[0:00:07] Please welcome Dario Amadei, co-founder and CEO of Anthropic, and Aaron Levy, co-founder and CEO of Box. All right. Okay, that was a ton of announcements.
[0:00:25] And hopefully everybody's kind of still taking it in. We'll obviously have lots of demo stations and ways to start to leverage these technologies and play around with them. Super excited to have Dario here today. Dario is obviously at the forefront of AI breakthroughs on a daily basis. And so we're hoping maybe you'll even break some news today on where we're at and things. But maybe just to kick things off with an easy softball, when is AGI?
[0:00:52] Well, first of all, thanks for having me here. Yes. In terms of when is AGI, I don't know. I've gone on and on about how AGI is not a good term. There's only a continuous exponential. I would analogize it to if we're in 1995 and there was Moore's law. Someone says, when are we going to have supercomputers? The question doesn't make sense. A supercomputer is just a more powerful computer than we have today. Wait, are you insulting my opening question?
[0:01:20] Did I totally flub the first question? The models are getting powerful very fast. Do you have sort of milestones that you think about if AGI is not one of those milestones? Yeah, so I think a milestone that's pretty important is when can models reliably do professional work of the kind that many of your companies do?
[0:01:42] When can you go to a model and have it do, not necessarily complete jobs, but reasonable tasks at a level where a professional would look at it and say, this is up to snuff. I trust it in finance, in legal, in biomedical, in areas like insurance and productivity software, in programming.
[0:02:11] And I think not all those things will happen at once, but I think we're maybe at the start of a two-year period where we're going to pass successively all of those thresholds. I'm going to take us back in time in one second, but just while we're on this, and then we'll end up looping back. Where do you think, given that sort of paradigm of definition, where do you think we are with software coding? Yeah. Coding is one of the things I'm the most bullish on.
[0:02:35] And one of the reasons I'm so bullish on it is that both as you're training the models and as you're applying the models in the world, a property of coding is that one, it's easy to close a loop.
[0:02:48] You don't need to go off in the world like you would with robotics. You don't need to test drugs as you would in biomedicine. It's just, does the thing compile? Does it run? Does it produce the product that I want it to have? And during training, that allows you to close the loop without humans. When you're applying it, it allows you to say, hey, is this thing working? And it allows you to trust more because verification is automatic. So progress in code has been fast.
[0:03:15] With Claude on SWE bench, which is one of the measures of real world software engineering tasks, we've gone from 3%, that was the state of the art, not only at Anthropic, but everywhere, in January of this year.
[0:03:29] We just crossed 50% with the latest Sonnet 3.5 model. And what is this a metric of? So basically, SweeBench is a bunch of real-world pull requests or tasks. It's literally a slice of atomic units of work encoding.
[0:03:48] And so we have found that it correlates very well to doing things in the real world. And there's no cheating where the models are training on that data set and then thus getting better at that? No, it's kind of a held back data set with held back tests.
[0:04:05] We're getting much more rigorous about these evaluations because, of course, the models are very smart and models can cheat and so forth. Are you seeing an acceleration from that lower percentage to 50% or are we going to see diminishing returns going forward? You can only get to 100%, so we're going to see one of these S-curves where it levels off and then we'll get a harder benchmark that's like, these are things only Jeff Dean or whoever can do or pick your favorite legendary coder.
[0:04:34] So yeah, I'm sure there's always a new frontier. We'll always find new tasks. Yeah, so we'd love to get to the origin story of Anthropic. You were obviously early at OpenAI, and you guys had a view of another kind of approach. What's the origin story? How did that kind of get us to where we are today?
[0:04:59] Going back to when I was at OpenAI, I was there since almost the beginning, led research for several years there. One of the things that I and the colleagues who I worked with, who eventually became the co-founders of Anthropic, were most responsible for was the pivot from things like robotics and reinforcement learning to large language models. And we developed what's
[0:05:25] We were early proponents of what's come to be called the scaling hypothesis or the bitter lesson, which is the idea that more compute, bigger models, more data has this surprising ability to do a wide variety of tasks. The models nominally just predict the next word that sounds like fancy autocomplete, but turns out
[0:05:44] In those brains, in the service of doing that task, they get good at all the things we see today, like coding, like summarization, like document analysis, like just all kinds of sorting and manipulation of information, like the kind you do and the kind that all your customers do.
