity. And then when we saw how popular it was, we kind of doubled down on it. My view is that coding is particularly interesting because A, the adoption is fast. And B, getting better at coding with the models actually helps you to develop the next model. So it has a number of, I would say, advantages. And now you're selling your AI coding through Claude Code. But it's very interesting. The pricing model has been confounding to some. You can spend $200 a month and get the equivalent. I spoke to one developer. They got the equivalent of $6,000 a month from your API. Ed Zitron has pointed out the more popular that your models get, the more money you're going to lose if people are super users of this technology. So how does that make sense? So actually, pricing schemes and rate limits are surprisingly complicated. So some of this is basically the result of when we released Claude Code in the max tier, which we eventually tied together, actually not fully understanding the implications of the ways in which people could use the models and how much they were

actually able to get. So over the last few days, as of the time of this, as of the time of this interview, we've adjusted that particularly on the larger models like Opus. I think it's no longer possible to spend that much with a $200 subscription. And, you know, it's possible more changes will come in the future, but we're always going to have a district. of users who use a lot and users who use some amount. And it doesn't necessarily mean we're losing money, that there are some users who get more, who if you were to measure via API credits, spend, get a better deal on the consumer subscription than they would on the API products, right?

There's a lot of assumptions there. And I can tell you that some of them are wrong. We are not, in fact, losing money. But I guess there's another question about whether you can continue to serve these use cases and not raise prices. So just to give you a couple stats, there are some developers that are upset because using Anthropics' newer models in Cursor is costing them more than it ever has. Startups that I've spoken with say Anthropics is down a bunch because they can't get access to the GPUs. At least that's what they imagine is happening. And I was just with Amjad Masad at Replit in an interview that we're going to air next week, who said there was a period of time where the price per token, price to use these models was coming down and it stopped coming down.

So is it the is what's happening that these models are just so expensive for Anthropic to run that it's hitting it's a wall of its own? Again, I think you're I think you're making assumptions here. That's why I'm asking the CEO. Yeah. You know, you know, the way I think about it is we think about the models in terms of how much value are they creating. Right. So as the models get better and better, I think about how much value they create. And there's a separate question about how the value is distributed between those who make the model, those who make the chips and, you know, those who make those who make the the the the the the the underlying applications.

So again, without being too specific, I think there are some assumptions in your question that are not necessarily correct. Can you tell me which ones? So I'll say this. I do expect the price of providing a given level of intelligence to go down. I expect the price of providing the frontier of intelligence, which will provide kind of increasing economic value. that might go up or it might go down my guess is it probably stays about where it is but again the value that's created you know goes way up so to you to two years from now my guess is that we'll have models that cost of the same order magnitude that they cost today, except they'll be much more capable of doing work, much more autonomously, much more broadly than they are capable of today. One of the things that Amjad mentioned was he thinks that the bigger models are not as intensive to run or more intensive to run given their size because of the architecture and some of these techniques that we talked about, that they're lighting up only certain sections of the model. So his idea, I'm hopefully conveying this truthfully, is that Anthropic can run these models without too much bulk on the back end, but is still keeping those prices where they are. And I think the line that I'm going to draw there

is maybe that to get to software margins, there were some reports that Anthropic is slightly below software gross margins. You're going to have to charge a little bit more for these models. So, yeah, again, I think larger models cost more to run than smaller models. You know, I think the technique you're referring to is maybe mixture of experts or something like that. So whether you're models or mixture of experts or not, like mixture of experts is like a way to run models more cheaply. They have a given number of parameters. It's a way to train models. But if you're not using that technique, then larger models that don't use that technique cost more to run than smaller models that don't use that technique.

