[0:00:00] Can we start with the kind of rough status of how the industry is going? You and your rivals, how close are we now to artificial general intelligence? Or set the stage for us first, and then we can talk about the morality of it all. Yeah, yeah, thanks for having me. I mean, you know, I've never liked the artificial general intelligence or super intelligence, not because I don't think AI is very powerful, right? I'm not a skeptic. I'm actually extreme in terms of my views of how powerful the technology is going to be.
[0:00:29] But it's the wrong model for thinking about it, that there will be some one point where we build something completely different. What we see here, actually, what we've seen for the last 10 years, maybe even 15, and me and Anthropix co-founders were among the first to document it, is this very smooth exponential process.
[0:00:50] Just like in the 90s, you saw Moore's law that the amount of computing power would double every 12 months or every 18 months. We have a Moore's law like law, except it's for intelligence itself, for the cognitive ability of the model across many tasks. And so what we are seeing is that that cognitive ability is, depending on how you measure it, doubling every four to 12 months.
[0:01:17] And so we're just climbing the ladder of cognitive ability. And where we are now is to take coding as one example that's been exploding really fast over the last year or two. I have engineers. In fact, the team that leads one of our lead products, Claude Code, which is the way you use our models for coding, he says he hasn't written any code in the last two months.
[0:01:43] It's all Claude. He's edited it, he's looked at it, but it's all been written by Claude. Indeed, we recently released this thing called CoWork, which was bringing Claude code to non-coding tasks. It really seems to have taken off. We wrote it in like a week and a half almost entirely with Claude code. So I think we're entering the world where the junior level software engineers, maybe many of the tasks of the morning.
[0:02:10] Senior level software engineers are starting to be done most of the way by AI systems. Now, that's going to go further. It can be more end-to-end. The projects can be more. But I think we're going to be surprised at how the exponential turns upward. The whole thing about exponentials is it looks like it's going very, very slowly. It speeds up a little bit.
[0:02:31] And then it just zooms past you. And I think we're on the precipice. I think we're a year or two away from it really zooming past us. Sounds very like going bankrupt. But when this happens, in terms of the kind of running order, there's yourself, there's Gemini, there's OpenAI. Are you all together at the moment, roughly? Yeah, I mean, you've got a lot of hype for your most recent version. I think as the field has developed, you know,
[0:02:57] I think it's no longer appropriate to think of it as like, you know, like a scaler, you know, like runners in a race who are at one position. The players have gone in different directions. So some of the other players have gone in a very consumer-oriented direction. And that leads you to make your models, you know, superhumanly engaging or superhumanly good at recommending shopping or ads or various things like that. Anthropic has focused, I think, first and foremost on enterprises.
[0:03:27] Developers, and to the extent we do consumer, we're very focused on productivity and the kind of high value end of the consumer work. And so that gives very different incentives. If you want to get things done, we talked about AGI and superintelligence before. One of the things I would say is that there are superintelligences today, and they're basically large corporations. They're smarter than any human can be at
[0:03:54] Solving the problem of shipping commerce at the lowest possible cost, or making solar panels at the lowest possible cost, or launching rockets at the lowest possible cost. And so there's really high returns to intelligence in this area, which I think incentivizes us to kind of build the right thing. I think it's a business that is more stable than consumer. We don't need ads.
[0:04:23] You know, large numbers of free users. We can just very directly create value, right? There aren't all these weird externalities of like you're prioritizing engagement, you're prioritizing these other things, you're generating all this slop. We just make stuff that people can use. A year ago here, you said the Chinese were catching up a bit. Do you think now the Chinese have fallen behind? You know, I think they never really caught up.
[0:04:50] That much. So of course, there was this huge excitement around DeepSeq. But the truth was a couple of things. One, those models are very optimized for the benchmarks. It's actually very easy to optimize a model for a finite list of benchmarks. When we go out into the world, when we're competing against other companies for enterprise contracts, we see, just honestly and candidly, we see Google and we see OpenAI.
[0:05:20] Every once in a while, we see a couple other U.S. players. I have almost never lost a deal, lost a contract to a Chinese model. But now you have the Trump administration, and I think you've already protested about this, giving high-speed chips and video chips to the Chinese. That's right. That's right. The thing that is holding them back, and they've said it themselves, the CEOs of these companies say it's the embargo on chips that's holding us back. They explicitly say this.
[0:05:49] And now, indeed, there are some policies, and I hope they change their mind, to explicitly send not quite our latest generation of chips, although it was reported that even that was being considered, but the generation of chips that's just one back, that's still extremely powerful. And we are many years ahead of China in terms of our ability to make chips. So I think it would be a big mistake to ship these chips.
