- Anthropic is experiencing "exponential" AI growth, with usage and revenue exceeding internal 10x predictions by far, creating challenges such as a high demand for compute resources.
- Developers are at the core of Anthropic's strategy, providing essential feedback and leveraging AI to build transformative applications and even entire businesses.
- The future of AI-powered development involves a shift towards multi-agent systems and organizational productivity, which necessitates addressing new bottlenecks highlighted by Amdahl's Law.
A conversation with Dario Amodei & Daniela Amodei
- Prepare for hyper-growth: The AI industry can experience explosive growth (e.g., 80x annualized in Anthropic's case), demanding rapid adaptation and securing resources like compute.
- Prioritize developers as critical users: Developers are considered the most important users, providing invaluable, honest feedback that drives product improvement and technology advancement.
- AI empowers individuals to create significant value: AI models are enabling single individuals or tiny teams to achieve unprecedented levels of productivity and build high-value businesses, potentially reaching "billion-dollar company" status with minimal human resources.
- Transition to multi-agent AI systems: Future development will move beyond single AI interactions to complex "multi-agent systems" where various AI entities cooperate and manage tasks within a hierarchical structure.
- Apply Amdahl's Law to AI development: As AI accelerates coding, identify and address new bottlenecks such as security, verification, design quality, and technical debt, requiring models to improve in these "soft" skills.
- Product innovation is model-capability driven: Product development in AI must be agile, constantly experimenting and revisiting ideas, as new model capabilities rapidly unlock previously impossible products.
- API-first strategy remains vital: The continuous evolution of AI ensures that the API market will remain a crucial avenue for developers to integrate new model capabilities into diverse applications across industries.
- Leverage AI to accelerate internal development: Use AI models to speed up your own product development processes, but be mindful of increased technical debt and develop strategies to manage it with AI.
exponential growth— A rapid, accelerating rate of increase where a quantity grows by a fixed percentage over successive periods.scaling laws— Empirical relationships describing how the performance of AI models (e.g., accuracy) improves predictably with increases in compute, data, and model parameters.Amdahl's Law— A formula that gives the theoretical speedup in latency of the execution of a task at fixed workload that can be expected of a system whose resources are improved. It highlights that the part of the system that cannot be improved limits overall performance.multi-agent systems— Software systems composed of multiple interacting intelligent agents that coordinate to solve problems that are difficult for a single agent or monolithic system.technical debt— The additional rework caused by choosing an easy solution now instead of a better approach that would take longer, accumulating costs over time.generative form factor— Refers to the design and interface of AI products primarily focused on creating new content, such as code, text, or images, rather than just processing existing information.API market— The ecosystem where Application Programming Interfaces are offered and consumed, allowing developers to integrate specific functionalities (like AI capabilities) into their own applications.
Thank you for joining us again. And I am so excited for this conversation with Dario and Daniella. Let's give them a round. Woo! It is a delight to have you here at Code with Claude. We've been having a great day of sessions, demos, customer sessions, all sorts of fun stuff. But I wanted to ask you, maybe just a little bit zooming out, we talked a lot about the exponential in some of our conversations this morning. And what it feels like to be on the exponential. And so as two people who are definitely on the exponential, we've talked a lot about growth and what it feels like. What does it feel like for you all? Well, first of all, it's great to be here. Thank you so much for having us. You know, at Anthropic, we have this little slack emoji of the roller coaster. You guys know the one I'm talking about. But it's an inflected roller coaster. So it's almost like it's going straight up. And I think of me and Dario as writing at the front and the back of that roller coaster. I don't know if you all have ever, maybe not recently, been on a roller coaster. I'm not always sure which one of us is in the front, and which one of us is in the back. But you get a different type of whiplash, depending on which side you're on. And I think that's probably the best encapsulation of what it has felt like. It's like we're having a lot of fun. There's a ton of adrenaline. We're not totally sure that the operator of the roller coaster isn't like a 15-year-old. It's your personal job of questionable level of sounds minds. But it's been great. It's fun. It's an adventure. There's a lot going on. Yeah. The way I always think about the exponential is, it was me and the other co-founders who first predicted it through the scaling laws over 10 years ago, and we wrote down these lines on graphs that say, well, first we're going to spend $1,000 in AIMON, then $10,000, and then $100,000. It's going to go all the way to hundreds of billions. And the model is going to be this good at this task, and this good at this other task, and that good at coding. And it is a remarkable experience to write down these lines on graphs and have the predictions come true. So in that sense, they are not surprising at all. And yet the actual experience of seeing what it's like is just so crazy that you're shocked anyway, even though what you wrote down on the graph is exactly what happened. I'm always reminded of like, there's this famous scene in the movie Interstellar where they go to this planet that's very close to a black hole. And so the planet has these waves that are like 2,000 feet high. And I was a physicist. I know the math, the general relativity, how much things can be assured. But actually seeing it on human scale, there's something deeply strange and unsettling about seeing it actually happen. And that's what it felt like every year at Anthropic. And I feel like this is the first year where the rest of the world is seeing it with us because we're so much in the spotlight. It applies to the internal growth of the company. It applies to our own work within the company. We're the first time we've seen the number of internal PRs inflict upwards due to the work that Claude is doing. And we've seen it externally because actually this is the first year we've grown faster than the exponential. So we tried to plan very well for a world of 10x growth per year. In the first quarter of this year, we saw if you were to annualize it, ADX growth per year, in revenue and usage. And so that is the reason we have had difficulties with compute. We've planned for anything from it only grows a little to it grows 10x. And yet we saw ADX. And so as you saw it today with the SpaceX compute deal, we're working as quickly as possible to provide more compute than we have in the past. We'll continue to do so. We'll pass that compute on to you as soon as we can do so. I guess I hope the ADX growth doesn't continue, because that's just crazy. And it's too hard to handle. I hope for some more normal numbers. A mere 10x. But we will manage it absolutely as best we can. And we're every day trying to obtain even more compute that we can pass on to you. We're sorry if sometimes it takes some time. But we're going to keep going to acquire as much as we can. Awesome. Yes, yay for compute right limits. You know, this is an audience