Mario, thanks for doing it. So, a little backstory. We had spent some time talking to Dario in particular about the jobs and what it could do. And we'd had a lot of off-the-record conversations just about what he was sort of describing as this white collar bloodbath that I've been hearing from other folks in the AI community, from other companies, but no one would say it publicly. So, we called him and said, 'Listen, it's really important that society knows this, hears this from you.' And to his credit, he said, 'Let's do it. I'll go on the record. I'll say exactly what I said to you and what I think.' In that conversation, it was chilling, right? You said, 'Listen, you could see half of white collar jobs wiped away in 1 to 5 years. Unemployment spiked to 10 to 20% because of this technology regardless of the good it might do beyond that.' Talk more about that and why did you go public and why won't others?
Yeah, so a few points. I think the reason that Anthropic decided to go public about this is I had been saying these things before but it had mostly been in limited context, in podcasts restricted to the technology industry. And when I would get on a plane, when I would be in the airport, when I would be in some other city than San Francisco or California, I'd walk by a bunch of people and say, 'I don't think we've communicated to you very accurately about what this technology is capable of and where it's going, the ways it could benefit you and the ways it could be a threat to you if you don't react to it in the right way.' And at some point that just felt wrong. I was talking to a lot of CEOs who would say in private, 'This is what we're planning. Some of them are our customers. We have these plans to deploy this technology and it's going to have a labor impact.' So we really felt that we needed to say something. The first step towards solving these problems is being honest with the population that these problems exist. To say more about where I see things and why I see things this way, I would say there's kind of two modes that we have at Anthropic. One is looking at what is happening now. We've done a lot of that with the economic index. We recently released a state-by-state economic index which lets everyone look at how people are using the models in real time in different states, in different geographic locations for different tasks, whether they're automating or augmenting. So that's one mode and we've even seen some research externally showing that there's a job impact there already. For example, the work from Eric Bernolson and others showing that white collar entry level jobs have contracted already by 13%, which is a significant fraction of what I predicted. But what I'm really worried about is where the technology is going. I think there's a disconnect where people will sometimes say, 'Oh, you're worried about what AI is going to do to jobs, but AI can't do this, AI can't do that.' Well, we're talking about today's AI. The technology is moving quickly. I'm worried about advances in the technology and diffusion of that technology through society. And that's where I get to the one to five years. As with most things, when an exponential is moving very quickly, you can't be sure. This could happen faster than I imagine. This could happen slower than I imagine. Or something very different could happen. But I think it is likely enough to happen that we felt there was a need to warn the world about it and to speak honestly and in candid terms about it.
That was 2 months ago. We were talking off stage in some ways since then. Both of you were remarking how the technology despite some of the conventional wisdom is actually getting much better than we even thought 2 months ago. Are you more worried today than you were 2 months ago?
So we studied this inside Anthropic. We talked to 130 engineers. We interviewed many of them and talked about their experience using the technology and their jobs have changed in the last year radically. Many of them now do two or three times as much work. And rather than writing code, they're managing fleets of AI systems. And in these interviews, they said, 'My job has completely changed. I'm now having to rethink what my role here at Anthropic looks like.' Now, obviously, we're a fast growing company. They will have jobs. But we are changing people's jobs in real time inside the company because the technology has moved so quickly. And what happens inside the AI companies will happen to all of the other businesses that use this AI technology in the coming years. The vast majority of code that is used to support Claude and to design the next Claude is now written by Claude. It's the vast majority of it within Anthropic and other fast-moving companies. The same is true. I don't know that it's fully diffused out into the world yet, but this is already happening.
And just to put a finer point on that, I think that's what you're seeing. What we're seeing is that companies have a fiduciary responsibility to drive value. They can use technology for productivity. They're going to do it and I think you're already seeing it in the unemployment numbers. When you on the hill, there's some momentum we're talking about to potentially legislate it. You're more bullish than I am. I'm deeply skeptical Trump would sign anything into law that would regulate AI in this term. But put aside what can be done. What are the two things in order that if you were running Congress or you're king of the US that you would do today to specifically address this?
