WHEN: Today, Tuesday, April 23, 2024
WHERE: CNBC’s “Squawk Box”
Following is the unofficial transcript of a CNBC exclusive interview with Anthropic Co-Founder & CEO Dario Amodei on CNBC’s “Squawk Box” (M-F, 6AM-9AM ET) today, Tuesday, April 23. Following are links to video on CNBC.com: https://www.cnbc.com/video/2024/04/23/anthropic-ceo-dario-amodei-on-ai-tech-race-were-not-beholden-to-one-particular-company.html and https://www.cnbc.com/video/2024/04/23/anthropic-ceo-dario-amodei-on-claude-3-model-ai-arms-race-and-big-tech-partnerships.html.
All references must be sourced to CNBC.
ANDREW ROSS SORKIN: Welcome back to “Squawk Box.” I want to get into the latest A.I. technology. In a rare and exclusive interview, I spoke with Anthropic’s co-founder and CEO, Dario Amodei. Anthropic is the maker of Claude 3. It is arguably now one of the most powerful AI models out there, if not the most powerful. I started out by asking him about how Anthropic differentiates itself from the other AI models.
DARIO AMODEI: You know, all of these models have some similarities and some differences, right? The similarities is that all of them are, you know, trained on large amounts of data, with very large models that use large amounts of compute. The basic formula for that is similar. But what’s different is what you do after, when you train the model how to be conversational, how to talk to humans, what areas it’s good at, how it answers questions. And so, we had a team called Claude Character, that was focused on the personality of Claude as a model. And that team focused on, you know, again, how to make the model for more warm, more human, more engaging. Of course, it also focused on these questions of safety and reliability. And so, we put a lot of effort into making sure that on one hand, the model doesn’t do genuinely bad things, doesn’t give false information, but that it does that while, you know, not refusing, you know, harmless queries, right? So, drawing that line between those things well. And we believe that that’s one of the things that Anthropic as a company specializes in, doing a very good job of that.
SORKIN: Where do you think you are on the journey of the various sort of levels of these models toward AGI? Meaning if you’re at 3 now, what does 4 and 5 and 6 and 7 look like? And how do you stack rank, given yourself now, against how you think about these other players?
AMODEI: I don’t think that AGI is a fully coherent concept. So, you know, 10 years ago, when almost no one was working on generalized models, and everyone was working on specialized models, I used to talk about AGI, because it was so different from what everyone was doing. But now I think we’re in a different place, where everyone is building models that in a way are general, right? There’s — they may be worse than humans at a bunch of things, but they try to do everything. And some of them, they’re very good at. So I don’t think about AGI, but I think what is real is the models are getting better and better at more tasks that more and more humans do.
SORKIN: So when Elon Musk comes out and says that he thinks that you know, in a year from now, that, you know, AI could be smarter than, you know, an individual, and in, you know, several years from now, could be smarter than all of the individuals collectively, you think what?
AMODEI: Yeah, so I mean, I think — I think, you know, another way to restate what I said before is that, you know, smarter isn’t a single thing, right? Right now, there’s some things the models are much smarter than humans at, right? Like — you know, like if I talked to Claude 3, it knows a lot more about, you know, the history of cricket or the history of samurai in Japan. But if I ask it to do some very simple things like, you know, like plan my day or something like that, it doesn’t even necessarily — it’s not even hooked up to the right things to necessarily do that. And there are other limitations, as well. So I think what is correct about that is the models are going to get smarter and smarter, and we will reach a point, probably relatively soon, where the models can do many tasks better than humans can. But there will always be strengths. There will always be weaknesses. And so, I think the picture we’re going to see —
SORKIN: Right.
AMODEI: — is much more — is much more complex than, you know, than some kind of one-line thing might indicate.
SORKIN: Right. When people talk about the next generation of this stuff and they talk about the limiting factors, they talk about processing power. They talk about the energy to produce that processing power. And they talk about data. How much data is there out there? People now have talked about using synthetic data to train on. Where — where are you going to get the data to train on ultimately?
AMODEI: Yes. So, one of the things is what you hit on, right? We’re working on several methods for developing synthetic data. So these are ideas where you can take real data that’s present in the model and have the model interact with real data in some way to produce additional or different data.