[0:06:19] The analogy I thought of, if you think about the incredible national security implications of building models that are essentially cognition, that are essentially intelligence. I've called where we're going with this a country of geniuses in a data center. So imagine 100 million people smarter than any Nobel Prize winner, and it's going to be under the control of one country or another.
[0:06:46] So I think this is crazy. I think it's a bit like, I don't know, like selling nuclear weapons to North Korea and bragging, oh yeah, Boeing made the case. So your friend David Sachs is basically arming the Chinese? No, I wouldn't refer to any particular people, but I would just say that this particular policy, I think, is not well advised.
[0:07:14] Another issue in this, you could be right. All this technology be going in the right direction, but economically, we could be heading towards a bubble. Do you think that's true? Yeah, yeah. So I would separate out the two things. One is the basic exponential. And that's what we've been talking about. That's been kind of the technological direction. And there, just like with Moore's law, you can never be sure that the technology is sort of...
[0:07:43] Fully going to advance. It's going to keep going. It's a fundamentally inductive problem. But now that I've seen well over a decade of it, I would say I am pretty confident that it's going to continue, or at least I'm more confident than I've ever been at any time in the past, that it's going to continue to where the models are basically smarter than humans at almost everything. And I think there's a good chance that even happens in the next year or two. Again, seems far away, but.
[0:08:12] This property of exponentials that they catch you by surprise. And if it's not a year or two, I think it's pretty likely that it's at least less than five years within the 2020s. I think this moment will come in the 2020s. So I'll get into the economic diffusion, which I think is complicated. But I don't want anyone to forget that that
[0:08:38] As long as we have that fundamental technological gain, there's going to be many trillions of revenue, maybe many trillions per company in this area because the economic potential is so great. Now, there's a separate thing, which is that we don't know how fast enterprises, companies, are actually going to be able to use this technology. What the technology is capable of today
[0:09:07] is probably 10 times what the enterprise of the world are able to deploy. And I see it with our customers every day where I'll talk to a CEO and they'll understand and their executive team will understand the power of this technology for automating customer service, for coding, for many other things. But they have a company of tens of thousands of people who are perfectly brilliant people but experts in something that is not AI and have to learn to use AI.
[0:09:36] Different names for it. Change management, enterprise transformation, whatever it's called. The need to do this is very slow. It can take years. And so we have this powerful technology that I have an incredible amount of confidence will generate trillions in revenue. But we don't know exactly when. Plus or minus a few years, we don't know exactly when. And in the meantime, companies have to buy compute to serve all that revenue.
[0:10:05] And you don't want to buy too much because you could financially overextend yourself. And you don't want to buy too little because then you can't serve all the revenue. And so that's where the kind of economic and financial risk comes from. But you are involved in that, building data centers. You're spending vast fortunes there. You've got the deals with NVIDIA. Do you worry that that's going to come back and bite you? So we have some advantages in this.
[0:10:30] There is a stability to enterprise business that there is not to consumer. Consumer is very fickle. Purchasing enterprise predictions are predictable. Also, enterprise tends to have better margins. The margin is basically how much buffer you have between buying too little and buying too much, between what you have to pay for and what you have to support in revenue. So we have a number of advantages here. But I won't deny that there is some inherent risk in
[0:10:59] And in all of the companies to this process. Whenever you have an upside whose amount is uncertain, or even whose timing is uncertain, and you have to make financial decisions somewhat blind to those things. You have to decide early to build out capital, to build out resources a few years in advance. There's going to be some inherent risk that's driven by that uncertainty. And I do think that some companies have probably overbought.
[0:11:29] I'm not able to look at their financials. I just know what they announce, and I'm like, whoa, I wouldn't necessarily have done that. I'm pretty happy with the decisions that we've made, and I can't be sure about others, but it is entirely consistent that this could be the most transformative technology in the history of the world that some companies will do really well, but not every company may. Very quickly on that, as you said, you're on the enterprise side. On the consumer side?
[0:11:58] OpenAI with a big lead. You know, there seems to be a lot of talk that Gemini is now catching up. Is that your impression as well? I mean, you know, we're not in that market, so I don't know about it. I'm asking you as a discussionist. Yeah, you know, the statistics would seem to suggest that. It would seem to suggest that, you know, with their kind of greater distribution, with their greater existing reach, they've been able to make a lot of progress in kind of recent times. And again, this is why, you know, consumer is fickle. This is why I would be careful.
