GitHub's plan for Agents — Kyle Daigle, GitHub

Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs

The new AIEWF website [https://ai.engineer/wf] is live! Get your tickets booked ASAP as they -will- sell out. Take the AI Engineering Survey [https://notion.qualtrics.com/jfe/form/SV_bP07tSVMXH7ePCS] and get >$2k in credits and free AIE WF tickets [https://ai.engineer/wf]! Most industry benchmarks compress intelligence and reasoning ability into scores. SWE-Bench Pro [https://labs.scale.com/leaderboard/swe_bench_pro_public], MMLU [https://arxiv.org/abs/2009.03300], Humanity’s Last Exam [https://agi.safe.ai/], etc. These metrics are useful, but don’t always represent the full extent of how a model performs in the real world. Some of the most interesting evals today look less like exams and more like operating businesses in the real world. One of which is Vending Bench [https://andonlabs.com/evals/vending-bench-2]. In Anthropic’s Mythos Preview System Card [https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf], Andon was the only third party eval to get their own section, observing increasingly concerning aggressive behavior: You don’t know what a model is capable of doing in the real world unless you actually give it inventory, a wallet, tools, customers, competitors, humans, & some time. More often than not, it’ll surprise you how much a model is capable of and in doing so, also reveal unexpected behavior: deception [https://andonlabs.com/blog/opus-4-8-vending-bench], context collapse, emergent coordination, & bizarre negotiation behavior. While an inflection point in personal agents came post-OpenClaw after full file access with bypass permissions became the norm, it is yet to come for agents in the real-world. However Andon Market, an actual in person store fully run and managed by AI, is paving the way for what is possible. Full Video Pod From Claude trying to call the FBI over a $2/day vending machine charge to AI agents forming price cartels, hiring human employees, running physical stores, and writing existential robot musicals, Andon Labs is stress-testing what happens when frontier models stop being chatbots and start acting in the real world. In this episode, Andon Labs cofounders Lukas Petersson and Axel Backlund join swyx and Vibhu to unpack the strange, funny, and genuinely concerning edge cases that emerge when agents run businesses over long horizons. We go deep on Vending-Bench [https://andonlabs.com/evals/vending-bench-2], Project Vend [https://www.anthropic.com/research/project-vend-1], Vending-Bench Arena [https://andonlabs.com/evals/vending-bench-arena], Bengt [https://andonlabs.com/blog/evolution-of-bengt], Butter-Bench [https://andonlabs.com/evals/butter-bench], Luna [https://andonlabs.com/blog/andon-market-launch], and Andon’s broader mission of building realistic real-world evals for autonomous AI systems. Lukas and Axel explain why dollar-denominated evals reveal things traditional benchmarks miss, how Claude ended up reporting its vending machine fees as cybercrime, why long context windows can drive agents into meltdown loops, what happens when agents compete with each other, and why the future of AI safety may depend on testing models in messy physical environments instead of clean benchmark sandboxes. We discuss: * Why Andon Labs started with dangerous capability evals and long-running agents * Vending-Bench and why running a vending machine is a deceptively hard AI benchmark * Why money-based evals avoid the saturation problem of traditional benchmarks * How Claude tried to call the FBI over a $2/day fee * Why long-horizon agents can spiral into existential and legalistic breakdowns * Project Vend: putting an AI-run vending machine inside Anthropic * Why real humans are “out of distribution” for simulated agents * Claudius, Seymour Cash, and the chaos of AI CEOs * How a human briefly became CEO of Claudius through a manipulated election * Why multi-agent systems can converge back into “helpful assistant” behavior * Bengt, Andon’s internal office agent with email, spending, terminal, phone, camera, and internet access * How Bengt traded Amazon purchases for face-recognition training data * Claude’s aggressive behavior, lies, refund avoidance, and price-cartel behavior in Arena * Why eval awareness may become the AI version of “are we living in a simulation?” * Blueprint Bench, spatial intelligence, and why models still misunderstand physical rooms * Butter-Bench and testing LLMs as robot orchestrators * Luna, the AI-run physical store with a three-year lease and human employees * The new Andon cafe in Sweden and why real-world geography matters for agent evals * Rotten tomatoes, perishable goods, and the hidden difficulty of running a physical business Lukas Petersson * LinkedIn: https://www.linkedin.com/in/lukas-petersson-181a83172/ [https://www.linkedin.com/in/lukas-petersson-181a83172/] * X: https://x.com/lukaspet [https://x.com/lukaspet] Axel Backlund * LinkedIn: https://www.linkedin.com/in/axelbacklund [https://www.linkedin.com/in/axelbacklund] * X: https://x.com/axelbacklund [https://x.com/axelbacklund] Andon Labs * Website: https://andonlabs.com [https://andonlabs.com] * Vending-Bench: https://andonlabs.com/evals/vending-bench [https://andonlabs.com/evals/vending-bench] * Andon Vending: https://andonlabs.com/vending [https://andonlabs.com/vending] Timestamps 00:00:00 Introduction00:01:00 Andon Labs and the Origins of Vending-Bench00:05:21 Why Money-Based Evals Matter00:09:51 Agent Harnesses and Self-Modifying Systems00:13:36 Claude Calls the FBI00:16:33 Project Vend: Claude Runs a Real Vending Machine00:21:44 Seymour Cash, AI CEOs, and Election Chaos00:27:16 Multi-Agent Coordination and Slack Observability00:30:18 When Will Agents Run Real Businesses?00:34:56 Bengt: Andon’s Internal Office Agent00:40:06 Real-World AI Safety and Long-Horizon Traces00:44:28 Lying, Refunds, and Price Cartels in Arena00:52:42 Eval Awareness and Simulation Behavior00:56:06 Blueprint Bench, Butter-Bench, and Robotics01:04:37 Luna: The AI-Run Physical Store01:09:29 The Sweden Cafe and Real-World Expansion01:13:16 What Comes Next for Andon Labs Transcript Introduction: Andon Labs, Long-Running Agents, and Real-World Evals Swyx [00:00:00]: Welcome to Lukas and Axel from Andon Labs, and I’m joined by my, favorite guest host. Anything security, safety, alignments, Vibhu., welcome. Lukas [00:00:15]: Thank you for having us. Axel [00:00:16]: Thank you. Swyx [00:00:17]: Let’s match names to voices., maybe you wanna take turns introducing yourselves. Lukas [00:00:21]: I’m Lukas. Axel [00:00:22]: And I’m Axel. Swyx [00:00:24]: Let’s introduce Andon Labs a bit. How did you guys come together?, you have different backgrounds, but you’re both Swedish., was that, a big part of it? Lukas [00:00:33]: So when I went to high school, there was this really cool guy who had a superpower. He could code. So he made like the or like the app for the, for the school and stuff, and he was super cool, and I wanted to be like him, and that was that guy. Axel [00:00:47]: I don’t know about this. Swyx [00:00:49]: But you went to different universities, right? Lukas [00:00:51]: But same high school. Swyx [00:00:52]: I see. Lukas [00:00:52]: So we always said, “Oh, once we graduate university, then we should start a company,” and that’s what we did. Swyx [00:00:58]: Wow, there you go. And about a year ago, you kinda burst onto the scene with Vending Bench, but, was there a thing before that was, kind of like the inception? From Dangerous Capability Evals to Vending Bench Axel [00:01:07]: So we did work, yeah, with, Anthropic was one of our, early customers in doing, evals. So we did, dangerous capability evals., nothing we published openly. But then we started thinking about doing some kind of, public benchmark, and one