We recorded this episode right after the release of OpenAI\u2019s GPT-5, which gave us an opportunity to talk about why he thinks progress on AI models has slowed. The work that David\u2019s team is doing is a big priority for Amazon, and this is the first time I\u2019ve heard him really lay out what he\u2019s been up to.<\/p>\n
I also had to ask him about how he joined Amazon. David\u2019s decision to leave Adept was one of the first of many deals I call reverse acquihire<\/a>, in which a Big Tech company all-but-actually buys a buzzy AI startup to avoid antitrust scrutiny. I don\u2019t want to spoil too much, but let\u2019s just say that David left the startup world for Big Tech last year because he says he knew where the AI race was headed. I think that makes his predictions for what\u2019s coming next worth listening to.<\/p>\n This interview has been lightly edited for length and clarity. <\/p>\n David, welcome to the show.<\/p>\n Thanks so much for having me on. I\u2019m really excited to be here.<\/p>\n It\u2019s great to have you. We have a lot to talk about. I\u2019m super interested in what you and your team are up to at Amazon these days. But first, I think the audience could really benefit from hearing a little bit about you and your history, and how you got to Amazon, because you\u2019ve been in the AI space for a long time, and you\u2019ve had a pretty interesting career leading up to this. Could you walk us through a little bit of your background in AI and how you ended up at Amazon?<\/p>\n First off, I find it absolutely hilarious that anyone would say I\u2019ve been around the field for a long time. It\u2019s true in relative terms, because this field is so new, and yet, nonetheless, I\u2019ve only been doing AI stuff for about the last 15 years. So compared with many other fields, it\u2019s not that long.<\/p>\n Well, 15 years is an eternity in AI years.<\/p>\n It is an eternity in AI years. I remember when I first started working in the field. I worked on AI just because I thought it was interesting. I thought having the opportunity to build systems that could think like humans, and, ideally, deliver superhuman performance, was such a cool thing to do. I had no idea that it was going to blow up the way that it did.<\/p>\n But my personal background, let\u2019s see. I led the research and engineering teams at OpenAI from 2017 to mid-2020, where we did GPT-2 and GPT-3, as well as CLIP and DALL-E. Every day was just so much fun, because you would show up to work and it was just your best friends and you\u2019re all trying a bunch of really interesting research ideas, and there was none of the pressure that exists right now.<\/p>\n Then, after that, I led the LLM effort at Google, where we trained a model called PaLM<\/a>, which was quite a strong model for its time. But shortly after that, a bunch of us decamped to various startups, and my team and I ended up launching Adept. It was the first AI agent startup. We ended up inventing the computer-use agent effectively. Some good research had been done beforehand. We had the first production-ready agent, and Amazon brought us in to go run agents for it about a year ago.<\/p>\n Great, and we\u2019ll get into that and what you\u2019re doing at Amazon. But first, given your OpenAI experience, we\u2019re now talking less than a week from the release of GPT-5<\/a>. I\u2019d love to hear you reflect on that model, what GPT-5 says about the industry, and what you thought when you saw it. I\u2019m sure you still have colleagues at OpenAI who worked on it. But what does that release signify? <\/p>\n I think it really signifies a high level of maturity at this point. The labs have all figured out how to reliably tape out increasingly better models. One of the things that I always harp on is that your job, as a frontier-model lab, is not to train models. Your job as a frontier-model lab is to build a factory that repeatedly churns out increasingly better models, and that\u2019s actually a very different philosophy for how to make progress. In the I-build-a-better-model path, all you do is think about, \u201cLet me make this tweak. Let me make this tweak. Let me try to glom onto people to get a better release.\u201d<\/p>\n If you care about it from the perspective of a model factory, what you\u2019re actually doing is trying to figure out how you can build all the systems and processes and infrastructure to make these things smarter. But with the GPT-5 release, I think what I find most interesting is that a lot of the frontier models these days are converging in capabilities. I think, in part, there\u2019s an explanation that one of my old colleagues at OpenAI, Phillip Isola, who\u2019s now a professor at MIT, came up with called the Platonic representation hypothesis<\/a>. Have you heard of this hypothesis?<\/p>\n So the Platonic representation hypothesis is this idea, similar to Plato\u2019s cave allegory, which is really what it\u2019s named after, that there is one reality. But we, as humans, see only a particular rendering of that reality, like the shadows on the wall in Plato\u2019s cave. It\u2019s the same for LLMs, which \u201csee\u201d slices of this reality through the training data they\u2019re fed.<\/p>\n So every incremental YouTube video of, for example, someone going for a nature walk in the woods, is all ultimately generated by the actual reality that we live in. As you train these LLMs on more and more and more data, and the LLMs become smarter and smarter, they all converge to represent this one shared reality that we all have. So, if you believe this hypothesis, what you should also believe is that all LLMs will converge to the same model of the world. I think that\u2019s actually happening in practice from seeing frontier labs deliver these models.<\/p>\n Well, there\u2019s a lot to that. I would maybe suggest that a lot of people in the industry don\u2019t necessarily believe we live in one reality. When I was at the last Google I\/O developer conference, cofounder Sergey Brin and Google DeepMind chief Demis Hassabis were onstage, and they both seemed to believe that we were existing in multiple realities<\/a>. So I don\u2019t know if that\u2019s a thing that you\u2019ve encountered in your social circles or work circles over the years, but not everyone in AI necessarily believes that, right?<\/p>\n [Laughs] I think that hot take is above my pay grade. I do think that we only have one.