Amazon Is Betting on Agents to Win the AI Race  ## Metadata - Author: [[Decoder with Nilay Patel]] - Full Title: Amazon Is Betting on Agents to Win the AI Race - Category: #podcasts - URL: https://share.snipd.com/episode/19dc45f1-9998-4ef3-ad0e-e9ff805bb430 ## Document Notes ## Highlights - **Models Are Becoming Factory Outputs** - Frontier labs are now building factories that repeatedly churn out better models rather than one-off models. - Different labs converge on similar capabilities as models learn more of the same underlying reality. Transcript: David Luan 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. 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 a professor at MIT, came up with called the platonic representation hypothesis. Have you heard of this hypothesis? No. 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. 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 ([Time 0:05:58](https://share.snipd.com/snip/daa3497e-0857-4261-99ec-8fd4b54ff5ef)) - **Train For Real Problems, Not Benchmarks** - Stop optimizing only for benchmark improvements and instead pick a concrete class of problems to solve. - Focus on building capabilities that handle broad, useful tasks beyond chat and code. Transcript: David Luan 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? ([Time 0:10:40](https://share.snipd.com/snip/fce48a57-d5b5-4e92-b719-c7184d053a4c)) - **AGI As A Universal Teammate** - Luan defines AGI as a model that can help a human do anything on a computer. - That framing prioritizes building universal teammates for knowledge workers over abstract self-improvement definitions. Transcript: David Luan 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 to. 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. ([Time 0:11:50](https://share.snipd.com/snip/93c0ec32-84ca-4f0e-baef-9617545742df)) - **Agents Need To Execute End-To-End Workflows** - Current 'agents' often behave like chatbots with a few extra steps and remain brittle. - The true promise is agents that execute multi-step real-world workflows end-to-end, e.g., running experiments in a lab until a goal is met. Transcript: David Luan 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, you know, a 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 and 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. 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 in 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 experiment, see the results, react to that, try again, et cetera, until it's actually achieved the goal for You. ([Time 0:16:25](https://share.snipd.com/snip/5a987027-9b87-4508-8bb5-ba13a29aefe0)) - **Behavioral Cloning Limits Agents** - LLMs trained by next-token prediction imitate behavior without learning causal consequences. - Agents must learn consequences via trial-and-error reinforcement learning to make reliable decisions. Transcript: David Luan So we started with a topic of how 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 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 his actions? 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. ([Time 0:21:08](https://share.snipd.com/snip/6327ff97-00ae-4065-ab13-45a412baba45)) - **Scale Self-Play In Many Gyms** - Train agents with large-scale self-play inside many task-specific simulators or 'gyms'. - Use rewards from task success to teach consequences of actions across diverse knowledge-work environments. Transcript: David Luan 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. 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 record 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 LLM 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, 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. ([Time 0:23:14](https://share.snipd.com/snip/7732d74f-74eb-47ae-9021-27bd8573f406)) - **Phone Tech Support Shows Chat Limits** - David described helping his dad with iPad tech support to illustrate chat's bandwidth limits. - He used the story to show why chat is a poor interface for action-heavy help. Transcript: David Luan And he calls me all the time for tech support help. And so he's like, David, something's wrong with my iPad. You got to 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? Oh, 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. ([Time 0:28:20](https://share.snipd.com/snip/b6f787ae-a6b5-4d8d-b8b1-be3816829e5c)) - **Build Shared Collaborative Interfaces** - Design agent products as shared collaborative canvases rather than one-on-one chat. - Build interfaces where user and AI act in parallel on the same workspace for better teamwork. Transcript: David Luan One of the real things we've always missed is this parallel interaction where both the user the AI actually have a shared canvas that they're jointly collaborating on. I think if you really think about building a teammate for knowledge workers or even just the world's smartest personal assistant, you would want to live in a ([Time 0:29:08](https://share.snipd.com/snip/a4d1c00a-1a53-40b9-9c5f-28f254d63410)) - **Alexa Plus Is A Production Web Agent** - Alexa Plus uses a remote browser powered by Luan's technology to perform tasks like booking a plumber via Thumbtack. - He noted it as an early production web agent shipped at Amazon. Transcript: David Luan That's a good question. So Alexa Plus has the ability to, for example, if your toilet breaks, it's like, oh 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. ([Time 0:30:28](https://share.snipd.com/snip/b3c87191-c0f4-4bd5-898e-f43c73a43f1e)) - **Amazon's