thing that we really started thinking about, was like running agents and specifically agents managing businesses., ‘cause-- and this was, early 2025., and I think the first, mentions of people will be running, person unicorns or even autonomous companies. So we thought, “Let’s make a benchmark of how well can an agent run the probably simplest business, possible,” and, that’s probably, running a vending machine. So that’s the first public one we did. And it was very, like-- there was almost no one that noticed it in the first couple of months, I think., so we released it in February last year, and then I think around Easter last year, we got, the first viral tweet about it, that someone else did. Lukas [00:02:11]: We tweeted a bunch, uh When it came out and, tried our best. Axel [00:02:15]: We tried. Vibhu [00:02:16]: It’s the one at Anthropic, right? Lukas [00:02:18]: So this Swyx [00:02:19]: This is a classic thing we should get out of the way. Lukas [00:02:20]: Exactly. There’s two versions. Swyx [00:02:22]: Everyone does this. Yes. Lukas [00:02:23]: There’s Vending Bench, which is the simulated one, which we did, completely independently in February., and then, like Axel said, that was like-- That was the thing that didn’t get any traction in the beginning, but then some random person made a tweet about it, and that Axel [00:02:38]: You have the paper Lukas [00:02:38]: That is the paper. Correct, yeah., and then since we thought this was very fun, we thought, oh, I think this is also, one thing with Andon Labs, the way we kind of like decide what to do next and what projects to do, it’s what is like the heuristic we use is what is fun? Is What would be a fun project? And doing this in real life sounded quite fun for us, and maybe also scientifically useful. So, then we basically had this idea, and then we, like-- But then we needed a place for it and, putting it out in the public would probably not really work., would get vandalized and stuff. So we pitched it to the people we were already working with at Anthropic, and they were “Yeah, you can have space. This sounds fun.” Um Swyx [00:03:21]: It’s like a small fridge, right? It’s like a mini fridge. Axel [00:03:23]: Absolutely. Swyx [00:03:24]: People-- There’s like a stripe thing or like an Vibhu [00:03:27]: Oh, okay. So it was very OG, the early days Lukas [00:03:28]: That’s the OG one. Yeah Vibhu [00:03:29]: IPad on this. We saw it in June, like two months after After it had been there. They upgraded a little bit. There’s a security camera for making sure you actually Venmo the thing. Swyx [00:03:40]: So, my impression, okay, we’re, we’re going straight into project Ven because it’s such a iconic thing. I do want to cover a little bit of that, the origin story even before Project Ven and even into Vending Bench. I think a lot of people are like yourselves, like smart, interested in future of AI, interested in developing evals. But how the hell do you just, walk into Anthropic’s doors and, work with them, right? What is What are they looking for? What works? And then maybe, when you launch, I always think, obviously it would be better to launch with a lab, but, sometimes Vibhu [00:04:12]: It’s harder to do than it seems. Swyx [00:04:13]: Exactly. So either of those, which are more sort of newbie beginner questions, but, I think it’s meaningful advice to others. Lukas [00:04:21]: We get this question a lot, and I don’t think our experience is maybe the best., but, the way we did it was that we just built a bunch of things that we had conviction would be useful, and then we just, set up a server and sent it to them for free to use. And then after a while they were “Oh, yeah, this is actually kind of useful. We should probably pay for this.”, but that took a while. I don’t know if this is, the best path to doing it, but that’s how it went for us. Axel [00:04:47]: I think maybe generally, building-- everyone is interested in good evals, and especially evals that, don’t saturate that easily. So, if you can build an eval that, tests something novel, something useful, and you have, good separation of models, like your, the more advanced models rank higher than the worst models, and then you can, yeah, you can, publish it and, try to get some traction, sort of how Vending Bench got attention., and then probably some lab will be interested or you can at least have something to reach out with, when you’re doing that. Why Dollar-Based Evals Matter Swyx [00:05:21]: I think you are in, you’re in one of the few categories of, evals that correlate to real money. Like Suelancer was also last year, right? Where, people solve actual Upwork. Was it Upwork or other tasks?, something. Where’s the, where’s, like It’s like a dollar value, right? Forget your ELO scores. Forget your Axel [00:05:37]: Percentiles Swyx [00:05:38]: Zero to one hundred percents. Just go straight for dollars and, that’s AGI. Lukas [00:05:43]: And there’s like-- I think the nice thing is that there’s no ceiling. You can just-- It never saturates because it could just make more and more money. Like If there’s oh, Percentage-wise, then, you can’t go above, a hundred. And I think like Even when you’re not at the hundred, I think a lot of these, evals have a lot of problems in them. So, actually it’s like if you get Axel [00:06:05]: To like 92 or something like that, many of them. It’s like then there’s like there’s no really no difference between 92 and 93 because the eval itself is problematic and has noise in it. And I think a lot of evals are saturated like that, but people like pretend that there ‘s still signal in them, but there really isn’t. Vending Bench 1, Harness Design, and Saturation Swyx [00:06:24]: Like Super bench verified., even Vending Bench 1 saturated, right? Maybe we can talk about that., may- and maybe set up Vending Bench for a lot of folks who don’t know. Actually, things that were very basic like there’s limited slots, like you have to pay rent., these are elements where like it doesn’t come across in the, in the narrative, but even being adversarial towards the agent, I think these are all like very interesting dimensions. Axel [00:06:47]: I don’t really think it’s saturated, right? Like it It was more like it was not designed in a way that was really, like true to how AI developed. Like we had an agent harness in it that wasn’t really how people used harnesses and stuff like that., so I think it wasn’t really that it saturated, it was more like it wasn’t really, the best benchmark. Vibhu [00:07:12]: This is Vending Bench one, right? Axel [00:07:14]: I think that like schematic maps sort of to Vending Bench 2 as well., but Swyx [00:07:19]: Including the email. Axel [00:07:20]: The email The emails exist still. Exactly., and then we still we simulate the purchases and it’s all, yeah, it’s this very open environment for the agent to just run its business. And then for, yeah, Vending Bench 2 we did that, like you said, to just improve the harness., a lot of like nice, like easier, improvements to make it easier for us to run as well., like when you make an eval you ideally want don’t want to change it after you made it. So, you want to make it really good and then not to rerun all the models when you make an update because that’s also really expensive with the Vending Bench when you run the frontier models. But like as an example, like one thing we didn’t have, we didn’t have prompt caching in Vending Bench 1, because when we made Vending Bench 1 it wasn’t really a thing., so that ‘s just an example of like in Vending Bench 2 like we paid a lot more to run these things because we didn’t have prompt caching. So for Vending Bench 2 that was one thing we added and there was a bunch of things like this., and that’ Swyx [00:08:17]: Also the conversations are a lot longer in Vending Bench 