<\/p>\n Yeah, we have too much to cover. We can\u2019t get into multiple realities. But to your point about everything converging, it does feel as if benchmarks are starting to not matter as much anymore, and that the actual improvements in the models, like you said, are commodifying. Everyone\u2019s getting to the same point, and GPT-5 will be the best on LMArena for a few months until Gemini 3.0 comes out, or whatever, and so on and so on. <\/p>\n If that\u2019s the case, I think what this release has also shown is that maybe what is really starting to matter is how people actually use these things, and the feelings and the attachments that they have toward them. Like how OpenAI decided to bring back its 4o model because people had a literal attachment to it as something they felt. People on Reddit have been saying, \u201cIt\u2019s like my best friend\u2019s been taken away.\u201d <\/p>\n So it really doesn\u2019t matter that it\u2019s better at coding or that it\u2019s better at writing; it\u2019s your friend now. That\u2019s freaky. But I\u2019m curious. When you saw that and you saw the reaction to GPT-5, did you predict that? Did you see that we were moving that way, or is this something new for everyone?<\/p>\n There was a project called LaMDA or Meena at Google in 2020 that was basically ChatGPT before ChatGPT, but it was available only to Google employees. Even back then, we started seeing employees developing personal attachments to these AI systems<\/a>. Humans are so good at anthropomorphizing anything. So I wasn\u2019t surprised to see that people formed bonds with certain model checkpoints.<\/p>\n But I think that when you talk about benchmarking, the thing that stands out to me is what benchmarking is really all about, which at this point is just people studying for the exam. We know what the benchmarks are in advance. Everybody wants to post higher numbers. It\u2019s like the megapixel wars from the early digital camera era. They just clearly don\u2019t matter anymore. They have a very loose correlation with how good of a photo this thing actually takes.<\/p>\n I think the question, and the lack of creativity in the field that I\u2019m seeing, boils down to the fact that AGI is way more than just chat. It\u2019s way more than just code. Those just happen to be the first two use cases that we all know work really well for these models. There\u2019s so many more useful applications and base model capabilities that people haven\u2019t even started figuring out how to measure well yet.<\/p>\n I think the better questions to ask now if you want to do something interesting in the field are: What should I actually run at? Why am I trying to spend more time making this thing slightly better at creative writing? Why am I trying to spend my time trying to make this model X percent better at the International Math Olympiad when there\u2019s so much more left to do? When I think about what keeps me and the people who are really focused on this agent\u2019s vision going, it\u2019s looking to solve a much greater breadth of problems than what people have worked out so far.<\/p>\n That brings me to this topic. I was going to ask about it later. But you\u2019re running the AGI research lab at Amazon. I have a lot of questions about what AGI means to Amazon, specifically, but I\u2019m curious first for you, what did AGI mean to you when you were at OpenAI helping to get GPT off the ground, and what does it mean to you now? Has that definition changed at all for you?<\/p>\n Well, the OpenAI definition for AGI we had was a system that could outperform humans at economically valuable tasks. While I think that was an interesting, almost doomer North Star back in 2018, I think we have gone so much beyond that as a field. What gets me excited every day is not how do I replace humans at economically valuable tasks, but how do I ultimately build toward a universal teammate for every knowledge worker.<\/p>\n What keeps me going is the sheer amount of leverage we could give to humans on their time if we had AI systems to which you could ultimately delegate a large chunk of the execution of what you do every day. So my definition for AGI, which I think is very tractable and very much focused on helping people \u2014 as the first most important milestone that would lead me to say we\u2019re basically there \u2014 is a model that could help a human do anything they want to do on a computer.<\/p>\n I like that. That\u2019s actually more concrete and grounded than a lot of the stuff I\u2019ve heard. It also shows how different everyone feels about what AGI means. I was just on a press call with Sam Altman for the GPT-5 launch, and he was saying he now thinks of AGI as a model that can self-improve itself. Maybe that\u2019s related to what you\u2019re saying, but it sounds as if you\u2019re grounding it more in the actual use case.<\/p>\n Well, the way that I look at it is self-improvement is interesting, but to what end, right? Why do we, as humans, care if the AGI is self-improving itself? I don\u2019t really care, personally. I think it\u2019s cool from a scientist\u2019s perspective. I think what\u2019s more interesting is how do I build the most useful form of this super generalist technology, and then be able to put it in everybody\u2019s hands? And I think the thing that gives people tremendous leverage is if I can teach this agent that we\u2019re training to handle any useful task that I need to get done on my computer, because so much of our life these days is in the digital world.<\/p>\n So I think it\u2019s very tractable. Going back to our discussion about benchmarking, the fact that the field cares so much about MMLU, MMLU-Pro, Humanity\u2019s Last Exam, AMC 12, et cetera, we don\u2019t have to live in that box of \u201cthat\u2019s what AGI does for me.\u201d I think it\u2019s way more interesting to look at the box of all useful knowledge-worker tasks. How many of them are doable on your machine? How can these agents do them for you?<\/p>\n So it\u2019s safe to say that for Amazon, AGI means more than shopping for me, which is the cynical joke I was going to make about what AGI means for Amazon. I\u2019d be curious to go back to when you joined Amazon, and you were talking to the management team and Andy Jassy, and how still to this day you guys talk about the strategic value of AGI as you define it for Amazon, broadly. Amazon is a lot of things. It\u2019s really a constellation of companies that do a lot of different things, but this idea kind of cuts across all of that, right?