A Cheeky Pint with Anthropic CEO Dario Amodei

I'm excited to finally learn what is it like to start a company with your sibling?

I don't know why you're asking me that question, because you know.

It's like the models want to learn, the models want to be extraordinarily successful in the market.

Yes, right, in addition to having this learning impulse, the models have this capitalistic impulse.

Sometimes people think of the API business and they say, "Oh, it's not very sticky," or, "It's gonna be commoditized." I run an API business.

I love API businesses. No, no, exactly, exactly.

I think we're gonna be in a world where the models will make mistakes much less often than humans, but they'll be stranger mistakes.

So we need to invent slurring for LLMs. And that should correctly pour it.

Oh wow.

Dario is CEO of Anthropic, one of today's frontier AI labs.

He's gone from being an AI researcher just a few years ago to now running one of the world's fastest growing businesses.

Cheers. Cheers.

So, I'm excited to talk a bit about the Anthropic business.

You studied physics and computational neuroscience.

Yes.

You then worked at Baidu, then Google Brain, then OpenAI, and then started Anthropic.

Yes.

And we'll get into the Anthropic business, but I'm excited to finally learn what is it like to start a company with your sibling?

I could ask the same question of you, but you know, it's almost like there's two things you need to do when you're running a company.

You need to operationally execute and you need to have a good strategy and kind of see the most important thing or the thing that no one else sees.

And so my job is the second, and Daniela's job is the first.

And we're both good at the things that we do.

And so I think it's allowed us each to spend most of our time on the thing that we're best at.

Presumably there's something about the trust side of things as well, where cofounder teams in general, in tech, I guess, and AI as well.

I mean they're unstable pairings and just having someone where you've a long running and deep trust— Yeah, yeah, where you have just total and complete trust.

I mean, I think even beyond that, Anthropic has seven cofounders.

When we founded it, basically, the advice from pretty much everyone was like, "Seven cofounders is a disaster.

The company will fall apart before you know it.

Everyone will be fighting with each other."

There was even more negativity on my decision to give everyone the same amount of equity.

But what we found, and I think it was because, obviously, me and Daniela are siblings, but then all seven of us, some of us knew each other for a long time or had a history of working, not just knowing each other, but working together in the past, and I think that really allowed us to always be on the same page.

And I think especially as the company grows, the idea that you have seven people who really carry the values of the company and project them to a wide set of people, it allows you to scale the company to a much larger size while holding onto the values and the unity that we have.

So I wanna ask about the Anthropic business, because again, it's an incredible story where it was reported recently that you'd blown through $4 billion in ARR, and so it's a lot of discussion correctly about the technology that you're developing, but also, this is just one of the fastest growing businesses in history.

And so I wanna talk a bit about the AI market, and maybe the place to start is, what is everyone doing with AI?

Like, there's coding, there's customer service work, but like, where does all this revenue come from?

Yeah, there's a wide range of things, and it's kind of changed over time.

I would say definitely the application that has grown the fastest, although it's not very far from the only application, we have a wide range of them, is definitely coding.

And my theory on why it's grown so fast, other than that we focused on coding and the models are good at coding, it's actually really a statement about kind of societal diffusion. Which is that if we look at today's AI models, I think in every area, there's a huge overhang in terms of what they could do compared to how they're actually being deployed today.

There's some friction, people at large enterprises are not familiar with the technology.

I look at what a bank does or what an insurance company does, and there's huge potential, even if the model stopped getting better, right?

Even if we stopped building products on top of the model, there's still huge, billion dollar potential in an individual enterprise.

And often, the CEOs of companies that I talk to understand that perfectly well.

But if the company is a 10,000 or a 100,000 person company.

Companies that size, they're set up to operationally do a certain thing a certain way, and it takes time to change them.

But in code, the people who write code are very socially and technically adjacent to the folks who develop AI models.

And so the diffusion is very fast.

They're also the kind of people who are early adopters, who are used to new technology.

And so I think the big growth in code, I would say the biggest cause of that is just that the people doing it and the startups devoted to it are fast adopters who understand the technology super well.

But it's by no means limited to code at all.

If you look at, there are a bunch of companies that do things like tool use.

There are, as you mentioned, customer service.

You know, we work closely with companies like Intercom.

We're starting to see some things on the biology side.

We're working both with pharmaceutical and healthcare companies, and we're working on the side of kind of basic scientific research.

We work with companies like Benchling, for example.

But we also work with some of the very large pharma companies.

There was something done a while back where we worked with Novo Nordisk to write clinical study reports.

So clinical study reports are like, you've done a clinical trial, and then you would kind of write up the results, and it's like, these are the adverse events, these are the statistics.

And the clinical study report normally takes like nine weeks.

Well, Claude could do it in like five minutes, and then it took a human a few days to check it.

And so you can really see the opportunity for acceleration, and as the models get better, you know, they'll reach into the deep research as well.

So, I guess a way to summarize it would be to say that kind of code is out in the lead, but we see a long tail of quite a lot of other stuff, including some some very, very significant use cases.

I think code is maybe an early indicator, like a premonition of what's gonna happen everywhere else.

It's the same exponential. It's just faster.

It's happening faster.

Right, so there are many places where there's significant AI uplift, but engineers are used to adopting...

Like, you think about Hacker News and people arguing over the best tools, people are passionate about adopting new tools.

Like two hours after we released Claude Code, there's some person out there who's tried 10,000 different things with it and plugged it into all the frameworks, and Twitter forms one opinion after two hours and then revises an opinion in two hours.

And you think of the speed of that as compared to the speed that a pharmaceutical company can use it in research, right?

Or that traditional retail company.

And we wanna bring everything— Some of the biggest benefits in the world are touching the physical economy, and we wanna get there, but it just intrinsically does not happen at the same speed.

How do you decide which verticals to do yourself versus which to allow platform...

Like, you have Claude Code, and obviously, there's also platform companies like Windsurf and Cursor and everyone like that.

You launch Claude for Financial Services.