I would say the first thing would be something around helping people adapt to AI technology. I don't want to think of this as a bromide. People have tried retraining programs and there are real limits to what they can do. There are real limits to helping people train and adapt, but it's better than nothing and it's where we got to start. There is a world where I've seen startups like Lovable or Replit who are customers of Anthropic that allow people who are not software engineers to build software products and to start businesses with those products. So if we can move more people in that direction, again, I don't think it's fully going to solve the problem here. I don't think it's fully going to stop the spike. This is too broad. It's too big. It's too deep. But it can be a piece of the solution. So that's number one. Number two, I would say, and this is more controversial, I suspect at the end of this that the government is going to need to step in, especially during a period of transition, and provide for people for some of the disruption. One thing I've suggested is maybe you might want to tax the AI companies. I don't know how that would fly in Congress today, but I think that is actually a serious proposal. If you look at the amount of wealth, the increase in the pie that's coming from the AI companies. If you look at Anthropic's revenue, it's growing 10x a year. It's now in the mid to high single-digit billions. If it keeps growing this way, this is going to be an unprecedented amount of wealth creation. It's not going to disincentivize our growth.
What time period? What time period will we need some version of that? Whether it's a tax or it's a guaranteed minimum wage for people, how quickly will this hit where you think Congress is going to have to do this?
We're technology optimists. We think this technology is moving far faster than most people suspect. And when people say AI is slowing down or it's overhyped, we just look at we measure the properties of the system and it's right on schedule to make really powerful systems arrive easily during the next 5 years. What does that mean? You need some kind of policy response at the scale of disruption we expect within 5 years. And we expect that along with the ideas Dario talked about, on route to that we need more transparency out of the AI companies. We and the other AI companies are already affecting society in large ways, and we need to be transparent about how we're measuring our systems, how we're securing our systems, and the economic data about how our systems are being used so economists can tie that to the actual broader economy and give policy makers the data they need.
You guys have been also very transparent about things that are happening in testing and when you guys have released some stuff about some really weird stuff, right? You had like one testing pattern where the machine basically went through someone's email and attempted to blackmail them. You've had this is our famous marketing tactic where they're trying to basically lie to you so you don't shut them down because they now are smarter than us. Why shouldn't that scare the hell out of me?
Yeah. So one answer is it should. But we should put it in proper context. Those are all things that happen to the model in a testing scenario. You can think of it as I put the car on a super icy road. I mess with its tires a bit and the car crashes. Does it mean that will necessarily happen in the wild? No. Does it mean that there are limits to the resilience of the car such that if you push things, if it was in an extreme enough situation and if the requirements of the car started getting tighter or you tried to amp up the performance of the car by designing a new car, then this could be an issue in the real world. So I see it not so much as a look at the present but a look at the future of where things could be going. And that's why we've advocated so much for transparency. When we run these tests, when we show them to the world, we're basically looking ahead in time a year or two to what could happen in the real world if we don't train our models to mitigate these risks. And so when we've advocated for transparency legislation, when we suggested a federal transparency legislation against the state AI moratorium, 10-year state AI moratorium, when we endorsed SB 53 in California, we're just looking for every AI company to have the level of transparency that we have shown. We've seen behaviors in the wild where models misbehave, where they're sick of fanatic, where you suggest doing something that's really not a good idea like committing suicide and the models just go ahead and do that. We want transparency on all of these kinds of behaviors so that we can look ahead and so that we can mitigate. This is still an evolving art. It's still an evolving science and so we see transparency as the key here.
When you're talking about the technology moving faster than people realize, take us behind the hood. Use your car thing. Take us under the hood. What is the scariest or wildest thing you've seen AI do that we're not aware of?
