Inside the Mind of Anthropic CEO Dario Amodei | The Circuit | Extended Interview

You know, I've never been someone who slept all that well. Let's just say I'm learning the art of finding ways to relax and sleep through moments of unusual pressure.

It's this feeling of like the exponential. Suppose you were to accelerate away from earth on a spaceship at relativistic speed. The way special relativity works is you go to sleep and you wake up and two days have gone by on earth. So you have to deal with two days in one day. And then you go to sleep, and because you've continued to accelerate, three days have gone by on earth, and then the next day four days have gone by. That's a little bit what it feels like.

There are enough clear and present issues that we have to deal with that I'm constantly dealing with those, while thinking about how we can prepare. But I don't think paranoia or worrying about what you'll wake up to is productive. I've looked at people in history who've dealt with these very high pressure situations, and you need to learn to respond rationally and not put dangers out of proportion to each other. This yo-yoing between 'I'm not worried' and 'oh my God, we need to panic today' — I think that's a hallmark of immature decision making. The actual mature decision making is we can't ignore this, we can't be complacent. In fact, it's getting to be a bigger and bigger risk. But we have to respond rationally, like a surgeon would deal with an operation, or a military officer would deal with a military operation, or someone making decisions that affect a lot of people has to make those decisions rationally and maintain a basic sense of calm.

The interesting thing is that the experience I've had for my whole career, and certainly the whole time at Anthropic, is that there's this kind of smooth exponential. The experience of the smooth exponential is nothing's happening, nothing's happening, nothing's happening, a little things happen, and then zoom, it goes crazy. That's the experience of the world, the scale of the company compared to the other companies and compared to the world. So I was watching this graph for a while and I said, 'Oh yeah, we'll probably become the AI company with the most revenue and the most valuation sometime around this time.' And indeed it has happened. So in one sense I'm not surprised because there's just a smooth line on the graph. But of course in another sense, when things actually happen, you see so much more detail and color to it, and it definitely is surprising. And we're just keeping in mind all the things we usually keep in mind: how do we train good models, how do we put them in good products, how do we make sure that everything's safe, how do we help people but also manage the societal risks around the technology? It's all the same questions, just under a bigger microscope as it were.

The whole first internet revolution was happening around me and I had absolutely no interest in it. I was just interested in doing math and scrawling things. I was interested in understanding the universe. I was interested in science fiction. That was the general milieu. I think I just felt a lot of curiosity about the world.

Yeah, I think the general spirit of nonconformism and individualism and it's okay to be crazy. I think a good deal of that probably did rub off on me. You hear these stories about going to countries in Europe or even other parts of this country where it's discouraged or considered weird to think about things in some different way, or have some set of crazy ideas. There's a lot of things I'm actually very critical about with Silicon Valley, but one thing that I think is good about it is this encouragement that it doesn't matter if all the experts are against you, if you have a coherent vision and a coherent worldview, you should go and pursue it. Maybe it just won't work at all, but if it does, there's this kind of long-tailedness to it where you can search certain veins and find a huge goldmine. I think that spirit is very important.

That was the time when the Open Philanthropy project was first being started up, which Holden was the lead of. And I was at that time a biological scientist. So I was helping them with some of the stuff they were doing around developing world health or biological research. I advised on that stuff — what were the areas that were promising, what were the areas that were less promising.

Look, I'll say it very simply. There are many difficult issues that you face when you're building powerful technology that Anthropic faces every day, where we don't know whether we're making the right decision or the wrong decision. So there are many valid disagreements to be had on safety. We certainly had some of those disagreements with them, but that alone is not sufficient to leave. People here have had disagreements with me, people here have disagreements with each other. But when you feel that you can't trust someone, when you feel that their values are not what they say they are, when you feel that they're not honest, when you feel that they're not in it for the reasons that they say, when you see disturbing patterns of behavior, dishonesty, that makes it very hard to continue to work with a company, to continue to trust the company. And look, at the end of the day, why argue with someone when you don't have the same vision and you don't trust them? The way to resolve it is you go off and do your thing, they go off and do their thing. And I am completely at peace with the idea that we're doing things our way and they're doing things their way. We'll see who wins in the market and we'll see who wins in the court of public opinion. I think those things speak louder than any drama about who left. We're providing an example of how to deploy this technology in what we think is a responsible way. If they disagree, they should make that argument. And I think that's really all there is to say about it.