<\/p>\n I think that if you look at it from the perspective of computing, so far the building blocks of computing have been: Can I rent a server somewhere in the cloud? Can I rent some storage? Can I write some code to go hook all these things up and deliver something useful to a person? The building blocks of computing are changing. At this point, the code\u2019s written by an AI. Down the line, the actual intelligence and decision-making are going to be done by an AI.<\/p>\n So, then what happens to your building blocks? So, in that world, it\u2019s super important for Amazon to be good specifically at solving the agent\u2019s problem, because agents are going to be the atomic building blocks of computing. And when that is true, I think so much economic value will be unlocked as a result of that, and it really lines up well with the strengths that Amazon already has on the cloud side, and putting together ridiculous amounts of infrastructure and all that.<\/p>\n I see what you\u2019re saying. I think a lot of people listening to this, even people who work in tech, understand conceptually that agents are where the industry\u2019s headed. But I would venture to guess that the vast majority of the listeners to this conversation have either never used an agent or have tried one and it didn\u2019t work. I would pretty much say that\u2019s the lay of the land right now. What would you hold out as the best example of an agent, the best example of where things are headed and what we can expect? Is there something you can point to?<\/p>\n So I feel for all the people who have been told over and over again that agents are the future, and then they go try the thing, and it just doesn\u2019t work at all. So let me try to give an example of what the actual promise of agents is relative to how they\u2019re pitched to us today.<\/p>\n Right now, the way that they\u2019re pitched to us is, for the most part, as just a chatbot with extra steps, right? It\u2019s like, Company X doesn\u2019t want to put a human customer service rep in front of me, so now I have to go talk to a chatbot. Maybe behind the scenes it clicks a button. Or you\u2019ve played with a product that does computer use that is supposed to help me with something on my browser, but in reality it takes four times as long, and one out of three times it screws up. This is kind of the current landscape of agents.<\/p>\n Let\u2019s take a concrete example: I want to do a particular drug discovery task where I know there\u2019s a receptor, and I need to be able to find something that ends up binding to this receptor. If you pull up ChatGPT today and you talk to it about this problem, it\u2019s going to go and find all the scientific research and write you a perfectly formatted piece of markdown of what the receptor does, and maybe some things you want to try.<\/p>\n But that\u2019s not an agent. An agent, in my book, is a model and a system that you can literally hook up to your wet lab, and it\u2019s going to go and use every piece of scientific machinery you have in that lab, read all the literature, propose the right optimal next experiment, run that experiment, see the results, react to that, try again, et cetera, until it\u2019s actually achieved the goal for you. The degree to which that gives you leverage is so, so, so much higher than what the field is currently able to do right now.<\/p>\n Do you agree, though, that there\u2019s an inherent limitation in large language models and decision-making and executing things? When I see how LLMs, even still the frontier ones, still hallucinate, make things up, and confidently lie, it\u2019s terrifying to think of putting that technology in a construct where now I\u2019m asking it to go do something in the real world, like interact with my bank account, ship code, or work in a science lab. <\/p>\n When ChatGPT can\u2019t spell right, that doesn\u2019t feel like the future we\u2019re going to get. So, I\u2019m wondering, are LLMs it, or is there more to be done here?<\/p>\n So we started with a topic of how these models are increasingly converging in capability. While that\u2019s true for LLMs, I don\u2019t think that\u2019s been true, to date, for agents, because the way that you should train an agent and the way that you train an LLM are quite different. With LLMs, as we all know, the bulk of their training happens from doing next-token prediction. I\u2019ve got a giant corpus of every article on the internet, let me try to predict the next word. If I get the next word right, then I get a positive reward, and if I get it wrong, then I\u2019m penalized. But, in reality, what\u2019s actually happening is what we in the field call behavioral cloning or imitation learning. It\u2019s the same thing as cargo culting<\/a>, right?<\/p>\n The LLM never learns why the next word is the right answer. All it learns is that when I see something that is similar to the previous set of words, I should go say this particular next word. So the issue with this is that this is great for chat. This is great for creative-use cases where you want some of the chaos and randomness from hallucinations. But if you want it to be an actual successful decision-making agent, these models need to learn the true causal mechanism. It\u2019s not just cloning human behavior; it\u2019s actually learning if I do X, the consequence of it is Y. So the question is, how do we train agents so that they can learn the consequences of their actions? The answer, obviously, cannot be just doing more behavioral cloning and copying text. It has to be something that looks like actual trial and error in the real world.<\/p>\n That\u2019s basically the research roadmap for what we\u2019re doing in my group at Amazon. My friend Andrej Karpathy has a really good analogy here, which is imagine you have to train an agent to go play tennis. You wouldn\u2019t have it spend 99 percent of its time watching YouTube videos of tennis, and then 1 percent of its time actually playing tennis. You would have something that\u2019s far more balanced between these two activities. So what we\u2019re doing in our lab here at Amazon is large-scale self-play. If you remember, the concept of self-play was the technique that DeepMind really made popular in the mid-2010s, when it beat humans at playing Go<\/a>.