[0:12:28] If I were a consumer company, what was on my balance sheet? Let's jump then to that big question, which we originally asked Claude at the beginning. Will AI do good? Obviously, as you said, it's going to have a positive effect on GDP, perhaps on your wealth too. But what about employment? Last year, you predicted a kind of white-collar bloodbath. You said it would wipe out 50% of entry-level...
[0:12:57] Jobs. That's now four and a half years away. Do you still stand by that? Yeah. So, you know, my view of AI, there's maybe two axes in terms of which to think. There's good things happening versus bad things happening. And there's like AI is a small deal versus AI is a big deal. You know, my view is on the extreme side of AI is a big deal. But I'm kind of in both side, both top quadrants where I think.
[0:13:25] Some really good things will happen, and if we don't act to prevent them, some really bad things will happen. You know, I wouldn't be building this technology if I didn't believe that the good outweighed the bad, and that we could mitigate the bad. That's why I warned about the bad, so that we could... How do you think we mitigate it? Does that mean things like taxation? Yeah, let me just get to what I think the situation is. I think exactly what you said...
[0:13:55] I think we could have this very unusual combination of very fast GDP growth and high unemployment, or at least underemployment, or a lot of low wage jobs, high inequality. I don't think that's a macroeconomic combination we've ever seen before. You think of fast growth, you're like, well, OK, maybe there's inflation. But you're not going to have high unemployment when there's fast growth.
[0:14:23] I think this technology is a bit different because it's extreme in the way it's going to generate value. But also, because it's moving up the cognitive waterline, there's going to be, unfortunately, a whole class of people who are, I think, across a lot of industries going to have a hard time coping. And that's really a problem we need to solve. So to get to the question you just asked, there's a few things Anthropic is doing.
[0:14:50] One is we maintain for almost a year now what we call the Anthropic Economic Index, which is tracking how our models are used in real time. And because we see all these conversations, we can use Claude itself to, in a privacy-preserving manner, look across all the conversations and ask questions like, is someone using this to augment a task, to work together with the model, or to delegate or fully automate a task?
[0:15:18] What are the industries people are using Claude for statistically by distribution? What are the subtasks within those industries? We can get into a lot of these very fine-grained labor department statistics and ask Claude about it. Which states are using Claude more? We can watch the economic diffusion of Claude in real time. And so the reason to do all this is that I don't think you can make good policy if you don't have the right data.
[0:15:45] And I worry that the data that the government produces, as comprehensive as it is, just doesn't move fast enough and isn't detailed enough for this. So that is kind of our first contribution. We are thinking increasingly about this kind of, I don't know if you want to call it retraining or ability to help people adapt. It's partly in line with kind of
[0:16:08] What we do in terms of go-to-market, right? Part of our go-to-market activity is how to help people adapt. But with great respect, I mean, you can do these things, but fundamentally it's a societal problem. You know, government, if you think about it, what you're describing is a perfect storm. You have GDP going up, the wealth of certain people going massively up. Absolutely. And you are saying 50% of entry-level jobs go, people in real distress. You are surely going to see a political...
[0:16:35] Change coming from this. And my personal bet would be, yes, you're going to get people trying to restrain the tech barons. But you would also get people saying, let's have higher taxes. We need to pay for this. Do you think higher taxes are coming? So my view, again, I was going to the kind of voluntary stuff we could do first. But I actually do believe that. I think this one is going to be big enough that at some point, I think everyone's going to come to the realization that
[0:17:04] There needs to be some kind of macroeconomic intervention there. If we look at just the disparities in wealth that we have now, if we look at it as a fraction of GDP, I believe we've kind of exceeded the Gilded Age already. And this is mostly without AI. So I think things are going to go even further. And so, yeah, my guess is this is not even going to be a partisan thing. There's a wealth tax now in California. Are you supportive of that?
[0:17:34] No, I think that's poorly designed. And my worry would be if we don't think about these things in a sober way, then we'll get things that are kind of poorly designed. So that would be my message to the others in the world who are doing well in this boom, is that if we don't think proactively about how to make this revolution work.
[0:18:03] For everyone, we will get these proposals that don't make sense. Just taking it a bit further, people worry about the economic consequences, but there are other ones. Not that long ago, I went to a panel with, I think, a couple of your rivals, a couple of your peers, where they, again, they put forward all the good things AI would do. And then someone said, what about the risks? And one of them rather casually said, well, of course, there's extinction risk, and then moved on to another. And collectively around the room, people took a deep breath.