Presumably there are other verticals where you say, "Well, we're not building a tool there."

we have things like Claude for Enterprise, which is not a vertical, but a general play to go with enterprise.

I think the way we like to think about it, I think we think of ourselves as a platform company first.

The analogy here would be maybe clouds or something.

If you think of a really large platform business of, the size we're trying to get to in hopefully small number of years, there are a number of reasons why you would also want to have things that are first-party and some verticals end up being more first-party heavy.

One is when you want to have direct exposure to the users.

The end user gives you some sense of how exactly are they using it, what are they most looking for?

If you're a pure platform and you don't have that direct connection, you can be disadvantaged in various ways.

It's hard to build the best products.

Yeah, it's hard to build the best products.

It may even be hard to know where the model really, really needs to go, right?

People say things like coding, but like, there are many models that seem to be good at coding, but they aren't good in the way that's actually relevant, right?

We've actually managed to make Claude good in a way that's relevant to what people actually use.

I think that's one reason.

Another reason goes back to the large enterprises, where building on an API, sometimes it's more challenging for a more traditional company to do that.

And you need to give them something that's a little bit easier to use, either a kit to help them build things, or you need to give them an app.

Enterprises have also liked Claude Code, and we're gradually developing Claude for Enterprise into, what we call a virtual coworker.

I find it hard to picture Anthropic developing Claude for oil and gas exploration, and why is that?

Why is it that you find it hard to imagine?

Yeah, or I mean, maybe in fact it's the next launch, but...

Yeah, we're not currently working on Claude for oil and gas exploration.

I would draw a distinction between things we just, things we just don't allow, right?

Things that are illegal, or things like that.

And there are a number of use cases that it's like, okay we're a platform, people are gonna do a bunch of things.

We're just not passionate about it.

But like, we're not passionate about it.

We're not gonna go out and make this happen before the other use cases.

I think there is a component of that where, probably we work on things like science and biomedical out of proportion to its immediate profitability.

Because you guys think it's worthwhile.

Because we think it's worthwhile.

We feel the same way about things in the developing world.

One I'll give you that's controversial, people think about it the opposite way.

So the work we do on defense and intelligence, people are often like, "Oh, these guys are selling out."

I think about it the opposite way, right?

So, you know, there was this contract with a ceiling of $200 million with, you know, the DOD and intelligence community.

People are like, "Oh, man, Anthropic's selling out."

It's exactly the opposite.

Getting another $200 million from some coding startup would take an order of magnitude less effort than getting that contract.

But you think defense is very worthwhile.

And we're doing it because we want to defend democracies.

And we do it within bounds.

There are some things we're concerned about.

I'm deeply concerned about abuse of government authority on the domestic side.

We think more on the kind of outward directed side.

But that's an example of the things we prioritize are things that we think are good, not necessarily things that feel good or that people will that we think the kind of external buzz will be positive.

We actually have conviction around some things, and we do them regardless.

You reference the kind of business you want to build.

What are your aspirations for the Anthropic business in, say, three to five years time?

AI is strange in, like, a number of ways.

I think one of the ways it's strange is that because it's an exponential, we have a hard time calibrating exactly how big the business will be.

We had the following experience.

In 2023 I'd never raised money from institutional investors before, and so our revenue was zero at the beginning of 2023 because we had not released a product.

So I was putting together something, and I'm like, "Oh, I think we can probably get $100 million of revenue in the first year."

This caused some investors to say, "This is crazy.

This has never happened in the history of capitalism.

You've lost all credibility." "You're just making up numbers." "Goodbye goodbye."

numbers." "Goodbye goodbye."

And then we actually did it.

And so then the next year, I was like, "I think we can go from $100 million to a billion."

And actually, having done it the first time, people were like, it was a little bit less dismissed as crazy, but still often dismissed as crazy, and then we did it again.

This year, we're halfway through the year.

We're, as you mentioned, well past 4 billion revenue in kind of logarithmic space, to add another order of magnitude.

So, there's a bunch of different futures.

There's one where once things get to a certain size, the curve slows down, but there's a provocative world where the exponential continues and in two or three years, these are the biggest businesses in the world.

And I think one of the fundamental experiences and uncertainties of working at or running something like Anthropic is you kind of don't know.

You make this exponential projection. It sounds crazy.

It might be crazy, but also, it might not be crazy, because that trend line has followed before.

And I've said much the same thing in the context of training AI models, in the context of the cognitive capabilities of AI models on the technological side, but now we're seeing the same kind of continuous lines on the business side.

So, what's the analogy to scaling laws here, where you scale up the relevant inputs for model quality in parallel, and you get kinda much better model performance?

Is there something where you put better models in, and I know the right organization?

I know what the inputs are. Yeah, yeah, yeah, there's something like, there's some curve where you spend 5x or 10x more to train a model, or you have 5x or 10x more data or whatever the scaling laws say, and there's some transfer curve for revenue, right, where I spend 10 times more on the model,

andthe model goes from being a smart undergrad to a smart PhD student.

Then I go to a pharmaceutical company, and I'm like, "Well, how much more is that worth?"

Often, they end up saying "That's worth about" More than 10x, yeah.

"That's worth about 10x," where these kind of power law distributions occur in a bunch of contexts.

Going on the technical side, when you train the model, there's a longer and longer tail of kind of correlations that you're capturing as you train the model, right?

Correlations in the structure of language, in the world, in patterns.

And that correlation is what's thought to lead to the scaling laws, because there's this kind of logarithmic distribution.

And then as you think of the model getting more and more capable in terms of cognitive tasks, there must be, or we're seeing empirically so far, if you think of the uses of the model in the economy, right?

If I think of the way that companies are organized, right?

There's a kind of power law, there's a power law structure of, like, the— The org chart.

The org charts of companies.

And it almost feels like you're climbing that power law distribution of value.

And then I guess the way I think about product and go-to-market is that the model wants to be on that exponential of revenue, and product and go-to-market, they're kind of a way to clean the window and let the light shine through.

A way to kind of open the aperture and let the exponential happen.

It's like the models want to learn, the models want to be extraordinarily successful in the market.

