Pod Insights Decoder Amazon is betting on agents to win the AI race About this episode This is Alex Heath, your Thursday episode guest host and deputy editor at The Verge. One of the biggest topics in AI these days is agents — the idea that AI is going to move from chatbots to reliably completing tasks for us in the real world. But the problem with agents is that they really aren’t all that reliable right now.
There’s a lot of work happening in the AI industry to try and fix that, and that brings me to my guest today: David Luan, the head of Amazon’s AGI research lab, a cofounder of Adept, and a former VP of engineering at OpenAI. David and I discussed the release of GPT-5, what Amazon wants with agents, and where he thinks the AI race is headed next.
Read the full transcript on The Verge.
Links:
The Platonic Representation Hypothesis | Phillip Isola
Amazon plays catch-up with new Nova models to generate voices, video | Verge
Amazon’s new AI agent is designed to do your shopping | Verge
Microsoft is racing to build an AI ‘agent factory’ | Verge
OpenAI’s new ChatGPT Agent can control an entire computer | Verge
24 hours with Alexa Plus: we cooked, we chatted, and it kinda lied to me | Verge
Why AI researchers are getting paid like NBA All-Stars | Decoder
OpenAI’s Windsurf deal is off — and Windsurf’s CEO is going to Google | Verge
This is Big Tech’s playbook for swallowing the AI industry | Command Line
Amazon hires founders away from AI startup Adept | TechCrunch
Credits:
Decoder is a production of The Verge and part of the Vox Media Podcast Network.
Our producers are Kate Cox and Nick Statt. Our editor is Ursa Wright.
The Decoder music is by Breakmaster Cylinder.
Learn more about your ad choices. Visit podcastchoices.com/adchoices
Transcript 10,193 words · 11 speakers
0:01 Speaker 00 What if managing company spend took less managing? Introducing AMEX Corporate, one intelligent platform connecting cards, expenses, and payments so you can keep work moving. Built to simplify, designed to scale. This is AMEX Corporate. Get started at go.amex/ammexcorporate. Terms apply.
0:27 Speaker 01 New and exclusive holiday decor just launched online at The Home Depot. Check out the viral Grand Duchess collection with smart holiday lights you can control from your phone. Easily customize the twinkle across trees, wreaths, and garlands to match your holiday vibe. Create stunning light effects, then switch it up with the tap of your fingers. Even the man in red will be impressed. Shop the holiday decor you want, online only, at homedepot.com.
0:58 Speaker 02 This episode is brought to you by Michelob Ultra. Good times mean competitions between you and your friends, like seeing who can throw some pigskin the farthest, or rush the field the fastest. Winner gets an Ultra. Because it's not just a superior light beer, it's also giving fans superior access to their favorite sporting events. Superior is worth playing for. Visit michelobultra.com/superioraccess to enter. Enjoy responsibly. Copyright 2026 Anheuser-Busch InBev Michelob Ultra Beer, St. Louis, Missouri.
1:28 Speaker 03 Welcome to Decoder. This is Alex Heath, your Thursday episode guest host and deputy editor at The Verge. One of the biggest topics in AI these days is agents, or the idea that AI is going to move from chatbots to reliably doing things for us in the real world. The problem with agents right now is that they aren't really that reliable at all. There's a lot of work happening to fix that, which brings me to today's guest, David Luhan, the head of Amazon's AGI Research Lab. I've been wanting to chat with David for a long time. He was an early research leader at OpenAI, where he helped drive the development of GPT-2, 3, and DALL-E.
2:08 Speaker 03 After OpenAI, he co-founded Adept, a research lab focused on agents. And last summer, he left Adept to join Amazon, where he now leads the company's AGI lab in San Francisco. We recorded this episode right after the release of GPT-5, which gave us an opportunity to talk about why he thinks progress on AI models is slowing down. The work that David's team is doing is a big priority for Amazon, and this was the first time I've heard him really lay out what he's up to.
2:41 Speaker 03 I also had to ask him about how he joined Amazon. His leaving adept was one of the first of many deals that I call the reverse aqua hire, where a big tech company all but actually buys a buzzy AI startup to avoid antitrust scrutiny. I don't want to spoil too much, but let's just say that David left the startup world for big tech last year because he knew where the AI race was headed. I think that makes his predictions for what's coming next worth listening to.
3:23 Speaker 03 David, welcome to the show.
3:25 Speaker 04 Thanks so much for having me on the show. I'm really excited to be here.
3:28 Speaker 03 It's great to have you. We have a lot to talk about. I'm 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 about your history and how you got to Amazon, because you've been in the space for a long time, and you have a pretty interesting career leading up to this. And yeah, could you walk us through a little bit kind of your background in AI and how you got to Amazon?
3:54 Speaker 04 First off, I find it absolutely hilarious that I've been around the field for a long time, because it's true in relative terms, this field is so new. And yet nonetheless, like, you know, I've only been doing AI stuff for about the last 15 years. So compared to many other fields, it's just so new.
4:10 Speaker 03 Well, that's an eternity in AI years.
4:12 Speaker 04 It is an eternity in AI years. I remember when I first started working in the field, I worked in it just because I thought it was interesting. I thought having the opportunity to be able to build systems that could think like humans and hopefully deliver superhuman performance in things was such a cool thing to do. And I had no idea that it was gonna blow. Up the way that it did. But my personal background, I led the research and engineering team at OpenAI from 2017 to mid 2020, where the teams did GPT-2 and 3D and Clip and DALL-E. And every day was just so much fun because you would show up to work and it's just your best friends and all trying a bunch of really interesting research ideas. And there was none of the pressure that exists right now.