<\/p>\n So for playing Go, what DeepMind did was spin up a bajillion simulated Go environments, and then it had the model play itself over and over and over again. Every time it found a strategy that was better at beating a previous version of itself, it would effectively get a positive reward via reinforcement learning to go do more of that strategy in the future. If you spent a lot of compute on this in the Go simulator, it actually discovered superhuman strategies for how to play Go. Then when it played the world champion, it made moves that no human had ever seen before and contributed to the state of the art of that whole field.<\/p>\n What we\u2019re doing is, rather than doing more behavioral coding or watching YouTube videos, we\u2019re creating a giant set of RL [reinforcement learning] gyms, and each one of these gyms, for example, is an environment that a knowledge worker might be working in to get something useful done. So here\u2019s a version of something that\u2019s like Salesforce. Here\u2019s a version of something that\u2019s like an enterprise resource plan. Here\u2019s a computer-aided design program. Here\u2019s an electronic medical record system. Here\u2019s accounting software. Here is every interesting domain of possible knowledge work as a simulator.<\/p>\n Now, instead of training an LLM just to do tech stuff, we have the model actually propose a goal in every single one of these different simulators as it tries to solve that problem and figure out if it\u2019s successfully solved or not. It then gets rewarded and receives feedback based on, \u201cOh, did I do the depreciation correctly?\u201d Or, \u201cDid I correctly make this part in CAD?\u201d Or, \u201cDid I successfully book the flight?\u201d to choose a consumer analogy. Every time it does this, it actually learns the consequences of its actions, and we believe that this is one of the big missing pieces left for actual AGI, and we\u2019re really scaling up this recipe at Amazon right now.<\/p>\n How unique is this approach in the industry right now? Do you think the other labs are onto this as well? If you\u2019re talking about it, I would assume so.<\/p>\n I think that what\u2019s interesting is this field. Ultimately, you have to be able to do something like this, in my opinion, to get beyond the fact that there\u2019s a limited amount of free-floating data on the internet that you can train your models on. The thing we\u2019re doing at Amazon is, because this came from what we did at Adept and Adept has been doing agents for so long, we just care about this problem way more than everybody else, and I think we\u2019ve made a lot of progress toward this goal.<\/p>\n You called these gyms, and I was thinking physical gyms, for a second. Does this become physical gyms? You have a background in robotics, right?<\/p>\n That\u2019s a good question. I\u2019ve also done robotics work before. Here we also have Pieter Abbeel, who came from Covariant<\/a> and is a Berkeley professor whose students ended up creating the majority of the RL algorithms that work well today<\/a>. It\u2019s funny that you say gyms, because we were trying to find an internal code name for the effort. We kicked around Equinox and Barry\u2019s Bootcamp and all this stuff. I\u2019m not sure everybody had the same sense of humor, but we call them gyms because at OpenAI we had a very useful early project called OpenAI Gym<\/a>.<\/p>\n This was before LLMs were a thing. OpenAI Gym was a collection of video game and robotics tasks. For example, can you balance a pole that\u2019s on a cart and can you train an RL algorithm that can keep that thing perfectly centered, et cetera. What we were inspired to ask was, now that these models are smart enough, why have toy tasks like that? Why not put the actual useful tasks that humans do on their computers into these gyms and have the models learn from these environments? I don\u2019t see why this wouldn\u2019t also generalize to robotics.<\/p>\n Is the end state of this an agent\u2019s framework system that gets deployed through AWS?<\/p>\n The end state of all this is a model plus a system that is rock-solid reliable, like 99 percent reliable, at all sorts of valuable knowledge-work tasks that are done on a computer. And this is going to be something that we think will be a service on AWS that\u2019s going to underpin, effectively, so many useful applications in the future.<\/p>\n I did a recent Decoder episode with Aravind Srinivas, the CEO of Perplexity<\/a>, about his Comet Browser. A lot of people on the consumer side think that the browser interface is actually going to be the way to get to agents, at scale, on the consumer side. <\/p>\n I\u2019m curious what you think of that. This idea that it\u2019s not enough to just have a chatbot, you really need to have ChatGPT, or whatever model, sit next to your browser, look at the web page, act on it for you, and learn from that. Is that where all this is headed on the consumer side?<\/p>\n I think chatbots are definitely not the long-term answer, or at least not chatbots in the way we think about them today if you want to build systems that take actions for you. The best analogy I have for this is this: my dad is a very well-intentioned, smart guy, who spent a lot of his career working in a factory. He calls me all the time for tech support help. He says, \u201cDavid, something\u2019s wrong with my iPad. You got to help me with this.\u201d We\u2019re just doing this over the phone, and I can\u2019t see what\u2019s on the screen for him. So, I\u2019m trying to figure, \u201cOh, do you have the settings menu open? Have you clicked on this thing yet? What\u2019s going on with this toggle?\u201d Chat is such a low bandwidth interface. That is the chat experience for trying to get actions done, with a very competent human on the other side trying to handle things for you.<\/p>\n So one of the big missing pieces, in my opinion, right now in AI, is our lack of creativity with product form factors, frankly. We are so used to thinking that the right interface between humans and AIs is this perpendicular one-on-one interaction where I\u2019m delegating something, or it\u2019s giving me some news back or I\u2019m asking you a question, et cetera. One of the real things we\u2019ve always missed is this parallel interaction where both the user and the AI actually have a shared canvas that they\u2019re jointly collaborating on. I think if you really think about building a teammate for knowledge workers or even just the world\u2019s smartest personal assistant, you would want to live in a world where there\u2019s a shared collaborative canvas for the two of you.