of developers and builders. And that's really what today is about is about how we're making our platform better. Because developers, if you help us close the gap between what the models can do and what they're actually doing for real people out there. And both of you talk a lot internally about the importance of developers and builders. Maybe Danielle, let me start with you. How do you think about supporting this ecosystem, supporting this community? Yeah, I mean, I'll start out by saying I think in many ways, developers are the most important users of Claude. I think for a variety of reasons. One is, anthropic ourselves are majority developers. If you think about how we develop this technology, what we're building, we learn so much from the developer community. It's the best partnership. Because we, first of all, I think it's a group that gives honest feedback. And so I think that is actually really hard to get. You know what I mean? It's like, you build a product and you're like, I see some numbers, those are nice. But the genuineness with which the developer community I think engages with us is something that is so special. We have tried really from day one, I think anthropic has always primarily built for developers for businesses. I think we're a little bit unique in the AI ecosystem for that reason. And I think we have been very fortunate to be able to benefit from the feedback, from the engagement, from the community development. Developers are, in my experience, they're very ecosystem and community oriented, which I think we are too. We're like, how do we build for this sort of broader ecosystem of people who are developing, by the way? Some of the most inspiring transformative technologies and building the most incredible companies in the entire world. There's been this renaissance of things in medicine and software development and financial service. I mean, it's like you pick the industry and there is an incredible developer or an incredible developer-based company that is transforming that industry, help leveraging our tools sometimes. And I think that's such a special, that's both a privilege and a responsibility that I think anthropocolds to say, developers are really the backbone of how we learn, how we build better tools for all of you. And I think that's a really special relationship that we feel like pride in and also a responsibility towards. Yeah, I mean. And this is an example of the developer community. I think part of what Danielle is saying is feedback as a gift. We hear the positive. We also hear the negative. Please keep it coming. It is part of how we know what's working and what isn't. And so we really value, and I'm sure as you're talking to people around, we really value all of that. And it helps us know what to do better. Sorry, Daria, back to you. Yeah, one thing I would say is that technology doesn't diffuse at kind of an even pace across the economy. Right? And I think there's a spectrum where the software engineers are the ones who are fastest to adopt new, kind of fastest to adopt new technology. That's why there's so much kind of focus on this area. And it's the beginning of it. But it's like, it's a foreshadowing of how things are going to work across the economy. And how the economy is going to be transformed by AI. So I think getting this right and really making it work for this community, it's kind of like a microcosm of how we have to, you know, of how we have to make it work across the world. And I think one thing that, you know, a dynamic we should watch, it was about, I think it was roughly a year ago, there was an event like this where Mike Krieger asked me, you know, when will there be the first billion dollar company with one person? And I said 2026. And I think we're actually on track to achieve. It hadn't quite happened yet. There's been like two person companies that are $1 billion built with AI. There's been like one person that's worth $700 million. But we got seven more months. So it's kind of, no, we do it. It's seven more months since 2026 or eight. And you know, that's, there's an eternity on the exponential. But what I'm trying to say with this though is that there's an enormous ability for one person or a tiny set of people to do a set of things that are incredible, right? Where, you know, before, if, you know, you just had an idea and you had a vision like there's so many resources you'd have to accumulate over several years in order to make that vision happen. And I think there's a very unique opportunity for single individuals or very tiny teams to do things that are incredible, right? Where I think we moved from the models are writing code to, you know, the models are helping us think of software engineering as a task. To the models are helping us think of like, how can I build a business or an economic unit as a task? And so there's an extraordinary amount of opportunity for people in this room to kind of take advantage of that. Yeah, it really feels like it's kind of removing barriers. You know, there were all these barriers to like creating that kind of value in the world. And now, I mean, I guess the gauntlet has been thrown. We've got an eight month timer. And I'm excited to see what comes of it. I would love to maybe hear what's gonna change for developers. So Dario, you talked a little bit about what they can do now and you've talked, you know, in the past about how you expect a Claude to build more with Claude. Can you just talk about how you expect things to move? Yeah, I mean, I think there's several, like maybe trends. One is going from single agents to kind of multiple agents. So the idea that you have a bunch of these Claude rights, like managing a team, right? You have a bunch of quads running and like, you know, you kind of farm bunch of things out to your quads and maybe some of the quads farm things out to other quads with like different press. So you have a kind of whole hierarchy or a whole, we're gradually making our way to like the country of geniuses in the data center. You know, we're starting with like a team of smart people in a room or something. We're working our way up to the different exponential in the country. So I think that's one trend that we're kind of already starting to see and we're already like offering tools that can help do that. I think a trend that's related to that is like, you know, what we've done so far with Claude Code is like, you know, it helps kind of individuals to be more productive. But I think increasingly we're going to start thinking at the level of whole teams and organizations and how can you make whole teams and whole organizations more productive in a way that is kind of more than just the sum of its parts. And then finally, you know, I think in this area as with everywhere else, if you want to think about what's next when something's working really well you should always think about Amdol's law which is you speed one thing up, what are the things you're not speeding up. And so I think there are a bunch of things like security, like verification, like just if you're living in a world where you can within an organization write three or four times as many PRs as you could previously. You start to understand there are all these other things that are holding you back or that will go wrong if you speed up just that and not everything else. And so working to speed up those kind of other things so that we can greatly increase people's productivity but we can do it smoothly and, you know, smoothly and productively and reliably. I think that's gonna be very important. Does that have any impact on how you think about training new models or, you know, what the future of models looks like? Yeah, I mean, you know, that's true on several levels. I mean, I've already said many times we're using clawed to speed up clawed, right? That's kind of already something that's happening. But I also wonder if the things we're trying