And if you're using that technique, larger models that use that technique cost more to run than smaller models that are using that technique. So I think that's sort of a distortion of the situation. Basically, I'm just guessing and I'm trying to find out what the truth is from you. Yeah, look, so I, you know, in terms of like the cost of the models, like one thing you'd be surprised by, people, you know, people kind of impute this thing to like, oh, man, it's going to be really hard to get the margins from like X percent to Y percent. We make improvements all the time that make the models like 50 percent more efficient than they are before.

we are just at the beginning of optimizing inference. Inference has improved a huge amount from where it was a couple years ago to where it is now. That's why the prices are coming down. And then how long is it going to take to be profitable? Because I think the loss is going to be like $3 billion this year. So I would distinguish different things. There's the cost of running the model, right? So for every dollar the model makes, It costs a certain amount that. is actually already fairly profitable. There are separate things. There's the cost of paying people and buildings. That is actually not that large in the scheme of things. The big cost is the cost of training the next model. And I think this idea of the companies losing money and not being profitable, it's a little bit misleading. And you start to understand it better when you look at scaling laws so as a thought exercise these numbers are not exact or even close for anthropic let's imagine that in 2023 you train a model that costs 100 million dollars um and then in 2024 you deploy the 2023 model it makes 200 million dollars in revenue but you spend a billion dollars to train um you know to train a new model in 2024 so and then you know and then in 2025 the billion

dollar model makes two billion in revenue and you spend 10 billion to train the next model So the company every year is unprofitable. It lost $800 million in 2024. And then in 2025, it lost $8 billion. So, you know, this looks like a hugely unprofitable enterprise.

Dario AmodeiBut if instead, I think in terms of is each model profitable, right? Think of each model as a venture. I invested $100 million in the model.

And then I got $200 million out of the model in the next year. So that model had 50% margins and made me $100 million. The next year, the company invested a billion dollars and made $2 billion in – or sorry, the next model, the company invested a billion dollars and made – So are your models profitable? So every model is profitable. They're all profitable. But the company is unprofitable every year. I'm not – this is a style. I'm not like claiming these numbers for Anthropic or claiming these facts for – but this general – Is it true for Anthropic? This general dynamic is, in general terms, the explanation for what is going on.

And so at any time, if the models stopped getting better or if a company stopped investing in the next model, you would have probably a viable business with the existing models. But everyone is investing in the next model. And so eventually, it'll get to some scale. But the fact that we're spending more to invest in the next model suggests that the scale of the business is going to be larger the next year than it was the year before. Now, of course, what could happen is like... The models stop getting better and there's this kind of one-time cost that's like a boondoggle and we spend a bunch of money, but then the companies, the industry will kind of return to this plateau, to this level of profitability, or the exponential can keep going. So I think that's a long-winded way to say, I don't think it's really the right way to think about things. Right. But what about open source? Because if you stopped, let's say you stopped investing in the models and open source caught up then people could swap in open source now i'd love to hear your perspective on this because one of the things people have talked to me about when it comes to the anthropic business is there is that risk eventually that open source

gets good enough that you can take anthropic out and put open source in yeah so you know people have i you know i think one of the things that's been true of this industry is that and you know i saw it early in the history of AI. Every community that AI has gone through, it has this set of heuristics about how things work. Like back when I was in, you know, AI back in 2014, there was an existing kind of AI and machine learning research community that like thought about things in a certain way. And we're like, this is just a fad. This is a new thing. This can't work. This can't scale. And then because of the exponential, all those things turned out to be false.

Then a similar thing happened with kind of like people deploying AI within companies to various applications. Then there was the same thought in the startup ecosystem. And I think now we're at the phase where kind of the world's business leaders, like the investors and the business, they have this whole lexicon of commoditization, you know, modes, which layer is the value going to, you know, which layer is the value going to accrue to. And open source is this idea that you can kind of see everything that's going on, you know, that it has a significance, that it kind of undermines the fact that, you know, the idea that it undermines business.

And I actually find as someone who didn't come from that world at all, who never thought in terms of that lexicon, this is one of these situations where not knowing anything often leads you to make better predictions than kind of the people who have their way of thinking about things from the last generation of tech. And, you know, this is all, I think, a long-winded way of

Dario Amodeisaying, I don't think open source works the same way in AI that it has worked in other areas.

Because with open source, you can see the source code of the model. Here, we can't see inside the model. It's often called open weights instead of open source to kind of distinguish that. But a lot of the benefits, which is that many people can work on it, that it's kind of additive, it doesn't quite work in the same way. So I've actually always seen it as a red herring. When I see a new model come out, I don't care whether it's open source or not. Like, if we talk about DeepSeek, I don't think it mattered that DeepSeek is open source. I think I ask, is it a good model? Is it better than us at, you know, the things that we, that's the only thing that I care about.