<\/p>\n Speaking of collaboration, I\u2019m really curious how your team works with the rest of Amazon. Are you pretty walled off from everything? Do you work on Nova, Amazon\u2019s foundational model<\/a>? How do you interact with the rest of Amazon?<\/p>\n What Amazon\u2019s done a great job with, for what we\u2019re doing here, is allowing us to run pretty independently. I think there\u2019s recognition that some of the startup DNA right now is really valuable for maximum speed. If you believe AGI is two to five years away, some people are getting more bullish, some people are getting more bearish. It doesn\u2019t matter. That\u2019s not a lot of time in the grand scheme of things. You need to move really, really fast. So, we\u2019ve been given a lot of independence, but we\u2019ve also taken the tech stack that we\u2019ve built and contributed a lot of that upstream to the Nova foundation model as well.<\/p>\n So is your work, for example, already impacting Alexa Plus<\/a>? Or is that not something that you\u2019re part of in any way?<\/p>\n That\u2019s a good question. Alexa Plus has the ability to, for example, if your toilet breaks, you\u2019re like, \u201cAh, man, I really need a plumber. Alexa, can you get me a plumber?\u201d Alexa Plus then spins up a remote browser, powered by our technology, that then goes and uses Thumbtack, like a human would, to go get a plumber to your house, which I think is really cool. It\u2019s the first production web agent that\u2019s been shipped, if I remember correctly.<\/p>\n The early response to Alexa Plus<\/a> has been that it\u2019s a dramatic leap for Alexa but still brittle. There\u2019s still moments where it\u2019s not reliable. And I\u2019m wondering, is this the real gym? Is this the at-scale gym where Alexa Plus is how your system gets more reliable much faster? You have to have this in production and deployed to… I mean, Alexa has millions and millions of devices that it\u2019s on. Is that the strategy? Because I\u2019m sure you\u2019ve seen the earlier reactions to Alexa Plus are that it\u2019s better, but still not as reliable as people would like it to be.<\/p>\n Alexa Plus is just one of many customers that we have, and what\u2019s really interesting about being within Amazon is, to go back to what we were talking about earlier, web data is effectively running out, and it\u2019s not useful for training agents. What\u2019s actually useful for training agents is lots and lots of environments, and lots and lots of people doing reliable multistep workflows. So, the interesting thing at Amazon is that, in addition to Alexa Plus, basically every Fortune 500 business\u2019s operations are represented, in some way, by some internal Amazon team. There\u2019s One Medical, there\u2019s everything happening on supply chain and procurement on the retail side, there\u2019s all this developer-facing stuff on AWS.<\/p>\n Agents are going to require a lot of private data and private environments to be trained. Because we\u2019re in Amazon, that\u2019s all now 1P [first-party selling model]. So they\u2019re just one of many different ways in which we can get reliable workflow data to train the smarter agent.<\/p>\n Are you doing this already through Amazon\u2019s logistics operations, where you can do stuff in warehouses, or [through] the robotic stuff that Amazon is working on? Does that intersect with your work already?<\/p>\n Well, we\u2019re really close to Pieter Abbeel\u2019s group on the robotics side, which is awesome. In some of the other areas, we have a big push for internal adoption of agents within Amazon, and so a lot of those conversations or engagements are happening.<\/p>\n I\u2019m glad you brought that up. I was going to ask: how are agents being used inside Amazon today?<\/p>\n So, again, as we were saying earlier, because Amazon has an internal effort for almost every useful domain of knowledge work, there has been a lot of enthusiasm to pick up a lot of these systems. We have this internal channel called… I won\u2019t tell you what it\u2019s actually called.<\/p>\n It\u2019s related to the product that we\u2019ve been building. It\u2019s just been crazy to see teams from all over the world within Amazon \u2014 because one of the main bottlenecks we\u2019ve had is we didn\u2019t have availability outside the US for quite a while \u2014 and it was crazy just how many international Amazon teams wanted to start picking this up, and then using it themselves on various operations tasks that they had.<\/p>\n This is your just agent framework that you\u2019re talking about. This is something you haven\u2019t released publicly yet.<\/p>\n We released Nova Act<\/a>, which was a research preview that came out in March. But as you can imagine, we\u2019ve added way more capability since then, and it\u2019s been really cool. The thing we always do is we first dogfood with internal teams.<\/p>\n Your colleague, when you guys released Nova Act, said it was the most effortless way to build agents that can reliably use browsers. Since you\u2019ve put that out, how are people using Nova Act? It\u2019s not something that, in my day-to-day, I hear about, but I assume companies are using it, and I\u2019d be curious to hear what feedback you guys have gotten since you came out with it.<\/p>\n So, a wide range of enterprises and developers are using Nova Act. And the reason you don\u2019t hear about it is we\u2019re not a consumer product. If anything, the whole Amazon agent strategy, including what I did before at Adept, is sort of doing normcore agents, not the super sexy stuff that works one out of three times, but super reliable, low-level workflows that work 99-plus percent of the time.<\/p>\n So, that\u2019s the target. Since Nova Act came out, we\u2019ve actually had a bunch of different enterprises end up deploying with us that are seeing 95-plus percent reliability. As I\u2019m sure you\u2019ve seen from the coverage of other agent products out there, that\u2019s a material step up from the average 60 percent reliability that folks see with those systems. I think that the reliability bottleneck is why you don\u2019t see as much agent adoption overall in the field.