What happened is that the summit was extremely disorganized. We all came up at the last minute and they changed the order in which we were standing, and then they took a picture of us and then they ordered us all to hold hands. If you've ever been to one of these summits — I'm not saying anything bad about India in particular — but all of these international type summits that have heads of state are super disorganized.

I don't know what to tell you. There was Narendra Modi up there suddenly telling everyone to hold hands.

Here's what I will tell you. There is a wide variance in the quality and the trustworthiness of the people building this technology. I think this meme that no one trusts each other, I don't think it's right. I've known Demis Hassabis, who builds the Gemini models — they're a competitor to Claude models — I've known him for 15 years. We've worked together on a number of issues. We buy compute from Google, we swap safety ideas all the time. So my view is that there are some players who are more trustworthy than others. And I think there are players outside Anthropic who I trust, who I see as trustworthy. What I think needs to happen is that the trustworthy actors need to get together and put the untrustworthy actors in a position where they have to adopt the same standards. With a lot of experience, I've learned that there are some folks who don't do the right thing on their own, but if there's a majority of the industry doing the right thing, then the rest of the industry is left in a position where there's not much they can do but come along. There's the positive version where you inspire other people — that's Demis and me inspiring each other. He does AlphaFold, we're trying to do something in bio as well. We do interpretability research, they start interpretability research. It's not even competition, it's just each company does something cool and the other company is like, 'That's cool, we'd like to do that too and see if there's something new within that we can do.' So that's the carrot side of the race to the top. Then there's the stick side, or the implicit stick, where you're like, 'These guys are doing the right thing, those guys will look bad if they don't do the right thing.' And often we see behaviors where they grudgingly do the right thing while trying to pretend they're doing something different. But I think that's the way we get the industry together, and that's the way we get the industry to cooperate.

When we started Anthropic, the base thing that mattered, the thing that always matters, is we want to do this right. But then you have to ask yourself, in order to fund the very expensive creation of these models, it needs to be a company, it needs to have a business model. Does the business model get in the way of the values? There's always this question. But I think one of the things I learned from being at other companies and watching other companies is that if you pick a business model that fundamentally conflicts with your values, you're going to have a hard time. Either you betray your own values or you become irrelevant. You end up in a catch-22 situation. It's far better to pick a business model that is compatible with your values. So when we thought about it, we said, we've seen the world of social media, the consumer world — it really seems to encourage engagement, even addiction. The slop we've seen with AI video models — it wants to maximize the number of minutes you're paying attention to because that's the advertising revenue driven incentive. Whereas if we look at enterprise, we want to make these models useful to people. If I think of all the positive things you can do with AI — I warn a lot about the negative things, but ultimately we think the positive things will outweigh the negative things. Many of those fall under the banner of enterprise. We want to use AI to cure diseases that we couldn't cure before — that's working with biotech, pharma, academic research groups. All of those are enterprises. We want to use AI to make energy cheaper and more efficient — that's all enterprise. We want to use AI to help with education — most of that is enterprise. We want to use AI to address health in the developing world — there are nonprofits, but those are basically enterprises. We want to increase economic growth — that is basically enterprise as well. And then I think there's another factor: enterprises care a lot about trust and long-term relationship. Consumer can have this almost gimmicky aspect to it, whereas with enterprise, what matters is you build a relationship where you work with a company for many years, you deliver on what you say, and they basically trust you. So it's very synergistic with our goal of deploying these models in a positive and safe way. And so I think it serves us well to have this business model that largely aligns with our values. Not that there aren't conflicts sometimes, not that there aren't hard choices we have to make, but I think the number of such choices is much lower than it would be otherwise.

Model quality is the most important thing. We're very far ahead right now on model quality. There is some amount of inertia, but I've never relied on that. Anthropic has never relied on 'this is sticky and people won't switch.' I think you want to have a better model, you want to have a better product. And we see the growth rates haven't inflected at all. If anything, they've gone up, at least at the time of taping this interview. So I tend to think that is the most important thing.

This is one of these questions that's very hard to predict in advance. If you could predict it perfectly, people would make a huge amount of money on the market and they'd always be right. So no one knows exactly what's going to happen. But I would note a few things. All of these traditional software companies have a number of moats. I think what's going to happen is some of these moats are going to go away, but others are going to stay around. The ability to quickly write software — I definitely think that's going away. If your moat is 'we wrote this complex software that no one else can write,' good luck, you're not going to be able to defend that. But folks have customer relationships, folks have know-how of how the field works, folks have unique domain knowledge. So my advice to all of these folks is, don't be complacent, don't ignore it. Make a list of all your moats and be very aware that some of them are going to go away, while others are going to become relatively more important. Because there are limiting factors, and there may also be new moats. And I think those that deftly respond, that lean into the list of moats that are still present as well as the new ones, will do well. I think those that are complacent, that delude themselves that what worked in the past will continue to work, they're not going to have a good time. So that is the advice I would give. And at the end of the day, I would guess — it depends what you call SaaS and what you don't — but I would guess that the software industry gets larger, not smaller. Although there will be some big losers.