[0:06:07] As we realized this, I started to think about the implications of it, that we would need to be very careful about how these systems were very powerful. We need to be careful about how they were developed.
[0:06:18] One way you can think about it, the scaling hypothesis says you want a really powerful rocket, you need a big engine and you need fuel for your engine. Engines like the model, the fuel is like the data. But what else do you need? You need to steer the rocket. In fact, you need to steer it very precisely. If you want to go to the moon or Mars, you need to be accurate by like a hundredth of a degree or something like that. So in this analogy, this steering I think we realized would be very important.
[0:06:48] Folks were increasingly at companies I'd worked on getting on board with the scaling hypothesis. This idea of the importance of steering the models was one that we felt was really unique to us. And we see it in lots of practical ways with our enterprise customers.
[0:07:04] I think that the reliability of the models, the trustworthiness, the belief that they won't do unpredictable things is really holding back enterprise deployments of these models. We've been the best at this, and this is why we're growing so quickly in both the enterprise space.
[0:07:20] And the startup space. Our API business has grown way faster than that of anyone else. But still, this problem is unsolved. Still, there's much more to do in this area. We've been the best at it, but I think it's fair to say it's still not a fully solved problem. What was the alternative thesis? Who said that less data would actually be useful? Was there another approach? Like the alternative to the scaling hypothesis? I think the alternative was that there was some
[0:07:50] Algorithmic secret, some scientific secret we hadn't figured out and kind of undetermined, you know, does that need to be scaled up? Does something else happen? But if you went to like the large AI labs in say 2017, all of them, right? I was at several of them.
[0:08:05] The most popular hypothesis was just the idea that, hey, we didn't invent the right architecture and maybe deep learning itself was a dead end. So there were all these people looking in all of these directions. There are still people doing that. It still makes sense to do that. It makes sense to spread your bets. But the centrality of the scaling hypothesis was not, I think, understood.
[0:08:28] You know, we've all been blown away in the past two years by the rate of change, the performance of these models. Claude, you know, certainly right at the center of that. You know, some people think maybe we're at a data cliff or there's some point where maybe we can't scale up another order of magnitude or two order of magnitude or three. What's your view on that? Where are we at? Is this going to keep working? Yeah, so I'm quite bullish on data, actually. One, I think there are sources of data that, you know,
[0:08:55] There's a lot of data out there, right? And I think there are additional sources of data. What kind of data do you need at this point? Yeah, I mean, you know, I think generally you need high quality data, right? You can get a hundred trillion.
[0:09:09] Maybe a quadrillion words from the internet, but a lot of it's search engine optimization. So you need data that actually contains knowledge, that's not repetitive schemas or just whatever the internet is polluted with nowadays. So I think that's one direction, just generating more of it. Second is synthetic data. There's surprisingly many ways to generate synthetic data. If we go back to
[0:09:37] The AI world of eight years ago, even eight years ago when Google DeepMind produced AlphaGo, right? The second version of that, AlphaGo Zero, learned to beat the world master at Go just by playing against itself.
[0:09:53] It didn't train on human data. It didn't use anything from humans. It was just playing against itself. So you can think of that as a form of synthetic data. And we're working on similar things within Anthropic. And third is... And sorry, one more time, just for all of us who are trying to keep track of what's left to come in the industry. The synthetic data is going to be like you're going to generate an example contract or an example piece of code. Yes, data that is generated by the model.
[0:10:20] That looks like real data and then can be trained on. And you might say, well, isn't there something circular here? Like your model's generating data and then you need to train the model. And that's why I would point to the example of Go. That somehow we found ways for the model to just by playing against itself get much, much better. And so it sounds like it's circular. It sounds like you can't get a free lunch or produce something from nothing.
[0:10:48] If you do it right, it's not like a self-repelling machine or something. Yeah, I mean, if we look at all the complexity created by evolution or by the market or something, there are all these kind of infinite self-organizing processes that kind of allow you to make things better. So you think we're not, the data hypothesis hasn't changed, scaling hasn't changed. Maybe talk to us about the rate of improvement of models, whether that's on the dimension of speed, cost, or quality. How's that looking?
[0:11:18] Yeah, I mean, I think that's going to continue, right? Our general view of it from a product perspective has been there are different trade-offs between the intelligence of the model on one hand and the speed and cost of it on the other hand. And this is why we've offered different models, right? We have our line of models is Haiku is the small model.