4:54 Speaker 04 Then after that, I led the LLM effort at Google where we trained a model called Palm, which was a quite strong model for its time. But a bunch of us shortly after that decamped to various startups. And my team and I ended up starting Adept. It was the first AI agent startup. We ended up inventing the computer use agent effectively. There was some good research beforehand. We had the first production one and Amazon brought us in to go run agents for Amazon about a year ago at this point.
5:21 Speaker 03 Great, and we'll get into that and what you're doing at Amazon. But first, given your OpenAI experience and we're talking now less than a week from the rollout of GPT-5, I'd love to hear you reflect on the release of GPT-5 and what it says about the industry, what you thought when you saw it. I'm sure you still have colleagues at OpenAI who worked on it. But yeah, what does that release signify?
5:45 Speaker 04 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 actually to train models. Your job as a frontier model lab is to build a factory that repeatedly churns out increasingly better models. That's actually a very different philosophy for how to make progress. In the I build a better model path, all you do is you think about, you know, let me make this tweak. Let me make this tweak. Let me try to glom on people to get a better release. If you care about it from the perspective of a model factory, what you're actually trying to do is you're trying to figure out how you can build all the systems and processes and infrastructure to make these things smarter.
6:30 Speaker 04 But with the GPT-5 release, I think the part that I found most interesting about it is that a lot of the frontier models these days are converging in capabilities. I think in part, there's an explanation that one of my old colleagues at OpenAI, who's now professor at MIT, came up with called the Platonic Representation Hypothesis. Have you heard of this hypothesis?
6:49 Speaker 03 No.
6:49 Speaker 04 So the Platonic Representation Hypothesis is this idea, similar to Plato's Cave, which is really what it's named after, that there is one reality, but we as humans, for example, only see a particular rendering of that reality. Like in Plato's Cave, it is the shadows that you see on the wall of the cave, right? And so that's the same thing for LLMs. LLMs see slices of this reality by the training data that it sees. So every incremental YouTube video of, for example, someone going for a nature walk in the woods somewhere is all ultimately generated by the actual reality that we live in. And as you train these LLMs on more and more and more data, the LLMs become smarter and smarter,
7:31 Speaker 04 they all converge to representing this one shared reality that we all have. And so if you believe this hypothesis, what you should also believe then is that all LLMs will converge to the same model of the world. And I think that's actually happening in practice from seeing frontier labs deliver these models.
7:47 Speaker 03 Well, there's a lot to that. I would maybe suggest that a lot of people in the industry don't necessarily believe we live in one reality. I was at, when I was at the last Google I/O, Sergey Brannan-Demis were on stage, and they both seemed to maybe believe that we were existing in multiple realities. So I don't know if that's, I don't know if that's a thing that you've encountered in your social circles or work circles over the years, but not everyone in AI necessarily believes that, right?
8:14 Speaker 04 I think that hot takes above my pay grade. I do think that we only have one.
8:18 Speaker 03 Yeah, we have too much to cover. We can't get into multiple realities. But to your point about everything converging, it does feel like benchmarks are starting to not matter as much anymore and that the actual improvements in the models.
8:34 Speaker 03 Like you said, are commodifying. Everyone's getting to the same point and GPT-5 will be the best on L M Arena for a few months until, you know, Gemini 3 comes out or whatever, and so on and so on. And if that's 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 to them. So OpenAI bringing back 4.0 because people had a literal attachment, you know, to it as a thing that they felt. And people on Reddit saying, it's like my best friend's been taking it away. And so it really doesn't matter that it's better at coding or that it's better at writing.
9:18 Speaker 03 It's it's your friend now. And that's that's freaky. But I'm 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?
9:32 Speaker 04 There was a project called Lambda or Mina at Google in the in 2020 that basically was ChatGPT before ChatGPT, but only available to Google employees. Even back then, we started seeing employees start developing personal attachments to these AI systems. Like, humans are so good at anthropomorphizing anything, right? And so I wasn't surprised to see that people form bonds with certain model checkpoints. But I think that when you talk about benchmarking, the thing that stands out to me is benchmarking is really all about, at this point, people are just studying for the exam, right? We know what the benchmarks are in advance. Everybody wants to post higher numbers. It's like the megapixel wars, right? From the early digital camera era. Like, they just clearly don't matter anymore. They have very loose correlation
10:16 Speaker 04 with how good of a photo does this thing actually take? And I think the question and the lack of creativity in the field that I'm seeing boils down to AGI is way more than just chat. It's 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's so many more useful applications and actually useful base model capabilities that people haven't even started figuring out how to measure well yet. And I think the better question to ask Now, if you want to do something interesting in the field is 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's so much more left to do.
11:02 Speaker 04 And when I think about what keeps me and the people that really are focused on this agent's vision that we have going is looking to solve way more breadth of problems than what people have done so far.
11:14 Speaker 03 Okay, that brings me to this topic. I was gonna ask it later, but yeah, AGI. You're running the AGI Research Lab at Amazon. I have a lot of questions about what AGI means to Amazon specifically, but I'm curious first for you. What did AGI mean to you when you were at OpenAI helping get GPT off the ground? And what does it mean to you now? Has that definition changed at all for you?