to do with the models could also influence how we build them. So when I talk about these things like, you know, kind of verification or kind of design quality or things like that, like one of the reasons training models for code and software engineering has gone so fast is that you have this verifiability, right? Where you train the model and it's like, you know, you're able to verify it by running unit tests and so that it has a lot of properties that make simplify the process of training. But what you discover is that there are these, of course, aspects of the job that are not verifiable, right? And, you know, some of the, you know, is this thing really right? Can we find errors or their security issues not quite as verifiable? And so training the models to be better at that, which I think will also make the models better at other things where they haven't made progress as fast as coding, like their ability to ride or their ability to kind of do, do, you know, to do, you know, less, less objective, scientific tasks. So I think it's gonna have benefits in many, in many other areas, but, you know, I think we find even within software engineering this, you know, these, these kind of soft or somewhat subjective skills and abilities are become surprisingly important because of M-Dullslaw. I would love to hear, you know, we talk a lot about our mission internally. Like, Daniela, as we just keep growing and the stakes of this whole industry keep getting higher, what should people know about our mission and about us as a company? You know, I think, when I think about what Anthropic is trying to do, there's these sort of like two, maybe two pillars, right? The first is around, how do we develop this transformative technology in a way that is good for everybody, right? And I think this goal of, Claude is this incredible tool. It has the power to really transform, you know, what people build and how they create and the ambition level of what they can develop themselves. And I think there comes a huge amount of opportunity there and there's also some risk, right? I think that we've talked about this a lot publicly. There's some risk to just labor disruption. There's risks to ensuring that the technology is developed safely, that it's good for people. And I think Anthropic's job or what we try to do is really think about how to balance these two things in equal measure. We have this internal cultural value called hold light and shade. And I think that it is such a good encapsulation of, you know, what we see about how the technology is being leveraged today and also just our approach to putting the technology out into the world, right? I think mythos and glass, we are a great example of this. The potential to build something incredible with a model that capable is so vast. And we want to be a little bit careful about how we release it because of some of the security vulnerabilities, right? And I think that this is this kind of complicated dance that we do where we're like, we really want to get stuff out as quickly as we can. We're trying to build the best products and release the most powerful models. And we're just trying to do it responsibly. I think that is really the underpinning of the majority of actions that we take is sort of grounded across those two pillars. I think that is, like, one of the things I find most meaningful about getting to work here is just thinking about how much everything is changing. And the fact that we're kind of all building in the industry right now, to me, it feels like we're getting the chance to have a vote in how everything unfolds and the trade-offs you're describing. That's just what I think of when I think of Hold Light and Shade is as things move so rapidly. We get to build experiences that other people use to understand what the future looks like. So that's something I'm always really personally excited about. Maybe I'll ask a little bit about product. Both of you kind of lean in quite a lot on the product side. One thing that we talk a lot about is building for the exponential. I always think about product as kind of a bridge between the technology that exists and the problems that people have. And it's just a very interesting time because the technology is changing much more rapidly than we're used to. Can you talk a little bit about how you think about product building in this world? So I love, I mean, I love your way of putting that. It's like, Dario and Daniela lean in a lot. And this cheap product episode, what Abby needs is like, you guys are up in my business all the damn time. Can you please leave me alone and let me do my damn job? I'll feed back as a gift. I'll enjoy every perspective. But no, Abby is right. I think you have a hard job that you wear incredibly well. But I think in all seriousness, Dario and I both, we care a lot about the product. It's a representation of what we are trying to build at Anthropic. We want it to be useful for people. We want it to be accessible. We want the product to be good, right? Part of why I think we're leaned in so much is we feel very invested in ensuring that people that are using, Claude, are getting out of it, everything that they can. And so I think our bothering you is really, I think our way of feeling like to the degree we can we're standing up for our customers. We're standing up for our users who are building sometimes their whole business around the premise of what these AI models are capable of doing. And I think the other thing, maybe the thing that makes product at Anthropic unusual or different than how I've seen it done, what other companies I've worked at is product is sort of one input and research is another. I'm sure you've felt this in your role. Sometimes we're like, man, this is a great place where we should just be able to build something better that's easier to use. Or here's a product idea we really want to enable people to be able to access it right out of the gate. But a lot of the time, product innovation is driven by what new capabilities emerge in the model. And I think coding is actually a great example of this. We didn't just sit out and say like, from day one, we're going to build a coding product. It was like once we saw that the models were able to write code at a reasonably accurate level, not perfect, we were like, huh, this is interesting. It seems like a lot of people that are kind of clawed files are developers, right? And they're using it to write code. This has always been a community that we've worked well with. We've wanted to lean in with and engage in support. Should we build a tool for them, right? Should we build something that actually is going to enable people to be better at doing their day-to-day work in this way? But I think that's just an interesting dance inside the company where there's a component that is sort of, I don't want to say traditional product because nothing about Anthropic is traditional. But I think there's a component that looks like a normal product organization. And then there's part of the organization that's like what is new from the models? What's happening on the research side? And how do we get to marry those two things together? Yeah, I would maybe take this from two lenses, which is building products for AI and building products with AI. And I think the internal experience at Anthropic, month by month and week by week, has given us lessons about both. So I think over the last few years, learned a lot of lessons about building products with AI. And in some ways, it's been an advantage that I was a researcher. I was never in the era of building products without AI. So it's like you can end up in a situation where there are things you don't have to unlearn. You can just learn the new world from scratch. I think you got the essential difference, which is that if you go back to the product there in the 2010s, you had a slowly changing technological background. You were