It actually doesn't matter either way. Because ultimately, you have to host it on the cloud. The people who host it on the cloud do inference. These are big models. They're hard to do inference on. And conversely, many of the things that you can do when you see the weights, you know, we're increasingly offering on clouds where you can fine tune the model. You can, you know, you know, we're even looking at ways to, you know, to to kind of, you know, investigate the activations of the model as part of like an interpretability interface. We did some little things around steering last time. So I think it's the wrong axis to think in terms of when I think about competition, I think about like which models are good.

at the tasks that we do. I think open source is actually a red herring. But if it's free and cheap to run. It's not free. You have to run it on inference and someone has to make it fast on inference. All right. So I want to learn a little bit more about Dario the person. Yes. So we have a little bit of time left. So I have some questions for you about early life and then how you became who you are. Yes. So what was it like growing up in San Francisco? Yeah. The city, when I first grew up here, had not really gentrified that much. When I grew up, the tech boom hadn't happened yet. It happened as I was going through high school. And actually, I had no interest in it. It was totally boring to me. I was interested in being a scientist.

I was interested in physics and math. And the idea of writing some website actually had no interest to me to me whatsoever or like founding a company like those weren't things that i was uh that i was interested in at all um you know i was interested in discovering fundamental scientific truth and i was interested in like you know how can i how can i do something that like makes the world better. Um, uh, so, so, you know, that was, that was kind of more and, you know, I watched the tech boom happen around me, but I, I feel like, you know, there was all kinds of things I probably could have learned from it that would have been helpful now, but I just actually wasn't paying attention and had no interest in it, even though I was like right at the center of it. So you're the son of a Jewish mother and an Italian father. That is true. From where I'm from in Long Island, we call that a pizza bagel. A pizza bagel. I've never, I've never heard that term before. So what was your relationship with your parents like? Yeah. I mean, you know, I was always pretty close with them. You know, I feel like they gave me a sense of, you know, of kind of right and wrong and what was important in the world. I feel like, you know, kind of imbuing a strong sense of responsibility

is maybe the thing that I remember most. You know, they were always people who felt that sense of responsibility and, you know, wanted to make the world better. And I feel like, you know, that's one of the main things that I learned from them. You know, it was always a very loving family, a very caring family. I was very close with my sister, Daniela, who, of course, became my co-founder. And, you know, I think we decided very early that we wanted to work together in in some capacity. I don't know if we imagined that it would happen at quite the scale that it has happened. But I think it really, that was something we decided early that we wanted to do.

The people that I've spoken with that have known you through the years have told me that your father's illness had a big impact on you. Can you share a little bit about that? Yes. Yes. Yes, he was, yeah, he was ill for a long time and eventually died in 2006. So that was actually one of the things that drove me to, I don't think we mentioned it yet in this interview, but before I went into AI, I went into biology. So, you know, I'd gone to, you know, I'd shown up at Princeton wanting to be a theoretical physicist. And, you know, I did some work in cosmology for the first few months of my time there.

You know, and, you know, that was around the time that my father died. And, you know, that did have an influence on me and kind of was one of the things that convinced me, you know, to... to go into biology, you know, to try and address, you know, human illnesses and biological problems. And so I started talking to some of the folks who worked on biophysics and computational neuroscience in the department that I was at at Princeton. And that was what led to the switch to biology and computational neuroscience. And then, you know, of course, after that, I eventually went into AI. And the reason I went into AI was actually a continuation of that motivation, which is that, you know, as I spent many years in biology, I realized that the complexity of the underlying problems in biology felt like it was beyond human scale. You know, in order to understand it all, you needed hundreds, thousands of, you know, human researchers. And, you know, they often had a hard time collaborating or sharing their, you know, combining their knowledge. And AI, which was, I was just starting to see the discoveries in it, felt to me like the only technology that could kind of bridge that gap, could bring us beyond

human scale to, you know, to fully understand and solve the problems of biology. So yeah, there is a through line there. Right. And I could have this wrong. But one thing I heard was that his illness was largely uncurable when he had it. And there have been advances that have been, can you share a little bit more? Yes. There are advances that have made it much more manageable today. Yes, yes. That is true. Actually, only in the maybe three or four years after he died, the cure rate for the disease that he had went from 50% to roughly 95%. Yeah. I mean, it has to have felt so unjust to have your father taken away by something that It could have been cured.