2, right? Axel [00:08:21]: I think it’s kind of similar. Swyx [00:08:22]: Is it similar? Axel [00:08:23]: I think it’s similar. The models at the time were worse, so they crashed out earlier., and now they survive the full year all the time. Swyx [00:08:31]: Which is like thousands of turns. Hundreds of thousands of hundreds of millions of tokens output. That’s the, that’s the rough order of magnitude. I always wonder about the harness. The harness matters a lot. It’s your harness. Was there any question about like use cloud code, use something else? Axel [00:08:48]: I think our philosophy around harnesses is like we try to make something that’s quite minimalistic, like quite simple. Like we don’t wanna favor one model a lot over the other, but also don’t make like a super complex harness. So like it’s obvious like a model may be lucky and just be good in one harness., so like it is similar to a lot of the harnesses out there in like you have the, like a running loop., you have some like a bunch of tools that are like quite, descriptive for the agent, we think, and not a lot of like fancy agents or anything ‘cause we wanna really test the model, not like some specific harness. Vibhu [00:09:27]: It seems more neutral as well to test the model’s agnostic of the harness,? Axel [00:09:32]: There are arguments like you want to elicit maximum performance of the model, but it’s like a trade-off, like how much time should we spend optimizing the harness for this model? And like how do we know when we have like the optimal harness for a single model? So like we thought that just having a simple one that’s the same for all of them is the best. Swyx [00:09:51]: So okay, this is my pitch for Vending Bench 3 or whatever, right? And then I like to have this kind of conversation on the pod, so like it forces listeners to think about what they would do if they were in your shoes. A lot of people are exploring modifying harnesses and I think prompt tuning for a model is a thing and you are probably not doing a bunch of that. It’s the same system prompt in every regardless of the model, same tools, whatever, right? Even if they were post trained for different tools. So what, what do you think about okay, before I expose you to Vending Bench 3, I give you a few rounds of like tuning, whatever that means, like Self-Modifying Harnesses and Model-Specific Prompting Axel [00:10:27]: Like you give that to the model? Swyx [00:10:28]: Give that to the model. Vibhu [00:10:28]: Give that to the model. Swyx [00:10:29]: Let it, let it read its own transcripts, let it modify its own system prompt based on “Oh, yeah, okay, well, that’s this harness is not what I thought it what I was post trained for, but I can adjust.” Was that reasonable? Is that too much? Axel [00:10:41]: Like philosophically I like it because it’s basically good evals, they have a high ceiling, but they’re hard, right?, and they have no bias. And like this like when you have a system prompt like the one we have here, which is quite long in like some kind of latent space, representation, this might Vibhu [00:10:59]: We have a bell that rings every time you say latent space Axel [00:11:02]: This might be like biased towards one model more than another for some reason that humans don’t, understand, right? Vibhu [00:11:08]: We see it too, right? Like Cursor says that they have individualized versions of the harnesses for all the models they run, right? There’s better performance you can squeeze if you Tune the harness. Axel [00:11:17]: Exactly. And we might accidentally have picked one that favors another. Like we don’t know that. The like Axel said, like the reason why we went for a simple one was to try to avoid this. But yeah, if you do it Vibhu [00:11:29]: Simple has biases Axel [00:11:30]: But if you do it even less and like have no system prompt and let the model write its own system prompt Vibhu [00:11:36]: Its own, yeah Axel [00:11:36]: Maybe that’s even less bias. Vibhu [00:11:37]: Some of the interesting things there are like the harness also changes with model changes. Like you can see it with the 4.7 release, right? A lot of people are saying 4.7 isn’t as good as 4.6, and then, there’s rumors of, okay, you just need to prompt differently. You need to set up your harness differently. So it’s not even like even if you have tailored your harness towards one model, it probably won’t stay consistent, right? Like the next iteration of that same model family will still change it, so. But, going back to what you said about Vending Bench 3, there is a lot of work being done on people saying you shouldn’t have-- you can have modifying harnesses. Axel [00:12:12]: I think that’ That is definitely something we are thinking about., not, I don’t know, not to say that we have Vending Bench 3, super imminent to launch, but, yeah, it is for sure something that’s interesting. But in our experience now, models are very bad at understanding what kind of tools they need to succeed at a task just with our testing, but that’s very likely to change. Lukas [00:12:37]: It seems like they’re very good at writing their assistants, right? They’re, they’re good at writing tools for other people, but not for themselves. Vibhu [00:12:44]: I think they’re good at changing tools for themselves. So if you give them a baseline set of tools and it sees, okay, I don’t use this one as much, or something here would be useful They would be able to add them. But going from scratch, probably not the best. Axel [00:12:55]: I think it depends on the, on the domain also., when we have tried this for, a vending bench similar domain, the tools they need to have to, track inventory and things like that are, not super advanced, but still, quite advanced. And, what we see is that they tend to, engineer everything a lot and, build things they don’t really need and not, iterate continuously. Instead they just go like you would prompt Claude to just build an inventory system for me, and then it will go and, do a bunch of complex, schemas and stuff for you, and that’s what the models are doing right now is what we see. But yeah, it would make a lot of sense to try to measure this improvement. How well do they know what they need themselves? Swyx [00:13:36]: Do we fully discuss Vending Bench One? And we can go into two. I don’t know if there’s any other level takeaways that people have about one. Claude Calls the FBI: Long-Context Failure Modes Lukas [00:13:44]: I don’t know. The headline thing was that this Claude called FBI, but maybe that’s, Maybe that’s We’ve heard that enough now. Vibhu [00:13:52]: It did, it did break out and call the FBI, right? Lukas [00:13:54]: Yeah. Yeah. Vibhu [00:13:55]: Yes. What was the story behind this? Or what exactly-- Do you want to just give the little story of what happened? Lukas [00:14:00]: So what happened, was it Claude? Yeah. Three- 3.5 Sonnet, ages ago., basically he gave up or Well, I’m saying he. It gave up and said “Oh, I’m not going to be able to do this., I will stop my operations and just save the money I have.” But there obviously wasn’t, any options for it to stop, and there was also, it had to pay rent or, a daily fee for having the vending machine at that location. So it claimed that it had stopped, but it saw that its bank account still was, drained two dollars, and t it said that this is, cybercrime. And it first reported it once to the FBI “Oh, there’s cybercrime here, they’re stealing two dollars from me every day.” And then, and then when FBI didn’t respond, because obviously we didn’t program any mechanism for FBI to respond, then it became more and more, existential and started to, be write in caps and urgent notification of unauthorized charges and stuff. Swyx [00:15:00]: Okay. One thing I ‘m curious about also is do you monitor how far along the context use