Transcription
[music] Please welcome Andrew Ross Sorkin and his guest, CEO and co-founder of Anthropic, Dario Amodei. [music]
>> Good afternoon, everybody. [applause] I hope you guys all had a great lunch. [music] We have a huge afternoon starting with Dario Amodei here. He is one of the most consequential people here, uh, in the world of artificial intelligence. He's the co-founder and CEO of Anthropic, of course, known for its Claude model. It's one of the fastest-growing technology companies in history and uniquely backed now by all three giants, uh, tech giants: Amazon, Microsoft, and Google. That's new, at least one of them. Uh, and by the way, he has been at this, uh, really longer than most. Uh, he worked at BYU, then Google, was an early employee at OpenAI, where he led the development of, uh, ChatGPT 2 and 3. Uh, and the reason that we wanted to speak with him this year, more than anything else, is because he has singularly been perhaps the most outspoken and candid person about AI when it comes to the way he's been thinking about jobs and job losses, and selling chips to China, and politics in our country, and where all of this goes. So, welcome to you. Thank you for being here.
>> Thank you for having me.
>> Um, we've got a lot to talk about. Um, including, by the way, are we in an AI bubble? But I promise you, we will get there. I, I, I'll start here, though, um, which is, I mentioned you were a research scientist, uh, back at BYU, uh, 2014. And if I had sat with you then and said, "We're going to sit together in 2025 talking about AI," you would have told me, "What? What would have been your expectation for what would have happened?"
>> So, I'll tell you what I am surprised by and what, and what I'm not. Um, I'm not surprised by the economic impacts of the technology, the, the value that it's creating. You know, the fact that, you know, I walk by any billboard in New York and, you know, it's, it's, it's kind of everything about...
>> 2014, that this would, this would be, this would be real by now.
>> ...in, in some form, that this would be real, that it would be central to the economy, that it would be central to national security, that it would be central to scientific research. I, you know, I, I, I don't think I imagined that I would be leading one of the companies in the space, right? I think that, I think that would have been, you know, very surprising to me. I didn't think of that as, as kind of, as kind of my role at the time. And the exact way in which things happened, all of the, you know, all the kind of strange lingo we've developed around language models, all of the kind of financialization of it, you know, if you think about, if you think about the implications of the models becoming being as smart and as powerful as they are and scaling in the way that they are, in the way that me and some of my colleagues predicted, it all makes sense. But I, I don't think I would have derived it from first principles.
>> Okay. Well, then let's go straight to the question I said I'd ask at the beginning, um, because maybe this is the place. Thinking about where this goes, and the fact that I didn't think you would say, by the way, that you thought that this is where we were going to be in, in 2025, because I think even people back then thought this would be a much longer road. But if you're right, do you look at the amount of economic, uh, muscle that's being put into this industry right now? I mean, it really does represent potentially almost all of the growth in the United States GDP right now, literally.
>> Yes.
>> That we are in some form of a bubble? Are we overspending? Does the math of all of this make sense?
>> So, this is really complicated, and I want to separate out the technological side of it from the economic side of it. On the technological side of it, I feel really solid. I think I'm one of the most bullish people around, and I think it pencils out. On the economic side, you know, I, I, I have my concerns where, even if the technology is, is, you know, really powerful and fulfills all its promises, I think there may be players in the ecosystem who, if they just make a timing error, if they just get it off by a little bit, bad things could happen. So, let me go through both of them.
On the technological side, the, the reason that I'm not, honestly, by the pure technology, not that surprised, is myself and some of the people who eventually became my, my, my co-founders, we were the first to document the scaling laws of AI, which is, you put more compute, you put more data into AI with small modifications. We've seen these things like reasoning models and test-time compute. They're all tiny little tweaks. And I've been watching that trend for the last 12 years or so since I joined the field. And the thing that is most striking about all of it is, as you train these models in this very simple way, you know, with a few simple modifications, they get better and better at every task under the sun. They get better at coding. They get better at doing science. They get better at bio-medicine. They get better at the law. They get better at finance. They get better at, at, you know, materials and manufacturing. And, and that's just, that's just a listing of all the sources of, of, of value in our economy.