trying to do new things with the kind of technology that was present. And of course, every once in a while, you'd have a new framework or a new way of doing it, but relatively slow. AI is moving lightning fast. And so there are a few consequences of that. One is that there are new products that are not possible with a given capability of model. But then when you take the next step, when you go far enough along the exponential, then suddenly they light up, suddenly they become possible. And so it puts a premium on internal experimentation. Because you always want to be trying something. Even if you tried something that didn't work, you want to revisit it a few months later, because it might work then when it didn't work before, which sounds a little bit crazy. But if we had tried to do Claude Code in like 2022, it wouldn't have worked because the models wouldn't have been strong enough. It was a frustrating experience. We did have some early things that were a little like Claude Code in 2022. And it was like, oh, this is interesting, but you couldn't actually derive value from it because the models were dumb. So that's, they were. I've been training these models since 2015. They were really dumb. They were really dumb back then. The second thing is that products reached their saturation. When models start to get too good. So I think this has happened with chatbots, right? Like, it's a big market. Lots of people use it. It's going to stay around. But the ways in which we're making model smarter today are much more evident in today's Claude Code form factor. And in more generally, in the genetic form factor, then they are in the chatbot code factor. So that's kind of the other side of the coin, which is that you always have to be thinking about what the new thing is, right? The way this business works isn't you make a product. It becomes very big. And then all the kind of stability set saying that you have to ask, you're constantly, not only are you able to make something new, you're constantly needing to make something new, or at least update the things that you've made. So it's kind of constantly an innovation laboratory. And I think the other thing it means of particular relevance for developers and software engineers is API never really goes away as a market. Because the fact that it's always possible to build new products that's true inside Anthropic, it's also true outside Anthropic. And so I actually think both within code as an application, but just coders writing code with Claude for medical applications, for law, for finance, or they're going to keep me new applications because the models get smarter and they enable them. So that's building in the age of AI. There's a newer thing we've seen maybe over the last year or six months, which is building with AI itself, like using AI to enable the process of doing product development faster. That one is interesting. And again, we would go back to our old friend, Amdol's Law, which is we've found with the internal model acceleration. You can write two times as many, four times as many, five times many, you see this within the company. But then you see what breaks. We've been able to ship a lot more products than we could a year ago and there of pretty high quality. Or I hope you think that. But it is possible to accumulate an extraordinary amount of internal technical debt when you ship that fast. And so then you have to say, well, can we also use the AI models to undo that technical debt or keep track of what it is that we're doing? And then you learn the team has to work together in a totally different way. And these revelations come month by month, and you kind of learn how to do things in a totally new way. So it somehow it increases the tempo, not just of building, but in which the way you have to change the way you build as a team. I really experience that. I think the hardest part of it is you can get really familiar with the problems because the problems don't change. That fast, the problems are about humans. Like we all, we're going to have similar problems. But a thing that's hard is to learn to be fresh eyed all the time about what the technology can do. And constantly scan for that. And then I also feel on a personal basis where your job just changes because you hit a new bottleneck. And the way you spend your day is just a little different. Maybe I'll just ask a couple more questions. Maybe Dario, very quickly, you talked a lot about model capabilities and how they change so often. Is there one thing that makes you most excited in the models that are coming in the next call it six months? I would say this idea of thinking at the level of the organization rather than just one person. And again, it ties to this one person, billion dollar business, which maybe that will turn out to be an underestimate. But just kind of the idea that you can both have AI do the work of many people but that when you have a team of humans, that AI is not just doing the work of many people working for one person, but that it does the work of many people many times over by operating within an organization of humans. Awesome. And maybe I'll just ask to close. Danielle, one thing we talked about is some of the kind of developers or use cases we see that are just so inspiring. I talked a bit in the keynote about some of my favorites, whether that's people using it for efficiency. It places like Stripe doing major info upgrades, but also people using it for these very personal and specific ways like connecting kids to foster families faster. And I would just love to hear, what are some of your favorite examples of how you see people using the tools? Yeah, I mean, I think even going back to the days when I was at Stripe, something that I've just loved about the developer community in general is like for every interesting challenge or problem there is in the world, there's like an incredible dev somewhere who's like trying to make that thing better. And I think getting to watch people build things with cloth that are just bringing so much utility and value and meaning to people around the world is really inspiring. I think for me personally, a pilot that I've seen with some developers who are building these interfaces basically for mobile doctors in the global south. So places where it's really hard to reach an actual doctor because of the access issues, right? You're down a dirt road somewhere, 50 miles from the nearest city, but people there still encounter health challenges. How do you actually like work with this really smart technology to build these like interfaces where people can just ask the question, it can give you some medical advice, right? In a way that's like sanctioned. I'm also just like in general blown away by the medical research field and what is coming out of cloth being able to accelerate biomedical research across many of the developers that work with us, that build on us. There are also just some really heartwarming individual examples we have this user happiness channel at work. So my favorite ones, there's a developer who used cloth to help retrieve their wedding photos that were corrupted like on a corrupted hard drive. I thought that was so sweet. And just like such a personal use case, but so like so meaningful. And then my last one was just so random is, someone is using cloth to chart the growth of their tomato plants in their garden. And I was like never in a million years would I have thought of it. But it's like, I was like, do you have a live camon? Can I switch out to this? Because I would really love to see it. But I think just the range of things that people are able to do is just astonishing. Well, Dario Daniela, thank you both so much for spending time with us here. It's a pleasure to see you. Thank you to all of you for joining us. We've got a great rest of the day coming. But please join me in giving around to applause. Please go find out. Thank you.