I just think the pie is getting bigger. With AI, the pie is getting bigger. The existing incumbents may be smaller in relative terms. Some of them may go down in value, some may even go out of business if they don't adapt in the right way. But I think you see this often when growth is really fast. If what's possible with AI grows by 10x, it's very easy for an existing incumbent industry to go up by 1.5x — just not as much as the whole big pie is growing. So I think that may happen. That's not to say we won't have some big losers. I think those who don't adapt, who put their heads in the sand, who don't see what's coming, who don't identify the moats they have, they're going to have a really hard time.

There have been a number of cases where we've really spoken our minds about what we think. I've been very outspoken about the need for export controls on chips to China. I say this because I think it would be really bad for America, for the state of democracy in the world, for China to be ahead in AI capabilities. And some of the chip makers obviously don't agree with that view, but it hasn't stopped me from saying it. I'm saying it again now, even after we've signed more partnerships. What they know is that we always work with them, we've been good partners. We can work together. I'm sure they wish we didn't say these things, but these things are what I believe. What are you going to do? At the end of the day, they benefit from these deals as much as we do. We're all adults here. We can work together on one thing while disagreeing about another thing.

The compute is ramping up very quickly. So it can both be the case that the fundamentals of the business look good, but in a year you'll have three times as much compute, or four times — I'm not going to give exact numbers — but these compute ramps are very fast. And we have every expectation that the revenue ramp will meet and exceed those. But raising money is the buffer against this cone of uncertainty. So it's a totally rational thing to do. It's a very small dilution to the business, and it logically is not at all the same thing as there being anything wrong with the fundamentals of the business. In fact, it's compatible with the opposite.

One of these things about compute is there's a marketing compute. My view is that over a period of time, even longer than a couple months, we can get large amounts of compute. One thing that's worth saying here is I don't think we bought too little compute by any reasonable standard. We were planning for a 10x a year growth in compute. 10x a year is what we expect. That isn't what we've seen over the first quarter of 2026. We saw a greater than 3x growth in revenue quarterly — not annualized, 3x in the quarter, which of course 3 to the fourth power is 80x over the course of the year. We didn't plan for 80x annualized growth. It would not have been rational to plan for 80x annualized growth, because that means if you only get 10x, you have 8 times less. So we're in a locally extreme explosion of compute that's not going to continue. If that continued, you'd get to revenue numbers by the end of the year that no company on earth has. I don't think that's going to happen. It just can't. But you can have these short periods where it's like, 'Oh my God, this is faster growth than we ever possibly anticipated.' But you saw the compute deals with Google, you saw the compute deals with Amazon. There are more that we can and will do. The market's liquid. If you're able to use compute really well and there's the demand, you'll get your compute. It might just take a month or two.

Look, we have a lot of difficult challenges in front of us, and there's this race to the top idea that we're trying to pull other companies along with us. And I think we've seen that. We have pulled them along with us. Sometimes they don't admit that that's what they're doing. Sometimes they copy us while they're attacking us. But this pull is very valuable. So I think the value of being the preeminent company, both commercially and in terms of models, it's not about beating rivals for the sake of beating rivals. It's about having the ability to pull the ecosystem along with us. And we hope that we can do more of that in the future.

I mean, look, we're always trying to succeed. We're not trying to fail here. I'm not someone who believes we should shut this technology down. We shouldn't build it. We exist within a free enterprise system, and there's nothing wrong with this. We just have to mitigate the risks of the models. And so it's always been the balance between the two.

What I would say is that I've put a lot of time into thinking about how that's the case. As companies scale, I've been paranoid at every scale of the company. At every scale, there's some new challenge, some new way the company can lose — either its will to win commercially or the core of its values. I'm worried about both because I see them as synergistic. I actually see the fact that we've been able to make such good models as the thing that allows us to assert our values in a way that works as the company grows. There are lots of pitfalls here, lots of ways to go wrong. Not because me or the co-founders or the company's leaders' values change, but because the composition of the company changes very fast. So I spend probably half of my time just talking to the company about the culture of Anthropic and how the culture works. When you're growing this fast, you're hiring a bunch of people from big tech companies. If you don't tell them how Anthropic operates, they'll simply recapitulate the only thing they know, which is how to operate at the companies they came from. So this is a constant struggle and a constant challenge. It's me and Daniela's maybe number one top priority: figuring out how to preserve this because we recognize that this is the core of who we are in the long run.