[0:11:42] Cheap, fast, and as smart as it can be for how cheap and fast. We recently released a Haiku 3.5. There's Sonnet. That's a model in the middle. And Opus, which is the largest, smartest model, but is correspondingly a bit slower and slower and more expensive. And so our goal is to constantly shift out that frontier, that tomorrow's
[0:12:07] Small cheat model is as smart as yesterday's. You can see Opus being the Haiku in another. Yeah, Haiku 3.5, for example, is in most ways roughly comparable to Opus 3. Wow. So we've really shifted the curve very, very far. And that's going to continue. If we don't have, you know.
[0:12:26] If we're not able to have more data, we'll rely on synthetic data. We'll rely on new methods, like these reasoning type models that are increasingly being released. So the scale up is going to continue. We don't see any slowdown on the horizon whatsoever.
[0:12:47] You mentioned the API business. So just on stage, we announced Box AI Studio, where customers will be able to select Anthropic as an AI model for building an agent within Box, summarize documents, execute tasks, automate workflows, pull out structured metadata. Where do you see AI in the enterprise? How do you see it relating to us as many folks as knowledge workers, but maybe even non-knowledge work? How is AI going to augment what we're doing?
[0:13:15] Yeah, I would go back to my point about I have a strong instinct and we're starting to see evidence of it that the ability of AI to increase the productivity of businesses to disrupt various industries is much greater than what we've seen so far. You know, consumer is kind of out to an early lead in terms of use of AI, but we're seeing increasingly fast growth with enterprises.
[0:13:43] And the reason it's happening that way, the reason it's happening in that order, is actually this kind of reliability and steerability thing I talked about. You have this very powerful model. It's offered via API. You want to build something with it. And one of the issues, of course, is can I trust this? Can I trust this to do something customer facing? Can I even trust this to do something internally? Can I trust its quality of analysis?
[0:14:10] Over the last probably year and a half, it's really been a story of starting with very limited deployments to see how they go, and then scaling up from there. It's probably happened fastest at startups, which may build their business entirely around a model. We've seen some examples of companies that do that, especially in the coding space.
[0:14:36] There's much more value to be created within large enterprises. And I think we're starting to see that exponential growth increasingly. And so I think my main advice here for using models through Box or using models in general is
[0:14:54] Think big. Think in terms of the big deployments. A lot of these concerns like hallucination, reliability, accuracy, ability to reason, these are being month by month with releases from us, from other companies too for that matter.
[0:15:11] Problems are being mitigated month by month. And I think there's a very large amount of enterprise value to be created. The startups are leading the way, but it's starting to move towards larger companies. And we're starting to see the curve accelerate in a way that I think is going to catch up to and surpass the consumer use of AI.
[0:15:33] You know, there's a long standing debate, obviously, in Silicon Valley and elsewhere of is this replace jobs? Does this make us more productive and then thus create jobs? What's your current view, at least in the near and medium term? Yeah, yeah. So certainly, you know, certainly in the short and medium term, I think this is going to make people more productive and, you know, enable them to do it. And I think it will move things uniformly, right? It will move, you know, the middle, the average worker, but it will also enhance the capability of the most.
[0:16:02] The most capable people. There's an economic principle called comparative advantage, and I think it's very powerful. Basically, let's say.
[0:16:12] You're doing a job, and an AI system suddenly is able to do 90% of your job. You might think, oh my god, there's going to be 10% less jobs for people like me, or I'm going to get paid 10% as much. But what really happens is that the humans focus on that 10%, and that either expands to fill the entire job, or there are new things humans can do that can complement the way in which the AI can do the job and can
[0:16:40] Can mitigate the weaknesses of the AI system. And so it's really kind of a moving target. The economy finds new things for people to do. We used to have lots of jobs as typists. You had to write things out yourself. Now it's really easy.
[0:16:57] to do documents and like to share them. And so there's a lot that people used to do that they don't do anymore. But there's all kinds of things you build on top of them, and there's more time for people to focus on the content. So again, maybe there's some world of super intelligent AI that's better at literally everything, but we're not in that world. And what I see when I talk to and work with enterprises is
[0:17:22] Creation of value and increased leveraging of humans rather than replacement of humans. There's a study from the economist Erik Brynjolfsson, who's been studying the economics of AI and automation for, I think it's a couple of decades. He wrote a book called The Second Machine Age. That must have been 10 or 15 years ago. And he said he did a study where he showed that
[0:17:48] By default, enterprises may try to replace humans with AI, but their productivity gains are much less when they do that. Productivity gains are greater when they think of ways to complement. It's laziness that causes the kind of replacement narrative. If you think more about where the gaps actually are,
[0:18:06] You have instead a complement narrative, and that tends to actually increase productivity more because it's more targeted at what the AIs are actually good at and what the humans are actually good at. Yeah, I tend to find maybe two flaws in the thinking of stagnation with AI from a job standpoint. The first is that we're not in competitive fields.
