All right, good afternoon, everybody. Are you all having a good time? Oh, fantastic. Amazing. Are you ready for Metallica? Are you ready for Benson Boone? Raise your hand if you're going to Metallica. Awesome. Who's looking forward to Benson Boone? All right. Well, so we got a it's kind of a mix. I don't know. Please welcome Dario.
Hey, Mark. Thanks for having me.
Great to have you, Dario. Welcome. Thank you. We're really excited to have you, Dario, and introduce you to the audience and everybody here. You know, we've had such a great partnership. We've had an opportunity to speak so many times and work with some customers together and also build some technology together. It's been exciting to use your model inside Slack with our new Slack bot. But, you know, as I've talked to so many of my friends about you, we've gotten to know each other, but a lot of people don't know you. So, maybe you could tell folks a little about yourself, kind of your beginnings, where you're from, you know, and also kind of I'll ask you a couple of core questions along the way, but to take it from the start.
Yeah, absolutely. So, I grew up right here in San Francisco. I was originally interested in science, not in tech per se. So, you know, I did my graduate work in neuroscience. I was a postdoc in computational biology at Stanford. And, you know, as I was working hard on solving the problems of biology, you know, trying to understand the diseases we face like cancer and to understand biology. One of the things I noticed is that the complexity of biology felt like it was beyond human scale, you know, and I was starting to despair a little bit that we would be able to make progress at any speed just because of, you know, if you've ever seen one of these diagrams of human metabolism, it's just beyond human scale. But almost exactly at the time that I was wondering about this, out of Google, out of University of Toronto came the first work on deep neural networks. And I looked at it, you know, I replicated some of the results myself. I said, 'My God, this stuff really works. This stuff really is going to go there.' So, I went into AI. I ended up working at Google. I joined OpenAI in the early days. I led research there for several years. And, you know, I was one of the people who kind of documented this trend of exponential scaling of the technology. You know, myself and the folks who eventually became the Anthropic co-founders were among the first to study the scaling of the technology. And so, you know, around 2019 or 2020 had really two realizations. One was wow, AI is really going to scale to the moon and it's really going to reshape the economy. It'll take a few years, but it's going to reshape the economy. And we're almost there now. And the second was look, these systems are different from ordinary software. You don't write every line of code, right? Building ordinary software, it's like building a skyscraper or something. You make the blueprint. You make everything to design. This is ironically a bit more like biological. It's organic or something. You're growing these models. You set a high-level recipe. It's like growing a plant or baking a cake or something like that. And that unpredictability combined with the incredible economic impact means that safety and responsibility are going to be very important to make sure that the models behave predictably as well as we deploy them and they have enormous impact on the world. We need to make sure that we deploy them responsibly. And so those were kind of the two pillars on which myself and the seven co-founders who left OpenAI to found Anthropic, you know, generally based our view. And, you know, the rest is history. Anthropic has been around for 4 and a half years. We're growing very quickly. You know, I think we're at this point by revenue have at least a plurality, perhaps the majority of enterprise customers and enterprise revenue while most of the other players are focused on the consumer side. And we find that, you know, as you know, I think you would know from running an enterprise focused company, these values of trust and responsibility are really synergistic with serving the enterprises, right? These values help us to serve the enterprise and having a business model that serves the enterprise helps to reinforce our values.
One of the things, and you went pretty fast into kind of starting the company and kind of shot out of Google and API to Anthropic, but you didn't mention your sister. And that was like one of the things that first really shocked me when we started to get to know you was you and your sister together are working on this company. And can you kind of tell us the origin story?
Yeah. Yeah. So, you know, there's seven co-founders and me and Daniela are two of them. As I mentioned, we both grew up in San Francisco. We went to Lowell High School. And, you know, even very early on, we really figured that we wanted to do something.
My mom went to Lowell and she's sitting right there. So, she's going to be very happy about that.
Amazing. No, no, it was a great place. I actually didn't overlap. Daniela was 4 years younger than me, so we were never there at the same year, but I was there for 4 and then, you know, the next 4 years she was there. So, yeah, you know, very early on, we had these ambitions to do something good for the world, to have some kind of grand project, and to work together. I think many people, you know, or many siblings, I guess, have these ambitions, but somehow things found their way to us actually being able to do it. So, she joined me at OpenAI in 2018. And then the two of us, along with some folks we worked with, many of whom had been long-time friends, decided to start Anthropic. So, me and Daniela essentially run the company together. She's responsible for the day-to-day running of the company, you know, how the company is put together, how the org runs, what its processes are, does a fair amount of the business on the structural side. You know, I just couldn't run the company without her and without her doing all those things. I tend to think about the strategic side, where the research is going, what our plans are as a company strategically, how we think about safety, how we think about the company's role in the world and the vision. And I find that's a very good split, that's a very good separation, because both of us can do both pieces, but it's really hard to switch between a kind of operational mode and a vision mode, and I think both people can be more effective if you split.