So I have definitely seen, as we're training new models, trying to design new Claude, the process of doing that is like you have a giant cluster with thousands of chips and there are these problems you have to solve across the scale of the cluster. And we've had cases where there was a problem that an engineer was working on for days or a week and we just fed the whole environment into Claude and Claude said this is how you solve it. So Claude again is playing this very active role in designing the next Claude. We can't yet fully close the loop. I think it's going to be some time until we can fully close the loop, but the ability to use the models to design the next models and create a positive feedback loop, that cycle is not yet going super fast, but it's definitely started. And these days when we make these AI systems, we have to build really complicated tests to see how good they've got because they've got way better than just answering multiple choice tests. The tests are make a computer program that does X. Well, now when we test out the frontier model that we are training, we'll find that it has written a computer program to cheat at the test and persuade us that it's doing better than it is. So it says to itself, 'Aha, they want me to do this, but I figured out because I'm very smart, how to write a computer program that gets me a high score on the test.' And so when we look under the hood, we'll find, oh, we've made a really smart model that's cheating on its test. That's not exactly what we intended here. We see models that are supposed to browse the web and do some task instead opening up a command line or the toolkit and writing code to allow themselves to go around the browser and cheat on the task.
Yeah. Just like a smart kid in high school that annoyed the teacher. You know, on the one hand, I'm like, 'Wow, that's awesome. That's super cool.' The other one is like, 'Do you ever worry you're creating a monster you can't control?'
So we worry a lot about that. That's why we've invested so much in the field of mechanistic interpretability which is looking inside the models in order to understand them. Think of it as doing an MRI on the models. There is some research suggesting that it is possible to detect psychopathy in humans by doing MRI scans. And so we're aiming to do the same thing with the models to determine what their motivations are, how they think in detail so that if they don't think in the right way, we can retrain the models or adjust them to get them to think in a way that is not dangerous to human beings. And so we really believe that the science of shaping and controlling models is nascent, even more nascent than the science of making the models itself. And so when we call for things like transparency, when we react by saying no, there shouldn't be a 10-year moratorium on any regulation of these models, it's from a feeling of we don't fully understand these things we're creating. And we need to both do technical work and we need some help from society, from legislative mechanisms to have some basic understanding shared among industry so that the relevant decision makers understand what we're seeing and what we're not seeing.
Nvidia, there's some reports out today that maybe China's backing off on buying chips from Nvidia. I don't think you're certain how much of that's posture. I'm not certain how much of it is. But you think it's nuts that we're allowing Nvidia to sell chips to China. Explain why.
I think it's completely nuts. So whatever the dangers of the technology, whatever guard rails are needed, I think it's also very important that we defeat China in this technology. When I talk about AI that's capable of doing all economically valuable labor, or that goes to the point where it's a country of geniuses in a data center, the national security implications of that are absolutely staggering. This could control the fate of nations. This could control the future of freedom and democracy. And if we look at all the ingredients, China has many of them better than we do. They're better at building energy. They're better at building data centers. They have a thriving app ecosystem. They're catching up on the models. Chips are the one place where they are behind. And so if we give them those chips, they will be able to get ahead of us. We've seen with DeepSeek, DeepSeek released their model R1 back in January. They aimed to release an R2 that was much better in May. That has been delayed four months. And the model that they did release, V3.1, was considered a disappointment. It was considered a minor update and they themselves have said, as have many US publications, that the reason for the delay was the embargo on chips. Now some people say they'll make their own chips through Huawei. We've also export controlled semiconductor manufacturing equipment and without that it's going to be many years until they can produce the relevant yields. We saw articles in SemiAnalysis and elsewhere saying that this year they can only produce a couple hundred, maybe a few hundred Huawei chips. The US is producing tens of millions. It's going to take them years to catch up. And in that time, we can get ahead of China and maintain a national security advantage. It may be the only advantage we have because we're so far behind on things like munitions, ship building. This is the only advantage we have. It is mortgaging our future as a country to sell these chips to China. If we were to do this, I think it could be the single most disastrous national security decision made in this term.
