Dario Amodei Interview Impact of AI

AI writing most of Anthropic's code

00:00you know, I would separate two things there. There's I think inevitably as we, you know, talk about AI in the future, this kind of sci-fi sheen on on on on, you know, the 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, you know, it just it just kind of conjures everyone's mind into this into this kind of uh, you know, into this kind of uh, you know, future dreamy mode. Um, uh, but, you know, I think there's some things that are real. One thing that's real is that, you know, every three months we make a model that's smarter than the month than the model we made before. And indeed, the models are slowly getting to where they're, you know, better than most

00:36humans at most things. Maybe at some point better all humans than 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 like making a specific copy of the the, you know, the the next model. you know, we don't have our self-replicating, you know, nanobots or robots or or whatever, but but what absolutely is happening is that within Anthropic and elsewhere, we are using Claude to help build products on top of quad and to help train the next Quad to help serve cla um you know, six months ago I I made this prediction that, you know, in in six months 90% of code would be written by by by AI models. Some

01:20people think that prediction is wrong but within Enthropic and within a number of companies that we work with that is absolutely true. Um now 90 you know >> so you're saying that 90% of all code at anthropics being written by the by the model today >> on on many teams you know not uniformly every >> something 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 the human kind of looks over the code or asks for help from another another model or another copy of Claude. Um uh the other thing is

02:01sometimes we have these long running bugs. you know, when we were training our our the 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 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 this was one area where it was like Claude is like having a very talented teammate. So if quad is now running writing 90% of all of the code in entropic do you need less engineers if 90% of all the code when will it be 95% or 99%.

Rebalancing, not replacing engineers

02:39>> Yeah. So you know I you know I think people can misinterpret it as you know you know like 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 >> you have made some very aggressive comments around firing executives >> that that so um >> I just want to put two and two together to get four. Okay. >> Um uh so uh um you know if >> 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. Yes. because you're go, you know, you you you're you're going fast. So trying to I'm trying to keep up with you and it's amazing, but I want to just make sure I understand exactly what

03:22you're doing. >> Yes, absolutely. Let me let me let me take them one by one. So um uh 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 they can then 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 a of AI models. And so what happens is, you know, you can just end up being 10 times more productive because the remaining 10% is is now a complement to the to the AI system that's that's getting much better at at getting much better at what it's producing. So So >> it's a rebalancing of the work is what

04:06you're really saying. >> It's it's not a replacement. >> It's not a replacement. It's a rebalancing. 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. >> It's very hard to say. Um, but you know, we're looking at it. We're looking at it closely. We're always looking for ways to encourage complimentarity. >> You see that in the short term or the long term? >> Um, I, you know, >> is that a 5year goal or a one year goal? >> I'm optimistic about 10ear goal. >> I'm optimistic about complimentarity in the short run. I you know I I I I I do worry as we get out to two years 5 years

Labor disruption in 2-5 years

04:45that across the economy not that there won't be anything for humans to do but that there will be enough disruption that that we'll have a you know a kind of labor disruption that is you know 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 >> you think that's in the next 2 to 5 years. >> Two to five years I would say. >> Okay. And at what point just coming back to my previous question, at what point is Enthropic the model writing 100% of the code? >> Yeah. Again, it's it's a tricky question because maybe even if the model's writing 100% of the code, the humans are still supervising the model or prompting

05:28it or clarif

05:34it's it's it's it's a surprisingly clever thing. I don't, you know, I I'm I I I am concerned particularly about people's ability to adapt as the jobs are changing fast. The jobs are changing so fast, but comparative comparative advantage is powerful. That's going to be one of our tools in this. >> Well, you've mentioned healthcare several times, but we've seen these kind of very conflicting studies recently in healthcare around accuracy, especially because these are word models. There seems to be some level of accuracy that's not 100%. and were you know you know and then when applied to some of these kind of other things like radiology is one where we saw these very interesting you know studies where we

AI in healthcare and radiology

06:12just didn't see the accuracy of a radiologist reflected in what the model was doing is that how you see it going forward or how did you reflect back on that >> yeah I I would say where we are now is I think of the models and I think a lot of people who use the models in practice >> you know where I'm going right where I'm trying to I'm trying to like understand Where are we really in the evolution of the world? >> Where's the arc? So, so I think a lot of you I I think and a lot of people who use the models in practice, I think think of the models as maybe a bit like a second opinion, right? Where you don't necessarily trust them alone. Just like if a doctor says something to me, I you know, I don't I don't necessarily take