Of course, of course. But it also tells you of the urgency of solving the relevant problems, right? That there was someone who worked on the cure to this disease that managed to cure it and save a bunch of people's lives, but could have saved even more people's lives if they had managed to find that cure a few years earlier than they did. And I think that's one of the tensions here, right? That, you know, I think AI has all of these benefits. And, you know, I want everyone to get those benefits as soon as possible. You know, I probably understand, you know, better than almost anyone how urgent those benefits are.

And so I really understand the stakes. You know, when I speak out about AI has these risks and I'm worried about these risks, I get very angry when people call me a doomer. I got really angry when, you know, when someone's like, this guy's a doomer, he wants to slow things down. You heard what I just said, like, you know, my father died because of, you know, cures that, you know, could have happened a few years later. I understand the benefit of this technology. When I sat down to write Machines of Loving Grace, you know, I wrote out all the ways that billions of people's lives could be better with this technology.

Some of these people, some of these people who on Twitter, you know, cheer for acceleration, I don't think they have a humanistic sense of the benefit of the technology. Their brain's just full of adrenaline. And they're like, they want to cheer for something. They want to accelerate. I don't get the sense they care. And so when these people call me a doomer, I think they just completely, completely lack any moral credibility in doing that. You know, it really makes me lose respect for them. And I've been wondering what this word impact has been because it's come up so often that those who have been around you have said you've been singularly obsessed with having impact.

In fact, I spoke with someone who knew you well who said you wouldn't watch Game of Thrones because it wasn't tied to impact, that it was a waste of time, and you wanted to be focused on impact. Actually, that's not quite right. I wouldn't watch it because it was so negative sum. People were playing such negative. It was like these people start off and they're partly the situation and partly because they're just horrible people. They, like, create the situation where at the end of it, everyone is, like, worse off than everyone was before. I'm really excited about, like, creating positive some situations.

I recommend you watch it. It's a great show. But I hear your complaints. I actually have watched some parts of it. I was just very reluctant and didn't watch it for a long time. Let's get back to the impact. Okay. Let's get back to the impact. So that's what impact is, is effectively your career has been this quest to have that impact, to be able, tell me if I'm going too far, to prevent other people from being in similar situations. You know, I think that's a piece of it. I mean, you know, I have looked at, you know, many attempts to help people. And, you know, some of them are more effective than others.

And, you know, I think I've always tried to, you know, there should be strategy behind it. There should be brains behind. um you know trying to trying to help people um you know which often means that there's a long path to it right it can run through a company and you know many activities that are technical and not immediately tied to the the kind of impact impact that you're trying to have but you know the the the arc is i'm always trying to bend the arc towards that i think i think that's my that's my picture of it that's that's really why i that's really why i got into this right you You know, I think, you know, similar to the reason to get into AI was that, you know, I saw the problems of biology as almost intractable without it or at least too slow moving.

You know, I think my reason to start a company was that I had worked at other companies and I just didn't feel like the way those companies were run was really oriented towards, you know, trying to have that impact. There was a story around it that was often used for recruiting, but it became clear to me over the years that story was not sincere. I'm going to circle around a little bit because it's clear that you're referring to OpenAI here. From what I understand, you had 50% of OpenAI's compute. I mean, you ran the GPT-3 project. So if anyone was going to be focused on impact and safety, it wouldn't have been you.

Yes, there was a period during which that was true. that was true. That wasn't true the entire time. That was, for example, when we were scaling up GPT-3. Yeah. So, you know, when I was at OpenAI, I and a lot of my colleagues, including the people who, you know, eventually founded Anthropic. The pandas. Pandas. That's the name you gave them. That isn't a name I gave them. The name they took. That isn't a name they took. That's the name other people called them. Maybe it's a name other people called them. That's not a name I ever used for my team. OK, sorry, go ahead. That's a good clarification.