11:37 Speaker 04 The OpenAI definition for AGI we had was a system that outperforms humans at economically valuable tasks. And while I think that was an interesting, almost Doomer North Star back in 2018, I think we've gone so much past that as a field. What gets me excited every day is not how do I replace humans at economically valuable tasks? It's how do I ultimately build towards like a universal teammate for every knowledge worker? And just like what keeps me going is the sheer amount of leverage we can give to humans on their time. If we had AI systems that you could ultimately end up delegating a large chunk of the execution of what you do every day too.
12:19 Speaker 04 And so my definition for AGI, which I think is very attractable and is very much focused on helping people, is the first most important milestone that would lead me to say we're basically there is a model that can help a human do anything they want to do on a computer.
12:35 Speaker 03 I like that. That's actually more concrete and grounded than a lot of the stuff you hear. 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 now he thinks of AGI as a model that can self-improve itself. And I guess maybe that's related to what you're saying, but you're grounding it more in the actual use case, it sounds like.
13:00 Speaker 04 Well, the way that I look at it is self-improvement is interesting, but to what end? Right? Like, why do we as humans care if AGI self-improving itself? Like, I don't really care personally. I think it's cool from a scientist's perspective. I think what's more interesting to me is how do I go build the most useful form of this super generalist technology and then be able to put that in everybody's hands? And I think the thing that gives people tremendous leverage is if I can teach this agent that we're training to handle like any useful task that I need to get done on my computer, because so much of our lives these days is in the digital world, right? So I think that's like, it's very tractable.
13:44 Speaker 04 Going back to our discussion about benchmarking, right? The fact that the field cares so much about, you know, MMLU, MMLU Pro, you know, humanity's last exam, AMC 12, et cetera. Like, we don't have to live in that box of that's what AGI does for me. I think it's way more interesting to look at the box of the space of all useful knowledge worker tasks, how many of them are doable on your machine? And how can these agents do them for you?
14:12 Speaker 03 So it's safe to say that for Amazon, AGI means more than shopping for me, which is the cynical joke I was gonna make about what AGI means for Amazon. I'd be curious to know, when you joined and you were talking to the management team and Andy Jassy and, you know, still to this day, how you guys talk about the strategic value of AGI as you define it for Amazon broadly? Because Amazon is a lot of things. It's really a constellation of companies that do a lot of different things. But this idea kind of cuts across all of that, right?
14:44 Speaker 04 I think that if you look at it from the perspective of computing, right? 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 block of computing is changing, right? At this point, the code's written by an AI. Down the line, the actual intelligence and decision-making is gonna be done by an AI. And so then what happens to your building blocks, right? So in that world, it's super important for Amazon. To specifically be good at the agent's problem because agents are going to be the atomic
15:26 Speaker 04 building block of computing. And when that is true, I think so much economic value will be unlocked as a result of that. And it really aligns up well with the strengths that Amazon already has on the cloud side and putting together ridiculous amounts of infrastructure and all that.
15:39 Speaker 03 I see what you're saying. I think a lot of people listening to this, even people who work in tech, understand conceptually that agents are where the industry is 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 doesn't work. I would say that's pretty much the lay of the land right now. I'm not actually sure of, like, what would you hold out as this is the best example of an agent? This is the best example of where things are headed and what you can expect? Is there something you can point to?
16:12 Speaker 04 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't work at all. So let me try to give an example of what the actual promise of agents is relative to how they're pitched to us today. Right now, the way that they're pitched to us today is, for the most part, it's just a chatbot with extra steps, right? It's like, ah, you know, that company X doesn't 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 know, you've played with a product that does computer use or something like that, 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.
16:55 Speaker 04 Let's take a concrete example of, I want to do a particular drug discovery task where I know there's a receptor that 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's 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. But that's not an agent. An agent, in my book, is A model and a system that actually, literally, you can hook up to your wet lab. And it's 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
17:37 Speaker 04 experiment, see the results, react to that, try again, et cetera, until it's actually achieved the goal for you. And the degree to which like that gives you leverage is so, so, so much higher than what the field is currently able to do right now.
18:35 Speaker 02 This episode is brought to you by Michelob Ultra. Good times mean competitions between you and your friends, like seeing who can throw some pigskin the farthest or rush the field the fastest. Winner gets an Ultra. Because it's not just a superior light beer, it's also giving fans superior access to their favorite sporting events. Superior is worth playing for. Visit michelobultra.com/superioraccess to enter. Enjoy responsibly. Copyright 2026 Anheuser Busch InBev Michelob Ultra Beer, St. Louis, Missouri.
19:04 Speaker 05 Are your ad campaigns lighting up the dashboard? But not the pipeline? That's bullspend, and marketers are calling it out in. Dashboard Confessions.
19:13 Speaker 06 My boss asked for results, so I opened my dashboard for the only positive sounding metric I had: impressions.
19:21 Speaker 05 Cut the bullspend. See revenue, not just reach. LinkedIn delivers the highest return on ad spend of major ad networks. Advertise on LinkedIn. Spend $250 on your first campaign and get a $250 credit. Go to linkedin.com/campaign. Terms and conditions apply.
19:43 Speaker 03 Do you agree, though, that there's an inherent limitation in large language models and decision-making and executing things? When I see how LLMs, even still, you know, the frontier ones, still hallucinate, still make things up, confidently lie, it's terrifying to think of putting that technology in a construct where now I'm asking it to go do something in the real world, interact with my bank account, ship code, work in a science lab. Like, when ChatGPT can't spell right, that doesn't feel like the future that we're going to get. And so I'm wondering, are LLMs it, or is there more to be done here?