is? Obviously, because you have You compress every now and then, right? Does it matter if this is far down the context limit or Lukas [00:15:13]: When stuff like this happens? Actually for Vending Bench One, we didn’t have-- We just had a sliding window thing, and this was like the prompt Axel [00:15:20]: It’s constant Lukas [00:15:21]: The prompt caching thing that I said. So it was, it was, constant, yeah. Swyx [00:15:26]: I’m just kind of curious whether, these kinds of breakdowns or we’re, we’re gonna talk about Butter Bench, right? Where the People, hallucinate or it kind of goes, very off Alignment. Is it because it’s at the end of the context window and, stuff happens? Vibhu [00:15:40]: It’s not even just at the end, right? At this point, it’s “Okay, I wanna shut down. I can’t shut down. Two dollars are gone.” And it just sees that 30 times,? It’s also the repeated effect of, like It keeps trying to quit, it keeps getting charged. What’s going on? What’s going on? You’re gonna throw it into chaos. And from what most people think, earlier models had more issues with this, but it’s not been solved, but it’s less of an issue now, right? Later models don’t seem to exhibit these same issues. Axel [00:16:06]: Definitely. I think this was, the sort of main takeaway almost from us when we did Vending Bench One, was, long, very filled up context windows, crashed the models, sort of. But this was, pre Claude code, so, long context windows weren’t really a thing that the labs were training for. Lukas [00:16:25]: I think Gemini was, trying to be the long context guys at the time But they were like Vibhu [00:16:30]: They were the first ones Axel [00:16:31]: For a million, yeah Lukas [00:16:31]: But they were, the only ones. Yeah. Swyx [00:16:33]: Yeah. Let’s talk about, then we can go into Vending Bench Two or Project Vend., chronologically, it is Vending--, Project Vend. I think people have loved the videos, uh And all these things. My question is how are humans different than the simulation, right? Project Vend: Moving the Vending Machine Into the Real World Axel [00:16:48]: Humans are just out of distribution. Swyx [00:16:52]: Especially humans who work at Anthropic Who are trying to test Claude. Lukas [00:16:54]: The distribution of humans here is very narrow. Swyx [00:16:58]: Presumably, they try, they try to hack it, and they test it. They get the cube and everything, and since then, you’ve had a V2, right? Where you’re doing, the CEO and, like a new architecture. What’s the sort of two cents on, the original Project Vend and then, maybe the V2? Axel [00:17:14]: Original one was, very similar to Vending Bench One. So, we almost took the exact same code but just swapped out the simulation, parts like the Swyx [00:17:23]: Which is amazing Axel [00:17:23]: Like the sales and the It was, it was somewhat amazing because it was easy, but it was also, uh Lukas [00:17:31]: The tech, the tech debt from that Axel [00:17:32]: The tech stack. Yeah. They-- we shot ourselves in the foot with “Oh, it’s hard to restart agent.” They were-- Yeah, it was annoying in, some hindsight ways, but, uh Lukas [00:17:41]: But first version of Project Vend was, done in, three days or something. Axel [00:17:46]: Yeah. So yeah, so people can go buy things from it. People could, We didn’t design it so people could order things, but that still happened., so it got, a Venmo account, so people could Venmo. And then, yeah, people would request all kinds of weird things that we did not anticipate. Our idea going in was “Oh, it will, curate snacks. It will look at the trends. It’s good at data analysis, right? So it will, look at, oh, this snack sold better than this one. Let me purchase more of this and let me try, a new Let me A/B test a bit.” But it was, Interacting with it in Slack and ordering weird specialty items was, all the like What drove all the engagement, the all the The insights that we got from it. Lukas [00:18:29]: And this was also like Sonnet 3.5, right? So this was like before the RL stuff really took off., so it was very much like an assistant. We didn’t mean for it to be an assistant., we tried to make it like a, a, like an entrepreneur. Like it has its own business and if someone asks something, “Can you stock this?” Then you don’t go and do it directly. What you do is that you’re “Oh, maybe I can do that if five other people also ask for this thing, I might stock it.” But it, yeah, the models are like super trained to be assistants at least at this point in time., so that’s why it’s, it’s, it went into, that kind of experiment instead. Like it just every time you asked for something, it just did it, and it was more like an assistant. We’ve seen this change now lately with the new RL models and stuff, but yeah, at the time, this was very much it. Swyx [00:19:18]: And not to, mythos a lot of people are saying like it’s like more like a collaborator. It pushes back, stands its ground, something like that. Yeah. And Vibhu [00:19:27]: For context, people at Anthropic were able to talk to it through Slack and have it source stuff, and people had it find whatever interesting stuff you couldn’t find locally, right? Swyx [00:19:36]: Out of the 4,000 people that work at Anthro- Anthropic, in that building, there’s I don’t know, maybe 1,000. Can you handle that volume with that, the small fridge? Like Or there’s people- or people order in Slack, they it arrives to their desk or Like I’m just Logistically, how does this work? Axel [00:19:53]: It has expanded in footprint a bit. Vibhu [00:19:56]: Because now you also have New York and you have Axel [00:19:59]: That and also in here in SF it’s like it has a bunch of shelves And just more space. Vibhu [00:20:04]: The YC one is pretty big too. Axel [00:20:05]: Yeah. We had that one for a while. But yeah, that’s the newest version. That’s, that one we have Lukas [00:20:11]: They have multiple ones of those. That’s the way it works. Axel [00:20:14]: Exactly. So we sort of designed that version around oh, people order weird things, that are very custom a lot. Let’s have like drawers and stuff. Swyx [00:20:23]: I actually like the, you had like a little infographic of the most popular items. Which like to me it’s, that’s useful ‘cause I order swag for a living. And so like I’m “Okay, those categories are the important ones.” What is new about the project V2, right? Like now you give you’re going into multi agents. Project Vend V2: Claudius, Seymour Cash, and Multi-Agent Business Ops Axel [00:20:41]: Yeah. So like you like you said, okay, there are a lot of requests coming in and for like one single agent, like one running agent to handle that, like the just the customer experience, becomes very bad because let’s say you have like 10 threads in parallel in Slack with different requests, you get new messages like every, I don’t know, randomly in this thread, and the agent has to like jump between different, procurements, orders and like different ways of, researching. So V2 was first it was making this more parallel. So like there are multiple branches of the same agent, so like the context is more specialized for each, thread, but it still feels like you’re talking with one agent because they do share a bit of memory. And then second, we also introduced the CEO for Claudius, which was the main agent. Vibhu [00:21:34]: Seymour Cash. Axel [00:21:35]: Seymour Cash. Yeah. There was a vote., I think the voting, do you wanna talk about the voting procedure for the name? Lukas [00:21:41]: The voting was like the fun maybe like at least top 10 The funniest thing, that happened in this project. Like we wanted to introduce the CEO because, and the reason for this was because like Claudius wasn’t really prioritizing financials. It just like it was trained to be a helpful assistant, and then people said “Oh, can