Thank you. So, yeah, you know, we were involved in scaling up these models. Actually, the original reason for building GPT-2 and GPT-3, it was an outgrowth of the kind of AI alignment work that we were doing, right? Where myself and Paul Cristiano and some of the Anthropic co-founders had invented this technique called RL from human feedback. And that was designed to help steer models in, you know, in a direction to follow human intent. It was actually a precursor to, you know, we were trying to scale up another method called scalable supervision, which I think is just starting to work many years later to help models follow more kind of scalable human intent.

But what we found is even with the more primitive technique, RL from human feedback, it wasn't working with the small language models with, you know, GPT-1 that we applied it to and that had been built by other people at OpenAI. And so the scaling up of GPT-2 and GPT-3 was done in order to kind of study these techniques, in order to apply RL from human feedback at scale. You know, this goes to one thing, which is that I think in this field, the alignment of AI systems and the capability of AI systems is intertwined in this way that always ends up being kind of more tied and more intertwined than we think.

Actually, what this made me realize is that it's very hard to work on the safety of AI systems and the capability of AI systems separately. It's very hard to work on one and not the other. I actually think the value and the way to inflect the field in a more positive way comes from organizational-level decisions. When to release things, when to study things internally, what kind of work to do on systems. And that was one of the things that kind of motivated me and some of the other you know to be anthropic founders to kind of go off and do it our own way but again like if you were driving if you think language uh if you think capabilities and safety are interlinked and you were the guy driving the cutting-edge models within open ai you know you if you left you knew they were going to be a company that was still doing this stuff that's right it seems like if you're driving the capabilities you'd be the one in the driver's seat to help it be safe the way that you want it again i will say um you know if there's a decision on releasing a model if there's a decision on the governance of the company if there's a decision on you know how the personnel of the company works um you know how the company represents itself externally

the decisions that the company makes with respect to deployment the claims it makes about how it operates with respect to society. You know, many of those things are not things that you control just by training the model. And, you know, I think trust is really important. I think the leaders of a company, they have to be trustworthy people. They have to be people whose motivations are sincere. No matter how much you're driving the forward, the company technically, if you're working for someone whose motivations are not sincere, who's not an honest person, who does not truly want to make the world better, it's not going to work. You're just contributing to something bad.

So, and I'm sure you've heard the criticism from people like Jensen who say, well, Dario thinks he's the only one who can build this safely. And therefore, speaking of that word control, wants to control the entire industry. What do you think about that? I've never said anything like that that's an outrageous lie that's the most outrageous lie I've ever heard um by the way I'm sorry if I got Jensen's words wrong but no no no the words were correct okay the but but but what he's saying but the words are the words are outrageous in fact I've said multiple times and I think anthropics actions have shown it that um you know we're aiming for something we call a race to the top you know I've said this on podcasts over the years, and I think Anthropic's actions have shown it, where, you know, with a race to the bottom, right, everyone is competing to like, you know, get things out as fast as possible. And so I say when you have a race to the bottom, it doesn't matter who wins, everyone loses, right? Because you make the unsafe system that, you know, helps your adversary or causes economic problems or, you know, is unsafe from an alignment perspective. The way I think about

Dario Amodeithe race to the top is that it doesn't matter who wins, everyone wins, right? So the way the race to the top works is you set an example for how the field works. You say, you know,

we're going to engage in this practice. So a key example of this is responsible scaling policies. We were the first to put out a responsible scaling policy. And, you know, we didn't say everyone else should do this or you're bad guys we didn't you know we didn't you know kind of try to use it as an advantage we put it out and then we encouraged everyone else to do it and many and and then we discovered in the months after that that you know there were people within the other companies were trying to put out responsible scaling policies but the fact that we had done it allowed you know gave those people permission right kind of kind of enabled those people to, you know, to make the argument to leadership, hey, Anthropic is doing this, so we should do it as well. The same has been true of investing in interpretability.

We release our interpretability research to everyone and allow other companies to copy it, even though we've... Seeing that it sometimes has commercial advantages, same with things like constitutional AI, same with the measurement of the dangers of our system, dangerous capabilities evals. So we're trying to set an example for the field, but there's an interplay where it helps to be a powerful commercial competitor. I've said nothing that anywhere near resembles the idea that this company should be the only one to build the technology. I don't know how anyone could ever derive that from anything that I've said.