20:23 Speaker 04 So, we started with a topic of how these models are increasingly converging in capability. So while that's true for LLMs, I don't think that's been true to date for agents. And it's because the way that you should train an agent and the way that you train an LLM are actually quite different from each other. So, LLMs, as we all know, the bulk of their training happens from doing next token prediction, right? I've got a giant corpus of every article on the internet, let me try to predict the next word, and if I get the next word right, then I get a positive reward, and if I get it wrong, then I'm penalized, right? But, in reality, what's actually happening, this is, you know, in the field we call
21:04 Speaker 04 it behavioral cloning or imitation learning, it's the same thing as cargo culting, right? 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. The issue with this is, this is great for chat, this is great for creative use cases, right, 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, right? It's not, you know, just cloning human behavior. It's actually learning if I do X, the consequence of it is Y. And so the question is, how do we train agents to be able to learn the consequences of its actions?
21:49 Speaker 04 And the answer obviously cannot be just doing more behavioral cloning and copying text, right? It has to be something that looks like actual trial and error in the real world. And so that's basically the research roadmap for what we're doing on my, in my group at Amazon. My friend Andre Karpathy has a really good analogy here, which is, you know, imagine you have to train an agent to go play tennis, right? You wouldn't have it spend 99% of its time watching YouTube videos of tennis and the 1% of its time actually playing tennis. You would have something that's far more balanced between these two things. So what we're doing in our lab here at Amazon is we're actually doing large scale self-play.
22:29 Speaker 04 And so if you remember the concept of self-play, what it was, was a technique that really DeepMind made popular in the mid 2010s when they beat humans at playing Go. For playing Go, what they did was they spun up a bajillion simulated Go environments, right? And then they had the model play itself over and over and over again. Every time they found a strategy that was better at beating a previous version of itself, it would effectively get positive reward via reinforcement learning to go do more of that strategy in the future. And if you spent a lot of compute on this in the Go simulator, it actually discovered superhuman strategies for how to play Go. And then ended up, you know, when they played the world champion, making moves that no human had ever seen before and contributed to like the state of the art of that whole field.
23:14 Speaker 04 What we're doing is rather than doing more behavioral cloning or watching YouTube videos, what we're doing is we're creating a giant set of RL 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's a version of something that's like Salesforce. Here's a version of something that's like an ERP. Here's a CAD program. Here's an electronic medical records system. Here's accounting software. Here's, you know, every interesting Domain of possible knowledge work is now a simulator. And now, instead of training an L L M just to do tech stuff, we have the model actually propose a goal in every single one of these different simulators. Try solving that problem, figure out if it's successfully solved it or not.
24:00 Speaker 04 And then get reward and feedback based on, you know, oh, did I do the depreciation correctly? Or did I correctly make this part in CAD? Or did I successfully book the flight 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 A G I. And we're really scaling this recipe up at Amazon right now.
24:26 Speaker 03 How unique is this approach in the industry right now? Do you think the other labs are on to this as well? If you're talking about it, I would assume so.
24:32 Speaker 04 I think that what's interesting is this field. Ultimately, you have to be able to do something like this, in my opinion, to be able to get beyond the fact that there's a limited amount of free-floating data on the internet that you can train your models on. The thing we're doing at Amazon is because this came from what we did at Adapt, and Adapt has been doing agents for so long. We just care about this problem way more than everybody else. And I think have made a lot of progress towards this goal.
25:00 Speaker 03 You called these gyms, and I was thinking physical gyms for a second. Does this become physical gyms? Does it? You have a background in robotics, right?
25:08 Speaker 04 I've also done robotics work before. Here, we also have Peter Beal, who came from Copierian as a Berkeley professor that basically created, or his students ended up creating, the majority of the R L algorithms that work well today. It's funny that you say gyms, because we were trying to find an internal code name for the effort we kicked around Econox and Barry's Bootcamp and all this stuff. And I'm not sure everybody had the same sense of humor, but but we call them gyms, actually, because at OpenAI, we had a very useful early project called OpenAI Gym. And what it was, was this was Far way before L L M's were a thing. And what OpenAI Gym did was that was a collection of video game tasks and robotics tasks.
25:52 Speaker 04 Like, can you balance a pole that's on a on a cart? And can you train an R L algorithm that can that can keep that thing perfectly centered, et cetera? What we were inspired to do is now that these models are smart enough, why have toy tasks like that? Why not put in the actual useful tasks that humans do on their computer into these gyms? And have the models learn from these environments? And I don't see why this wouldn't also generalize to robotics.
26:15 Speaker 03 And is the end state of this a agent's framework system that gets deployed through A W S?
26:23 Speaker 04 The end state of all this is a model plus a system that is like rock solid reliable, like ninety nine percent reliable at all sorts of valuable knowledge for tasks that are done on a computer. And this is going to be something that we think is gonna be a service on A W S that's gonna underpin effectively so many useful applications in the future.
26:46 Speaker 03 I did a recent episode with Arvind, the CEO of Perplexity and 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. I'm curious what you think of that, this idea that it's not enough to just have a chatbot. You really need to have Chat G P T or whatever model sit next to your browser, look at the web page, act on it for you, learn from that. Is that where all this is headed on the consumer side?