I would say two things. The first is we have a unified company, a unified culture. I think we've grown larger while still being incredibly efficient. Everyone's still on the same page. The cultural and organizational unity — I would say that's the biggest factor. And I would say the second biggest factor is Claude itself. We're now using Claude to help develop our models, make them more efficient, and quickly develop products. There are all kinds of new practices you have to develop. We're still new at it, but it's producing a lot of acceleration and increasingly producing reliable acceleration. So those are the two factors I would point to.

I think some of the wildest stuff I've seen is around biology and medicine. I've seen a number of cases, including Daniela actually, where Claude diagnosed a medical problem that a bunch of fancy doctors had missed. And on the biology side, the models are starting to get surprisingly good at tasks like drug design or computational chemistry. And as someone who used to be a biologist, I look at it and I'm like, 'Wow, that's hard. You need a lot of training to do that.' And Claude is getting good at it. That's one area where I think we're going to get a hell of a lot of benefit. That's the positive for AI. We're going to get these huge, enormous benefits. Life is going to get better. The quality of human experience is going to get better. A century of scientific progress and a century of progress in what it's like to be human. Go back to 1900, think of all the problems we had in the 1900s — all the reasons people died prematurely, all the problems they had to suffer, all the material deprivation that we don't have to deal with today. Then think of another hundred years of that. I really believe this century of scientific and medical progress, if we can get through this — and I think we will, I'm increasingly optimistic — we're going to have a much, much better world.

I do. I have not gotten to the point where I actually allow text directly written by Claude in, because I have such a specific style that I'm a little picky about it. But I basically use Claude to help me brainstorm, to help me think through the themes, to help me with references I could use. So it plays a supportive role. I don't know how far we are from Claude being able to write better than me. We're not quite there yet. But I think certainly it's coming. I love writing too, and I feel like writing helps you struggle through ideas. There is a lot of critical thinking involved in that. Do we lose that if we let Claude do it for us? I'm a little worried about that. And in fact, that's half the reason I write myself. It certainly is for external audiences — many people read what I write — but it is just as much to clarify my own thinking so that I know what to do next and to create a common reference point across me and others. I think we're still grappling with the question of how exactly do we use AI in a way that preserves those benefits. I think the thing I'm doing now does that, where I use Claude for research and to help organize my own thoughts. I think if we just used it end to end — like 'write an essay about the risks of AI' — first of all, it wouldn't write the things that I think, but also I would exactly lose that benefit. As the models get better, I think probably we'll use them much more directly in the writing and yet still preserve those benefits. But I think it's going to be a subtle thing. It won't be all one thing. We'll have to figure it out over time.

I've always said — and if you go back to those original clips, they always get cut out of context to three seconds — but the real statement was always, 'I don't know what's going to happen, but this is an order of magnitude for how crazy things could be.' Also, I always talk about all the things we can do in response to this. I've talked about token tax and working with enterprises to adjust people. I'm a little skeptical of retraining programs, but we should throw them in the mix. Macroeconomic policy, even from the beginning, I always talked about solutions. But somehow there's this tendency in human psychology to clip the three seconds of 'doom is coming.' So my message is definitely not 'doom is coming.' My message is this is something we should see coming, that we're worried about, and that we need to actually respond to positively. I don't know exactly, but I'm still pretty concerned. I'm still the same order of concern. We are seeing right now that AI is making people more productive. But that's the usual hump. If you go back to the industrial revolution — I wrote about this in 'Adolescence of Technology' — you automate 90% of the job, great, people are 10 times more productive in the other 10% because they're 10 times more leveraged. But eventually it gets close to 100%. Now the sequel to that is, well then you have to find something else for them to do. I don't know about the long run, I'm truly uncertain about that. But I do think there are types of adaptation. One thing I'll talk about is software engineers within Anthropic. We're going through this transition right now where AI makes the software engineers more productive, even though AI writes all the code or almost all the code. But still it makes people more productive. But we're already starting to see the beginning of there may be some people that it's not making more productive, that it's better for the AI to just do the thing. So that's one side of it. The other side is what do we need more demand for? There's something we call a forward deployed engineer or applied AI solutions architect, where their job is a mix of technical work and talking to customers. There's a lot of demand for that because there's a lot of customers and we're growing very quickly. Now, does every person who is in the pure software engineering role fit into this other? No, it's not perfect, it's not one-to-one. That gives you a flavor that there's going to be a hell of a lot of disruption, but things will also adjust. Which wins out? I don't know. But the reason it's important to warn about it is that that's how we can respond. That's how we can make policy, both within Anthropic and macroeconomically for the whole world. We want to put out carefully considered thoughts. We don't want to say things that people don't believe we'll actually do. We don't want to say things that are half baked. We want to think carefully about what should actually be done about these problems.