TL;DR
- Anthropic is experiencing "exponential" AI growth, with usage and revenue exceeding internal 10x predictions by far, creating challenges such as a high demand for compute resources.
- Developers are at the core of Anthropic's strategy, providing essential feedback and leveraging AI to build transformative applications and even entire businesses.
- The future of AI-powered development involves a shift towards multi-agent systems and organizational productivity, which necessitates addressing new bottlenecks highlighted by Amdahl's Law.
Takeaways
- Prepare for hyper-growth: The AI industry can experience explosive growth (e.g., 80x annualized in Anthropic's case), demanding rapid adaptation and securing resources like compute.
- Prioritize developers as critical users: Developers are considered the most important users, providing invaluable, honest feedback that drives product improvement and technology advancement.
- AI empowers individuals to create significant value: AI models are enabling single individuals or tiny teams to achieve unprecedented levels of productivity and build high-value businesses, potentially reaching "billion-dollar company" status with minimal human resources.
- Transition to multi-agent AI systems: Future development will move beyond single AI interactions to complex "multi-agent systems" where various AI entities cooperate and manage tasks within a hierarchical structure.
- Apply Amdahl's Law to AI development: As AI accelerates coding, identify and address new bottlenecks such as security, verification, design quality, and technical debt, requiring models to improve in these "soft" skills.
- Product innovation is model-capability driven: Product development in AI must be agile, constantly experimenting and revisiting ideas, as new model capabilities rapidly unlock previously impossible products.
- API-first strategy remains vital: The continuous evolution of AI ensures that the API market will remain a crucial avenue for developers to integrate new model capabilities into diverse applications across industries.
- Leverage AI to accelerate internal development: Use AI models to speed up your own product development processes, but be mindful of increased technical debt and develop strategies to manage it with AI.
Vocabulary
exponential growth— A rapid, accelerating rate of increase where a quantity grows by a fixed percentage over successive periods.scaling laws— Empirical relationships describing how the performance of AI models (e.g., accuracy) improves predictably with increases in compute, data, and model parameters.Amdahl's Law— A formula that gives the theoretical speedup in latency of the execution of a task at fixed workload that can be expected of a system whose resources are improved. It highlights that the part of the system that cannot be improved limits overall performance.multi-agent systems— Software systems composed of multiple interacting intelligent agents that coordinate to solve problems that are difficult for a single agent or monolithic system.technical debt— The additional rework caused by choosing an easy solution now instead of a better approach that would take longer, accumulating costs over time.generative form factor— Refers to the design and interface of AI products primarily focused on creating new content, such as code, text, or images, rather than just processing existing information.API market— The ecosystem where Application Programming Interfaces are offered and consumed, allowing developers to integrate specific functionalities (like AI capabilities) into their own applications.
Transcript
Thank you for joining us again. And I am so excited for this conversation with Dario and Daniella. Let's give them a round. Woo! It is a delight to have you here at Code with Claude. We've been having a great day of sessions, demos, customer sessions, all sorts of fun stuff. But I wanted to ask you, maybe just a little bit zooming out, we talked a lot about the exponential in some of our conversations this morning. And what it feels like to be on the exponential. And so as two people who are definitely on the exponential, we've talked a lot about growth and what it feels like. What does it feel like for you all? Well, first of all, it's great to be here. Thank you so much for having us. You know, at Anthropic, we have this little slack emoji of the roller coaster. You guys know the one I'm talking about. But it's an inflected roller coaster. So it's almost like it's going straight up. And I think of me and Dario as writing at the front and the back of that roller coaster. I don't know if you all have ever, maybe not recently, been on a roller coaster. I'm not always sure which one of us is in the front, and which one of us is in the back. But you get a different type of whiplash, depending on which side you're on. And I think that's probably the best encapsulation of what it has felt like. It's like we're having a lot of fun. There's a ton of adrenaline. We're not totally sure that the operator of the roller coaster isn't like a 15-year-old. It's your personal job of questionable level of sounds minds. But it's been great. It's fun. It's an adventure. There's a lot going on. Yeah. The way I always think about the exponential is, it was me and the other co-founders who first predicted it through the scaling laws over 10 years ago, and we wrote down these lines on graphs that say, well, first we're going to spend $1,000 in AIMON, then $10,000, and then $100,000. It's going to go all the way to hundreds of billions. And the model is going to be this good at this task, and this good at this other task, and that good at coding. And it is a remarkable experience to write down these lines on graphs and have the predictions come true. So in that sense, they are not surprising at all. And yet the actual experience of seeing what it's like is just so crazy that you're shocked anyway, even though what you wrote down on the graph is exactly what happened. I'm always reminded of like, there's this famous scene in the movie Interstellar where they go to this planet that's very close to a black hole. And so the planet has these waves that are like 2,000 feet high. And I was a physicist. I know the math, the general relativity, how much things can be assured. But actually seeing it on human scale, there's something deeply strange and unsettling about seeing it actually happen. And that's what it felt like every year at Anthropic. And I feel like this is the first year where the rest of the world is seeing it with us because we're so much in the spotlight. It applies to the internal growth of the company. It applies to our own work within the company. We're the first time we've seen the number of internal PRs inflict upwards due to the work that Claude is doing. And we've seen it externally because actually this is the first year we've grown faster than the exponential. So we tried to plan very well for a world of 10x growth per year. In the first quarter of this year, we saw if you were to annualize it, ADX growth per year, in revenue and usage. And so that is the reason we have had difficulties with compute. We've planned for anything from it only grows a little to it grows 10x. And yet we saw ADX. And so as you saw it today with the SpaceX compute deal, we're working as quickly as possible to provide more compute than we have in the past. We'll continue to do so. We'll pass that compute on to you as soon as we can do so. I guess I hope the ADX growth doesn't continue, because that's just crazy. And it's too hard to handle. I hope for some more normal numbers. A mere 10x. But we will manage it absolutely as best we can. And we're every day trying to obtain even more compute that we can pass on to you. We're sorry if sometimes it takes some time. But we're going to keep going to acquire as much as we can. Awesome. Yes, yay for compute right limits. You know, this is an audience of