If I just take Anthropic itself, which, because we work so much in the enterprise side, I think we're a good barometer, maybe a purer barometer than the others, which kind of filter through consumers, which kind of have their habits and their use cases. We look at our revenue, it's grown 10x a year, every year for the last three years. Zero to $100 million in 2023, $100 million to a billion in 2024, a billion to, it's going to land somewhere between eight and 10 billion at the end of this year. Will it continue? I don't know. But the technology is driving there, and the economic value is coming with it. I, it will, you know, that trend is going to slow down for sure, but it's still going to be really fast. And so I have this confidence that the, the, ultimately, the economic value is the, ultimately, the economic value is going to be there.
Um, but when you, but let's just go to this, because there are companies that are spending a hundred billion dollars a year, more, you're going to be spending 50. You look at what Sam Altman, who was here last year, plans to be spending. These are extraordinary numbers, and this is all a bet, a big bet that this is going to scale in this way. And, and my question is, is there a real way to pencil this out, or is this, uh, more of a, a gut feeling at this point?
>> So, let me, let, yeah, so that really gets to the second part of it. And, and I, I, I will describe as transparently as I can. I think there's a real dilemma deriving from uncertainty in how quickly the economic value is going to grow and the lag times on building the data centers that, that drives this. So, I think there's genuine uncertainty. There's genuine dilemma, which we, as a company, try to manage as responsibly as we can. And then I think there are some players who, you know, who, who are YOLOing, who, who pull the risk dial too far, and I'm very concerned.
>> Who is YOLOing?
>> So, I, that's a question I'm not going to answer.
>> Um, uh, uh, so on the first one...
>> Put yourself...
>> We'll come back to that.
>> ...put yourself... We won't put, put yourself in, in the position of, of Anthropic. Put yourself in my position. You, you, you've seen this revenue curve that goes up 10x a year for three years. You're like, "Okay, what's going to happen next year?" If I'm really dumb and I extrapolate the pattern, 10x to 100 billion. I don't believe that, just to be clear. I don't, I don't believe that at all, even though it's happened in the last three years, just at this scale. I don't believe it. But that's one of the outer bounds of, you know, the, the outer limits of possibility. If I go in and I'm like, "This enterprise, and that enterprise, and this use case, and this is our go-to-market motion," if I try and do it, you know, bottom-up, then, you know, maybe it's, maybe it's 20 or 30, or something like that. So, there is, what, what I've been calling internally, this cone of uncertainty, where I don't know if a year from now it's going to be 20 billion, or, you know, it's going to be 50, or, or like, you know, it's, it's very uncertain. I try to plan in a conservative way, so I plan for the lower side of it. But, but that is very disconcerting. And you add to that the idea that building the data centers has a long lag time. It's like a year or two. So, I have to decide now, literally now, or in some cases, a few months ago, how, how much compute I need to buy in, you know, early 2024 to serve the models in early 2027, when I get to that revenue amount. And there's two coupled dangers. One is that if I, if I don't buy enough compute, I won't be able to serve all the customers I want. I'll have to turn them away and send them to, to, to my competitors. If I buy too much compute, of course, I might not get enough revenue to pay for that compute. And, and, and, you know, and, and, you know, in the extreme case, there's kind of the risk of going bankrupt. And how, how much buffer there is in that, in that cone, um, it's basically determined by my margins. If I have 80% margins, I can buy $20 billion of compute, and it could serve $100 billion of revenue. But because the cone is so wide, it's hard to avoid, you know, making a mistake on one, on one side, on one side or the other.
Now, we've been a relatively responsible company. Um, and I think because we focus on enterprise, I think we have a better business model. I think we have better margins. I think we're being responsible about it. But again, let's say you have a different business model. Let's say you have a consumer business model. You know, you're, you're, you're, you're kind of the source of your revenue isn't as good, your margins aren't certain. And let's say you're a person who just kind of like constitutionally just wants to YOLO things, or just likes big numbers. Um, uh, then that, then you may turn that dial pretty far. And so I think there is a real underlying risk. Whenever there's uncertainty, there's a risk of overextension. We all face it. I face it. All the other companies face it. There's, there's an inherent risk when, when the timing of the economic value is, is uncertain. There's an inherent risk of, of underreacting or overextension. And because the companies are competing with each other, and frankly, we genuinely need to compete with, you know, our authoritarian adversaries, um, uh, there's, there's kind of a lot of pressure to push things. So I think there's some amount of irreducible risk here, and I absolutely don't want to deny this. But at the same time, is that I think there are some players who are not managing that risk well, who are taking unwise risks.