Thanks for allowing me to come back to that, because I think it's such a great part of the company and the culture and how you're working. And last year she was here and now this year you're here. And we've known you all from the beginning. I mean what I think Salesforce owns like some little point or something of Anthropic's. So we're very happy to be an investor.
Very early investor. Very early investor, believe in you.
I think you invested when the valuation was 4 billion. Is that correct?
I don't know, but it was too high whatever it was.
Before you say it's too high, you should calculate how many times it's gone up since then.
Okay, we made money so far.
Well, no, I'm thrilled obviously that we're part of your journey and, you know, we love working with you. We love working with Daniela so much and, you know, I think kind of getting back to the story of you and your sister and your friends, you know, starting this company out in San Francisco. And then you came to us, we're starting to consolidate some of our real estate here and then you had a conversation about you taking over the Slack building. We had just bought Slack.
Which we have done. And now you're in the Slack building. So you're right on the park with us looking at the tower from your window. I believe my office is, I was told it was formerly Stewart's office.
Formerly Stewart's office, very good. And I would just, you know, it just occurred to me like here you are, you're the San Franciscan. You're right on the park with us. You're right on the tower with us. We're an investor in the company. It's obviously a critical part of Slack today, right? In terms of how you couldn't use Slack without it. But you're at the very bleeding edge of these models. It's really incredible and you've really been visionary right from the very beginning. Now when I didn't know you when you were at OpenAI, I only met you after you started the company. So just now kind of take us through your expectation. What has been your biggest surprise, you know, since it's only been what four and a half five years since you kind of walked out the door, said goodbye to Sam, right? And then you just said okay, Daniela, we're going to get this going. What was your biggest shock right there?
Yeah, I mean, you know, there were so many moments because I think the last four and a half years were just so surprising overall. You know, I think at a high level there's a dichotomy where on one hand everything we've seen over the last four and a half years was in some sense predicted by those straight lines on graphs that I made with the scaling laws, right? These enormous data center build-outs, the way that our revenue is going up, you know, 10x per year, the use across the economy. You could see it all in these scribbles on the page. But actually seeing it play out is wild. And a lot of the individual things that I'm seeing happening, I just never would have expected. I guess if I were to pick one most surprising thing, it is, you know, we left one company and started another company. And something you see over time is that as you build a company, you really do see through the people you hire, through the choices that the company makes, you kind of see the paths diverge, and you see differentiation, right? OpenAI's gone very much in the consumer direction. We've gone very much in the enterprise direction. We've taken, you know, just the way we approach things is very different. And so you have this market structure where there are these two companies that are like it's sort of a natural experiment. They're just taking a very different approach. Both exist in the ecosystem, but the approach they take is very different.
And then you have to reflect back at your time at Google as well. They're also very much with Gemini and, you know, very much a top model as well. How do you look at them? Do you look at them as a blend then?
Yeah, I mean, you know, Google's a partner to us. They're an investor in us. Obviously, their models compete with us and compete with us at the same time. But this co-opetition dynamic is becoming more and more common, right? As we have all these layers, we have like the chip layer, the cloud layer, the model layer, the go-to-market and distribution layer, and various companies are involved in more than one of these layers. And so, I think it's not uncommon for companies to cooperate in one area and compete in another area. It's fairly normal. I think it's also been helped by, you know, I know quite well some of the folks behind Google. I've been friends with Demis Hassabis for something like 15 years, right? I think he was maybe one of the only people who was earlier than me in seeing the potential of the technology. He also has a neuroscience background like me. So, I think that helps us to get along even as there are some competitive aspects between the companies.
What was your reaction when he won the Nobel Prize this year?
I was very happy for Demis. Actually, you know, I was writing this essay about the potential of AI in biology called Machines of Loving Grace. And 2 days before I was about to publish it, of course, Demis and his colleagues AlphaFold won the Nobel Prize. So, I ended up dedicating the essay to the 2024 winners of the Nobel Prize in chemistry for AlphaFold because it was an example of the future that I was talking about.
Very exciting. So, just getting back to you and Daniela and you're working on and growing the company and up it goes. And you're doing it right here in San Francisco. You know, what are some of the characteristics of the models that you're looking at now that are most surprising? And what's really exciting about, you know, as you see the next generations of these models unfold?
Yeah, you know, I think broadly it's the jump to being able to do agentic abilities. Being able to do things.
It? I didn't hear the word.
Oh, sorry. Agentic abilities. Turning the models into agents that can.
Oh, oh, oh. Oh, I see you've been. I don't worry about it Dario don't worry. I wasn't. I just wanted. I'm a little weak in my right ear that's all.