SORKIN: But it all depends on the original data being good data, I assume. And could we ever get to the situation where if we’re all just training off of synthetic data, that data becomes bad, it becomes a “garbage in, garbage out” situation?
AMODEI: Yeah, I mean, this is what a lot of the research that’s focused on synthetic data is focused on, right? So, if you — if you don’t do this well, you don’t get much more than you started with. But it actually is possible, by injecting very small amounts of new information to get more than you started with. If we go back to systems of eight years ago, so if you remember AlphaGo, which is the system that was used to play Go, note that the model there just trains against itself, with nothing other than the rules of Go to adjudicate. And those little rules of Go, that little additional piece of information is enough to take the model from no ability at all, to smarter than the best human at Go. And so, if you do it right, with just a little bit of additional information, you can — I think it may be possible to get an infinite data generation engine.
SORKIN: You’re in an unusual position where you have major partnerships and investments from two of the major tech players in Amazon and in Google. Both of them also compete with you in that Google is creating Gemini and Amazon is creating something called Olympus, which they haven’t released yet. How do you think about this frenemy, friend — what is this?
AMODEI: Yes. So, I don’t know, maybe if we — if we back up a little bit like, why do these cloud partnerships exist in the first place? Ultimately, they exist because of economic complementarity, right? Anthropic is a company that focuses on producing AI. We don’t focus on provisioning chips. And we don’t focus on, you know, provisioning cloud and, you know, building data centers of provisioning cloud instances to, you know, to serve the models on. And so, these other companies that we’re partners with, the cloud providers, provide something that’s complementary to us in an economic sense.
SORKIN: What do you make of the argument that big tech is trying to hoover up and control this new technology, whether Microsoft is partnering with OpenAI, or Google and Amazon are partnering with you?
AMODEI: Yeah. So, you know, I can’t, I can’t speak to other companies and other partnerships. But I think the mere fact that we’re partnering with two different companies to put our models on the cloud, and, you know, various other aspects of the partnership, you know, shows that, you know, we’re not beholden to one particular company, right? We’re an independent actor, neither of these companies have board seats — board seats — board seats at Anthropic. You know, we offer our models on our own — on our own first party. So, you know, we have Google as a — Google as a cloud where we offer our models, Amazon is a cloud we offer our models, and we also offer them first party. So, there are a number of ways in which we’re simply an independent company that operates differently. Again, the fact that these clouds provide something that’s complimentary to what we provide makes these partnerships make sense.
SORKIN: Do you think that if the regulatory environment was different, that these big tech companies would want to ultimately buy you? We talked to Andy Jassy about whether he would want to buy you, and he said, go look at the iRobot deal that was bought by the government. Meaning, that in this environment, even if you wanted to buy somebody, you couldn’t.
AMODEI: Yes. So, I can’t speak to whether another company would want to buy us. What I — what I can say is that, you know, right now, Anthropic is more than happy operating as an independent company. You know, I think — I think our goal is to, you know, pursue our — pursue our safety priorities and the values of our company. You know, those values have been set up in a very specific way and our view is that’s best served by being — that’s served by being an independent company.
SORKIN: So another player, by the way, is Meta. They just introduced their new Llama 3. They say it is the most advanced open-source model. And some of the reviewers suggest it is pretty close, it is getting very close to a Claude 3 or to a ChatGPT 4. What do you make of that?
AMODEI: Yeah. So, you know, there are a number of kind of questions about open-source models. I mean, some of these are kind of around, you know, safety, where people worry about the safety of open-source models. You know, I actually think that’s kind of not really the right focus. You know, I’ve said today in terms of safety, I’m more worried about large models versus small models. And in particular, models that are so large and so powerful that they’ve never been built before. So, in terms of safety, I’m worried about the models that we’re going to build in 2025 or 2026.
SORKIN: How much do you worry about election interference, given that we’re in the year 2024?
AMODEI: Yes, yes. So, this has been, you know, this has been something that we thought about a great deal. So, first of all, we — you know, as relates to Anthropic itself, we’ve banned the use of Claude for, you know, electioneering, political campaigning. And we have —
SORKIN: And you think there’s no way to get around it?