First-Party Data Is An Agent Advantage** - Amazon's massive internal operational data and private environments are unique training assets for agents. - Those 1P workflows let Amazon train agents on reliable multi-step real-world tasks at scale. Transcript: David Luan And what's really interesting about being within Amazon is going back to what we were talking about earlier, web data is effectively running out and it's not useful for training agents. What'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 there's 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 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. ([Time 0:31:36](https://share.snipd.com/snip/ca021e55-fea2-4677-8c50-32b88172752b)) - **Prioritize Reliability Over Flash** - Prioritize reliability over novelty when deploying agents; aim for 95%+ success in workflows. - Focus on normcore, high-accuracy automations rather than flashy but brittle demos. Transcript: David Luan 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 other agent products out there, is like a material step up from the average 60 percent level of reliability that folks see with those systems. And I think that reliability bottleneck is why you don't see as much agent 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 Nivan, which is formerly Trip Actions, which uses us basically to automate a lot of back-end bookings for travel for their customers. We've got companies that basically have like 93-step QA workflows that they've automated with a single-act script, etc. So I think the early progress has been really cool. ([Time 0:34:18](https://share.snipd.com/snip/b74aa7ea-2aa5-48a6-a0e2-5f5395b1c6c1)) - **Agents Are The Next S-Curve** - When one S-curve of progress slows (e.g., pretraining), another rises; agents are the next S-curve. - Large-scale RL on many environments could unlock the next wave of rapid AI progress. Transcript: David Luan 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 whole cycle that we're doing for this training loop seems to be going a little bit faster every day. Going back to GPT-5, people have said, does this portend to slow down in AI progress? And 100%, I think the answer is no, because when one S-curve peters out, the first one being pre-training, which I don't think is petered out, by the way, either, but it's definitely At this point less easy to get gains than before. And then you've got RL with verifiable rewards. 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. ([Time 0:36:04](https://share.snipd.com/snip/c3f0ceb3-c6f3-41d5-a920-53d5d601adc2)) - **Adept Pioneered The Agent Startup Story** - Adept started as the first startup explicitly focused on agents and produced an 'action transformer' product. - The company later popularized the 'agent' term as the product category caught on. Transcript: David Luan Yeah, actually, we were the first startup focusing 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. 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 played with things like large action models and action transformer. So our first product was called action transformer. ([Time 0:41:35](https://share.snipd.com/snip/d05fe05a-51c0-4058-a286-d72e051a4454)) - **Frontier Work Needs Frontier Compute** - If you aim to build frontier AI, secure massive compute resources or join an organization that already has them. - Don’t expect small startups to scale frontier training without multi-billion-dollar clusters. Transcript: David Luan So the idea of turning Adept into, you know, an 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 appeal to me. Like, I want to figure out, here are the four crucial remaining research problems left to AGI. How do we nail them? Every single one of them is going to require, like, two-digit billion-dollar clusters to go run at. ([Time 0:45:23](https://share.snipd.com/snip/13d36ceb-0bd7-4bdc-be05-6ee725c08a60)) - **Frontier Talent Is Extremely Concentrated** - The number of people you fully trust to run frontier model projects is very small, roughly under 150. - A slightly larger group (~500) can contribute meaningfully when paired with that core talent. Transcript: David Luan Think it depends on how generous or tight you want to be. I would say probably- Who can actually build and conceptualize training a frontier model holistically. The number of people who I would trust with a giant dollar amount of compute to go do that is probably sub 150. 150? Alex Heath Yes. David Luan 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. Alex Heath But the total market, that's still less than a thousand people. David Luan I'd say it's probably less than a thousand people. But again, I don't want to, I don't want to trivialize. I think junior talent is extremely important. ([Time 0:48:33](https://share.snipd.com/snip/2a7befcc-2f48-410f-a472-7026411469bb)) - **Choose Tiny Teams With Product Focus** - Join small, high-compute teams and keep strong product co-design skills to maximize impact. - Avoid large bureaucratic orgs where you can't run experiments quickly or shape product-model co-design. Transcript: David Luan 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 SF 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 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 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? 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. ([Time 0:52:25](https://share.snipd.com/snip/eb0abd90-9606-4f68-bc9e-72faed51ed79))