I thought you didn't hear. I see I understand. Um, yeah. It was very funny wasn't it? Yeah okay. It's definitely a buzzword and you could perhaps contribute more than most to making it a buzzword but I think it's a real thing which is being able to do tasks end to end in a way that was not possible before and I think we see it more in code than we see it in any other area but I think code is just a leading indicator. Honestly in other tasks in tasks like financial services in tasks like health and medical in tasks like insurance or manufacturing many of the regulated industries that we're working together on our partnership agents are just as valuable you just as much want to do things end to end right you think of the sales and go to market and customer management workflows you want to string a bunch of things together and you want the models to pull in a bunch of things in the context and learn this do that access this information. So it's just as valuable it's just the coders are early adopters so they've given us a glimpse of the future because they adopt things so quickly but I think we're going to see a similar spike in other areas.
So I want to ask you I have a few questions in that area. So the first, you know, when we first started getting in probably one of our very first conversations we were starting to talk about this theory and also AGI especially and what the potential in the future and AGI at that time and this is kind of now we go back several years was kind of being defined as yeah the model is almost kind of recursive. It's building itself. It's growing. It's evolving. It knows it is coding itself. It can code and then it's enhancing and it's often going off and, you know, we had Brett Adcock on the stage earlier today and he was talking about how kind of like out of scene out of, you know, we are Bob. We are Legion that the robot had taken off and was now, you know, self-replicating into the future and, you know, an infinite number of robots and well, it was quite incredibly compelling would go off and that was that. You know, to that point a few years ago when we were talking it was like, well, now the model is kind of building itself. It's growing itself. It's developing. It didn't turn out exactly like that though, right? Or we're not at that point yet, are we?
Yeah, so, you know, I would separate two things. There's inevitably as we talk about AI in the future there's kind of sci-fi sheen on the whole thing and that's, you know, in a way it's helpful as a metaphor but also it can be distracting because it just conjures everyone's mind into this kind of future dreamy mode. But, you know, I think there's some things that are real. One thing that's real is that, you know, every 3 months we make a model that's smarter than the model we made before and indeed the models are slowly getting to where they're better than most humans at most things. Maybe at some point better than all humans at all things. The other thing I think is real is, you know, the model is not building itself in the sense that like I have a specific copy of the model that's making a specific copy of the next model. You know, we don't have our self-replicating nanobots or robots or whatever. But what absolutely is happening is that within Anthropic and elsewhere we are using Claude to help build products on top of Claude and to help train the next Claude, to help serve Claude faster. You know, 6 months ago I made this prediction that in 6 months 90% of code would be written by AI models. Some people think that prediction is wrong, but within Anthropic and within a number of companies that we work with, that is absolutely true.
So, you're saying that 90% of all code at Anthropic's being written by the model today?
On many teams, you know, not uniformly everywhere.
That you've kind of provided something kind of something auditable around that or like shown how that's happened?
Yeah, the teams play an editing and supervisory role, but, you know, Claude can both, you can say, write this feature and Claude will do it, and then the human kind of looks over the code or asks for help from another copy of Claude. The other thing is sometimes we have these long-running bugs, you know, when we were training our most recently released Claude, it was like, you know, the cluster was broken, there was this bug, engineers were spending a few days on it, and, you know, we just went to Claude and we said, 'I don't know, just kind of play around with the cluster and see if you can figure out what's wrong.' And, you know, Claude found this just super obscure bug that the engineers had missed. So, this was one area where it was like Claude is like having a very talented teammate.
So, if Claude is now writing 90% of all of the code at Anthropic, do you need less engineers if 90% of all the code? When will it be 95% or 99%?
Yeah, so, you know, I think people can misinterpret it as, you know, Claude writing 90% of the code means we're firing 90% of the software engineers, but, you know, if you're familiar with the principle of comparative advantage.
Well, you have made some very aggressive comments around firing executives.
So, um, I just want to put two and two together to get four. Okay.
One thing I've learned in knowing you for so long is I have to come back and kind of like get that qualifying question again because you're going fast so I'm trying to keep up with you and it's amazing but I wanted to just make sure I understand exactly what you're saying.
Yes, absolutely. Let me take them one by one. So, you know, under comparative advantage if Claude is writing 90% of the code what that means usually is you need just as many software engineers. You might need more because they can be more leveraged. They can focus on the 10% that's editing the code or writing the 10% that's the hardest or supervising a group of AI models and so what happens is you can just end up being 10 times more productive because the remaining 10% is now a complement to the AI system that's getting much better at what it's producing.
So it's a rebalancing of the work is what you're really saying.
It's not a replacement. It's a rebalancing.
Oh, okay. Okay. Now as that 90% goes to 95, 99, totally automated end-to-end, there may be some point at which all bets are off.
What's your time frame for that? It's a pretty big jump there.
Very hard to say. But, you know, we're looking at it closely. We're always looking for ways to encourage complementarity.
In the short term or the long term?
A five-year goal or a one-year goal?
About a 10-year goal. I'm optimistic about complementarity in the short run. I do worry as we get out to two years, five years that across the economy not that there won't be anything for humans to do but that there will be enough disruption that will have a kind of labor disruption that is not fundamentally different from what we've seen with previous technologies but that operates faster, that is broader across the economy because it relates to all kinds of jobs.
In the next two to five years?
Two to five years, I would say.