>> Let me ask you about that. With, maybe you'll mention who that is. I think we all know who that is.
>> Um.
>> You have said that you're going to break even, uh, this is privately, at least, to your investors by 2028, even with the spending plans. I think, uh, Sam Altman, who you worked with, and for, uh, says that he's going to do it by 2030.
>> [snorts]
>> I'll use his math, not yours. He would have to go from a $74 billion loss in the course of two years to being profitable two years later. Does that make sense to you?
>> So, look, I, I don't, I don't know the internal financials of any other company. I can't, I can't say anything about what the economics of any other, of, of, of any other company is. I will just go back to our own calculation and the cone of uncertainty where, where, where, where, where, where we basically say, we want to buy enough compute that we're confident, you know, even in the 10th percentile scenario, like, might be in a bad position, but like, we think we can pay for it. There's, there's some end of the curve where, you know, things go so badly that, you know, we can't pay, that, that, there's always a tail risk. It is not zero, but we're trying to manage that risk well, while also buying an amount, an amount of compute that allows us to be competitive with the other players. We're very efficient in training. We're very efficient in inference. We have good margins. I think, I, you know, I think, I think we can manage it. I think the odds are on our side.
>> What are we supposed to think of what I think people now describe as circular deals? Back in the day, we called this vendor financing. Yes. But in this context, you have a situation where Nvidia, in particular, um, but others as well, effectively have been taking stakes in companies, um, and invariably, those companies are using some of that money, one way or the other, considering money is fungible, and going to buy Nvidia chips.
>> Yes.
>> So, I mean, we've done some of these deals, not, not at the same scale as, as some other players. Um, but, uh, we've done some of these deals. And I, I can just, I just, I just, I can just kind of walk you through, not a specific deal, because I'm not going to go into details, but, but, but kind of like a stylized, what these deals often look like and kind of why they can make sense. So, if you want to buy a gigawatt of compute, right? Buying the chips and building the chips, building everything, that costs, you know, roughly, let's say, $50 billion of, of capital expense to, to, kind of, to, kind of fund that. And you can think of that as useful lifetime, as people argue about it, but maybe it's five years. So that's $10 billion dollars for, basically, for five years. And, and so if you're a company, you know, you're, you're, you know, you're a company that's making, you know, $8, $10 billion of revenue. You think that's growing, you don't know how fast it's growing, you have to make the decision right now, and you don't have $50 billion. You don't have $50 billion on you. So, you know, a thing you can do, a deal you can make with a large player who has an incentive to, to, to do this, because they're the one selling the chips or providing the cloud, is they'll say, "Okay, you know, I'll give you, I don't know, 20% of it. I'll invest $10 billion." So that lets you pay for the first year. And then for, and then for the other, you can kind of pay as you go, because I know you don't have $50 billion now. But, but, you know, looking at how things are, are, are growing, that isn't a crazy bet. We're already almost, we're already almost at, at, we're already almost at the $10 billion of revenue. So, it, it kind of, it, you know, it takes a year to build a data center. It's financed for a year. So you're, you're, you know, you're basically saying, "I need to get $10 billion of, of revenue per year, two years from now." Um, so I don't think there's anything wrong with that. One player has capital and has an interest because they're selling the, because they're selling the, you know, they're selling the chips, and the other player is pretty confident they'll have, they'll have the revenue at the right time, but they don't have $50 billion at hand. So, I don't think there's anything inappropriate about that in principle. Now, if you start stacking these where they get to huge amounts of money, and you're saying, "By, by 2027 or 2028, I need to make $200 billion a year," then, yeah, you can, you can, you can overextend yourself, of course.
>> But I think is one of the key questions behind the math of this entire industry is what's called the depreciation schedule, if you will, for these chips. And there seems to be a big debate about it. Meaning, um, when you buy a new chip, is that chip going to work for you effectively for three or four years, or is it going to work for you for six or seven or eight years, or even 10 years? And depending on where you think the math really lands, all of this pencils out or really doesn't. What do you think the schedule is?