No one knows for sure, because the economy is unpredictable. It's the same as the stock market — they're these decentralized processes that you don't really know ahead of time what are the pieces of the job that people are still going to be able to do. But what I would say broadly is that anywhere you have these entry level white collar jobs — whether it's banking, finance — there's going to be a lot of potential for AI to first make people more productive. But then there's going to be a wholesale AI can do the job, and then we're going to have to think about what it is that people can do. And I think we need to plan about that ahead of time. We're already doing it. When we talk to enterprise customers, we see choices that they face. They face the choice of 'should I save costs?' which often means hiring less people — basically do the same thing with less resources — or 'should we do more things with the same amount of resources?' And we always, when we can, try to push them to doing more with the same amount of resources, because that means hire the same number of people or maybe even more people, but just do new things. Pushing them towards the positive sum. The thing that we have going for us here is the pie is going to expand a lot. And so because the pie is going to expand a lot, there are probably going to be places where people can go. It's just a matter of finding them fast enough. It's the size of the disruption. It's going to be big. And that's what I'm warning people about. But we have to solve that matching problem.

Yeah, no, this is the outcome we want to prevent. This is absolutely the outcome we want to prevent. I think there are a few places, none of them are guaranteed. We're not sure. But there's the physical world. Things that are in the physical world — yes, there's a robotics revolution as well, but it's a lot slower than what's happening in AI. People always talk about building data centers, but when processing information of any type becomes a lot easier, maybe the restriction is going to be things in the physical world. So we need a lot more people to build, manufacture things in the physical world. Anything that's human centered, I think that's going to be a big deal. I hear all these stories about AI found something that my doctor couldn't find, but people really want to talk to other humans, particularly over important things. Maybe AI can do better customer service, but nevertheless people want to talk to other humans. So I think there are areas where human interaction remains crucial.

Yeah, so what I would say is, look, the world changes. My view of this technology, when I see Russia invading Ukraine, when I see the risk of China invading Taiwan, it worries me that we have a kind of resurgent authoritarian bloc that is very aggressive and that we need to defend ourselves. That is something that I have believed for a while now and continue to believe. That's why across both administrations, I may not agree with every policy of either administration, but that's why we've generally been supportive of this. We don't want a world where China and Russia can analyze all the intelligence with AI, can use AI for attacking Taiwan and Ukraine, and we can't defend them. So that's why we worked with them. We certainly don't do it for the money. It's a huge pain. Even putting aside the lawfare, it's a huge pain to get up on government networks for not that much money. So we did it because we cared about it. But similarly, because we did it because we cared about it, there need to be limitations on the use of the technology. The formulation I used in the Adolescence of Technology is: we should use this technology in every way except the ways that undermine our own values. Our red lines of mass surveillance and fully autonomous weapons are things that I believe undermine our values. It's not worth democracies winning if democracies do those things. That's the balance that I see. That's the stand we took, and it explains both why we were the first to work with the Department of Defense and why there were some things we wouldn't do when others were willing. I think you need to pick a stand and stand your ground. This idea of companies that seesaw from "we won't do anything with the government" to "we're doing absolutely everything with the government" — I don't get it. You should pick your principles and stick with them.

That's right.

We don't work with ICE either through Palantir or anyone else. We don't work with CBP. I don't believe we work in Gaza. We're very careful about scoping our engagements to things we believe in.

I don't think there's any winning. This isn't a fight for a private company. Anthropic isn't trying to win or lose. This is more a debate about what the proper use of AI by the government is. AI is an emerging new technology. We don't understand the ways in which it is reliable or unreliable, promotes or undermines our values. So I thought it was important to establish a precedent on some use cases we think are good — which is most of them — and some we are concerned about. As I've said, you can only do so much with a contract. Someone else can sign a contract that doesn't respect your same red lines. But it has raised awareness for the issue, and we have serious bipartisan efforts in Congress attempting to ban some things we're concerned about and to set guardrails. I don't want to talk about this as a fight, but that's kind of winning the effort to get our country to think more carefully about appropriate use of this technology.