It's just, yeah, yeah, it's just an incredible and bad faith distortion. All right, let's see if we can lightning around like one or two before I ask you the last one, which we'll have five minutes for. What happened with SBF? What happened with SBF? I mean, he was one of the—go ahead. I couldn't tell you. What didn't you answer? I probably met the guy four or five times. Okay. So I have no great insight into the psychology of SBF or why he did things as stupid or immoral as he did. I think the only, you know, the only thing I had ever seen ahead of time with SBF was, you know, a couple people mentioned to me that he was like hard to work with, that, you know, he was like a bit of a move fast and break things guy.

And I was like, OK, you know, there's like plenty of people. Welcome to Silicon Valley. Yeah, like welcome to Silicon Valley. And so I remember saying, OK, I'm going to give this guy non-voting shares. I'm not going to put him on the board. He sounds like a bad person to deal with every day, but he's excited about AI. He's excited about AI safety. He's a bull on AI, and he's interested in AI safety, so seems like a sensible thing to do. In retrospect, that move fast and break things turned out to be much, much, much more extreme and bad than, you know, than I ever imagined. Okay, so let's end here.

So you found your impact. I mean, you're working the dream pretty much right now. I mean, think about all the ways that AI can be used for biology just to start. You also say that this is a dangerous... technology. And I'm curious if your desire for impact could be pushing you to accelerate this technology while, you know, potentially devaluing the possibility that it could, that controlling it might not be feasible. So, you know, I think I have more than anyone else in the industry warned about the dangers of the technology, right? We just spent 10, 20 minutes talking about, you know, the frightening, you know, the large array of, you know, people who run, you know, trillion dollar companies criticizing me for, you know, for talking about the dangers of these technologies, right? You know, I have U.S. government officials. I have people who run four trillion dollar companies criticizing me for talking about the dangers of the technology, right? Imputing all these bizarre motives that bear no relationship to, you know, to anything I've ever said, not supported in anything I've ever done. And yet I'm going to continue to do it.

I actually think that, you know, as the revenues, as the economic business of AI ramps up and it's ramping up exponentially, you know, if I'm right, in a couple of years, it'll be the biggest source of revenue in the world, right? It'll be the biggest industry in the world. And people who run companies already think it. So we actually have this terrifying situation where, you know, hundreds of billions to trillions to, I would say, maybe 20 trillion of capitals on the side of Accelerate AI as fast as possible. We have this company that's very valuable in absolute terms, but looks very small compared to that, right? $60 billion. And I keep speaking up, even if it makes folks, there have been these articles, some folks in the US government are upset at us, for example, for opposing the moratorium on AI regulation, for being in favor of export controls for chips on China, for talking about the economic impacts of AI.

Every time I do that, I get attacked by many of my peers. Right, but you're still assuming that we can control it. That's what I'm pointing out. But I'm just telling you how much effort, how much persistence, how much despite everything that's stacked up, despite all the dangers, despite the risk that it has to the company of being willing to speak up i'm willing to do it and and and that's that's what that's that's why i'm saying that look if if i thought that there was No way to control the technology, right? If I thought, even if I thought this is just a gamble, right? Some people are like, oh, you think there's a 5% or 10% chance that AI could go wrong.

You're just rolling the dice. That's not the way I think about it.

Dario AmodeiThis is a multi-step game, right? You take one step. You build the next step of most powerful models.

You have a more intensive testing regime. As we get closer and closer to the more powerful models, I'm speaking up more and more. And I'm taking more and more drastic actions because I'm concerned that the risks of AI are getting closer and closer. We're working to address them. We've made a certain amount of progress. But when I worry that the progress that we've made on the risks is not fully aligned with the – is not going as fast as we need to go for the speed of the technology, then I speak up louder. And so, you know, you're asking, why am I talking, you know, you started this interview by saying, what's gotten into you?

Why are you talking about this? It's because the exponential is getting to the point that I worry that we may have a situation that our ability to handle the risks is not keeping up with the speed of the technology. And that's how I'm responding to it. If I believe that there was no way to control the technology, which I see absolutely no evidence for that proposition. We've gotten better at controlling models with every model that we release, right? All these things go wrong, but like you really have to stress test the models pretty hard. That doesn't mean you can't have emergent bad behavior.