<\/p>\n We\u2019ve been having a lot of really good luck, specifically by focusing extreme amounts of effort on reliability. So we\u2019re now used for things like, for example, doctor and nurse registrations. We have another customer called Navan, formerly TripActions, which uses us basically to automate a lot of backend travel bookings for its customers. We\u2019ve got companies that basically have 93-step QA workflows that they\u2019ve automated with a single Nova Act script.<\/p>\n I think the early progress has been really cool. Now, what\u2019s up ahead is how do we do this extreme large-scale self-play on a bajillion gyms to get to something where there\u2019s a bit of a \u201cGPT for RL agents\u201d moment, and we\u2019re running as fast as we can toward that right now.<\/p>\n Do you have a line of sight to that? Do you think we\u2019re two years from that? One year?<\/p>\n Honestly, I think we\u2019re sub-one year. We have line of sight. We\u2019ve built out teams for every step of that particular problem, and things are just starting to work. It\u2019s just really fun to go to work every day and realize that one of the teams has made a small but very useful breakthrough that particular day, and the whole cycle that we\u2019re doing for this training loop seems to be going a little bit faster every day.<\/p>\n Going back to GPT-5, people have said, \u201cDoes this portend a slowdown in AI progress?\u201d And 100 percent I think the answer is no, because when one S-curve peters out\u2026 the first one being pretraining, which I don\u2019t think has petered out, by the way, but it\u2019s definitely, at this point, less easy to get gains than before. And then you\u2019ve got RL with verifiable rewards. But then every time one of these S-curves seems to slow down a little bit, there\u2019s another one coming up, and I think agents are the next S-curve, and the specific training recipe we were talking about earlier is one of the main ways of getting that next giant amount of acceleration.<\/p>\n It sounds like you and your colleagues have identified the next turn that the industry is going to take, and that starts to put Nova, as it exists today, into more context for me, because Nova, as an LLM, is not an industry-leading LLM. It\u2019s not in the same conversation as Claude, GPT-5, or Gemini. <\/p>\n Is Nova just not as important, because what\u2019s really coming is what you\u2019ve been talking about with agents, which will make Nova more relevant? Or is it important that Nova is the best LLM in the world as well? Or is that not the right way to think about it?<\/p>\n I think the right way to think about it is that every time you have a new upstart lab trying to join the frontier of the AI game, you need to bet on something that can really leapfrog, right? I think what\u2019s interesting is every time there\u2019s a recipe change for how these models are trained, it creates a giant window of opportunity for someone new who\u2019s starting to come to the table with that new recipe, instead of trying to catch up on all the old recipes.<\/p>\n Because the old recipes are actually baggage for the incumbents. So, to give some examples of this, at OpenAI, of course, we basically pioneered giant models. The whole LLM thing came out of GPT-2 and then GPT-3. But those LLMs, initially, were text-only training recipes. Then we discovered RLHF [reinforcement learning from human feedback], and then they started getting a lot of human data via RLHF.<\/p>\n But then in the switch to multimodal input, you kind of have to throw away a lot of the optimizations you did in the text-only world, and that gives time for other people to catch up. I think that was actually part of how Gemini was able to catch up \u2014 Google bet on certain interesting ideas on native multimodal that turned out well for Gemini.<\/p>\n After that, reasoning models gave another opportunity for people to catch up. That\u2019s why DeepSeek was able to surprise the world<\/a>, because that team straight quantum-tunneled to that instead of doing every stop along the way. I think with the next turn being agents \u2014 especially agents without verifiable rewards \u2014 if we, at Amazon, can figure out that recipe earlier, faster, and better than everybody else, with all the scale that we have as a company, it basically brings us to the frontier.<\/p>\n I haven\u2019t heard that articulated from Amazon before. That\u2019s really interesting. It makes a lot of sense. Let\u2019s end on the state of the talent market and startups, and how you came to Amazon. I want to go back to that. So Adept, when you started it, was it the first startup to really focus on agents at the time? I don\u2019t think I had heard of agents until I saw Adept.<\/p>\n Yeah, actually we were the first startup to focus on agents, because when we were starting Adept, we saw that LLMs were really good at talking but could not take action, and I could not imagine a world in which that was not a crucial problem to be solved. So we got everybody focused on solving that.<\/p>\n But when we got started, the word \u201cagent,\u201d as a product category, wasn\u2019t even coined yet. We were trying to find a good term, and we played with things like large action models, and action transformers. So our first product was called Action Transformer. And then, only after that, did agents really start picking up as being the term.<\/p>\n Walk me through the decision to leave that behind and join Amazon with most of the technical team<\/a>. Is that right?<\/p>\n I have a phrase for this. It\u2019s a deal structure that has now become common with Big Tech and AI startups: it\u2019s reverse acquihire<\/a>, where basically the core team, such as you and your cofounders, join. The rest of the company still exists, but the technical team goes away. And the \u201cacquirer\u201d \u2014 I know it\u2019s not an acquisition \u2014 but the acquirer pays a licensing fee, or something to that effect, and shareholders make money. <\/p>\n But the startup is then kind of left to figure things out without its founding team, in most cases. The most recent example is Google and Windsurf<\/a>, and then there was Meta and Scale AI before that<\/a>. This is a topic we\u2019ve been talking about on Decoder a lot. The listeners are familiar with it. But you were one of the first of these reverse acquihires. Walk me through when you decided to join Amazon and why.