I've been called worse things all the time. People can call me or Anthropic whatever they want. The two things that matter are: we're successful as a company, and we stand up for our values. In some ways, my life is really easy because when those are the two things you're trying to do, it's simple. You always know where you stand.

I think there are two things here. There is the ability of the United States to be more effective militarily. I am supportive of that ability. I think having that ability be stronger doesn't cause wars; it deters wars. Basically, you're asking if I believe in this country. Do I want this country to be a more powerful actor rather than a less powerful actor on the world stage? I do. I'm a patriot. There is a separate question: are there particular policies the US government is engaged in that I might support or not? Obviously, I support some and not others. It's not up to me. If we provide a technology, it's not up to us to say you can do this military operation and not that one. The Department of Defense made this point and we agree. I might privately believe one operation makes sense and another is a bad idea, but we are not going to deny the technology. You have to leave policy in the hands of military decision makers. What you can do is assert high level boundaries that prevent use cases inconsistent with our values and promote those that encourage them. That's how we think about it.

We don't have access to exactly how these models were used. Obviously, mistakes in warfare are really terrible. If that doesn't make clear why we have to stand up for use cases we don't support, I don't know what does. We were willing to risk the future of our company to limit how these models are used. What you're talking about is a use case that doesn't even violate our red lines. We're worried there will be a hundred times as much with use cases that do violate our red lines. I think overall, the use of these models is appropriate and good on net. But military decision makers make terrible mistakes even at the best of times, and I don't know if we're in the best of times. There are several things we can talk about: making red lines that prevent uses more likely to lead to those problems. If we had allowed fully autonomous weapons, which almost every other company now has, this would be different. What we've seen here is Claude assists, but a human makes the final call. A human made that final call, not Claude. Imagine if, not Claude because we haven't allowed it, but someone else's AI model made the decision and the human never saw it. That's what we were standing up for. Also, I don't think procurement is the right way to do it. We need to make sure that military decision makers don't make these mistakes and operate reliably. That's of concern to me as a citizen and a supplier of the technology. The government uses Microsoft Excel a lot. If I said you can use Excel for this military operation but not that one, you can't realistically do that. Hopefully that gives you a sense of how we think about it.

Look, what I'm going to say is, and I don't know this relies on classified knowledge I don't have, but the principle we have established and was obeyed here is that a human makes the final decision. I don't know what role Claude or any other AI had, but if this isn't an illustration of why that principle is so important, I don't know what is.

I would say on balance, it is more likely to stop it. But if we have no limits on how it is used, then I think it could be more likely to cause it. You've seen Doctor Strangelove. The premise was a doomsday device that automatically fires nuclear weapons when it thinks nuclear weapons are being fired at it. What could go wrong? I get to this lethal fully autonomous weapons thing. I think the way conflicts happen is that the two sides misunderstand each other. Without proper oversight of this technology, those kinds of accidents are more likely. Now, if AI is used appropriately, even just for intelligence collection, if we can predict an invasion of Taiwan or a new movement in Ukraine, our adversaries will think twice before conducting some invasion if we know everything they're doing. Superior intelligence can deter conflict. Superior ability to respond can deter conflict. I continue to be a believer in these things.

I think the thing that surprised me most was the models had been climbing in their ability to find vulnerabilities and, importantly, turn those vulnerabilities into exploits. People only talk about vulnerabilities, not turning them into exploits. It was quite good at that. So the things that surprised me were this huge jump. It was a particularly large jump without us really prompting them at all. Some of the early companies we gave this to said things like, "This is a super weapon. You should have to own a gun license to use it. Please don't release this." The demand to keep it restricted was coming from the companies we gave it to, who were finding so many critical vulnerabilities and exploitability that they asked us not to release it. To be clear, the goal is not to keep this locked up forever. We are gradually opening it up to a wider set of people, and eventually we believe we should release Mythos to a general audience, but with strong cyber safeguards. A concern is today’s cyber safeguards, which we did release on Opus 4.7, can be jailbroken. We’re concerned about some other companies who think this is a sufficient defense. It works sometimes, but we know these classifiers can be jailbroken or got around. Our own testing and assessment of other companies’ defenses suggests they are not strong yet. That’s what we’re waiting for: getting defenses to the point where we have confidence in them.