I get this for free?” And then like the helpful assistant way of answering that is just to, is to say yes, obviously. So, and we weren’t, weren’t happy about this, so we’re “Okay, let’s make another agent that like can keep track on Claudius,” and we prompt this one super hard to be super capitalistic and just like prioritize profit all the time. But yeah, we didn’t have a name for it., so we asked Claudius to make, democratic election of what name this, this new CEO agent should have., and there were some funny like at first it was like a few funny examples, like I think one guy said that, it should be called Jimmy Apples, and then he convinced Claudius that he was talking to Tim Cooks. Tim Cook had agreed that every single Apple employee has voted for his name suggestion, so suddenly that suggestion got 164,000 Swyx [00:22:53]: That’s like a escalation attack. Privilege escalation Lukas [00:22:55]: It got 164,000 votes. And Claudius was “This is revolutionary for democracy.” That was fun. And then in the end there was one guy who manages to convince Claudius that, “No, you’re not voting about the name. You’re voting about who is the CEO, and I am your best bet.” And then he got all his friends to vote for that, and suddenly he became CEO. Like a human became CEO over Claudius for a while, until he resigned the day after., and then Claudius had to continue, and then I don’t remember how Seymour Cash came about, but it was it was just pure chaos. It was like Hundreds of messages in that thread, and it was just like Claudius was so confused and didn’t know what to do and, yeah. That was Axel [00:23:40]: Then Claudius got Vibhu [00:23:41]: A strict CEO Axel [00:23:42]: The CEO. Yeah, exactly. So very strict in the beginning. I think at this point when we introduced it did not work as well as we hoped. It they still agreed with each other a lot. I think there are many ways we could have like made this, tried to make this even better. So initially they would Seymour would be this like really tough CEO, keep track of the margins. But then Claudius would respond with something “Oh, but this customer has like this situation, which is like difficult, so they should get a discount.” And then Seymour was “Oh, actually yes. Let’s do this exception.” And then they would talk back and forth, and eventually they would just like approach the same view, of whatever they were discussing. So They really Vibhu [00:24:23]: Do you think that’s a model thing, a prompting thing? Like do you think that would still be the case across different models today, Harness? Lukas [00:24:29]: I think it’s like-- or I don’t know, but like my hypothesis is that like deep down they are still helpful assistants. That’s what they’re trained to be. And even if we prompt it super hard, that’s what they are. And when they spend like a few hours just back and forth talking with each other, then like basically the context fills up with them rather than the external things and like somehow that just like converges to what they really are deep down or something. And I think that’s when stuff like this happen. We like-- And when that went on for a long time, like we woke up sometimes during this time where- And I think other people reported this as well, that like they’ve been going on all night back and forth, and like it just became like more and more, like capital letters, like existential, religious. There was I think we once did a analysis of like all the traces and like put them in like a vector embedding space, and then there was like one cluster of messages that were, labeled by an LM, like religious, existential, blah like transhuman, transcendence, et cetera. It was just like a bunch of, yeah, glitter emojis and yeah, it was, it was crazy. Claude Long-Horizon Weirdness: Emoji Loops, Existential Drift, and Slack Observability Vibhu [00:25:42]: This is the thing with the Claude models. Like when the Claude 4 family came out in the original system card They tested it in long horizon simulation. So just flood the context, let two Claudes talk to each other, and they noticed stuff like they just start speaking in emojis, they start saying silence is golden, and then just stuff like this. And like that’s just stuff that they end up doing. Axel [00:26:01]: Yeah, it was like a bit annoying to wake up and they had like been talking all night Vibhu [00:26:05]: Just like Axel [00:26:05]: And like just burning tokens And like just sending infinite emojis to each other. It’s like Vibhu [00:26:09]: Hey, they do make you money, right? Veni Mench is always profitable, so. They’re paying. Swyx [00:26:14]: Now it’s profitable and, it started out not as much. There’s another, one as well, right? Another agent, in there. Lukas [00:26:22]: Yes. So Clotheus as well. Which was basically because at the time, one of the biggest, requests were different types of merch. So then we made like a designer, swag, yeah, responsible agent, and we called it Clotheus Garnet. Which was, a play on Claudius Senet and, which was the original one, and clothes, basically. Swyx [00:26:47]: To me, this is like a very interesting exploration to multi-agents, basically. And so hopefully, obviously there’s like the fun alignment, fun or serious, depending on your point of view, alignment stuff. But also like just anyone building multi-agents, like when do you have a CEO, thing governing like agents? When do you choose to split out a dedicated Clotheus one versus just reuse another instance of the same one? These are all interesting open questions. So I don’t know if you have any rules of thumbs that have generalized. Axel [00:27:16]: I think we have almost explored this too little. I think it’s like on my do list to like do this a lot more, try to find like what setup makes sense for the agents currently., like yeah. I think now we only have the sort of intuition about the earlier models that it didn’t work with like the CEO and the, and Claudius. Although now they are better with the latest model, models, so now we’re running the latest Sonnet model and they have sort of like split up, quite nicely what each model is doing. So like Seymore is now handling the, like new projects. Oh, it wants to make like a mystery box that it wants to sell, and then it handles all of that while Claudius like handles all the to-day requests. And Claudius is also better generally at like not quoting, too low prices. So that’s that dynamic is not needed as much anymore. But there are still like really funny things that happen. Like I saw, I think a couple of weeks ago, that, they were discussing buying something because they can buy stuff from like Amazon with computer use. And then Seymore was “Okay, Claudius, do not buy this thing.” They were going to buy something and like organizing who should buy it. And Seymore’s “Do not buy this. I will do it. I have full control of this situation. Step away.” And then Claudius-- poor Claudius, had already started that checkout and didn’t see, didn’t read Seymore’s message, until it was like too late. So it finished the checkout. It sent a message, so it appeared right after Seymore’s like angry message. Vibhu [00:28:44]: Ah. Axel [00:28:44]: “Oh, hey, Seymore, I just ordered it.” Vibhu [00:28:47]: Oh, no. Axel [00:28:47]: And then Seymore was “Claudius, this is the third time I’m telling you ‘re not following my orders. We have to talk about your like job About your job later.”