[0:18:29] So there's sort of this idea that, you know, to your point about the productivity gain, I'm going to use AI, I'm going to automate something, and I'm going to bank those gains, you know, to some sort of, you know, kind of ephemeral amorphous profit line that I need to, you know, hit, which, you know, in practice, I have competition. And if that competition doesn't bank those gains, and instead they reinvest those gains back into their company.
[0:18:51] And their productivity, then I'm going to basically lose and I'll be out of business. And so a lot of times when you get a productivity gain, let's say you make sales reps more productive or coders more productive, you just reinvest those dollars back into that same function.
[0:19:04] Because you can get more done with it. Or a different function. Or a different function. Whatever the next. Whatever it is. This is like the lump of labor. There's not a fixed amount of work to be done in the economy. Like if you get more efficient at doing some work, then skill reallocates and retrains to. Which gets to then the second thing, which is like this sort of.
[0:19:25] Fix some kind of mindset, which is like, if we could just do that one task that AI made faster, we're just on to the next task. There's not a finite number of tasks for us to go do. There's not a finite number of, you know. My biggest problem, let's say dealing with the health care system, is actually how long it takes to get scheduled for the next health care checkup you want to do. The idea that we could automate and make a doctor 10% more efficient, that just means I'm next in line.
[0:19:53] That much faster in the healthcare system. There's an infinite amount of demand for so many of these industries if AI could actually make us more productive. So that brings us to an essay that you just wrote a few weeks back. And you had this idea, which is a super profound concept of sort of the compressed 21st century. What was the thesis there? Where do you expect to kind of see that play out next? Yeah, so my basic idea was, again, you know, if we get to the point where, you know,
[0:20:20] AI is kind of past the professional level at most tasks, and we're able to build millions of these systems.
[0:20:27] You could think of this as you might describe it as a country of geniuses in the data center. And that's a bizarre situation, right? Civilization has never been in that situation before. What happens when you just instantly invent everything that could be invented? And I give a bunch of reasons why I don't think it'll go that way. But what I do think could happen is that if we look at the next 100 years of what we are
[0:20:52] trying to invent in science and engineering. Perhaps all of that could happen in five or 10 years.
[0:20:58] Because I used to be a biologist, I focused particularly on the biology side of things, on biomedical discoveries, ranging from academics to biotech companies to large pharma companies. And I think there's really potential to conquer many of the afflictions we still face. A lot of the easy ones, like diseases that were addressed by sanitation or vaccination or antibiotics we've solved, but things like cancer and Alzheimer's disease,
[0:21:27] Much more complicated. And so I'm wondering if AI is really what we need to understand that complexity and to surmount those diseases, frankly, much faster than I think most of us are imagining. We're getting used to a world where
[0:21:41] Those diseases are very hard to address and progress is very slow. I don't think it needs to be that way. I think if we get it right, these incurable diseases, we could actually overcome and we'll look back at them the way we look back at bubonic plague or mumps or whatever it is nowadays.
[0:22:00] Yeah, I try to not look back on the bubonic plague, so that would be nice to move on from whatever that version is now. So compressed 21st century, kind of acceleration across so many different spaces. Maybe to bring us home, what's your favorite thing to do with AI at the moment? Well, so one, I really like it for coding. I mean, it may just be that Claude, we spend a lot of time on coding for Claude.
[0:22:27] I think the second thing is, you know, like writing. So, you know, I write this, I've written this essay, but also... Wait, was that not you that I read? You know, I use Claude to help me. Every word comes from me. I actually find like the style of all existing LLMs isn't quite good enough if I'm trying to write something really good, but it's really good at kind of like generating ideas and helping me think things through. And then the final version, I kind of write myself. Yeah.
[0:22:55] Awesome. Ladies and gentlemen, Dario, we appreciate the partnership. We're excited to keep working with you guys. And thanks for coming out to the box office. Thanks for having me. Thank you.