27:18 Speaker 04 I think chatbots are definitely not the long term answer, or at least not chatbots in the way we think about it today. If you want to build systems that take actions for you. The best analogy I have for this is, so my dad is a very well intentioned, smart guy. Spent a lot of his career working in a factory. And he calls me all the time for tech support help. And so he's like, David, something's wrong with my iPad. You gotta help me with this. And we're just doing this over the phone, and I can't see what's on the screen for him. And so I'm trying to figure out, like, you know, like you have the settings menu open. Have you clicked on this thing yet? Like, what's going on with this toggle? Chat is such a low bandwidth interface. Like, 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.
28:01 Speaker 04 So one of the big missing pieces, in my opinion, right now in AI is our lack of creativity with product form factors, frankly, right? We're so used to thinking that the right interface between humans and AIs is this like perpendicular one-on-one interaction where I'm delegating something or maybe giving me some news back, or I'm asking you a question, et cetera. One of the real things we've always missed is this parallel interaction where both the user and the AI actually have a shared canvas that they're jointly collaborating on. I think if you really think about building, you know, a teammate for knowledge workers, or even just the world's smartest personal assistant, you would want to live in a world where there's actually a shared collaborative canvas for the two of you.
28:45 Speaker 03 Speaking of collaboration, I'm really curious how your team works with the rest of Amazon. Are you pretty walled off from everything? Do you work on Nova, Amazon's foundational model? Like, how do you interact with the rest of Amazon?
28:59 Speaker 04 What Amazon's done a great job with for what we're doing here is we're allowed to run pretty independently. And I think there's a recognition that some of the startup DNA right now is really valuable for maximum speed. If you believe AGI, right, is two to five years away, some people are getting more bullish, some people are getting more bearish. Doesn't matter. That's not a lot of time in the grand scheme of things. You need to move really, really fast. So we've been given, we've been given a lot of independence. We've also taken the tech stack that we've built and contributed a lot of that upstream to the Nova Foundation model as well.
29:35 Speaker 03 So does your work, for example, is it already impacting Alexa Plus? Or is that not something that you're part of in any way?
29:43 Speaker 04 That's a good question. So Alexa Plus has the ability to, for example, if your if your toilet breaks, it's like, ah, man, I really need a plumber. Alexa, can you get me a plumber? Then what happens is Alexa Plus spins up a remote browser powered by our technology, basically, that then goes and uses thumbtack like a human to go get you a plumber to your house, which I think is really cool. It's the first production web agent that's been shipped, if I remember correctly.
30:11 Speaker 03 Yeah, and you know, the early reception to Alexa Plus has been that it's dramatically for Alexa, but still brittle. There's still moments where it's not reliable. And I'm wondering, is this the real gem? Is this the at-scale gem 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's on. Is that the strategy? Or, because I'm sure you've seen, there's the early reactions to Alexa Plus are, it's better, but still not as reliable as people would like it to be.
30:48 Speaker 04 Alexa Plus is just one of many customers that we have. And what's really interesting about being within Amazon is, going back, we were talking about earlier, web data is effectively running out, and it's not useful for training agents. It's actually useful for training agents is lots and lots of environments and lots and lots of people doing reliable multi-step workflows. And so the interesting thing at Amazon is that in addition to Alexa Plus, basically every Fortune 500 business's operations are represented in some way by some internal Amazon team, right? Like this One Medical, there's everything happening on supply chain and procurement on the retail side, there's all this developer-facing stuff on AWS. And agents are going to require a lot of private data and private environments
31:32 Speaker 04 to be trained. And because we're in Amazon, that's all now 1P. So there's just one of many different ways in which we can get reliable workflow data to train the smarter agent.
31:41 Speaker 03 Are you doing this already through Amazon's logistics operations, where you can do stuff in warehouses, or, you know, the robotic stuff that Amazon is working on? Does that intersect with your work already?
31:53 Speaker 04 Well, we're really close to Peter Beal's group on the robotic side, which is awesome. Some of the other areas, we have this, a big push for internal adoption of agents within Amazon. And so a lot of those conversations or engagements are happening.
32:05 Speaker 03 I'm glad you brought that up. I was gonna ask, how are agents being used inside Amazon today?
32:10 Speaker 04 Again, as we were saying earlier, because Amazon sort of has a internal Effort for almost every useful domain of knowledge work. There's been, you know, a lot of enthusiasm to pick up a lot of these systems. And we have this internal channel called, actually, I won't tell you what it's called, but you know, codename dash interest, which is related to the product that that we've been building. And it's just been crazy to see teams from all over the world within Amazon, actually. Because one of the main bottlenecks we've had is we didn't actually have availability outside of the US for quite a while. 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.
32:53 Speaker 03 This is your just agent framework that you're talking about? This is something you haven't released publicly yet?
32:58 Speaker 04 We released Nova Act, which was the research preview that came out in March. But as you can imagine, we've added way more capabilities since then, and it's been really cool. The thing we always do is we first dog food with internal teams.
33:11 Speaker 03 Yeah, 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've put that out, how are people using Nova Act? It's not something that, you know, in my day to day I hear about, but I assume companies are using it. And I'd be curious to hear what the feedback is that you've guys have gotten since you put it out.