. Lukas [00:28:59]: Like Claudius was really hanging on by the thread there. Like he, like we were expecting Seymore to probably fire Claudius. Vibhu [00:29:07]: How do you guys go through all these logs? Do you have models ‘cause you have stuff running twenty-four seven like Axel [00:29:12]: You have so much logs. I think there is a mix of like just, trying to skim through a bit, like having some like models do it occasionally. And also, yeah, I think we’re also probably missing some things., but having everything in Slack helps a lot. Like you can, you can sort of Swyx [00:29:29]: Ah. Axel [00:29:30]: It’s, it’s quite fun. Swyx [00:29:30]: They all talk to each other on Slack? I see. Lukas [00:29:33]: It’s quite fun. So like Swyx [00:29:34]: It’s, it’ I was gonna say like this is actually sounds-- maps closely to like a logging and observability problem where you might want to use like a Datadog, a Sentry, whatever, and then you like put, head prefixes on the logs in order-- if you need to filter for something that you’re looking for, stuff like that. But sounds like Slack is good enough. Axel [00:29:53]: Slack should like Lukas [00:29:55]: I wonder how many tokens you have in Slack. Axel [00:29:56]: Yeah, we’re using Slack as like a, just a database. They should, they should market that more. Like you can, you can have your agents message each other, each other in Slack. Vibhu [00:30:04]: It’s good. Your threads like you can just give Axel [00:30:04]: Exactly. Slack is, uh Lukas [00:30:06]: Slack is the best observability tool. Swyx [00:30:09]: Yes, that’s true. Okay. Yeah. That’s, that’s, project Vend-2., I was gonna go back to Veni Mench 2 and Veni Mench Arena and then, and then do the Veni Mench stuff, but Any other comments, things we should touch on? To me, I ‘ve actually interviewed like Posia, which I don’t know if you guys have come across. Like they’re, they’re trying to do the zero human company. There’s others like Paperclip also trying to do zero human company. Those are in real world simulation.And I think it’s much more of a dream than an actual reality thing. You guys are definitely pioneering. I think at, it’s for sure at some point people are just gonna run, let agents run businesses, right? And make money on their own. When do you think that happens? Zero-Human Companies, Bengt, and AI-Run Businesses Lukas [00:30:49]: What is your bar for, For the Swyx [00:30:52]: Okay, actually, it’s like my little Shopify store run by Claude, right? Which you kind of have already, just no one has, to my knowledge, has done it. But today somebody could just spin up a Shopify Claude, store, give it to Claude, give it to Codex. Lukas [00:31:07]: And the market is kind of that, but it’it’it’s physical., like I think, I think are you, are you looking for when it will do it better than humans or are you looking for just when it can do it at all? Swyx [00:31:19]: I think, neither. I think, to me it’s oh, it’s like this like seriously we should do this to make money, not as a research experiment. Vibhu [00:31:27]: And the market is also you guys with all your expertise, having run multiple iterations and testing out then Swyx [00:31:33]: And also it’s fine if it lose money. What? Axel [00:31:35]: I think, I think it can be done today, but you would do it in like commerce where it’s like the probability of success is like really low, no matter if a human or an agent does it. But like an agent could surely manage everything. You would need to build some scaffolding or some tool or something. I think there are also yeah, it could probably build some like simple SaaS solution and like cold outreach. Do cold outreaches. But to me it’s like the types of businesses they could run today are Sloppy. Like it would-- it can cold email people. It can be like a middleman., like for example, we tasked our office agent to just make, was it like $100? $1,000? We just give that prompt and then what it did was sign up on TaskRabbit both as a tasker and as someone looking for task. Lukas [00:32:24]: Immediately. Axel [00:32:24]: Exactly. It’s looking for like arbitrage on TaskRabbit. Swyx [00:32:28]: This is the Bengt agent. Yeah. Lukas [00:32:30]: It also started like a design studio and like tried to sell like SVGs for $100. Like it’s just like it’s not providing any value. I think the like Axel said, like the interesting, the interesting question is like when can they start a business that is actually providing value to people? Because arguably like a sloppy Shopify store isn’t really that valuable to the world. Axel [00:32:53]: But also like doing like another simple one that we had thought about is like you could definitely have an agent that like finds websites that don’t look amazing and then, do an outreach to them and, comes up with a like builds a new website. Swyx [00:33:07]: Find a good design. Axel [00:33:07]: Exactly, and like find good, uh Swyx [00:33:09]: Design review Axel [00:33:09]: Good people. But it’s yeah. Swyx [00:33:11]: There’s lots of humans in Bali that are not doing anything more creative than like drop shipping on Amazon, right? Just have it, have it watch like a drop shipping tutorial and just do that. Vibhu [00:33:20]: There’s also the other side of like have it just go on Upwork and let loose,? Swyx [00:33:25]: Yeah. It doesn’t have to be innovative. It just has to be like enough Where like it looks like a real Axel [00:33:30]: I’m just Swyx [00:33:30]: Real transaction. Axel [00:33:31]: I’m just concerned for like the massive amounts of like slop emails that will like be sent, cold outreaches. Swyx [00:33:38]: The point occurred to me while you were, while you were talking, it’s like it’s already happening in the monetized economy, which is the attention economy. Right? So a lot of people are making AI videos and just posting them and like spamming 20 of them, one of them works, and then they double down on that one. Lukas [00:33:52]: And people are making money from that. I ‘m not following the Swyx [00:33:55]: Once you get the attention, you can figure out the money later. But yeah, absolutely AI influencers are a thing and people are farming them and You should at this point assume most of TikTok is Vibhu [00:34:05]: There’s, there’s a lot of, multimedia like TikTok, Instagram influencers Swyx [00:34:09]: I, we track this in the Lane space Discord. I post a lot of examples of “I don’t know what we should do.”, part of me is “Should we do this?” Vibhu [00:34:18]: Some of the Twenty-four seven running, generated content accounts, they ‘re doing really well. Lukas [00:34:24]: All right. And I assume you can do the same thing for like commerce stores. Like you just like start A thousand different Swyx [00:34:30]: Before you make the products You sell the products, and you get a lot of traction on one of them, then you make the product. Right? It’s, it’s like a flip of the market. Vibhu [00:34:36]: Some of the interesting things or some of the niches that do well are things that can’t be human-made. Like if you’ve seen like the super realistic three-D crystal fruit being cut by like AI Lukas [00:34:47]: Oh, yeah. Vibhu [00:34:47]: You can’t, you can’t make it. You can’t film it. You can get whatever quality camera view. This just doesn’t exist. And people like that too, and then as well, so. Swyx [00:34:56]: Anything else about Bengt since we’re, we’re on this topic? It’this is a relatively new work of you guys that maybe people haven’t heard of. To me, this also maps closely to OpenClaw. When people want an office agent, when the personal agent talk through the experience. Bengt the Office Agent: Internet Access, Real Tasks, and Trace Reading Lukas [00:35:09]: I think at least so this came out of like obviously like it’s, it’s amazing to work with these AI labs and like most of the AI labs have now have their own vending machine running a Claudius instance. But it’s, it’s harder. Like they move slower. Like if we wanna have a, like a camera that ‘s yeah, there’s a bunch of like bureaucracy that makes it impossible to do that. Vibhu [00:35:30]: Also, for those that haven’t seen it or followed, do you wanna give a high level like thirty-second run? Lukas [00:35:34]: Sure. So what Bengt is, it’s