Yes, right, in addition to having this learning impulse, the models have this capitalistic impulse that they want to embody, unless they're given a bad product or bad sales to go with them.

Because they're really useful, that intelligence is really useful to people, and so it kind of gets pulled out of you.

Yes, yes, yes. That is a way to think about it.

What is the terminal market structure here?

Like, is there a few large scaled players, or do we kind of keep seeing new upstarts for kind of specific use cases?

It's very hard it's hard to tell for sure, and I think there was quite a lot of uncertainty two or three years ago.

But I think we might be relatively close to the final set of players, if not necessarily the final market structure or the roles of the players.

I would say there's probably somewhere between three and six players, depending on how you count, and those are the players that are capable of building, the players that are capable of building frontier models and have enough capital to plausibly bootstrap themselves.

I would love to understand how the model business works, where you invest a bunch of money upfront in training, and then you have this fast-ish depreciating asset, though maybe with kind of a long tail of usefulness, and hopefully kind of you pay that back, Thus far, like, I think the image people have from the outside world is ever larger amounts of CapEx and how does all— Get kind of burned.

There's kind of two different ways you could describe what's happening in the model business right now.

So let's say in 2023, you train a model that costs $100 million, and then you deploy it in 2024 and it makes $200 million of revenue.

Meanwhile, because of the scaling laws, in 2024, you also train a model that costs $1 billion.

And then in 2025, you get $2 billion of revenue from that $1 billion, and you spend $10 billion to train the model.

So, if you look in a conventional way at the profit and loss of the company you've lost $100 million the first year, you've lost $800 million the second year, and you've lost $8 billion in the third year, so it looks like it's getting worse and worse.

If you consider each model to be a company, the model that was trained in 2023 was profitable.

You paid $100 million and then it made $200 million of revenue.

There's some cost to inference with the model, but let's just assume in this cartoonish cartoon example that even if you add those two up, you're kind of in a good state.

So, if every model was a company, the model is actually, in this example is actually profitable.

What's going on is that at the same time as you're reaping the benefits from one company, you're founding another company that's much more expensive and requires much more upfront R&D investment.

And so the way that it's gonna shake out is this will keep going up until the numbers go very large and the models can't get larger, and, you know, then it'll be a large, very profitable business, or at some point the models will stop getting better, right?

The march to AGI will be halted for some reason, and then perhaps it'll be some overhang, so there'll be a one-time, "Oh man, we spent a lot of money and we didn't get anything for it," and then the business returns to whatever scale it was at.

Maybe another way to describe it is, the usual pattern of venture-backed investment, which is that things cost a lot and then you start making it, is kind of happening over and over again in this field within the same companies.

And so we're on the exponential now.

At some point, we'll reach equilibrium.

The only relevant questions are, at how large a scale do we reach equilibrium, and is there ever an overshoot?

Right right.

You referenced the cloud companies as a point of comparison, but I don't know, there's something about the cloud companies where it feels like their data center CapEx is more continuous, they're just always doing new data centers, whereas there's something about how discrete these generations are that maybe it's like the way the engine manufacturers, they keep coming up with new technologies— Yeah, yeah, it's like the F-16 or something, orit might be a little bit like drug development.

Exactly.

Kind of an R&D heavy thing.

Yes, and when do you actually go to the effort of training a model?

Yeah, yeah it's almost like a drug company, where it's like you develop one drug, and then like if that works, you develop 10 drugs, and then if that works, you develop 100 drugs.

The drug development market does not work like that numerically, but it is as if it did.

Right, so we can look at each of these models as individual programs and look at their individual P&Ls, and you're saying that the payback math on those, at least in the models we've seen to date in the industry, is not actually that challenging.

Where I think most...

When you're acquiring a customer, if you have a nine-month payback on acquiring a customer, you'll do that all day long.

That's a very easy payback to underwrite.

And you were saying the paybacks are kind of nine months, 12 months, so like, they're very easy to write. I don't wanna make any specific claims— Sure yeah yeah.

But qualitatively, if you look at the business this way, model by model, it looks very viable.

Yes, because the ever-growing CapEx is masking the underlying quality of the model businesses.

Yes. In 2023,

everyone is talking about the data wall.

Is this how we solved our way out of the data wall?

Yeah, so I don't know, people talk about things in public, and sometimes they're rumors or suppositions or whatever.

I wouldn't even necessarily assume that there's a data wall.

One thing I will say is that the idea of using RL has been around for a while, right?

If we go all the way back to when Google DeepMind won, beat the world Go champion with AlphaGo, it was RL first.

And then we built these language models, and now we're kind of uniting the two together by putting RL on top of the language models.

That's all chain of thought or reasoning is, it's just a fancy way of saying RL, where the RL environment is that the model writes a bunch of things and then gives an answer.

There's nothing more to it than that.

It just kind of has a fancy name.

And so I think of these as kind of the two key ways of learning, right?

I think of base LLM training as learning by imitating, and RL as learning by trial and error.

I think those are the two styles of learning.

If I'm like a child, there's two ways for me to learn.

I look at my parents and I'm like, "Oh, they do something," and I try and learn what they do, or I can just kind of experiment with the world and learn things, and it's very clear in developmental psychology that people use both.

And so we're now seeing that recapitulated in the language models, and so we have a stage where we do the imitative learning, and we have a stage where we learn by trial and error.

So it seems very natural to me.

The other thing that's obviously notable to people not in the AI industry looking at it, is all of the talent wars and the fact that your IP walks out the door each evening.

You referenced in a recent interview you gave, $100 million secrets that were a few lines of code.

And obviously I think you were talking about that in a national security context, but you could also think about it in a talent context.

Of course. And so how does one...

Like, in the pharma industry, they protect their secrets with patents.

In Wall Street, where also they have $100 million secrets that are just a very simple idea, Renaissance Technologies, the hedge fund, just very successfully locked up its employees.

How do you make keeping a commercial lead work in kind of the current AI environment?

One thing I will say is that there are some things that are like that, but I think more and more as the field matures, it starts to be more about know-how and ability to build complex kind of objects.