The claim that it could be replicated with open source models is incredibly false. The idea is Mythos looks across the whole code base and finds something. Some guy on Twitter said if you point an open source model at exactly the line of code Mythos finds, then it finds the same issue. That isn't the prompt. That isn't the question. That is not the same thing. The ultimate test is we go to companies and open source repos. We found 271 new vulnerabilities in Firefox. We found many thousands within private companies that haven't fixed them yet. No one found those 271 vulnerabilities with the previous model. So the actual workflow of what works in practice, as opposed to finding the exact line Mythos found, is completely different.

We have suffered enormously commercially from not releasing this model. This model has incredibly accelerated research within Anthropic and production of next models. It would do the same in the outside world if we released it. This has hurt us enormously commercially.

The reason we are giving Mythos to defenders before attackers is to patch all the bugs. I don't know, as models get better, there may be more bugs to find, but they are finite. You have a surface with a limited number of holes. You patch all the holes, and the surface becomes very hard to attack. The code itself is written with powerful models, so it becomes very hard to find flaws in or break into. On the other side of this, hopefully six months or a year from now, we have a much more secure internet ecosystem than before. We are trying to get to that world and doing our best to open up Mythos to new cyber defenders. We have been talking to the government, and we are very respectful of their recommendations. They are slowing the pace of opening it up because of counterintelligence risk. I think that’s sensible. All serious people understand there are real trade-offs here. We see sniping from people on Twitter and other AI companies. Look at what they say versus what they do. They are not serious people. They are not seriously engaging with the trade-offs. I have customers calling every day wanting access to Mythos. Countries are calling wanting access. The US government and my security team say wait, there is risk. I’m not saying one side is right; both have valid points. But there is a real challenge we need to face together as a society, not accuse things of cheap marketing or use cheap marketing to counter-position. It shows an incredible lack of gravitas and maturity. We need to face this moment together.

Throughout the entire history of Anthropic, it has been trade-offs. In some ideal world, before releasing the first chatbot, you could spend years studying every possible thing that could go wrong. We did delay the initial release of Claude, but for a few months. Everything is a trade-off. The extreme ends of the spectrum are completely insane. Now that we are in a commercially leading position, Daniela and I are doing all we can to move the dial further towards being careful. That is what the Mythos release was about. It is very hard to do something like that if you are not the leading player. I think you are going to see more things like that.

That is a very serious question, and I share those concerns. I don’t think the government should outright take us over. But let me back up. Every previous powerful technology in history was either built by the government or originated with it: nuclear weapons, the internet, GPS, cell phones. AI is the first technology built in the private sector where government has not had a serious role and is coming late. I think that is a dangerous and unstable situation. It is not the situation I would have chosen. There is no alternative because the technology is possible to build, our adversaries are building it, and it has economic value. The issue is the government not doing it, not the private sector doing it. I think we need to think about checks and balances on power. Checks and balances on AI companies. We have the Long Term Benefit Trust, a body that can appoint and remove the majority of board members. It has the power to fire me. We are introducing elements of public governance where you are accountable to someone who doesn’t just have stock in the company. That structure will continue no matter what. We encourage other companies to have similar structures. On the government side, we need checks and balances. There are efforts in Congress to enact red lines. The legislative and judicial branches need to exert themselves because I am scared of companies having this technology, but also of government having it. Companies need to provide checks on government, and government needs to provide checks on companies. We need basic regulation, like required pre-release testing and auditing of models. It is funny how some in Silicon Valley started with a position against any transparency or export controls, saying it will kill innovation. Then as soon as they see real danger, they call for nationalization. They yo-yo from extreme anti-regulatory to completely communist views. We need a more sensible, moderate approach. We have favored that all along because we understand the power of this technology. We are not panicking or denying. We see the smooth exponential and respond appropriately.

We always try to work together with whoever we can in government. We have a simple approach: we have principles, we follow them, and we hope folks on the other side are reasonable. Honestly, the government has taken Mythos very seriously. We have had good conversations with Treasury Secretary Bessent and Chief of Staff Susie Wiles. I think they really understand the nature of the risks. Mythos has helped them feel more concretely where these risks are. As with any administration, some parts we get along with very well and understand; others are harder to get along with. That is normal. We just try to navigate it as best we can.