AMODEI: Well, I was — I was going to say that. You know, with these models, you never reduce things to literally zero. But we have a — we have a trust and safety, whole trust and safety infrastructure. We’ve also done red-teaming in line with White House commitments that a number of companies have made to test for use of the models for disinformation, influence operations. The point I would make is, look, these models are ultimately statistical systems. You’ll never quite get it down to zero. There’ll always be some way that someone can, every once in a while, you know, produce something that’s bad. But our goal is to prevent or at least detect things like scaled influence operations, right? Things that operate at scale and are serious threats.
SORKIN: Within the context of elections, open-source, do you worry that that’s going to be used and misused? I mean, when you see the Llama 3 comes out and you can get it — you’re not on Instagram —
AMODEI: I’m not on Instagram.
SORKIN: But it’s on Instagram right now. You can — you can ask it whatever you like. Does that scare you?
AMODEI: Yeah. I mean, on one hand, you know, I would repeat the point I made before that, you know, I think, currently, the level of models that we have, and, you know, that includes the open-source models and the closed source models, are not really at a level of, you know, capability where I kind of seriously worry about them destabilizing the world. But I think, you know, they’re starting to get closer to there, right? And so, I think there are potential — you know, there are potential concerns that, again, come from open-source and closed source models, for — you know, for interfering with elections or generating disinformation at scale.
SORKIN: Right.
AMODEI: So, you know, I do have — I do have some concern in that area. And I think the only difference between the two is, you know, with a model — with a model like Claude, we can — we can monitor for it.
SORKIN: How much do you worry these models become commoditized? If OpenAI’s got one and now they have an inflection with Microsoft on one end, you, we’ll say that Amazon has its own, eventually, you say that Google has its own eventually, Meta sounds like they want to have their own eventually. I don’t know if you believe in — a believer in X.AI? Do you think that Elon Musk is a competitor of yours?
AMODEI: I can’t comment on that in particular.
SORKIN: OK. But you see where I’m going. There’s going to be a lot of these models. The sort of large language models that can do a lot of things or — and do they become just a feature set, a sort of feature of everything that we touch, and does that commoditize it?
AMODEI: Yeah, so a couple of points on that. First, I would say, you know, I think I’ve said this elsewhere, but you know, the models of today cost about $100 million, plus or minus, factor of 2 — 2 or 3. I think we’re going to see models trained in the next year are going to be about $1 billion. And then 2025, 2026, we’re going to go to $5 billion or $10 billion. And I think there’s a chance it may go beyond that to $100 billion.
SORKIN: Wow.
AMODEI: So, I think actually the number of players that can train models of those size — and again, we’re talking about models that are — you know, that are professional level across a wide variety of tasks and can even surpass human level, I think the number of companies that are going to be able to train models at that scale is going to be relatively small to start with. So that’s one factor fighting commoditization. The other factor is this thing that we talked about earlier, which is the models having different personalities and skills, right? Would you describe humans as commoditized? I mean, it’s a bit of a funny question, right? Humans have different personalities, different skills. They then — you know, they then amplify that by educating themselves in different directions and, you know, walking different life paths, gaining different career experience. This thing I mentioned about, you know, you’re going to have this model that like, you know, that, you know, talks to the biochemists for years, and that model that becomes an expert in the law or a national security. I think that force is going to lead to different model providers specializing in different things, even as the base model they made is the same, right? We as humans, we all have — our brains are all basically designed the same, but we’re very different from one another, and I think models will be the same.
SORKIN: What do you hope the future of your company looks like? We’re here at the New York Stock Exchange. I think right over there is where they ring the bell when someone goes public and oftentimes, a CEO is up there with the gavel. Is that something that’s in your future?
AMODEI: Well, so, you know, I think Anthropic is very happy being a private — private company at this time. You know, I think there are a number of — there are a number of benefits of this. You know, we have a kind of unique — we have a unique social mission, we have a unique structure. You know, who’s — you know, it’s hard to say, you know, what we’ll do five or ten years from now.
SORKIN: Right.
AMODEI: But, you know, I think — you know, at least on its — at the present time, in its present way of thinking about things, you know, Anthropic is very happy being a private company.
SORKIN: Dario clearly one to watch. I think the man in the middle of the AI game in really big way, and given the partnerships with Google and Amazon, and sort of their independence unlike OpenAI’s, I think it’s going to be very interesting to watch what ultimately happens to this company.