06:52that as fact. They're an expert, but and they're often right, but they're not always right. Um uh so you know I'll tell you >> especially in this one category radiology it seems strange like it just can't get its >> I was going to say can't get its head around it but that's not really the right metaphor but doesn't have >> it does there is something about either it doesn't have enough data or it's not getting it's too nuanced. What is it? >> Yeah. So let me let me actually tell you a story about about one case where it where it where it did work. Um so you know we're talking about you know founding the company with with my sister. She recently had her second child while she was pregnant. Um, you

07:28know, pregnancy can suppress your immune system. Uh, was was viral. For some reason, it didn't occur to to any of them that it could be a bacterial infection. She took all her data. She uploaded it to Claude and, you know, Claude, you know, systematically listed the possibilities in the way that the doctors had somehow missed. >> That is amazing, right? How many of us have already used the models in some way to look at our healthcare? Raise your hand. >> Yeah, I think we're all having exactly that. >> That's that's a lot of people. It's amazing, right? Where you can kind of say especially how many of you are the hypochondrics that are doing that? Raise your hand. Yeah, I got a few. Okay, I know I'm

08:04>> But but but here's the point, which is Claude listed possibilities. If she had just blindly taken the advice of Claude, that would not have gone well. Um Claude was good for saying, "Hey, have you considered this? Have you considered that?" So, so it has this assistive augmentative ability >> especially in regards to words, right? It's very good on words. >> Words. Yeah. Yeah. I mean, you know this specialized stuff of reading images like I think that's more about specialized models like if you don't have the right data maybe it's not doing well again. It's less reflective of the >> is that the crux of the issue like in radiology because like when we're saying it's going to improve health care but we

08:44also saw this weird study where doctors who were using the models some in some cases were relying too much of the models and then ending up with bad outcomes. >> On the other side then we saw models that were like trying to read radiology but then were not able to achieve accuracy. There was like this kind of strange bifurcation between the two things. Yeah, you know, I think there's a lot still left to work out in what the role of the human is and what the role of the model is. And of course, that's changing as the models improve rapidly. But, you know, I think it's something you often see in new technologies where there's this great technology and then these practical issues of exactly how do

09:24you use it, exactly how do you interact with it? They take surprisingly long to work out. And you know if you look back at you know the early days when people you know kind of you know deployed deployed new technologies I don't know it's it's just it's often >> you think this is going to be about humans and agents working together. >> Yeah. No no no exactly just just the collaboration model needs to be worked out a bit more than it has in these practical areas. code is good to point to like code happens so fast right developers are early adopters and so that cycle we've gone through it in like months right you can see people talking in blog posts about oh you know what's the best way to use quad code do I want

Coding benchmarks and the tooling ecosystem

10:06it to suggest and I make edits or I write code and it makes edits and these people riff off each other and you know over the course of three or four months they work out what the best way to use quad code is I think that same thing is happening in healthcare and radiology technology, but because there are a lot of barriers in these industries, because they're high stakes, that that process is working itself out much slower. >> So, that's really kind of getting down to it, right? That it's a collaborative thing. Like, here's the best way to use cloud code. Cloud code is writing the code, but here's really the best way as a developer now to work on cloud code. It's not writing 100% of the code. It's

10:45writing code now. I'm like checking here's here's and now now I'm going to have to maintain the code. I'm going to have to provide all the capabilities the whole infrastructure around the code. Is this the right way to think about it? I I think that's the right way to think about it right now. Um uh you know as as the models get stronger then the collaboration may change to more distant supervision by the human. Um so it's it's very hard to say what will happen even six months from now. But but I think what you've said is a very accurate description of where we are now as of you know our our you know as of Sonnet 4.5 which was the latest model released at >> which is probably one of the best coders