developers and builders. And that's really what today is about is about how we're making our platform better. Because developers, if you help us close the gap between what the models can do and what they're actually doing for real people out there. And both of you talk a lot internally about the importance of developers and builders. Maybe Danielle, let me start with you. How do you think about supporting this ecosystem, supporting this community? Yeah, I mean, I'll start out by saying I think in many ways, developers are the most important users of Claude. I think for a variety of reasons. One is, anthropic ourselves are majority developers. If you think about how we develop this technology, what we're building, we learn so much from the developer community. It's the best partnership. Because we, first of all, I think it's a group that gives honest feedback. And so I think that is actually really hard to get. You know what I mean? It's like, you build a product and you're like, I see some numbers, those are nice. But the genuineness with which the developer community I think engages with us is something that is so special. We have tried really from day one, I think anthropic has always primarily built for developers for businesses. I think we're a little bit unique in the AI ecosystem for that reason. And I think we have been very fortunate to be able to benefit from the feedback, from the engagement, from the community development. Developers are, in my experience, they're very ecosystem and community oriented, which I think we are too. We're like, how do we build for this sort of broader ecosystem of people who are developing, by the way? Some of the most inspiring transformative technologies and building the most incredible companies in the entire world. There's been this renaissance of things in medicine and software development and financial service. I mean, it's like you pick the industry and there is an incredible developer or an incredible developer-based company that is transforming that industry, help leveraging our tools sometimes. And I think that's such a special, that's both a privilege and a responsibility that I think anthropocolds to say, developers are really the backbone of how we learn, how we build better tools for all of you. And I think that's a really special relationship that we feel like pride in and also a responsibility towards. Yeah, I mean. And this is an example of the developer community. I think part of what Danielle is saying is feedback as a gift. We hear the positive. We also hear the negative. Please keep it coming. It is part of how we know what's working and what isn't. And so we really value, and I'm sure as you're talking to people around, we really value all of that. And it helps us know what to do better. Sorry, Daria, back to you. Yeah, one thing I would say is that technology doesn't diffuse at kind of an even pace across the economy. Right? And I think there's a spectrum where the software engineers are the ones who are fastest to adopt new, kind of fastest to adopt new technology. That's why there's so much kind of focus on this area. And it's the beginning of it. But it's like, it's a foreshadowing of how things are going to work across the economy. And how the economy is going to be transformed by AI. So I think getting this right and really making it work for this community, it's kind of like a microcosm of how we have to, you know, of how we have to make it work across the world. And I think one thing that, you know, a dynamic we should watch, it was about, I think it was roughly a year ago, there was an event like this where Mike Krieger asked me, you know, when will there be the first billion dollar company with one person? And I said 2026. And I think we're actually on track to achieve. It hadn't quite happened yet. There's been like two person companies that are $1 billion built with AI. There's been like one person that's worth $700 million. But we got seven more months. So it's kind of, no, we do it. It's seven more months since 2026 or eight. And you know, that's, there's an eternity on the exponential. But what I'm trying to say with this though is that there's an enormous ability for one person or a tiny set of people to do a set of things that are incredible, right? Where, you know, before, if, you know, you just had an idea and you had a vision like there's so many resources you'd have to accumulate over several years in order to make that vision happen. And I think there's a very unique opportunity for single individuals or very tiny teams to do things that are incredible, right? Where I think we moved from the models are writing code to, you know, the models are helping us think of software engineering as a task. To the models are helping us think of like, how can I build a business or an economic unit as a task? And so there's an extraordinary amount of opportunity for people in this room to kind of take advantage of that. Yeah, it really feels like it's kind of removing barriers. You know, there were all these barriers to like creating that kind of value in the world. And now, I mean, I guess the gauntlet has been thrown. We've got an eight month timer. And I'm excited to see what comes of it. I would love to maybe hear what's gonna change for developers. So Dario, you talked a little bit about what they can do now and you've talked, you know, in the past about how you expect a Claude to build more with Claude. Can you just talk about how you expect things to move? Yeah, I mean, I think there's several, like maybe trends. One is going from single agents to kind of multiple agents. So the idea that you have a bunch of these Claude rights, like managing a team, right? You have a bunch of quads running and like, you know, you kind of farm bunch of things out to your quads and maybe some of the quads farm things out to other quads with like different press. So you have a kind of whole hierarchy or a whole, we're gradually making our way to like the country of geniuses in the data center. You know, we're starting with like a team of smart people in a room or something. We're working our way up to the different exponential in the country. So I think that's one trend that we're kind of already starting to see and we're already like offering tools that can help do that. I think a trend that's related to that is like, you know, what we've done so far with Claude Code is like, you know, it helps kind of individuals to be more productive. But I think increasingly we're going to start thinking at the level of whole teams and organizations and how can you make whole teams and whole organizations more productive in a way that is kind of more than just the sum of its parts. And then finally, you know, I think in this area as with everywhere else, if you want to think about what's next when something's working really well you should always think about Amdol's law which is you speed one thing up, what are the things you're not speeding up. And so I think there are a bunch of things like security, like verification, like just if you're living in a world where you can within an organization write three or four times as many PRs as you could previously. You start to understand there are all these other things that are holding you back or that will go wrong if you speed up just that and not everything else. And so working to speed up those kind of other things so that we can greatly increase people's productivity but we can do it smoothly and, you know, smoothly and productively and reliably. I think that's gonna be very important. Does that have any impact on how you think about training new models or, you know, what the future of models looks like? Yeah, I mean, you know, that's true on several levels. I mean, I've already said many times we're using clawed to speed up clawed, right? That's kind of already something that's happening. But I also wonder if the things we're trying to