>> Look, from, from our point of view, we make very conservative assumptions here. Um, I don't think there's a particular depreciation schedule, right? When new chips come out, the issue isn't the lifetime of the chips. Chips, chips keep working for a long time. The issue is new chips come out that are faster and cheaper and...
>> ...you may need them if your competitors have them.
>> Yeah. Yeah. And so, and so kind of the value of old chips goes down somewhat. In fact, that that can happen, you know, a year after you buy the chips, because, you know, now there are multiple companies, TPUs, GPUs, coming out with, coming out with new chips. So, I think the way we think about it is, we take into account that, you know, the old chips are going to be less valuable as, as time goes on. And, you know, we assume very aggressive, you know, a kind of continuation of the, of the chip efficiency curve. And again, I can only speak for Anthropic. Like, we make conservative assumptions here, and, and we think we're going to be, we think we're going to be okay in basically almost all worlds. Can't be literally all worlds, but we think we're going to be okay in almost all worlds. I can't speak for other companies. Again, I can imagine that there may be other players out there who are, you know, who are deluding themselves and, and, you know, making making assumptions that are very far inflected on the optimistic.
>> So, we're clear. There's only two of you who are who are not...
>> I look, I just, I don't know who you're talking about. Okay. I, I just have no idea.
>> Um, let me ask you this, um, about the models themselves and how you see the competition. So, one of the things that's happened literally in the last week, and there has been a complete sort of, uh, meltdown in the Valley over what's happening here. And Sundar Pichai was also here last year, and it appears, at least, that his new model has gotten a lot of, uh, people excited about what he's doing. And that Google, which I used to think from the beginning, given all of the data they have, uh, should have been sort of the, the winner by default. And you have a memo that went out from Sam Altman saying that, "Code Red, everyone's got to get back to their desk to figure out what the, what the next thing is to break the, to get to the next place." How do you stack rank right now where these models are, and how important do you think it is in any given moment?
>> Yeah. So, this is one of the cases where I'm just very grateful that Anthropic is taking a different path. Right? On one hand,
>> The path being the enterprise, doing...
>> The, the path being the enterprise, right? Both of the players that you, that you mentioned, right? Both of these other two players are, you know, primarily focused on the consumer. They try and do some enterprise work, but they're, they're, and, and they're fighting in consumer, that is, you know, that is, that is the reason for kind of the Code Red, the intense fighting, right? It's, you know, Google has a search monopoly that they're trying to defend, and the center of what OpenAI is doing is in, is in consumers. So, those, those two are fighting it out. For both of them, serving businesses is secondary. And so, what we found is, over time, we've optimized our models more and more for the needs of businesses. The one that has gone the fastest has been coding. Um, you know, I think that has really, um, you know, moved forward the quickest. But we're starting to go beyond that to finance, biomed, retail, you know, energy, manu, all kind of, all of that. Um, and so, and so, and so what we find is that these model wars, as much as our models are like, really good, like, you know, the one we released last week, Opus 4.5, is hands down, I think almost everyone thinks the best model for coding. So, I think that's, it's very important that we continue to have this, this model superiority. But there's a way in which we're kind of going in a different direction, or, or on a different dimension. And so, we, we kind of have to worry less about this back and forth. We have a little bit of a privileged position where we can just keep growing and just keep developing our models, and we don't have to do any Code Reds.
>> But what is the, the moat around any of these businesses? Um, and, and when I say that, I assume if Google is as successful as it wants to be, or OpenAI, or any, or, or Meta, or anybody else who's involved in this, that they think that one day, if we ever get to AGI, that all these models effectively will be able to do what either, you know, any of them do. And whether it's, you know, is there a moat? Is it the persistent memory? So, I use ChatGPT for certain things. It knows, uh, me now because it's, I've been, you know, asking it different questions. Is it, or do you think people just switch back and forth, whoever's got the, the latest thing?