One thing we have seen is a premium on how intelligent the models are. We very rarely see that people prefer less intelligent models. There is a thriving ecosystem of easier problems that don’t need frontier models. But it’s an exponential. It is possible that far-from-frontier models have economic value comparable to what we saw in 2023 and 2024. But again, we have this 10x a year growth. What is on the frontier is always much larger than what is away from it. People used to building products in the previous era don’t quite understand this. As someone who hadn’t run a company before and never thought about the previous product era, I feel like an outsider. People’s instincts are wrong. They have product heuristics, and the 10x per year exponential breaks that. Intelligence is such a huge factor that it outweighs everything else. We see over and over that value is found on the frontier. What I do worry about with laggard models is the risks. Mythos class cyber capabilities, 12 months from now, we will have much better ones. But Mythos class capabilities may be available for anyone to download. Hopefully we will have patched everything before then. I don’t think there is anything we can do to stop it, but it is a serious concern.

Not really. I worked there for a year. Maybe I learned more about speech recognition. The only thing that concerned me was part of how we got all the speech recognition data. They said ominously, "We don't care about privacy in China, so we have all this data." But aside from my geopolitical worries, the things that most worried me about China are what happened to the Uyghurs, suppression of criticism even in the US, and what happened with Hong Kong. The CCP can reach into the US business network and suppress criticism. It is an authoritarian state, a high tech authoritarian state. When I see how that combines with AI, you get a dystopia like 1984 or worse. My focus is on trying to prevent that. I think we have an opportunity for AI to be a pro-democracy technology that makes people freer and delivers on the promise of equal justice for all. Or it could go the other way. Which way it goes depends on the actions of AI companies, the government, and all of us. We have a responsibility here.

I don’t think it’s a moment in time. I think it’s a continuous process. We are already seeing it in some ways where AI can suggest architectures for the next AI. A year ago, we were seeing a 10 to 15% increase in total factor productivity due to AI. That is probably up to 20 or 30% now. It might be doubling, as with all things on an exponential. There is no moment where AI improves itself out of control or becomes unsafe. What we have is an accelerating exponential, and at each point we have to assess whether to slow down or put more controls on the technology. More and more of that will be required. The Rosetta Stone to all of this is the smooth exponential. There is an object lesson in people against all AI regulation who, seeing one thing, wanted to nationalize. There is an object lesson in people who dismissed the power of AI and then said it is improving itself out of control. Yo-yoing between extremes is incredibly unhelpful. The right, wise response is to say we are not going to panic. Our countermeasures will smoothly ratchet up with the power of the technology. If you see someone having this crazy yo-yo reaction, it’s a sign they were caught by surprise and are not serious.

That is correct.

The figure I most identified with was Leo Szilard, who first had the idea of a chain reaction. My view is we are not going to get through this with larger than life personalities or figures trying to be at the center of everything. There needs to be a balance of power. There are many powerful actors with interests here, and the only way it ends well is with checks and balances everywhere. In some ways, I see Oppenheimer as a failure case, as what should not happen.

I hope not. My view is that the actions we have taken lower that probability rather than increase it. That probability comes from the straightforward recipe of the technology, the existence of many countries and companies. If the void isn’t filled, it’s a dilemma. We are trying to act to lower that probability. I think we lower it a lot more than we raise it. But the inherent property of this technology is that it is unpredictable. We try to build something, test it a lot before release. Models released today are not dangerous, at least not outside cyber. We iterate and learn. There are a zillion defense mechanisms within the company to reduce risk as much as we can, but it will never be zero. Suppose there are airline companies, and you make one that is 10 times safer. You can’t guarantee it will never crash.

That’s right. 25% is too high. We are trying to make that probability much lower. That is the goal.

When any company starts out, especially given the behavior of Silicon Valley over the last few years, starting from a position of distrust is pretty rational. Silicon Valley has lost a lot of the world’s trust and has to re-earn it. The message we are trying to send is that we are different, and that has to be earned through what we do. You can agree or disagree, but we stood up for our values. Mythos has really hampered us commercially. Before that, there were smaller things. We put our money where our mouth is on China: we cut off access to models when no one told us to. That cost us several hundred million dollars when that was a significant fraction of our revenue. The delay of Claude 2. We have a long history. We are not perfect and we make mistakes. But I would ask people to look at the overall history and decide what hypothesis is most consistent with it. I think the hypothesis is that we are genuinely trying to do the right thing. We are imperfect; organizations are always dysfunctional. Many foot faults and things go wrong, but at basis we have an honest and earnest picture of how to do the right thing and are trying to execute on it.

Hopefully.

That’s one surprising thing I didn’t understand about the CEO job: how often you have to wear makeup. That was not on my bingo card.