Some of the ideas we work with are simple, but I would say the simple ideas, the ones that are like, oh, yeah, twiddle this element of the transformer or something, those tend to be independently discovered or anyone knows them before too long.

But there are things like, oh man, this thing is actually really hard to implement from an engineering sense, and we have it implemented, or this thing is just kind of a pain to do, or there's a know-how to doing it.

And those tend to be more collective things that are more difficult to leak.

And so I think those things are substantially more defensible.

That said there's still leakage, and we still don't want it to happen again, both for commercial competitive reasons and for national security reasons.

Both are problems. Yes.

And so a few things we do.

One is, you know, we tend to compartmentalize information.

So if you talk to any intelligence agency, that's how they operate.

You're only told what you need to know, and I think everyone within Anthropic.

But that's probably quite different to a normal Silicon Valley culture, where everything's just flying around the company.

Yes, we actually do that at the same time as we have a very open culture.

I say things to the company that maybe another person would put it in kind of PR speak or, you know what I mean?

But when there is a secret, then I think that actually leads to people trusting that it's something that you actually need to know.

And then finally, having better retention rates and losing less people is one of the most important things here.

We have the highest retention rate of all the AI companies.

I think the differences are even starker, because everyone has kind of a non-regretted attrition rate that's maybe constant, so if you subtract that off, then the difference is even larger.

Sometimes when people leave, they come back.

Yes. I saw that recently.

If you look at you can see publicly the list of people who went to the Meta Superintelligence lab.

Even if you normalize for our size, it's not, and many turned them down.

So in the crazy $100 million comp wars that everyone's been talking about, you guys have not had too hard a time of that.

I think relative to other companies, we've done well.

We even have been relatively advantaged.

It's like a mixture of true belief in the mission and belief in the upside of the equity.

Anthropic has developed a reputation for doing what it says it will do for, in some cases, making less promises, but keeping those promises that we make.

And being very clear on what we stand for, and being consistent over the years and standing for it. That creates

a unity around the company, and I think it's a good guard against cynicism.

When you're talking about the upside of the equity, when you're pitching investors, or maybe candidates, how do you pitch the Anthropic business?

Like, we're building a very large business.

That's a good start, but what else goes into it?

Often I'll talk about the platform and the importance of the models.

For some reason, sometimes people think of the API business and they say, "Oh," you know, "It's not very sticky," or, "It's gonna be commoditized." I run an API business, I love API businesses. No, no, exactly, exactly.

There are even bigger ones than both of ours.

I would point to the clouds again.

Those are $100 billion API businesses, and when the cost of capital is high and there are only a few players...

and relative to cloud, the thing we make is much more differentiated.

These models have different personalities.

They're talking to different people.

A joke I often make is, if I'm sitting in a room with 10 people, does that mean I've been commoditized?

There's like nine other people in the room who have a similar brain to me, they're about the same height, so who needs me?

But we all know that human labor doesn't work that way.

And so I feel the same way about this.

I think the API business is a great business, but we wanna go broader than that.

The way I think about it is other players such as OpenAI and existing incumbents such as Google are very focused on the consumer side.

The idea of providing AI to businesses is something that we are trying to get better and better at.

I think we're out to an early lead in that.

I'm not sure, because I don't know for sure what the revenues of the other players are, but I think we probably at this point have the plurality of the API market, most likely, and AI for business market, perhaps.

It's funny when you talk about kind of the commodity argument, where we obviously grew up facing this as a skeptical argument, and I remember finding it so striking when AWS finally had to break out their numbers in 2015.

Remember they used to be wrapped up in Amazon's numbers, and they broke them up?

Yes yes yes.

And people had been talking about, pundits had been saying, "Oh, cloud is a commodity, it's uninteresting," and then they broke it out and it was one of the greatest businesses of all time.

And there's something where a business can have competitors and it can have buyers who care about price, but that's very different from being a commodity, and as you say— Yeah yeah yeah exactly.

All these products work differently.

No exactly.

I mean, we're like one of the biggest customers of the clouds.

And we use more than one of them.

I can tell you, the clouds are much less differentiated than the AI models. For sure.

Because it feels like, one, the behavior is non-deterministic, which not by design trying to make it hard, but that just naturally means that, oh, we get the customer service answers we prefer with this model versus that model, and I don't know why. No, no, no, exactly.

You don't know why.

It's a little like baking a cake, right?

You put in the ingredients— It just works.

It just, it kind of comes out a certain way.

One chef makes it this way, and the other chef makes it this way.

And if you're like, "Make it exactly like that chef makes it," you can't, right?

You just can't.

And presumably, it's striking to me, none of the AI products are that personalized right now, but it feels like personalization will be a huge deal.

Will be a huge deal.

And will be a big source of stickiness for the...

Like, 'cause you won't want to switch products.

And I don't know exactly what that looks like, but given the amount of...

For both the consumer and the business use case.

Absolutely absolutely.

You know, I think we've just started to scratch the surface in terms of models that are customized in various ways for working with a particular business or particular person within the business.

So I think we're just seeing the beginning of the API business.

But I don't think AI for business is just about API.

With things like Claude Code, we're selling that to not just individual developers, but enterprises as well.

And they find it some useful.

Claude for Enterprise, that is selling to a lot of enterprises.

I actually see it, and you see this with some of the clouds, where they have a bunch of different services, right?

Some of them are apps, some of it is the underlying cloud itself, and what they are is the way that AWS or GCP or Azure will present themselves, and the way that we are starting to present ourselves is, hey, we wanna be your one stop shop for AI or for cloud, and you can buy all of these things, and you can talk to us about which to use for what.

And so I think that starts to create the outlines of a more durable business.

If you think about a typical Fortune 500 company, how...

They're probably playing with AI for customer service or engineers maybe have AI powered coding tools, how AI adopted are they compared to how much they should be?

Well, certainly much less than there should be.

Yeah. But is it like 5%, 30%?

What I would say is there is almost, there's very often conviction at the top.