<\/p>\n So I hope, in 50 years, I\u2019m remembered more as being an AI research innovator rather than a deal structure innovator. First off, humanity\u2019s demand for intelligence is way, way, way higher than the amount of supply. So, therefore, for us as a field, to invest ridiculous amounts of money in building the world\u2019s biggest clusters and bringing the best talent together to drive those clusters is actually perfectly rational, right? Because if you can spend an extra X dollars to build a model that has 10 more IQ points and can solve a giant new concentric circle of useful tasks for humanity, that is a worthwhile trade that you should do any day of the week.<\/p>\n So I think it makes a lot of sense that all these companies are trying to put together critical mass on both talent and compute right now. From my perspective on why I joined Amazon, it\u2019s because Amazon knows how important it is to win on the agent side, in particular, and that agents are a crucial bet for Amazon to build one of the best frontier labs possible. To get to the level of scale, you\u2019re hearing all these CapEx numbers from the various hyperscalers. It\u2019s just completely mind-boggling and it\u2019s all real, right?<\/p>\n It\u2019s over $340 billion in CapEx<\/a> this year alone, I think, from just the top hyperscalers. It\u2019s an insane number.<\/p>\n That sounds about right. At Adept, we raised $450 million<\/a>, which, at the time, was a very large number. And then, today is…<\/p>\n [Laughs] It\u2019s chump change.<\/p>\n That\u2019s one researcher. Come on, David.<\/p>\n [Laughs] Yes, one researcher. That\u2019s one employee. So if that\u2019s the world that you live in, it\u2019s really important, I think, for us to partner with someone who\u2019s going to go fight all the way to the end, and that\u2019s why we came to Amazon.<\/p>\n Did you foresee that consolidation and those numbers going up when you did the deal with Amazon? You knew that it was going to just keep getting more expensive, not only on compute but on talent.<\/p>\n Yes, that was one of the biggest drivers.<\/p>\n And why? What did you see coming that, at the time, was not obvious to everyone?<\/p>\n There were two things I saw coming. One, if you want to be at the frontier of intelligence, you have to be at the frontier of compute. And if you are not on the frontier of compute, then you have to pivot and go do something that is totally different. For my whole career, all I\u2019ve wanted to do is build the smartest and most useful AI systems. So, the idea of turning Adept into an enterprise company that sells only small models or turns into a place that does forward-deployed engineering to go help you deploy an agent on top of someone else\u2019s model, none of those things appealed to me.<\/p>\n I want to figure out, \u201cHere are the four crucial remaining research problems left to AGI. How do we nail them?\u201d Every single one of them is going to require two-digit billion-dollar clusters to go run it. How else am I \u2014 and this whole team that I\u2019ve put together, who are all motivated by the same thing \u2014 going to have the opportunity to go do that?<\/p>\n If antitrust scrutiny did not exist for Big Tech like it does, would Amazon have just acquired the company completely?<\/p>\n I can\u2019t speak to general motivations and deal structuring. Again, I\u2019m an AI research innovator, not an innovator in legal structure. [Laughs]<\/p>\n You know I have to ask. But, okay. Well, maybe you can answer this. What are the second-order effects of these deals that are happening, and, I think, will continue to happen? What are the second-order effects on the research community, on the startup community?<\/p>\n I think it changes the calculus for someone joining a startup these days, knowing that these kinds of deals happen, and can happen, and take away the founder or the founding team that you decided to join and bet your career on. That is a shift. That is a new thing for Silicon Valley in the last couple of years.<\/p>\n Look, there\u2019s two things I want to talk about. One is, honestly, the founder plays a really important role. The founder has to want to really take care of the team and make sure that everybody is treated pro rata and equally, right? The second thing is, it\u2019s very counterintuitive in AI right now, because there\u2019s only a small number of people with a lot of experience. And because the next couple of years are going to move so fast, and a lot of the value, the market positioning, et cetera, is going to be decided in the next couple of years.<\/p>\n If you\u2019re sitting there responsible for one of these labs, and you want to make sure that you have the best possible AI systems, you need to hire the people who know what they\u2019re doing. So, the market demand, the pricing for these people, is actually totally rational, just solely because of how few of them there are.<\/p>\n But the counterintuitive thing is that it doesn\u2019t take that many years, actually, to find yourself at the frontier, if you\u2019re a junior person. Some of the best people in the field were people who just started three or four years ago, and by working with the right people, focusing on the right problems, and working really, really, really hard, they found themselves at the frontier.<\/p>\n AI research is one of those areas where if you ask four or five questions, you\u2019ve already discovered a problem that nobody has the answer to, and then you can just focus on that and how do you become the world expert in this particular subdomain? So I find it really counterintuitive that there\u2019s only very few people who really know what they\u2019re doing, and yet it\u2019s very easy, in terms of the number of years, to become someone who knows what they\u2019re doing.<\/p>\n How many people actually know what they\u2019re doing in the world from your definition? This is a question I get asked a lot. I was literally just asked this on TV this morning. How many people are there, who can actually build and conceptualize training a frontier model, holistically?<\/p>\n I think it depends on how generous or tight you want to be. I would say the number of people who I would trust with a giant dollar amount of compute to go do that is probably sub-150.<\/p>\n Yes. But there are many more people, let\u2019s say, another 500 people or so, who would be extremely valuable contributors to an effort that was populated by a certain critical mass of that 150 who really know what they\u2019re doing.