And I think, you know, if we got to much more powerful models with only the alignment techniques we have now, then I'd be very concerned. Then I'd be out there saying everyone should stop building these things. Even China should stop building these. I don't think they'd listen to me, which is one reason I think export controls is a better measure. But if we got a few years ahead in models and had only the alignment and steering techniques we had today, then, you know, I would definitely be advocating for us to, you know, to slow down a lot. The reason I'm warning about the risk is so that we don't have to slow down, so that we can invest in safety techniques and can continue the progress of the field.

it would be a huge economic effort even if one company was willing to slow down the technology you know that doesn't stop all the other companies that doesn't stop our geopolitical adversaries to whom this is a existential fight fight fight fight for survival So, you know, there's very little, you know, there's very little latitude here, right? We're stuck between all the benefits of the technology, the race to accelerate it, and the fact that that is a multi-party race. And so I am doing the best thing I can do, which is to invest in safety technology to speed up the progress of safety. I've written essays on the importance of interpretability, on how important various directions in safety are.

We release all of our safety work openly because we think that's the thing that's a public good. That's the thing that everyone needs to share. So if you have a better strategy for balancing the benefits, the inevitability of the technology, and the risks that it face, I am very open to hear it because I go to sleep every night thinking about it. Because I have such an incredible understanding of the stakes in terms of the benefits, in terms of, you know, what it can do, the lives that it can save. I've seen that personally. I also have seen the risks personally. We've already seen things go wrong with the models.

You know, we have an example of that with Brock. And, you know, people dismiss this, but they're not going to laugh anymore when the models are taking actions, when they're manufacturing, and when they're in charge of, you know, medical, medical, medical interventions, right? People can laugh at the risks when the models are just talking, but I think it's very serious. And so I think what this situation demands is a very serious understanding of both the risks and the benefits. These are high-stakes decisions. They need to be made with a seriousness. And I think something that makes me very concerned is that, on one hand, we have a cadre of people who are just doomers.

People call me a doomer. I'm not. but there are doomers out there, people who say they know there's no way to build this safely. You know, I've looked at their arguments. They're a bunch of gobbledygook. The idea that these models have dangers associated with them, including dangers to humanity as a whole, that makes sense to me. The idea that we can kind of logically prove that there's no way to make them safe, that seems like nonsense to me. So I think that is an intellectually and morally unserious way to respond to the situation we're in. I also think it is intellectually and morally unserious for people who are sitting on 20 trillion dollars of capital who all work together because their incentives are all in the same way, there are dollar signs in all of their eyes, to sit there and say, we shouldn't regulate this technology for 10 years.

Anyone who says that we should worry about the safety of these models is someone who just wants to control the technology themselves. That's an outrageous claim, and it's a morally unserious claim. We've sat here and we've done every possible piece of research. We speak up when we believe it's appropriate to do so. We've tried to back up, you know, when we make claims about the economic impact of AI, we have an economic research council. We have an economic index that we use to track the model in real time. And we're giving grants for people to understand the economic impact of the technology. I think for people who are far more financially invested in the success of the technology than I am to just, you know, breezily lob ad hominem attacks. You know, I think that is just as intellectually and morally unserious as the doomer's position. I think what we need here is we need more thoughtfulness. We need more honesty. We need more people willing to go against their interest, willing to not have, you know, breezy Twitter fights, hot takes. We need people to actually invest in understanding the situation actually do the work actually put out the research and and actually add some light

and some insight to the situation that we're in i am trying to do that i don't think i'm doing that perfectly as no human can i'm trying to do it as well as i can it would be very helpful if there were others who would try to do the same thing well dario i said this off camera but i want to make sure to say it on as we're wrapping up. I appreciate how much Anthropic publishes. We have learned a ton from the experiments, everything from red teaming the models to vending machine Claude, which we didn't have a chance to speak about today. But I think the world is better off just to hear everything going on here. And to that note, thank you for sitting down with me and spending so much time together. Thanks for having me. Thanks everybody for listening and watching. And we'll see you next time on Big Technology Podcast.

No one's journey is the same. That's why Delta SkyMiles moves with you. Whether it's earning miles from coffee runs to flights, to redeeming for vacations your way, or unlocking exclusive experiences, a SkyMiles membership fits into your lifestyle,