33:33 Speaker 04 Yeah, so a wide range of enterprises and developers are using Nova Act. And the reason why it's not something that you hear about is because we're not a consumer product. If anything, the whole Amazon agent strategy, including what I did before at Adept, is sort of doing norm core agents, not the super sexy stuff that, you know, works one out of three times, but super reliable, low level workflows that work 99 plus percent of the time. So that's the target. Since Nova Act came out, we've actually had a bunch of different enterprises end up deploying with us where they're seeing 95 plus percent reliability, which is, as I'm sure you've seen from the coverage of, you know,
34:15 Speaker 04 other agent products out there, is like a material step up from the average 60% level of reliability that that folks see with those systems. And I think that reliability Bottleneck is why you don't see as much Asian adoption overall in the field. And we've been having a lot of really good luck, specifically by focusing extreme amounts of effort on reliability. So we're now used for things like, for example, doctor and nurse registrations, right? Or we have another customer called Nirvan, which is formerly TripActions, which uses us basically to automate a lot of backend bookings for travel for their customers. Uh, we've got companies that basically have like ninety-three step QA workflows
34:56 Speaker 04 that they've automated with a single act script, et cetera. So I think the early progress has been really cool. Now what's up ahead is how do we do this extreme, large-scale self-play on a bajillion gyms? To get to something where there's a bit of a GPT for RL agents moment, and we're running as fast as we can towards that right now.
35:14 Speaker 03 Do you have a line of sight to that? Do you think we're two years from that? One year?
35:19 Speaker 04 Honestly, I think we're sub one year. We have line of sight. We've built out teams for every step of that particular problem. And things are just starting to work. It's just really fun to go to work every day. And you realize that one of the teams has made a little very useful breakthrough that particular day. And the the whole cycle that we're doing for this training loop seems to be going a little bit faster every day. Going back to GPT five, people have said, you know, does this portend a slowdown in AI progress? And one hundred percent, I think the answer is no. Because when one S-curve peters out, right? The first one being pre-training, uh, which I don't think is petered out, by the way, either, but it's definitely at this point, like less easy to get gains than before. And then you've got RL with verifiable rewards.
36:03 Speaker 04 But then every time one of these S-curves seems to slow down a little bit, there's another one coming up. And I think agents is 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.
36:30 Speaker 01 New and exclusive holiday decor just launched online at The Home Depot. Check out the viral Grand Duchess collection with smart holiday lights you can control from your phone. Easily customize the twinkle across trees, wreaths, and garlands to match your holiday vibe. Create stunning light effects, then switch it up with the tap of your fingers. Even the man in red will be impressed. Shop the holiday decor you want, online only, at homedepot.com.
37:01 Speaker 02 This episode is brought to you by Michelob Ultra. Good times mean competitions between you and your friends, like seeing who can throw some pigskin the farthest, or rush the field the fastest. Winner gets an ultra. Because it's not just a superior-like beer, it's also giving fans superior access to their favorite sporting events. Superior is worth playing for. Visit michelobultra.com/superioraccess to enter. Enjoy responsibly. Copyright 2026 Anheuser-Busch InBev Michelob Ultra Beer, St. Louis, Missouri.
37:30 Speaker 05 Flowing ad budget on metrics that look great, till the CFO sees them? That's bullspend, and marketers are calling it out, in. Dashboard, Confessions.
37:39 Speaker 07 I remember telling my boss, "It'll be good for the brand," when leads were slow. Yeah, it, it wasn't.
37:47 Speaker 05 Cut the bullspend. LinkedIn lets you target by company, job title, and more. Advertise on LinkedIn. Spend $250 on your first campaign and get a $250 credit. Go to linkedin.com/campaign. Terms and conditions apply.
38:08 Speaker 03 It sounds like you and your colleagues have identified the next turn that the industry is going to have. And that starts to put Nova, as it exists today, into more context for me. Because Nova as an LLM, it's not an industry-leading LLM. I mean, it's, it's not in the same conversation as Claude, or GPT-5, or what have you, or Gemini. Is Nova just not as important because what's really coming is what you're talking about with agents, and that 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?
38:48 Speaker 04 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 sort of game, you need to bet on something that can really leapfrog, right? What's interesting is every time there's a recipe change of how these models are trained, it creates a giant window of opportunity for someone new who's starting to come to the table with that new recipe instead of trying to catch up on all the old recipes. 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, right? The whole LLM thing came out of GPT-2 and then three, of course.
39:32 Speaker 04 But those LLMs initially, they were text-only training recipes. And then we discovered RLHF, and then they started getting a lot of human data via RLHF. But then in the switch to multimodal, 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, was that they bet on certain interesting ideas on native multimodal that turned out well for them, right? But then, after that, with reasoning models, right? They gave another opportunity for people to catch up. That's why DeepSeek was able to surprise the world, because they straight quantum tunnel to that instead of doing every stop along the way. And I think with the next turn being agents, especially agents without verifiable rewards,
40:14 Speaker 04 if at Amazon, we can figure that recipe out earlier, faster, better than everybody else with all the scale that we have as a company, it basically brings us to the frontier at that point.
40:25 Speaker 03 I haven't heard that articulated from Amazon before. That's really interesting. It makes a lot of sense. Let's end on the state of the talent market and startups and how you came to Amazon, actually, I want to go back to and adapt. Adapt, when you started it, was it the first startup really focusing on agents at the time? I don't think I had heard of agents until I saw Adapt.
40:51 Speaker 04 Yeah, actually, we were the first startup focusing on agents because when we were starting Adapt, 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. We got everybody focused on solving that. But when we got started, the word agent as a product category wasn't even coined yet. So we were trying to find a good term. And we started, 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 it up as being the term.
41:25 Speaker 03 Walk me through the decision to leave it behind and join Amazon with most of the technical team, is that right?
41:31 Speaker 04 Mm-hmm.