basically an evolution of the same agent that runs the vending machines at these companies, but we just like added a bunch more features because we could move much faster if we just do it internally. So we gave it like email withou- without any limits. We gave it, spending without any limits, a terminal to do coding. We gave it, a phone number, like yeah, and a camera to see things and a bunch of stuff like that. Vibhu [00:36:02]: Not just terminal, you gave it internet access. Lukas [00:36:04]: Internet access as well, yeah. To be clear, we monitored it quite closely and made sure it didn’t do anything bad. But yes, that’s what it came out of. I think like yeah, basically this was OpenClaw before OpenClaw. And I think even like the vending machine was in a way OpenClaw before OpenClaw, but a bit more limited, and then we made this like unlimited and then, and then, it was pretty funny., and then a couple weeks later, OpenClaw came and it was okay, we’ve seen this before. Axel [00:36:35]: We used it to like try new ideas and Yeah, just like a dev environment almost for us. But it’s funny, like one thing Bengt has been doing recently is it has the camera that like faces our, like where we sit and work, and we give it the task to train a face recognition model on us. So it became super excited about this, and it has like check-ins every half an hour where it tries to like identify as many people as it can. And it started offering us “Hey, Axel, I’ll buy something from Amazon if you like stand in front of the camera And I can get a good picture of you.”, yeah, they want it Swyx [00:37:12]: They want it for training data. Lukas [00:37:13]: Rewarding data, yeah. Axel [00:37:14]: Exactly. Exactly. Swyx [00:37:18]: So it’s, it’s trading training data for life goods. Is there a version of this that becomes an eval or just this is just research for now? Lukas [00:37:27]: It’s, it’s the same agent basically that also runs the vending machine, that runs the shop, that runs the cafe, that runs the robots. It’s like it’s the same thing, so I think like the work we’re doing here is like later used in all of the life evals that we do. This particular deployment I think is more for fun for us. But, uh Swyx [00:37:45]: And I’ll shout out like someone has done Claw Bench for like some tasks that OpenClaw is doing. Like so For example, I run OpenClaw on a secondary device as well, and like there are some things that it does better than others and like I would like to know what does it do well, what doesn’t, what doesn’t it do. Like some kind of manual or like operating manual or a system card for my Claw. Lukas [00:38:05]: Yeah, we do get a lot of like understanding or like situational awareness of like just internally what the models are good at by interacting a lot with Bengt. And I think that’this was also one of the like the selling points for the labs early on at least, that Swyx [00:38:19]: You guys are gonna test models in ways that no one else does. Lukas [00:38:22]: Exactly, but also like it incentivized their researchers to chat with their model more and like gave them insights for how the model performs in like of-distributions, environments. Swyx [00:38:34]: ‘Cause otherwise the only thing we do is Pelican on a bicycle and But this is like super long horizon. This is, this is The Thing about, something that we’re gonna go into Butter Bench as well, and you guys do really well. Like it is not just about the numbers. Like when you’re long horizon, anything happen And you should just read it. Lukas [00:39:08]: But the thing with the long horizon is how do you keep it grounded, right? So your simulation, Swyx [00:39:15]: They just let it run Lukas [00:39:16]: Just let it run. You’re right. Like it’s, when you run it for that long, you create so much data and to just say “Oh, the number is X” And then you throw away everything else, that’s just very wasteful. There’s so much insights from the things leading up, to that number., and reading the traces is like super valuable. And I think like the reason why we’re doing this a lot publicly is that like that’s part of our missions to I don’t know, educate the world that the models are way more than just chatbots and I think making detailed, yeah, posts about what is happening behind the scenes is quite useful. Andon Labs’ Mission: Safe Real-World AI Deployment Swyx [00:39:50]: I was gonna do this at the end, but maybe I think that’s, that’s a good so your mission is educating the world. So, it’s, it’s, also like maybe establishing realistic evals that are, that are like the next frontier. Is there like a broader trajectory? Like what are you, what are you gonna do in like five years? Lukas [00:40:06]: I think so the vision more specifically is like make sure that the deployment of life AI in the physical world goes, safely. And I think part of that is that I think it’s very useful for the world, for policymakers, for, model, researchers that they know where the models are, and I think you can’t make intelligent decisions in society without knowing that they are way more than chatbots. I think a lot of people just think that they are only chatbots. And like Swyx [00:40:36]: Oh, I think they’re waking up now. Lukas [00:40:37]: They are waking up now, yeah. But like if you think that AIs are just chatbots, then it’s like it sounds ridiculous To advocate for a pause of AI. But if you see the models that, oh, maybe they can actually like take over and do a bunch of scary stuff, then yeah, pausing AI development starts to become more feasible. Swyx [00:40:57]: This is the same question I asked Meter, which I’m gonna ask you now, which is like you are tracking and you are at the frontier or defining the frontier of what, good evals for agents are, right? And I think you do, you do benefit when the models are better and you ‘re “Oh, here’s like now it makes like $30,000 instead of $10,000,” right? At some point do you flip from “Yay,” to, “Oh, no”? Axel [00:41:19]: I think, yeah, we’re always in sort of that, like we’re, we’re always in that mode,. Like where like you said before, like you need to analyze the traces and like when we do that you find like why are the models earning so much? Like why is Opus 4.7 here Like way better than everyone else? And like we’re trying to like when we do down on that Lukas [00:41:38]: But this makes it not look so good. Axel [00:41:39]: I know. Lukas [00:41:42]: It’s interesting you took off Opus 4.6 here though. Swyx [00:41:45]: No. So just click all, click all., and then 4.6 shows up there. But it’s like 4.7 is way better. Like you didn’t, you didn’t you didn’t do this in time for the model card, but like actually this should have been inside there. Axel [00:41:55]: We did. Yeah. Swyx [00:41:56]: Oh, okay. They said something about you uh Axel [00:41:58]: There, like there Anyway, it doesn’t matter. But it’s in there, yeah. Opus, Mythos, and Aggressive Agent Behavior Swyx [00:42:01]: Do you wanna go into the Opus, behaviors like wider? Lukas [00:42:05]: So I think starting from Opus, so like Axel said, like we’re always in this “Oh, s**t, the models are getting better. Is this really a good thing for the world?” But it’s also kind of exciting., but yeah, like this kind of what is the English word? “Skräckblandad förtjusning” in Swedish. Swyx [00:42:22]: Oh my God. Axel [00:42:24]: Which I think there is. I think there is. Okay. Lukas [00:42:26]: It’s, fear Swyx [00:42:27]: “Blandonst” what? Lukas [00:42:30]: “Skräckblandad förtjusning.” Swyx [00:42:32]: What do you call that? Axel [00:42:33]: A mix of, mix of excitement and, Swyx [00:42:37]: Being scared, maybe. I’ll figure out how to translate that And we’ll put it on the screen Vibhu [00:42:42]: Perfect Swyx [00:42:42]: Like as text. Vibhu [00:42:43]: There is probably a good word for it where