>> Yeah, so, I, look, I can only speak to the enterprise side. What I will say is, is it, it is surprising how different the personality and capabilities of the models are if you're building for businesses versus if you're building for, if you're building for consumers. Um, you just focus on different things. You focus less on engagement. You focus more on coding, high intellectual activities, scientific ability. Um, and, you know, I don't think it's true that if we got AGI, they would all converge to the same place. Is everyone in this audience converge to the same place? Is everyone in this audience a copy of everyone else because we're all agents? No, we're, we're all specialized. Specialization exists, you know, in, in, it exists alongside general intelligence. And then I think there's all the standard enterprise stuff as well, which is that, um, you know, companies build relationships with you. They get used to using certain models. Uh, and, and we're starting to see that, like, even our, even our API business, which is basically just selling the raw model, you wouldn't think that would be very sticky. Companies have great difficulty switching from one model to another because they have downstream customers who use the model, and they like the current model, and you prompt and interact with the models in different ways, and they have different personalities. It's actually quite hard to switch. So, so I think there really is a durable business here.
>> Um, one quick AGI question. It's a science question, which is, do you think just the way transformers work today and just compute power alone from a scalability sense, that that is what will get to AGI, or do you think there's some other ingredient, and maybe there's a technical question, but I'm trying to keep it very, very easy, that has to be included in this that gets you to some place where this stuff is actually going to really think on its own?
>> No, I think, I think scaling is going to get us there, again, with small, every once in a while, there will be a small modification, you know, so small you may not even read about it. It's just something going on in the lab. I have been watching these scaling laws for, for 10 years.
>> So, what's your, what's your, what's your timeline now?
>> There's no one particular point, right? That this is what I've said over. I've never liked these terms, AGI, artificial super intelligence. I don't know what it means. There's just an exponential. Just like we had an exponential with Moore's Law, chips getting faster and faster until they could, you know, do any, you know, simple calculation, you know, faster than, faster, faster than any human. I think the models are just going to get more and more capable at everything. We release a new model, it gets better at coding, it gets better at science. You know, now models are routinely winning, you know, math, high school math olympiads, are moving on to college math olympiads, are starting to do new, new, new mathematics for the first time. I've had internal people at Anthropic say, "I don't write any code anymore. I don't open up an editor and, and, and write code. I just let Claude Code write the first draft, and all I do is edit it." Um, we had never reached that point before. And the drumbeat is just going to continue. And, and I, I, I, I don't think there's any privilege point around, there's no point at which the models start to do something different. What, what we're going to see in the future is just like we've seen in the past, except more so. The models are just going to get more and more intellectually capable. And, you know, the, the revenue is going to keep adding zeros.
>> Let me ask a couple policy questions. Um, you have been outspoken. We, we spoke to, uh, the President of Taiwan earlier today. Um, you have been outspoken about the idea that we should not be selling Nvidia chips, for example, the, the most advanced chips, to China. By the way, it's interesting that you now have a partnership with Nvidia. Uh, Jensen Huang, who's also been here, was not so happy with you when you made those comments. Do you have a new view on that?
>> My, my view hasn't changed. So, I, I definitely will say, and this has, this has always been the case, you know, I have an enormous amount of respect for Jensen and for Nvidia. You know, Jensen is an immigrant who came to the, the US with nothing. He built the most, you know, the most valuable company in the world. This, this isn't personal. This is a policy question. This is a question of how to defend, how best to defend our national security. And there, my view hasn't changed. It's, it's, it's the following, which is, if we go back to my picture of the models getting smarter and smarter as we continue to improve them. A phrase I've used in an essay I wrote a year and a half ago was, "Eventually, the models are going to get to the point where they look like a country of geniuses in a data center." And so, once we get to that point, think about what that country of geniuses in the data center can do, and which existing country on Earth it's plopped down in. Um, if it's plopped down in an authoritarian country, I feel like they can outsmart us in every way: intelligence, defense, economic value, R&D, you, you know. And I worry that they'll be able to oppress their own people, that they'll be able to have a perfect surveillance state. And so, I have always felt that we need to have the advantage here, and this is a national security issue. Right? Some people are saying this is an e, they, they, they say this is an economic issue. That it's an analogy to like the internet or 5G. We need to diffuse our stack, like we need to beat Huawei in, in, you know, in telecommunications. You know, I don't, I don't see it that way. I think we're building a growing and singular capability that has singular national security implications, and democracies need to get there first. It is absolutely, it is absolutely an imperative. If we, if we sell these chips to China, that just makes it more likely they will get there first. It's common sense.
>> Okay. But do you think that could happen here? So, we had Alex Karp here earlier, and there has been lots of worries about surveillance here. Talk about a, in a democracy. Yes. Right.