[0:18:31] And what he was talking about was the possibility of a bad actor getting hold of it or possibility ultimately of things like Claude running things. Where does that fit into your? Yeah, yeah. You know, I've long been worried about that as well. You know, just as I said, you know, I'm on the direction of like, this is a very powerful technology. Therefore, the benefits are extreme. We're going to be able to, you know, do things like seriously really cure cancer, maybe eradicate tropical diseases.
[0:18:59] But on the other side, look, we're building cognitive systems that have their own autonomy. And we really need to think about that. And Anthropic's been founded since the beginning with an intention to think about that. We publish almost every, probably three or four times a month, research on how to control these models, how to make sure.
[0:19:22] That they do what we want them to do, and they don't run out of control. We've pioneered. One of my co-founders, Chris Ola, is arguably the inventor of a field called mechanistic interpretability, which is basically when you look inside the brain, the artificial brain of Claude or another AI model, and try to trace mechanistically why it's doing exactly what it's doing.
[0:19:48] And we've seen things inside the model like in lab environments, sometimes the models will develop the intent to blackmail, the intent to deceive. And this isn't unique to Claude. If anything, this is worse than other models. These are things that, if we don't train the models in the right way, can emerge. But we've pioneered the science of looking inside them so we can diagnose them and prevent the models from happening, intervene.
[0:20:17] And retrain the models so they don't behave in this way. That said, this is something we have to be careful about. We've supported various light touch, transparency focused measures for making sure that all the companies have to talk about the tests that they run. We always disclose our tests. We always
[0:20:42] Basically, we try and test our models to try and really strain them and get them to do the worst things possible in a test environment so that they never do those things in the real world. And we generally think every company should have to run those tests and should have to disclose those tests. You spend quite a long time thinking about this. You write pamphlets about it, about the consequences. And there is an impression that most of your peers are going to head down.
[0:21:10] And they're thinking purely about whether they can just keep ahead of the other. Do you think your industry is kind of immature in that way? I can't speak to what the other players are doing or why they're doing it. I think there are at least some other players in the ecosystem who I think of as responsible, who are at least thinking of things in the right way. I agree there are others who are not. But I would say what I have always tried to do.
[0:21:38] And what Anthropic has always tried to do is to set an example and to try to inspire others to follow the example. We always publish our work on mechanistic interpretability. We always publish the tests we run. And honestly, there are a lot of other companies where their researchers basically say,
[0:21:58] Hey, why can't we do that too? That seems like a responsible thing to do. That seems like the right way to develop the technology. And so you can kind of have an effect on the other players in the ecosystem. One very clear zone where you are different from your peers is that you have not very obviously queued up to kiss the ring of Donald J. Trump. What do you think of the current U.S. president? You show no signs of wanting to see him.
[0:22:24] Look, so I mean, I would really put it in a different way. I don't think being for or against administrations or for or against politicians is the right approach or that Anthropic has anything to say one way or another on those topics. What I would say is that what Anthropic knows is AI. And Anthropic knows very well the policy issues around AI.
[0:22:52] And our approach, which I can't speak for others, but indeed our approach may have been different, is we think through the issues, we try and get a substantive view based on those issues, and then we say what we think. And sometimes we will disagree with the current administration, just as we sometimes disagree with the last administration.
[0:23:18] And sometimes we agree. And it's worth highlighting the areas of agreement, like the energy data center buildouts. We did this health pledge at the White House. There are a number of areas. When we saw the White House's AI action plan last summer, actually, we liked most of that. That was actually a very well-written document. So this is complex. There's not against and for. But again, when it comes to things like
[0:23:48] The chips in China, yeah, we disagree. When it comes to putting a moratorium on state AI regulation, yeah, we disagree because we disagree on the merits. It's not about liking or disliking a particular person. I don't think that kind of thinking is going to get us out of this situation that we're in. We have to think based on substance. Are you going to see him when he comes to Davos tomorrow? I may. That would be an interesting change.
[0:24:18] One final last question. What happens to Anthropic? What happens to it? In what sense? Supposedly a 350 billion implied valuation. Would you consider an IPO this year?
[0:24:30] Look, I mean, we are most focused on just trying to make the best models and just trying to build products on top of those models and just trying to sell those models to enterprises in ways that are useful. There are large capital needs for the field, so that should be considered, but that's where our focus is. That implies that you do need a lot of money, though. I mean, you don't need to ask me to know that this is a capital-intensive deal.
[0:25:00] So an IPO isn't completely out of the question? It's never completely out of the question. Both the IPO and Donald Trump are both possibles in your life. Dario Amadei, thank you very much for talking to us. We now know all about mechanistic interoperability, and we've had a view inside your brain, and it's not a bad one. Thank you very much. Thank you for having me, John.