11:23that we've seen right >> it it you know every every you know we have these benchmarks that are like you know what fraction of you know a random sample of GitHub PRs can this thing write from a description and that's over the last year and a half that's gone from like 5% which was basically the thing can't code at all to something like you know 77%. Uh so it's it's it's really >> and it's also yielding also this whole other ecosystem of things on top of it as well like even we showed agent force fives yesterday that I think our customers will benefit from or other you know other tools lovable these things have been built on these things right has that been a surprise to you to see a whole ecosystem on top of the model

12:05>> I mean that's been that's been great to see there are there are many different models right if you look at the GitHub co-pilots of the world you know that's kind of complimentary to GitHub if you look at the cursors of the world. That's kind of an IDE model. If you look at the lovables and replets, that's more of a consumer model. It allows people who are not traditionally software developers to write code. Um, you know, there's like a, you know, versel is like a different a different way of looking at it. Augment code, wind surf, you know, there's a whole ecosystem. And then there's the command line tools like like claude code. Uh, so I think it's too early to say, you know, what's going to

12:42be the winner there or or even if there is one winner, it's just possible there's just different things that work for different people or or different >> there's a whole civilization. >> There's a whole eos capability, right? It's a whole civilization like it's an eco it's even it's broader like it's really coming across the whole span. It it's really impressive, >> you know? It's it's it's so fundamental that it's able to yield all this incredible innovation from others. >> Yeah. No, this I mean >> which one of those is the most like the one you did not expect? >> Which one is the one that I most did not expect? Um I think the model of kind of allowing people to write code end to end like allowing non-coders to write code

13:24it totally makes sense once I see it and it's doing really well as a business model. But but I I I just when I was conceptualizing all the >> which product do you love the best? >> Uh yeah this is this is one of those questions that you know you know it's C your customers are like your children. You you you you really shouldn't give an answer. So, I'm not going to give an answer.

13:56All right. So, are you getting a sense for Dario, by the way? Are you getting a sense for Dario? Okay. So that's I'm trying to like you know help them to understand you and how you're thinking about this and where we're going. So now here you are, you know, you're entrepreneur, you've left Open AI, this thing is taking off. It's you and Daniela, you know, um you're in the Slack building, you're get where is it going to be in 5 years. What does it look like? We've seen zero to five years now. You don't have to tell me as an investor and there's nobody hold gonna hold you accountable for it, but just if you just fantasized out and now it's not 2025, but it's 2030 and you know where's Enthropic? What does it

Anthropic's five-year vision

14:43look like? Forget about all the other companies and all the other competitors, whatever. Nobody cares. Where's Enthropic 5 years from now? What's your dream big what's your big dream? So, you know, I I think I think maybe maybe one of the best ways to think about this is, you know, I wrote this essay called Machines of Loving Grace where where it's really it was written from a societal perspective, but you know, of course, the company is trying to enact that and and then contribute to it and so you can back out from that where we where we hope the the company will be. So, you know, I I really thought of it as like the upside of AI, right? There's lots of concerns with AI. I I you know I'm one of the loudest in voicing them

15:19but but I am you know I am I am I am incredible believer in the upside of AI you know I think I think more so than people who code or describe themselves as more optimistic I have just a really expansive view I am perhaps most excited about what it can do for curing diseases um you know going back to my background in biology like you know we look at these tough diseases like cancer like Alzheimer's um you know like like you know various kinds of metabolic disease like I I I just think AI is the right thing to cure them and so I think of all the work we can do working with pharmaceutical companies working with scientific research institutions and working with the health care system to kind of enact it out and you know you

16:03again our partnership is about regulated industries and one of the regulated industries I think especially is is healthcare but you know that is one of the most positive applications but I think that we will see applications across the industry I'm very excited about financial services. I'm very excited about insurance. I'm very excited about taking basically every enterprise um and learning to serve every enterprise with smarter and smarter models. And the skill ceiling is very high. One way to say it is what is an enterprise? It's kind of a super intelligence, right? It it is an entity that acts in the world in in in in in in a way with a level of strategy and and knowledge and power in the world that

16:45goes far beyond what any individual human can do. So, you know, even as these models get smarter than than than any than any individual human, there's a ceiling substantially higher than that in terms of doing the things that a company tries to do in the world. Um, you know, most of which are good for the world, most of which create economic value. Um and and and so then my my vision of Anthropic is a company that serves the enterprise of the world that transforms every every enterprise puts AI at the center at the center of every enterprise. Um uh you know and and and and and you know I think I won't give specific numbers but like if you if you think about the actual market potential of that it's very very high right our