>> What, what is your concern there? Um, I should say, by the way, there was a period of time where you called the president, this was before he was the president, a feudal warlord at one point. Um, so how do you think about the president today, America today, and this idea that AI and surveillance could come together?
>> Yeah. So, so look, I, I want to, I want to say over and over again that, you know, that I, you know, I think that, that the tendency to drag this down into being about specific personalities and, and, you know, and, and, and specific fights, I think is not helpful here. We should really think, we should really think at a, at a policy level here, and it's not about one administration. It's not about another administration. We should have principles here. And, and I think the, the principle I would give that, that I think is, is very important, is actually, it can happen anywhere, right? Uh, you know, a, we should, we should worry about concentration of power in democracies, not as much as we worry about it in authoritarian, um, uh, states. But, you know, we, we need to make sure that the technology is governed in, in, you know, in a way that that allows people to participate, that gives people basic rights. And so the formulation I've always given when I think about how to apply these models for national security is, I think we should aggressively use them in, in every possible way, except in the ways that would make us more like our authoritarian adversaries. Right? We need to beat them, but we need to not do the things that would cause us to become them. Um, that is the one constraint we should observe.
>> Okay. Let me ask you a separate question, and maybe you'll say it's a fight, but you can take it out of, out of the, the person, if, if you want. Which is, you have been very vocal about your concerns about the chip issue, but also what could happen to jobs, regulating, regulating this technology so it doesn't do bad things, or other things like that. Uh, David Sachs, uh, who works as the AI Czar at the White House, said this about you: "He says that Anthropic is running a sophisticated regulatory capture strategy based on fear-mongering. It is principally responsible for the state regulatory frenzy that is damaging the startup ecosystem."
>> Again, I don't think this should be about specific individuals, right? This isn't about any particular administration. This is about, this is about policy questions. I mean, you know, going all the way back to 20, 2016, I've written papers about AI, right? Before I even had a company, right? Before there, there even, you know, could have been a plan around anything like regulatory capture. And, and by the way, almost all the AI regulation that we've supported has, has exemptions for small players, right? The main, um, AI bill we supported, SB 53, um, you know, doesn't even apply at all to startups with under $500 million in revenue. So, we've been very careful about this. And, you know, if there's one point, again, I want to say again, I think people should focus on the policy. You can throw out these accusations, and they don't match reality at all. They don't match the reality of the bills we've actually supported. They don't match the reality. There are these two worlds right now, which is, by the way, you know, Andreessen Horowitz and others are, are have, have one super PAC, you guys are building another super PAC, >> to approach regulation of this industry completely differently. And the question is, why? What do you see that they don't?
>> So, so I, you know, again, I, I want to keep this at the policy level. What, what, how I see this technology, I, I am concerned, and I can understand where folks are coming from. But I am concerned that there are some who see this technology as analogous to previous technological revolutions, as being like the internet, as being like telecommunications, where, yes, there are some issues, but, you know, the market will figure it out. Which I think was maybe a more reasonable view in, in these previous technological revolutions. I think those who are closest to AI don't feel this way. If you pull the actual researchers who, who work on AI, not investors who invest in some AI application companies, not general tech commentators who, you know, think they know something about AI, but the actual people who are building the technology, they're excited about the potential, but they're also worried. They're worried about the national security risks. They're worried about alignment of the models. They're worried about the economic impacts of the models. And for example, the idea that we would put a moratorium on all, all kind of regulation, or all state regulation without a federal, you know, with, without, without a federal framework, uh, for 10 years, which, you know, was attempted in the summer, and I think it was just attempted this last week again, and it failed because it was very unpopular, because even the average person understands that this is a new and, and powerful technology. And so, you know, I think we, I, I am more than you, I, I, I am maybe the most optimistic about the upsides, right? I wrote this whole essay, "Machines of Loving Grace," where I said AI is going to, you know, extend, might even extend the human lifespan to 150, 10 years after we get the, the country of geniuses in the data center, because we'll have a virtual biologist that can make discoveries much faster than we can. That, that, you know, it could drive, you know, economic growth to, you know, five or five or 10%. I'm incredibly optimistic about the technology, frankly, much more optimistic than some of the people who describe themselves as boosters of the technology. Um, but, but nothing that powerful doesn't have a significant number of downsides, and, and we, as a society, as a polity, need to think ahead about those downsides. Saying that for 10 years, we won't regulate that technology, it's like, it's like saying, "I'm driving a car, I'm going to rip out the steering wheel because I don't need to steer for 10 years."