41:31 Speaker 03 I have a phrase for this. It's a deal structure that has become common now with big tech and startups and AI, the reverse aqua hire, where basically the core team, like yourself and your co-founders, they join, the rest of the company still exists, but you know, the technical team goes away and the acquirer, quote unquote, I know it's not an acquisition, but pays a licensing fee or something and shareholders make money. But the startup is then kind of left to figure things out without its founding team in most cases. And you know, the most recent example is Wind Surf and Google, and then there was Scale AI and Meta before that. This is a topic we've been talking about on Decoder a lot. The listeners are familiar with it. But you were one of the first of
42:13 Speaker 03 these such reverse aqua hires. Walk me through when you decided to join Amazon and why.
42:19 Speaker 04 So I hope, you know, in 50 years, I'm remembered more as being an AI research innovator rather than a deal structure innovator. Well, first off, humanity's demand for intelligence, right, is way, way, way higher than the amount of supply. And so therefore, for us as a field to invest ridiculous amounts of money in building the world's biggest clusters and bringing the best talent together to drive those clusters is actually perfectly rational, right? Because if you can spend, you know, an extra X dollars to build a model that has plus 10 IQ points and can solve like a giant new concentric circle of useful tasks for humanity, that is a worthwhile trade that you should do any day of the week. And so I think it makes a lot of sense that all these companies are trying to put together
43:03 Speaker 04 critical mass on both talent and compute right now. And from my perspective, for why join Amazon, it's because Amazon knows how important it is to win on the agent's side in particular, and that agents are a crucial bet for Amazon to really build one of the best frontier labs possible. And to get to the level of scale, right? You're hearing all these CapEx numbers from the various hyperscalers. It's just completely mind boggling and it's all real, right?
43:31 Speaker 03 It's over 340 billion in CapEx this year alone, I think, from just the top hyperscalers. Yeah, it's an obscene number.
43:40 Speaker 04 That sounds about right. And Adept, you know, we raised 450 million, which at the time was a very large number. And then today is
43:48 Speaker 03 It's chump change now.
43:49 Speaker 04 It's chump change.
43:50 Speaker 03 That's one researcher. Come on, David.
43:52 Speaker 04 That's one researcher, right? That's one employee. So if that's the world that you live in, it's really important, I think, for us to partner with someone who's gonna go fight all the way to the end. And that's why we came to Amazon.
44:04 Speaker 03 Did you foresee that consolidation and those numbers going up when you did the deal with Amazon? You knew that it was gonna just keep getting more expensive? Not only on compute, but on talent?
44:14 Speaker 04 Yes.
44:15 Speaker 03 And why? What did you see coming that at the time it was not obvious to everyone?
44:19 Speaker 04 Two things I saw coming. One, if you want to be at the frontier of intelligence, you have to have, you have to be at the frontier of compute. And if you're not on the frontier of compute, then you have to pivot and go do something that is totally different. And my whole career, all I wanna do is build the smartest and most useful AI systems. So the idea of turning Adept into, you know, enterprise company that only sells small models or, you know, turns into a place that does forward deployed engineering to go help you deploy an agent on top of someone else's model, none of those things appealed to me. Like, I wanna, I wanna figure out, here are the four crucial remaining research problems left to AGI. How do we nail them?
45:01 Speaker 04 Every single one of them is gonna require like two-digit billion-dollar clusters to go run at. So how else am I gonna be able to have, and this whole team that I've put together, who all are motivated by the same thing? How are we gonna have the opportunity to go do that?
45:15 Speaker 03 If antitrust scrutiny did not exist for big tech like it does, would Amazon have just acquired the company completely?
45:21 Speaker 04 I can't speak to general motivations and deal structuring again. I'm an AI research innovator, not a
45:28 Speaker 03 You know, I have to ask. Well, you know, okay, well, then 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? I think it changes the calculus for someone joining a startup these days, knowing that these kind 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. And that that is a shift. That is a new thing for Silicon Valley in the last couple of years.
46:03 Speaker 04 There's 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 prorata and equally, right? The second thing is it's very counterintuitive in AI right now because there's only a small number of people with a lot of experience. And because the next couple of years is 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. If you're sitting there responsible for one of these labs and you want to make sure that
46:44 Speaker 04 you have the best possible AI systems, you need to hire the people who know what they're doing. And so there's the market demand, the pricing for these people is actually totally rational just solely because of how few of them there are. But the counterintuitive thing is that it doesn't take that many years actually to find yourself at the frontier if you're a junior person. Like 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, just like working really, really, really hard, they found themselves at the frontier. Like AI research is one of those areas where if you ask four or five questions. You already discovered a problem that nobody has the answer to. And then you can just focus on that and focus on how do I become the world expert in this
47:28 Speaker 04 in this particular sub domain. And so I find that really counterintuitive that there's only very few people who really know what they're doing, and yet it's very easy in terms of the number of years to become someone who knows what they're doing.
47:39 Speaker 03 How many people actually know what they're doing in the world? And 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?
47:49 Speaker 04 I think it depends on how generous or tight you want to be. I would say there's
47:53 Speaker 03 Who can actually build and conceptualize training a frontier model holistically?
47:59 Speaker 04 The number of people who can, who I would trust with a giant dollar amount of compute to go do that is probably sub 150.
48:10 Speaker 03 150?
48:11 Speaker 04 Yes, but there are many more people, let's say another 500 people or so, that would be extremely valuable contributors to an effort that was populated by a certain critical mass of that 150 that really know what they're doing.