You talk to the CEO, the CEO gets it.

You talk to the CTO, the CTO gets it.

The struggle they have is that they have 100,000...

Pick a company that has 100,000 people, who their job is to do something else.

Their job is to do banking or insurance or drug development.

And they've heard about this AI stuff, but they're not like— It's not what they're an expert in.

And so the challenge is often we are working with the leadership of the company to get the 100,000 people in the company really familiar with and using the technology.

I think, again, the code stuff goes the fastest, because the developers are the ones who are most adjacent and most watching the trend.

Some of the kind of customer service and process stuff is next to go, but you really have the instinct that even with today's models, it could be, 100 times bigger than it is.

You really get that sense.

My intuition is sort of that we will see the patterns of AI adoption from startups, because they're unconstrained by existing organizations, so they can kind of do whatever makes sense, versus large organizations are somehow calcified, because they have all these people whose job is to do X and need to be consulted and everything like that, so we'll see the new behaviors from the small startups, and then large companies, as you say,

the CEOs and CEOs are switched on and they're smart, they say, "Hey, we should be doing that," and they'll kind of port the new ideas from...

Kinda like the adoption we saw of cloud or many of these other techs.

Is that what you're seeing?

Is that your intuition?

They'll port the new ideas from the small companies, or the small companies will become threatening to them and disrupt them.

Sure yeah.

And that will give them the urgency to kind of drive things through and make them happen.

A pattern I've seen that works pretty well that I actually recommend if you're a large company is to kind of make a strike team or strike force that's separate from the rest of the company and kind of develops these prototypes.

And then basically, you can get momentum behind something, and then there's always this hard work of integrated into the rest of the company, but if you have a lot of momentum and you've done the hard work and you've shown the thing works, then it's easier to do that.

Dwarkesh, did you read his recent blog post on his AI timelines?

Oh, on continual learning, yeah?

Yeah.

He talked about how his fundamental issue at many of the AI models for productivity is that they're like the virtual co...

They're like the super smart virtual coworker who started five minutes ago, but they remain the coworker that started five minutes ago.

You know, they don't learn over time.

How will we solve that?

Yeah, so, you know, the pattern that I've seen in AI on the kind of research and technical side is that what we've seen over and over again is that there's what looks like a wall.

You know, it looks like AI models can't do this, right?

It was like, "AI models can't reason."

And recently, there's this, "AI models can't make new discoveries."

A few years ago, it was like, "AI models can't write globally coherent text," which of course now they obviously can.

You go back a few more years, and you know, it was like this Chomsky thing of like, you know, they can get syntactics right, but they can't get semantics right.

And every one of those has been blown through.

Sorry, what have we blown through on the new discoveries?

This is a thing that people have said recently.

Actually, my view on this, like many of the other things, on new discoveries, is that it's not really a binary, right?

They don't get to have their name in the paper.

Yeah, yeah, they don't get to have their name in the paper, but like, what is a new discovery?

What is genius?

You know, I think I remember this developmental psychology book, but they were saying something like we kind of lionize genius, but let's say that a table's wobbly, and I'm like, "Oh," I take the coaster and I put it under the table, and it's not wobbly anymore.

That's an idea in a way that's like a new discovery.

Even if I've never seen someone do that before, you know, that's like a new discovery.

And you know, the difference between that and the Nobel Prize winning discovery, it's a matter of degree, not a fundamentally different matter.

And so I would say that the AI models make discoveries all the time.

I've had family members where they had a medical problem, and the AI model diagnosed their medical problem when doctors missed it.

That's not a big, new, but that's a new discovery.

And you could say, "Oh, they're just pattern matching to things that happened before," but new discoveries are like that.

You think of writers who have written novels or something that are totally new, and you're like, "Well, what are your influences?"

And they're remixing together and adding a new element.

It's all more continuous.

And that was the thing I was gonna say about continual learning.

I think this idea that it isn't present is...

I would say it's present a little bit.

Yes. It's comfortable practice.

We're gonna find a way to get more of it.

So for instance, the models learn within the context.

You talk to them and they absorb the context.

Eventually, the context is gonna be 100 million tokens, and maybe we'll train the model in such a way that it is specialized for learning over the context.

You could even, during the context, update the model's weights.

There are lots of ideas that are very close to the ideas we have now that could perhaps do this.

I think people are very attached to the idea that they want to believe there's some fundamental wall, that there's something different, something that can't be done.

It kind of reminds me of— Do you think it's a coping mechanism deep down?

Yeah. You know what it reminds me of?

So, you know the 19th century notion of vitalism?

This was the idea that the human body and organisms that are alive are made of a fundamentally different material than inanimate matter.

Which of course we know scientifically now is not true, but it's something people very much want to believe, and your common sense seems to suggest it.

I'm not very much like a table.

I'm made of very different materials than metal or glass or whatever, but when we actually go down to the fundamental units, of course, we're all made of the same thing.

But you think if people now have this kind of modern concept of vitalism in whatever the fundamental humanity is, and they're saying, "Oh, you know, models can't do that."

I think there's some tendency to believe it.

And I think as with vitalism, the way around it is to recognize that a mind is a mind no matter what it's made of.

The notion of the dignity or the specialness of cognition or sentience, it's not that it isn't special, it's that it can be made out of anything.

You referenced the medical use case, which I think is a very cool use case.

Obviously, one, because of all the people who have fixed medical issues as a result, but another one is you talked in your "Machines of Loving Grace" post, which I really enjoyed, I thought it was very well done, about the marginal returns to intelligence.

What are places where intelligence is the limiting factor?

And my read of the popular medical use case is obviously it's kind of a charismatic use case, but also, for most normal people, they have some kind of medical issue, low level or serious or something like that, and actually, society is very just intelligence limited there.

Not that you don't have access maybe to a smart doctor, hopefully you do, but they give you very limited time.

You know, they think for 10 seconds about your problem.

Yeah yeah yeah exactly.

And it turns out test-time compute was actually what we needed there on the medical stuff.

But is that kinda your take on this?

That is how I think about it as well.