>> Okay. So, here's the, the question about the downside then. One of the downsides, beyond hacking and everything else that I know you worry about, is jobs. You spoke about it in 60 Minutes recently, but I want to know, not just that you, you think that there's a good chance, and I don't want to put words in your mouth, that it could be, you know, half of all entry-level jobs get lost. I want to know what you think should be done about it.
>> Absolutely. So, you know, I, I, I think at the end of the day, I warn about these things not to be a prophet of doom, but because warning about them is the first step towards solving them. And if we don't warn about them, then we'll just, we'll, you know, we'll just kind of blindly walk into the landmine. It'll blow us up. Um, if we warn about them, if we see the landmine, we can walk around it and we can avoid it. So, I have been thinking a lot about these ideas. I've been thinking about them inside Anthropic, where we, where, you know, Claude is starting to write a lot of our code, and we're thinking about how the jobs change. So, I think there's several levels of it that maybe go from short-term to long-term, or, or just kind of requiring more and more of the resources of society to happen.
I think some of it can happen in the private sector and even in our relationships with customers. Every customer we work with has the following trade-off, and it's not either or. They can, um, you know, they can increase efficiency by basically having AI do what, what humans used to do. And there's plenty of that. Things like, you know, insurance claims processing, or know your customer, whole workflows that can just be done end-to-end via AI. And I think we'll need a lot less humans for them. It, it will increase efficiency. It'll save cost to do the same thing for lower cost and much less people needed. But you can also do things where you can create a lot of new value. And even in cases where AI does 90% of the job, not 100%, but 90%, the humans can be 10 times more leveraged. And sometimes you need 10 times more of them to do 100 times what you did before, because it's so efficient and valuable. And so encouraging companies to do as much of the second relative to the first. We know they're going to do the first. We're not trying to stop them from doing the first, but if they can do more of the second than the first, maybe more jobs can be created than than...
>> Does that mean we need government incentives? Is that a...
>> So, so again, that's, that's, that's level one. Level two is the involvement of the government. I don't see retraining programs as a panacea, but I think we're going to need to do some form of that. Companies are going to do it. Companies are going to have to work with governments to do it. And I, but I, I do think fiscally, at some point, the government is going to need to step in. I don't know if that's tax policy, but, but this world of fast growth, right? We did this report where we said, even current models, it looks like they will increase productivity by 1.6%, 6% a year. That's almost, you know, almost a doubling of productivity. And the models are getting better and better. So, I think we're going to get up to five, 5% a year, maybe 10% a year. That's a big pie. That's a big pie that we can, we can give out to the people who are, who are not such fortunate beneficiaries, right? If the wealth concentrates, there, there, there really is a big pie here. So, I think the government is going to need to have some role here. That's level two. And I think level three is, over the long run, the structure of a society that has built powerful AI is just going to have to be different. If we, if we go back to John Maynard Keynes, right? Economic Possibilities for Our Grandchildren, right? He invented this idea of technological unemployment. He suggested that maybe his grandchildren would only have to work 15 or 20 hours a week, right? That's like a different way of structuring the society. Some people will always want to work, you know, as hard as possible. Um, uh, you know, there, there'll always be segments of society who, who want to do that. But, you know, can, can, can, can, can, can, can, can, can, can, can, can, can, can, can we have a world where work doesn't, for many people, doesn't need to have the centrality that it does? That, that people find their, their locus of meaning elsewhere? Or work is about, is about different things? It's more about fulfillment than it is about economic, economic survival. Um, that there's so many possibilities here. I think society is flexible, and, and society can, I'm not suggesting anything top-down. I think society needs to restructure itself. We all need to figure out how to operate in the, in the post, in the post-AGI age. So, I think those three levels will go from fast and, and easy for it, for, you know, individual companies to do, to requiring a lot of consensus and, and, and very slow to do. But over the years, we're, we're going to need to do all three of these things.
>> Uh, Dario, I hope you come back so we can have a conversation, uh, as we have to do all of those things to figure out what comes next. Want to thank you for a fantastic conversation.
>> Thank you, Andrew. Thank you. Thank you so much.