do with the models could also influence how we build them. So when I talk about these things like, you know, kind of verification or kind of design quality or things like that, like one of the reasons training models for code and software engineering has gone so fast is that you have this verifiability, right? Where you train the model and it's like, you know, you're able to verify it by running unit tests and so that it has a lot of properties that make simplify the process of training. But what you discover is that there are these, of course, aspects of the job that are not verifiable, right? And, you know, some of the, you know, is this thing really right? Can we find errors or their security issues not quite as verifiable? And so training the models to be better at that, which I think will also make the models better at other things where they haven't made progress as fast as coding, like their ability to ride or their ability to kind of do, do, you know, to do, you know, less, less objective, scientific tasks. So I think it's gonna have benefits in many, in many other areas, but, you know, I think we find even within software engineering this, you know, these, these kind of soft or somewhat subjective skills and abilities are become surprisingly important because of M-Dullslaw. I would love to hear, you know, we talk a lot about our mission internally. Like, Daniela, as we just keep growing and the stakes of this whole industry keep getting higher, what should people know about our mission and about us as a company? You know, I think, when I think about what Anthropic is trying to do, there's these sort of like two, maybe two pillars, right? The first is around, how do we develop this transformative technology in a way that is good for everybody, right? And I think this goal of, Claude is this incredible tool. It has the power to really transform, you know, what people build and how they create and the ambition level of what they can develop themselves. And I think there comes a huge amount of opportunity there and there's also some risk, right? I think that we've talked about this a lot publicly. There's some risk to just labor disruption. There's risks to ensuring that the technology is developed safely, that it's good for people. And I think Anthropic's job or what we try to do is really think about how to balance these two things in equal measure. We have this internal cultural value called hold light and shade. And I think that it is such a good encapsulation of, you know, what we see about how the technology is being leveraged today and also just our approach to putting the technology out into the world, right? I think mythos and glass, we are a great example of this. The potential to build something incredible with a model that capable is so vast. And we want to be a little bit careful about how we release it because of some of the security vulnerabilities, right? And I think that this is this kind of complicated dance that we do where we're like, we really want to get stuff out as quickly as we can. We're trying to build the best products and release the most powerful models. And we're just trying to do it responsibly. I think that is really the underpinning of the majority of actions that we take is sort of grounded across those two pillars. I think that is, like, one of the things I find most meaningful about getting to work here is just thinking about how much everything is changing. And the fact that we're kind of all building in the industry right now, to me, it feels like we're getting the chance to have a vote in how everything unfolds and the trade-offs you're describing. That's just what I think of when I think of Hold Light and Shade is as things move so rapidly. We get to build experiences that other people use to understand what the future looks like. So that's something I'm always really personally excited about. Maybe I'll ask a little bit about product. Both of you kind of lean in quite a lot on the product side. One thing that we talk a lot about is building for the exponential. I always think about product as kind of a bridge between the technology that exists and the problems that people have. And it's just a very interesting time because the technology is changing much more rapidly than we're used to. Can you talk a little bit about how you think about product building in this world? So I love, I mean, I love your way of putting that. It's like, Dario and Daniela lean in a lot. And this cheap product episode, what Abby needs is like, you guys are up in my business all the damn time. Can you please leave me alone and let me do my damn job? I'll feed back as a gift. I'll enjoy every perspective. But no, Abby is right. I think you have a hard job that you wear incredibly well. But I think in all seriousness, Dario and I both, we care a lot about the product. It's a representation of what we are trying to build at Anthropic. We want it to be useful for people. We want it to be accessible. We want the product to be good, right? Part of why I think we're leaned in so much is we feel very invested in ensuring that people that are using, Claude, are getting out of it, everything that they can. And so I think our bothering you is really, I think our way of feeling like to the degree we can we're standing up for our customers. We're standing up for our users who are building sometimes their whole business around the premise of what these AI models are capable of doing. And I think the other thing, maybe the thing that makes product at Anthropic unusual or different than how I've seen it done, what other companies I've worked at is product is sort of one input and research is another. I'm sure you've felt this in your role. Sometimes we're like, man, this is a great place where we should just be able to build something better that's easier to use. Or here's a product idea we really want to enable people to be able to access it right out of the gate. But a lot of the time, product innovation is driven by what new capabilities emerge in the model. And I think coding is actually a great example of this. We didn't just sit out and say like, from day one, we're going to build a coding product. It was like once we saw that the models were able to write code at a reasonably accurate level, not perfect, we were like, huh, this is interesting. It seems like a lot of people that are kind of clawed files are developers, right? And they're using it to write code. This has always been a community that we've worked well with. We've wanted to lean in with and engage in support. Should we build a tool for them, right? Should we build something that actually is going to enable people to be better at doing their day-to-day work in this way? But I think that's just an interesting dance inside the company where there's a component that is sort of, I don't want to say traditional product because nothing about Anthropic is traditional. But I think there's a component that looks like a normal product organization. And then there's part of the organization that's like what is new from the models? What's happening on the research side? And how do we get to marry those two things together? Yeah, I would maybe take this from two lenses, which is building products for AI and building products with AI. And I think the internal experience at Anthropic, month by month and week by week, has given us lessons about both. So I think over the last few years, learned a lot of lessons about building products with AI. And in some ways, it's been an advantage that I was a researcher. I was never in the era of building products without AI. So it's like you can end up in a situation where there are things you don't have to unlearn. You can just learn the new world from scratch. I think you got the essential difference, which is that if you go back to the product there in the 2010s, you had a slowly changing technological background. You were