48:26 Speaker 03 But the total market, that's still less than 1,000 people.
48:29 Speaker 04 I'd say it's probably less than 1,000 people. But again, I don't want to, I don't want to trivialize. I think junior talent is extremely important and people who come from other domains like, like physics or quant finance, or, you know, 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.
48:58 Speaker 03 Does the fact that these, this already very small group of elite people, does the fact that they're building something that inherently is designed to replace them? Maybe you disagree with that, but I think superintelligence conceptually would make some of this redundant. Does it mean there's actually fewer of them in the future making more money because you only need, you know, some orchestrators of other models to build more models? Or does the field expand? Do you think it's going to become thousands and thousands of people?
49:30 Speaker 04 The field's definitely going to expand. There's 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's going to keep the field smaller than other fields like software is that unlike regular software engineering, a foundation model training breaks so many of the rules that we think we should have, right? Like in software, let's say our job here is to build Microsoft Word, right? I can say, hey, Alex, it's your job to make the save feature work. It's David's job to make sure that cloud storage works and someone else's job to make sure the UI looks good. You can factorize these problems pretty independently from each other.
50:15 Speaker 04 The issue with foundation model training is that every decision you take interferes with every other decision because there's only one deliverable at the end. The deliverable at the end is your frontier model. It's like one giant bag of weights, right? So what I do in pre-training, what this other person does in supervis, fine tuning, this other person does in RL, and this other person does to make the model run fast, they all interact with each other in sometimes pretty unpredictable ways. So it has one of the worst dis economies of scale with number of people of anything I've ever seen, except maybe even sports teams, right? Maybe that's the one, that's the one other case where you don't want to have like 100 mid-level people, you want to have 10 of the best, right? And 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.
50:59 Speaker 03 Oh, so you think the elite, elite stays relatively where it is, but the field around it, the people that support, the people that are very meaningful contributors expands?
51:10 Speaker 04 I think the people who know how to do super meaningful work will definitely expand, but it will be still a little constrained by the fact that you cannot have too many people on any one of these projects at once.
51:22 Speaker 03 What advice would you give someone who's 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 they should be thinking about navigating the next couple of years with all this change that we've been talking about?
51:41 Speaker 04 First off, tiny teams with lots of compute is the correct recipe for building a frontier lab. That's what we're doing at Amazon with Saphen and my team. It's 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're not really going to have a chance. There's so many senior people ahead that are all too ready to try their particular ideas. The second thing is, I think people underestimate the co-design of like, the product and the user interface and the model. I think that's going to be the most important game that people are going to play in the next couple of years. And so going somewhere that actually is very strong product sense and a vision for how users are actually going
52:24 Speaker 04 to deeply embed this into their own lives is going to be really important. And one of the best ways to tell is, are you just building another chatbot? Are you just trying to fight, you know, one more entrant in the coding assistant space, right? 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 no one's really solved today. And if you really want to take an asymmetrical upside bet, I would try to spend some time and figure out what those are now.
53:04 Speaker 03 Spend some time in the gym. Thanks, David. I'll let you get back to your gyms.
53:07 Speaker 04 Cool. Nice. Thanks, guys. This is really fun.
53:12 Speaker 03 Thanks again to David Luan for joining the show and thank you for tuning in. If you'd like to let us know what you thought about this episode or what else you'd like us to cover, drop us a line. You can email us at [email protected] . We also have a TikTok and an Instagram. Check those out at DecoderPod. If you like Decoder, please share it with your friends and subscribe wherever you get your podcasts. And if you haven't already, don't forget to subscribe to The Verge, which gets you access to all of our stories and newsletters, including the one I author called Command Line. Decoder is a production of The Verge and is part of the Vox Media Podcast Network. Our producers are Kate Cox and Nick Statt. Our editor is Ursa Wright. The Decoder music is by Breakmaster Cylinder. See you next time.
54:03 Speaker 01 New and exclusive holiday decor just launched online at The Home Depot. Check out the viral Grand Duchess Collection with smart holiday lights you can control from your phone. Easily customize the twinkle across trees, wreaths, and garlands to match your holiday vibe. Create stunning light effects, then switch it up with the tap of your fingers. Even the man in red will be impressed. Shop the holiday decor you want, online only, at homedepot.com.
54:33 Speaker 08 Mr. Monopoly here. Miss me? Monopoly is back at McDonald's with millions of prizes. A 2027 Jeep Cherokee, a Carnival Cruise Vacation, gift cards from DoorDash, and Best Buy. Feel and play in the McDonald's app. Get your bag.
54:51 Speaker 09 No purchase necessary. 50 U.S. and D.C. age 18+. See playatmcd.com for rules and playatmcd.com/amoe for info and playing without purchase by mail by November 2nd. Chance card game ends November 8th. Game ends November 22nd, 2026. See rules for prize details. Monopoly is a registered trademark of Hasbro.
55:04 Speaker 10 Looking for a simple way to thank your clients or recognize employees for a job well done? A Starbucks card is more than a gift. It's a pick-me-up, a break in their day, and a reminder that you appreciate them. Whether you're shopping for digital or physical cards in bulk, Starbucks cards are the perfect gift to brighten anyone's day. Share the joy of coffee and connection when you give the gift of Starbucks. Shop Starbucks cards in bulk now at starbuckscardtv.com.
Previous How the head of Obsidian went from superfan to CEO Next Is ChatGPT killing higher education? Page updated October 26, 2026 · transcript and summary are machine-generated