17:27revenue has been going up 10x per year. It's currently, I think it was reported this morning, which I'm happy to confirm is correct, is approaching approaching 7 billion run rate. Um, that's gone up 10x a year for the for the last three years. It can't continue like that for much longer, of course, but uh you know, >> as an investor, I'm looking forward to the $70 billion year next year. >> You should not feel bad investing at 4 bill. You should not feel bad. >> Um uh uh uh so, you know, but I I I think the potential is is is very large. But but I think even even more important than that or perhaps even necessary to that success, we want to be known for, you know, the company that is trustworthy, the company that behaves

Revenue growth and trustworthiness

18:09responsibly, the company that has a that has a reputation for, you know, doing what it says, saying what it does, and building products that are trustworthy that, you know, if you if you if you build on us, you won't get any surprises. You won't get anything crazy. You won't get the model, you know, saying it's Hitler. you won't you won't get, you know, all kinds of um you won't get all kinds of crazy stuff. Um and and I think that's very important to us being good stewards of this. >> Are there certain companies that you're trying to imply that that might be >> None whatsoever. >> There's an entrepreneur down in Palo Alto. >> None at all. I don't know what you're talking about. >> Okay, I see. That's a certain letter in

18:50the alphabet. >> I don't I don't I I have no idea what you're talking about. Um >> Okay, it's before the letter Y. I sorry I still don't know what you mean. >> Okay, one more thing. I'm going to take a couple I'm going to use moderator privilege here and take a couple extra minutes with you. So number one, we see right now a tremendous fever on building data centers and you know we've seen a tremendous fever around the chips, the financing, the size and scale of the data centers, the energy of the data centers. you haven't really addressed it in your comments per se besides the scaling laws but let's talk about it just in part and parcel to anthropic and when you think about obviously one of your core partners has been Amazon it's

Data centers, compute and power needs

19:35been very important for your growth very important company um even though you do have you know obviously relationship with others as well but that one has been especially important to you know just give us a vision of what's happening with these data centers right now what are we going to see in the next few years with these these data centers and how do you think about it as an entrepreneur who's going to need data centers to grow? Who will you go to? How will you do it? Will you build your own? Will you just continue to partner? What what is your plan? >> Yeah. So, you know, my my my view is is that there is going to be an enormous need for compute for for all the reasons that uh you know that that we've laid

20:13out, right? The the scaling laws suggest that we're going to need more and more compute to train smarter and smarter models. And then as the revenue goes up, we're gonna we're, you know, we're going to need to serve more. We're going to need to serve more of our models. So when that revenue is not seven billion, but a hundred billion or more. Um uh and you know, when when model training requires millions of chips, you know, we're we're going to need many many gigawatts of power to to train the models. So we're trying to secure that power basically wherever we can. Um you know, and I'm confident in the company's uh I'm confident in the company's ability to do so. I I I do get a bit worried when the press focuses too much

20:52on the building of the data centers and deals to build data centers. Some of these deals seem a little bit fishy. They seem like there may be double counting. And I I also worry again >> deals are you ask talking about? >> I don't know what you're talking about. I have no idea. Um uh uh uh >> who is double counting? These are all public companies. I I >> we're going to want to know who's double counting. >> I don't I don't I don't know. I don't know anything more than you. >> What about the ones who are doing triple counting? >> I I don't There might be some of that, too. >> Oh, wow. All right. >> I don't know. Um uh uh uh but but but look look look look at the end of the day all this data center stuff is

21:30spending money, right? Um I think the thing that's most important is making money. Um and and so this is why I focus on, you know, the revenue and the actual work that we're doing with businesses. ROIC is growing faster than any other company in the space. The revenue is growing faster and you know I think I think basically if if you know if the if the demand is there if customers want the money and they're willing to pay you then you are willing to pay your compute providers and you'll get the compute you need at the end of the day. >> Daria, we could not be more grateful for you to be here to be in San Francisco to be a native son and to Daniela as well and for everything that you're doing with anthropic congratulations. Please

22:13give Dario a round of applause.