trying to do new things with the kind of technology that was present. And of course, every once in a while, you'd have a new framework or a new way of doing it, but relatively slow. AI is moving lightning fast. And so there are a few consequences of that. One is that there are new products that are not possible with a given capability of model. But then when you take the next step, when you go far enough along the exponential, then suddenly they light up, suddenly they become possible. And so it puts a premium on internal experimentation. Because you always want to be trying something. Even if you tried something that didn't work, you want to revisit it a few months later, because it might work then when it didn't work before, which sounds a little bit crazy. But if we had tried to do Claude Code in like 2022, it wouldn't have worked because the models wouldn't have been strong enough. It was a frustrating experience. We did have some early things that were a little like Claude Code in 2022. And it was like, oh, this is interesting, but you couldn't actually derive value from it because the models were dumb. So that's, they were. I've been training these models since 2015. They were really dumb. They were really dumb back then. The second thing is that products reached their saturation. When models start to get too good. So I think this has happened with chatbots, right? Like, it's a big market. Lots of people use it. It's going to stay around. But the ways in which we're making model smarter today are much more evident in today's Claude Code form factor. And in more generally, in the genetic form factor, then they are in the chatbot code factor. So that's kind of the other side of the coin, which is that you always have to be thinking about what the new thing is, right? The way this business works isn't you make a product. It becomes very big. And then all the kind of stability set saying that you have to ask, you're constantly, not only are you able to make something new, you're constantly needing to make something new, or at least update the things that you've made. So it's kind of constantly an innovation laboratory. And I think the other thing it means of particular relevance for developers and software engineers is API never really goes away as a market. Because the fact that it's always possible to build new products that's true inside Anthropic, it's also true outside Anthropic. And so I actually think both within code as an application, but just coders writing code with Claude for medical applications, for law, for finance, or they're going to keep me new applications because the models get smarter and they enable them. So that's building in the age of AI. There's a newer thing we've seen maybe over the last year or six months, which is building with AI itself, like using AI to enable the process of doing product development faster. That one is interesting. And again, we would go back to our old friend, Amdol's Law, which is we've found with the internal model acceleration. You can write two times as many, four times as many, five times many, you see this within the company. But then you see what breaks. We've been able to ship a lot more products than we could a year ago and there of pretty high quality. Or I hope you think that. But it is possible to accumulate an extraordinary amount of internal technical debt when you ship that fast. And so then you have to say, well, can we also use the AI models to undo that technical debt or keep track of what it is that we're doing? And then you learn the team has to work together in a totally different way. And these revelations come month by month, and you kind of learn how to do things in a totally new way. So it somehow it increases the tempo, not just of building, but in which the way you have to change the way you build as a team. I really experience that. I think the hardest part of it is you can get really familiar with the problems because the problems don't change. That fast, the problems are about humans. Like we all, we're going to have similar problems. But a thing that's hard is to learn to be fresh eyed all the time about what the technology can do. And constantly scan for that. And then I also feel on a personal basis where your job just changes because you hit a new bottleneck. And the way you spend your day is just a little different. Maybe I'll just ask a couple more questions. Maybe Dario, very quickly, you talked a lot about model capabilities and how they change so often. Is there one thing that makes you most excited in the models that are coming in the next call it six months? I would say this idea of thinking at the level of the organization rather than just one person. And again, it ties to this one person, billion dollar business, which maybe that will turn out to be an underestimate. But just kind of the idea that you can both have AI do the work of many people but that when you have a team of humans, that AI is not just doing the work of many people working for one person, but that it does the work of many people many times over by operating within an organization of humans. Awesome. And maybe I'll just ask to close. Danielle, one thing we talked about is some of the kind of developers or use cases we see that are just so inspiring. I talked a bit in the keynote about some of my favorites, whether that's people using it for efficiency. It places like Stripe doing major info upgrades, but also people using it for these very personal and specific ways like connecting kids to foster families faster. And I would just love to hear, what are some of your favorite examples of how you see people using the tools? Yeah, I mean, I think even going back to the days when I was at Stripe, something that I've just loved about the developer community in general is like for every interesting challenge or problem there is in the world, there's like an incredible dev somewhere who's like trying to make that thing better. And I think getting to watch people build things with cloth that are just bringing so much utility and value and meaning to people around the world is really inspiring. I think for me personally, a pilot that I've seen with some developers who are building these interfaces basically for mobile doctors in the global south. So places where it's really hard to reach an actual doctor because of the access issues, right? You're down a dirt road somewhere, 50 miles from the nearest city, but people there still encounter health challenges. How do you actually like work with this really smart technology to build these like interfaces where people can just ask the question, it can give you some medical advice, right? In a way that's like sanctioned. I'm also just like in general blown away by the medical research field and what is coming out of cloth being able to accelerate biomedical research across many of the developers that work with us, that build on us. There are also just some really heartwarming individual examples we have this user happiness channel at work. So my favorite ones, there's a developer who used cloth to help retrieve their wedding photos that were corrupted like on a corrupted hard drive. I thought that was so sweet. And just like such a personal use case, but so like so meaningful. And then my last one was just so random is, someone is using cloth to chart the growth of their tomato plants in their garden. And I was like never in a million years would I have thought of it. But it's like, I was like, do you have a live camon? Can I switch out to this? Because I would really love to see it. But I think just the range of things that people are able to do is just astonishing. Well, Dario Daniela, thank you both so much for spending time with us here. It's a pleasure to see you. Thank you to all of you for joining us. We've got a great rest of the day coming. But please join me in giving around to applause. Please go find out. Thank you.