I have talked to Nobel Prize winning biologists who say, "I will only"...

I mean it sounds a little elitist, but they'll say, "I'll only go to the top 1% of doctors, because the rest of the 99%, I can get better advice from an LLM."

It really is true.

Doctors are busy, they're overworked.

And just the nature of medical data and medical information, it's a lot of pattern matching.

It's a lot of the same things.

The level of consistency and the ability to put together many different facts, I think it's something that LLMs are quite good at.

So you talked in this "Machines of Loving Grace" post about some of the big humanity-level areas where we're intelligence-limited, but again, the personal medical use case is a good example of one where society's intelligence-limited, and if you give lots of people much more intelligence on their specific issues, it's very valuable.

What are other areas, either in the consumer use case or in the business use case where you think we're just very obviously intelligence limited?

Yeah, the places where at least the AI models of today can help the most, the characteristic quality is something is repetitive, but every example is a little different.

Automation before AI, if you could program exactly how it happened, you could do it.

So if you were doing the same thing over and over again.

But customer service is like, you know, just to take customer service as an example, there's like a long tail of stuff, but a lot of it is like, you get a bunch of calls, each call is different, but each call is basically about one of 10 things, and it's like a different person in a different voice saying basically one of these 10 things in a different way.

And that situation, where things are repetitive and similar but not the same, and each has its own thing, that's where AI can come in the most.

Dwarkesh had in the same blog post the prediction that you can't yet give an existing AI all of your financial data and forge all the emails and have it do your taxes, and his prediction for the year in which you can, what is the year where your first tax return is done by just emailing everything to whatever AI you use, his prediction is 2028.

What do you make of that prediction?

Probably sooner than that.

Okay.

I don't know if it's '26 or '27.

Some of that is model, mostly accuracy.

I think the model could do that today, but it would make too many mistakes.

And so working on ways to have the model check its own work and do less mistakes is one part.

There's kind of an interface part of it as well, but I would be surprised if it takes that long.

Okay, '26 or '27.

And when you say about mistakes, actually, you're running through the list of things that people thought we would never solve in AI.

It feels like hallucinations should be on that list.

Not that they're totally solved, but they've gotten a lot better.

They've gotten a lot better, and I think people have gotten more used to.

They kind of know what to trust the model for and what not to trust the model for.

The models have also been grounded in citations.

I mean, we've done that with Claude.ai, we've done that with Enterprise Claude.

So I think part of the solution is citation, part of the solution is algorithmically, the models hallucinate less now, and part of the solution is people have adapted and understand the weaknesses of the model.

My view on things like hallucinations has always been, there's a certain class of critic who points to something where models are weird or worse than what humans do, and say, "See, they're not like us at all," or, "They'll never get there."

And I kind of get where the instinct comes from, where, maybe they're looking to see if we've matched the human brain exactly.

They're saying, "Oh, this is so different.

It can't be like a human brain."

But I basically, I just think it's a fallacy.

There's a notion of kind of general intelligence, but it's made up of a bunch of different things, and you can simply have most of the things and be much worse on some and much better on others.

If we look at humans that are— Yeah, have you met humans?

Yeah, yeah, have you met humans, right?

Like if you look at humans who are autistic, versus humans that are schizophrenic.

If you look at the optical illusions that humans face that machines are not fooled by, it's very clear that we have some of these weaknesses, just like just similar to the model's hallucinations, it's just that they look very different and we're much more used to them because we're surrounded by humans all day.

The autonomous vehicle double standard feels like the kind of clearest example of this.

Clearest example of this, yes.

Yeah, where people have much higher standards.

People have much higher standards.

But I think it's gonna be a feature of this technology, and it has implications on the business side.

I think we're gonna be in a world where the models will make mistakes much less often than humans, but they'll be stranger mistakes.

And actually, that takes some adaptation, because imagine you're an end user.

If you work with humans, you get used to it, and you have some notion, right?

So if a human makes a mistake 5% of the time, you might have a good understanding of why, you know?

Like, let's say I'm talking to a customer service agent, and they kind of sound incoherent and they're slurring their speech.

They've probably had too much of this and they're not doing their job very well.

That's a bad mistake to happen, but also, if I'm talking to this person, I kind of know what's going on and I know not to trust what they're saying.

Whereas an LLM might make a mistake five times less often, but it's more deceptive.

The model sounds just as erudite, just as coherent as it does when it's saying something that's right.

But that's an adaptation thing.

You know, that's not a fundamental thing.

And that's something that when we talk to our customers, we tell them about that, we tell them they need to get used to that.

So we need to invent slurring for LMS, is what you're saying. Right, right, right, exactly.

So, you started out as a researcher, but now you're the CEO of a company and you're in the business of selling AI.

And so what have you had to learn about go-to-market or dealing with customers, all the rest of the stuff?

Absolutely.

I think my view on this was I started a company not because I was initially excited about selling things or business or any of that, I'd seen the way that some of the other companies had run, and the magnitude and gravity of what they were trying to build, and I was just a bit concerned

that the people and the motivations were maybe not the best ones.

And I knew that there would be a number of players in this space, but it felt like having at least one player that kind of had a strong compass in how we do things could have positive effects on the ecosystem.

We would build things in a different way, we would deploy them in a different way, and above all, we would have a, again, sometimes short list of principles, but we would stick to them as well as we could.

So I think that was the initial motive.

And of course, I was excited about building the technology.

And you know, I think as that has happened, of course I and the other cofounders have kind of had to learn how to think about, kind of the business and the strategy.

I think I've been very naturally interested in the business side of it.

Actually, I was surprised at how quickly I became interested in it.

And actually, the primary reason was that I was curious about all the industries that are customers of us.

Somewhat like the clouds, and perhaps like your business, the businesses that we serve, they run across every possible industry.

And so you just learn these things about parts of the economy that you've never thought about.

And even in areas where I nominally know a lot.

I used to be a biologist, so in a way, I know a lot about the pharmaceutical business, but I'd never thought about it beyond the science.

I never thought about the portfolio side of it.