<\/p>\n But for the total market, that\u2019s still less than 1,000 people.<\/p>\n I\u2019d say it\u2019s probably less than 1,000 people. But again, I don\u2019t want to trivialize this: I think junior talent is extremely important, and people who come from other domains, like physics or quant finance, or who have just been doing undergrad research, these people make a massive difference really, really, really fast. But you want to surround them with a couple of folks who have already learned all the lessons from previous training attempts in the past.<\/p>\n Is this very small group of elite people building something that is inherently designed to replace them? Maybe you disagree with that, but I think superintelligence, conceptually, would make some of them redundant. Does it mean there\u2019s actually fewer of them, in the future, making more money, because you only need some orchestrators of other models to build more models? Or does the field expand? Do you think it\u2019s going to become thousands and thousands of people?<\/p>\n The field\u2019s definitely going to expand. There are going to be more and more people who really learn the tricks that the field has developed so far, and discover the next set of tricks and breakthroughs. But I think one of the dynamics that\u2019s going to keep the field smaller than other fields, such as software, is that, unlike regular software engineering, foundation model training breaks so many of the rules that we think we should have. In software, let\u2019s say our job here is to build Microsoft Word. I can say, \u201cHey, Alex, it\u2019s your job to make the save feature work. It\u2019s David\u2019s job to make sure that cloud storage works. And then someone else\u2019s job is to make sure the UI looks good.\u201d You can factorize these problems pretty independently from one another.<\/p>\n The issue with foundation model training is that every decision you take interferes with every other decision, because there\u2019s only one deliverable at the end. The deliverable at the end is your frontier model. It\u2019s like one giant bag of weights. So what I do in pretraining, what this other person does in supervised fine-tuning, what this other person does in RL, and what this other person does to make the model run fast, all interact with one another in sometimes pretty unpredictable ways.<\/p>\n So, with the number of people, it has one of the worst diseconomies of scale of anything I\u2019ve ever seen, except maybe sports teams. Maybe that\u2019s the one other case where you don\u2019t want to have 100 midlevel people; you want to have 10 of the best, right? Because of that, the number of people who are going to have a seat at the table at some of the best-funded efforts in the world, I think, is actually going to be somewhat capped.<\/p>\n Oh, so you think the elite stays relatively where it is, but the field around it \u2014 the people who support it, the people who are very meaningful contributors \u2014 expands?<\/p>\n I think the number of people who know how to do super meaningful work will definitely expand, but it will still be a little constrained by the fact that you cannot have too many people on any one of these projects at once.<\/p>\n What advice would you give someone who\u2019s either evaluating joining an AI startup, or a lab, or even an operation like yours in Big Tech on AI, and their career path? How should they be thinking about navigating the next couple of years with all this change that we\u2019ve been talking about?<\/p>\n First off, tiny teams with lots of compute are the correct recipe for building a frontier lab. That\u2019s what we\u2019re doing at Amazon with its staff and my team. It\u2019s really important that you have the opportunity to run your research ideas in a particular environment. If you go somewhere that already has 3,000 people, you\u2019re not really going to have a chance. There\u2019s so many senior people ahead of you who are all too ready to try their particular ideas.<\/p>\n The second thing is, I think people underestimate the codesign of the product, the user interface, and the model. I think that\u2019s going to be the most important game that people are going to play in the next couple of years. So going somewhere that actually has a very strong product sense, and a vision for how users are actually going to deeply embed this into their own lives, is going to be really important.<\/p>\n One of the best ways to tell is to ask, are you just building another chatbot? Are you just trying to fight one more entrant in the coding assistant space? Those just happen to be two of the earliest product form factors that have product market fit and are growing like crazy. I bet when we fast-forward five years and we look back on this period, there will be six to seven more of these crucial product form factors that will look obvious in hindsight but that no one\u2019s really solved today. If you really want to take an asymmetrical upside bet, I would try to spend some time and figure out what those are now.<\/p>\n Thanks, David. I\u2019ll let you get back to your gyms.<\/p>\n Thanks, guys. This was really fun.<\/p>\n Questions or comments about this episode? Hit us up at [email\u00a0protected]<\/a>. We really do read every email!<\/p>\n Decoder with Nilay Patel<\/p>\n A podcast from The Verge about big ideas and other problems.<\/p>\n SUBSCRIBE NOW!<\/a><\/a>Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.Alex HeathClose Posts from this author will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All by Alex Heath<\/a><\/p>\n AICloseAI<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All AI<\/a><\/p>\n AmazonCloseAmazon<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All Amazon<\/a><\/p>\n Amazon AlexaCloseAmazon Alexa<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All Amazon Alexa<\/a><\/p>\n BusinessCloseBusiness<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All Business<\/a><\/p>\n DecoderCloseDecoder<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All Decoder<\/a><\/p>\n PodcastsClosePodcasts<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n See All Podcasts<\/a><\/p>\n TechCloseTech<\/p>\n Posts from this topic will be added to your daily email digest and your homepage feed.<\/p>\n PlusFollow<\/p>\n![\"Alex](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2025\/08\/1755791170_75_alex_heath_blurple_headshot_720.0.png\"\/)
Alex Heath<\/p>\n