I never thought about how clinical trials work and how they could be made cheaper.

I never thought about the defense and intelligence business in any great detail.

And so you run through those, and I just find it super interesting to understand what people's problems are and how AI can help with those problems. I feel like I took very naturally to that.

Actually, the product side was one where I was initially more reluctant.

I felt like I just had a natural interest and curiosity in the business side of it, but like, building apps, somehow, initially it was never like, a thing that drew me in even after I started the company.

But I think more recently, as I've seen what products have succeeded and what products haven't, I think this idea of how to design products so that they're, what we call AGI pilled.

So that the direction of the product is durable and is kind of a bridge to things that are useful in the future.

We've all heard this idea of wrapper companies or wrapper products.

The idea is you make Claude N, and someone makes a product that basically addresses the deficiencies of Claude N, but then you come out with Claude N+1 and it just kind of eats it.

The advice I always give that I think all the AI, all the folks at the AI companies give, you know, don't make that.

See the direction of the field and try to make something that's complementary.

And I think thinking about how to make products in a new way, in a kind of AGI pilled way, that actually has caught my interest a great deal.

Okay, so, glad you brought this up.

Doesn't it feel like we have no AI UIs right now?

Like, we still enter text into text boxes, literally same as terminals from the 1970s.

I mean, a bit more rounded corners and everything.

Like, we still talk into voice companion modes that are manually triggered, which is the same as pre-transformer Siri.

UIs are just completely the same.

Yeah, yeah, there's something not quite right about it.

I basically agree with you.

Could have generative UIs.

It reminds me a little bit of, you know in the early days of the internet, it was like people would make these websites that had structures that looked like they were in the physical world, open the closet and do the...

They're from the horseless carriage era.

There's some term for this, I forget.

There's some word for this, I forget what it is.

Skeuomorphism?

Skeuomorphism yes.

It feels like there's some of that going on here.

A thing I would say is that as we move more towards agents, we're gonna be in a world where the AI model can do something end to end.

Like, we're almost there with Claude.

Can do something end to end and get it right most of the time.

And a human's main job is to kind of check.

Or check sometimes.

But interestingly, checking often means getting really into the details of what happened.

And so there's some kind of impedance mismatch here, that some product or interface is the solution to, where you want something that's as slick as possible and just goes off and does something, and you don't wanna have to pay attention most of the time, but when something's wrong, you might actually need to get quite involved.

And I don't feel like any products or interfaces operate on this principle now or handle this problem now.

I don't know if that makes sense, but— No, it does. Like, agreed.

I think what you want is your agent to go away and do really good work for you, and then come back with its work product to let you review, steer it, decide everything.

But you can't be overwhelmed, because it's gonna do so many more things than you have time to look at, and if you're always looking at it, you know, it can be slower than if you just did it yourself.

And so it actually strikes me as an interface problem.

Yes yes.

The generalization of this is it feels to me, one of the most exciting things about AI is we have such an overhang of current capabilities, turning them into good products, where even if AI progress was frozen right now, we'd have like 10 years of good products, because— Oh, oh, I completely agree.

And actually, the way that products are being built, I think by everyone in the industry, but we've thought about it this way, is very different because the progress is continuing.

If the progress in models stopped, the way we built products would change instantly.

The reason is I don't think we've ever had before a situation which the technology is changing under you so fast as you're building the product.

And so this idea of long-term product roadmaps or the usual way of product planning, I've started explicitly...

Again early in Anthropic, I was like, "I don't know anything about product.

I'm a doofus."

But now I always try to talk to people when they come in and they say, "This is not like building products in the non-AI space," right?

Because they need to be more AGI pilled?

Yeah, you may be the expert at building these, but like, the technology is moving under you, so like, these ideas about fast iteration, they're even more true than they are normally.

What's a specific example of this?

I think that if you're like, "We're gonna make something, and it's gonna be ready in six months," I think that makes even less sense here than it makes, the building in isolation.

So you need to just have tighter ship schedules and more iteration? You need to have tighter ship schedules, you need to try things.

It's very hard to tell, even harder to tell what's gonna catch on, because a new model may have come out, a new model may suddenly be good at something that makes a product possible, and so much more than anything else, you're trying something that's never been tried.

There's a new model.

It's only available within the company.

So the thing you ought to do is just build something on it, let people internally try it.

There's this eternal September vibe to it, right?

Where it's like, it's as if you discovered database technology for the first time, and you're like, "Well, what could you build on this?"

And it's always the first day.

That's what is different.

You mentioned database technology, and maybe that provides an interesting analogy, and as we think about open source, the first relational databases that were successful in terms of adoption were proprietary, but then the open source guys caught up.

How do you keep the gap with the open source options?

Yeah, open source I think has a different meaning in AI models than it has in other areas, right?

And for this reason, some have called it open-weights models to distinguish.

I think the main difference is that if you see the weights of the models and you look in, you can't understand what's actually going on.

There's not that kind of composability.

I can't read the source code, I can't— You can't produce a trivially different version.

Yeah, yeah, you know, I can't produce a trivially different version of it.

Now, you know, Anthropic is actually working on mechanistic interpretability, which allows you to see inside the models, and so we're actually working on things that would allow some properties, but we're not there yet.

We're not anywhere close to there.

There are some things you can do.

For example, if you have access to the model, you can fine-tune the model.

We're now through interfaces kind of allowing people to fine-tune the model.

There is a question of how valuable access to the actual model weights is over and above some thick API that lets you do something.

There's some question of economics, but note that it costs a significant amount to run the models on the cloud.

Someone has to host it, someone has to run fast inference, and then you're back to the margin or some portion of the margin.

So you think open-weight models are not that useful, and fully open source models, there's just a big gap?

I guess what I would say is that the analogy to previous technologies is only partial, right?

It's kind of a different thing that we're still discovering.

But I can say from our perspective that when a new model comes out, when a competitor model comes out, we don't really think about whether it's an open-weights model or not, we think about whether it's a strong model.

If someone makes a strong model that's good at the things that we do, that's competition, that's bad for us, whether it's an open-weights model or not.

There's not a huge difference between the two.

How is Anthropic more AGI pilled than other organizations?

So one is faster, like a tighter product release cadence.

But maybe more broadly across the organization, not just within product development.

I mentioned this thing that every couple of weeks, I get up in front of the organization, and kind of describe my vision, and I think one of the purposes of that is to keep people kind of focused on the mission.

It's a strange state of the world, and I always express uncertainty about it, but I say if I were to bet, I would bet in favor of this, that, in one or two or three years, I don't know exactly how long it's gonna be, we'll have what I've described as like a country of geniuses in the data center.

And like, this is weird.

It's gonna change the economy, it's gonna accelerate the pace of science.

It's gonna pose global alignment and national security risks.

It may pose economic problems. The upside is huge.

The potential for disruption is also huge.

And I think what I'm trying to fight against is the idea of employees who joined and they're like, "Oh, you know, I worked in this industry, you know, I worked at this kind of company, and I'm gonna work at an AI company, and maybe a couple years later, I'll go to this," you know?

This is very categorically different from previous experiences that have happened.

This is a really different thing, and I think up and down the organization, we want to make sure that when our finance people think about financial projections, they understand this, not that there's certainly going to be an exponential, but like, wild outcomes are possible.

When our recruiting thinks they're like, "Oh yeah this crazy cop stuff could happen because"...

You know, and when the product people think they make AGI pilled products, when the policy people interact, they understand the stakes of what may happen.

And so I think a big part of my job is keeping the coherence of the organization around this central thesis.

Not that everyone has to believe the thesis.

There's not an indoctrination and people chanting with robes or anything.

But like, the basic idea that the company is built around this hypothesis that it is possible and perhaps likely that these large changes will happen, and every aspect of the business, as well as the things the company is doing for social benefit, should be constructed around

strong possibility that this may happen.

To put numbers on this, you've talked about the potential for 10% annual economic growth powered by AI.

Doesn't that mean that...

When we talk about AI risk, it's often harms and misuses of AI, but isn't the big AI risk that we slightly misregulate it or we slow down progress, and therefore there's just like a lot of human welfare that's missed out on because you don't have AI?

I've had the experience where I've had family members die of diseases that were cured a few years after they die, so I kind of truly understand the stakes of not making progress fast enough.

I would say that some of the dangers of AI have the potential to significantly destabilize society or threaten humanity or civilization.

And so I think we don't want to take idle chances with that level of risk.

Now, I'm not at all an advocate of like, stop the technology, pause the technology.

I think for a number of reasons, I think that's just...

It's just not possible.

We have geopolitical adversaries.

They're not gonna not make the technology, the amount of money.

I mean, if you even propose even the slightest amount of reg...

I have, and you know, I have many trillions of dollars of capital lined up against me for whom that's not in their interest, so that shows the limits of what is possible and what is not.

But what I would say is that instead of thinking about slowing it down versus going at the maximum speed, are there ways that we can introduce safety security measures, think about the economy, in ways that either don't slow the technology down or only slow it down a little bit?

If instead of 10% economic growth, we could have 9% economic growth and buy insurance against all of these risks.

I think that's what the trade off actually looks like.

And precisely because AI is a technology that has the potential to go so quickly, to solve so many problems, I see the greater risk as like the thing could overheat.

And so I basically want...

I don't wanna stop the reaction. I wanna focus it.

That's how I think about it.

You said, "If we hit December, 2025, and there's no AI law, I'll be really worried."

How are you feeling?

There is actually something in California.

There's a bill out, SB53.

Last year, we had the whole SB1047 thing.

We had mixed feelings on SB1047.

There was initial version that I think was too aggressive.

And when I say that, what I mean is the technology is moving fast, and it's kind of unhelpful if you're too prescriptive about it.

It ends up actually not contributing to safety.

And I was worried a little bit, if something like this passes, it's like the tests that were prescribed to run will end up looking stupid, and then all the people in the AI industry will be like, "Oh, this is what regulation for safety and security looks like, it's really stupid," and they won't take it seriously.

They'll kind of do everything they can to comply in letter and not in spirit.

And so, as an advocate of thoughtful regulation, I was actually a bit concerned about this.

We offered some changes to the bill to a point where kind of we felt good about it, and we tried to make a compromise between kind of industry and the safety advocates.

We didn't really succeed, as you saw, but this year, I think we're making a bill that is something more moderate.

It's focused particularly on transparency of practices, transparency of safety and security practices, which is something that Anthropic has been very forward about, and that I think other companies are starting to do, but not all the companies do it, and there's no way to tell if folks are telling the truth about what they're revealing.

And California regulation is enough, because all the companies have a nexus here.

Yeah, yeah, I mean, I think most of these bills are organized around, doing business in California, and so it would be difficult to shut off.

People are very AI-pilled here.

People are very AI-pilled here.

So, you know, we'll see what happens.

I'm not sure what's gonna happen.

But we've always had this approach that we kind of are in favor of guardrails, including legislative guardrails on the technology, but we recognize the need to be careful.

We don't wanna kill the golden goose, we just wanna stop it, from overheating or running off the road.

Yeah maybe something like modern bank regulation for all the people complained is a good example, where there's an inherently very risky activity, you know?

Yeah, no, no, the dangers are pretty clear.

I mean it's like the bank runs are not- Right, but it all works pretty well in the modern era once we figured out the regulatory environment.

Last question. What is your personal AI stack?

How do you use AI differently to maybe other people in tech?

Yeah interesting.

You know, I basically write a lot.

Perhaps I have too much pride in my own writing.

I use Claude to generate lots of ideas.

You know, I kind of use it as research.

But so far, I've done the writing myself.

Claude is actually maybe closer than the other ones, but it's still not there.

I'd be comfortable with it for business emails, but if I'm kind of, like, writing an essay or something that I wanna really get right, it's not quite there yet.

But maybe it will be in, you know, a year or so.

Yeah, very cool. Well, Dario, this was awesome.

Thanks for coming by.

Thank you for having me.