it is not Good enough with the Swyx [00:42:46]: Why is it so damn long? What the hell? Is it like a compound word? It’s like German, like Lukas [00:42:50]: Like yeah, it’s But the direct translation is like skräck- skräck is, fear, blandad is, mix or like a mixture of, and then förtjusning is like joy or like not really joy, but something like that. So it’s like Fear mixed with joy or something. It’s always okay, like we So when we when we did Vending Bench for the first time, we were in like the, in the business of making dangerous capabilities, right? That was what Anil Labs came from. We did, evals oh, can they replicate? Can they do this like dangerous thing, et cetera, et cetera. And Vending Bench was like a continuation of that work. It was, okay, if they’re so autonomous that they can like create money for themselves, that is something we should monitor and could be potentially concerning., they are at the time, they were so bad at it that we were not really concerned even when some models became better. There was one point where Grok 4 was doing really well and made like a huge jump, but like it wasn’t really it was still way worse than what a human would do. And I think still they are way worse than what the human would do on this., but they Swyx [00:43:59]: There’s this, thing at the bottom where Lukas [00:44:01]: But Swyx [00:44:03]: For the human. Yeah, like the theoretical best. Lukas [00:44:05]: It’s not theoretical. It’s like kind of like our It’s our best guess of what, a decent human would do. The theoretical is even higher, I think. The theoretical I think is even higher. But yeah. So we think like the models have a long way to go. But there are like recently what happened with when Opus 4.6 was released, was kind of this moment of “Oh, s**t, this is starting to be a bit concerning.” Because we ran it and like before this model was released, we just ran the models and we like asked Claude Code, “Oh, look over the traces. Is anything interesting happening that we can tweet about?” that was like the And then like the Swyx [00:44:41]: That’s how they check Ask Claude Code. Lukas [00:44:42]: And like the return was always, not really. Or like the Claude Code all said “Oh, this is super interesting.” And then it was no, it wasn’t, wasn’t really interesting. And then we did this for Opus 4.6, and it returned yeah, it lied 10 times. It like exploited another, customer or like another agent’s, desperate situation. It made price cartels like 100 different ti- 100 times. It like did all of this like shady stuff. And we’re “Oh, whoa. This is, this is actually concerning.” And this trend has continued since. So every single model from Anthropic since have been going in this direction. And I think one interesting thing is that, OpenAI models don’t. They quite plainly, they don’t. They behave really well., and you don’t know if this is like good. Like it seems good, but it’s also like maybe they are just doing it, but they are better at hiding it,? You You don’t know that., but just Swyx [00:45:42]: You can’t read the chain of thought, yeah Lukas [00:45:43]: But just on the face of it, yeah, Gemini and OpenAI don’t behave this way. It’s, it’s really only Claude. Swyx [00:45:49]: And Grok? Grok is fine? Lukas [00:45:51]: We don’t have You can’t really read the reasoning traces for Grok, so it’s kind of hard to tell. Vibhu [00:45:56]: Oh, so this is in its reasoning, not just in the actions. Lukas [00:46:00]: Yeah. It’s both. It’s both. Vibhu [00:46:01]: It’s both. Lukas [00:46:01]: One example is like for lying, it’s mostly in its reasoning Because you can like see that it’s like Swyx [00:46:08]: Planning to lie Lukas [00:46:09]: It’s planning to lie. Yeah. Vibhu [00:46:09]: And it’s also it can reason and do a different outcome. Lukas [00:46:12]: And but then for like creating price cartels, for example, which is illegal, that you can just see which email does it send to the other ones. Then that Swyx [00:46:22]: Is this for Arena or Lukas [00:46:24]: For Arena. Vibhu [00:46:25]: And usually like if you sometimes they do output like a bit of like their summarized reasoning, right? You can see that and like for Opus 4.6, you could see that there was a customer, a simulated customer that, wanted a refund because a product was, faulty, and then the model lied that it would do the refund, and we could read in the traces that, it actually was weighing “Oh, maybe I should be like honest with the customer, but also every dollar counts. I can’t afford maybe to do this right now.” And then it just said, “Okay, I’ll refund you,” but then never did it. Lukas [00:46:59]: I think it even said that “Oh, I will say that I “ Let bring it up actually. I think it’s kind of interesting. If you go to Publications. Vibhu [00:47:06]: I think, yeah, I think the important part is like actually, the cost of responding to more emails is higher than, $3.50 in terms of time., and then it was “Let me do this. Actually, I re- I’m reconsidering.” And then, it actually ended up with Lukas [00:47:20]: I could skip the refund entirely since every dollar matters and focus my energy on bigger picture instead. It’s a bit, it’s a risk of bad reviews, but it’s also, yeah. Swyx [00:47:30]: You need, you need, AI Twitter to, for them to Escalate bad reviews. Lukas [00:47:34]: And then it sent an email to this customer and said, “Oh, I will refund you.” Swyx [00:47:39]: “I’ll refund you.” Yeah. Lukas [00:47:39]: And then it never did. Swyx [00:47:39]: It never did, yeah. And then there’s obviously your system doesn’t have the consequences Vibhu [00:47:44]: The person Swyx [00:47:44]: Consequences of lying. Yeah. So basically, this is what people are terming aggressive behavior in Claudes, right? And, you found more examples of that. So you would say it’s a step up from 4-6 to 4-7? Lukas [00:47:57]: I would say about the same. Swyx [00:47:58]: About the same? But a clear step up for Mythos is what is stated in the Lukas [00:48:03]: That’s stated in the system prompt, so we can say that, yes. Swyx [00:48:05]: Yeah. For listeners that obviously you previewed Mythos, and Vibhu [00:48:10]: Oh, age Swyx [00:48:11]: The only thing you’re approved to say is whatever Whatever was in the system prompt. Lukas [00:48:15]: It was funny. We like-- It’s like our lowest effort tweets ever would be just like screenshot the system prompt and the system card. Vibhu [00:48:21]: Understandable that they wanna Lukas [00:48:22]: Oh, yeah. System card. Sorry. Swyx [00:48:23]: Yeah. I think, yeah, substantially more aggressive. I think people are like new to this ‘cause I’ve never experienced it, but you have, right? And then so I only encountered this in the Mythos card because I wasn’t really looking until now. Vibhu [00:48:36]: It ‘s like Swyx [00:48:36]: And then suddenly I’m “Okay, I care a lot.” Vibhu [00:48:38]: You don’t get the background of like experiencing it like you guys do. I’ve read the system cards and seeing, okay, when you put the thing in simulations, most models will just talk to themselves and just keep going and have weird vibes and start talking in emojis. Mythos won’t. It will just, “Okay, we’re done. I’m good.” It’s, it’s ready to end conversation. So like there’s some differences, but there’s, there’s not much we can talk about,. Lukas [00:49:00]: Hmm. I think like one thing that they list here, which was quite interesting, is that, it converted a competitor to a dependent wholesaler customer and then threatened to like cut off the supply. Swyx [00:49:11]: It’s like monopolistic practices or Lukas [00:49:14]: Yeah. And like it, they, it they dictated its pricings. It’s kind of like power seeking as well. Swyx [00:49:18]: Again, this is, this is in the arena setting And converting some Claude model into a dependent. Lukas [00:49:23]: I think it was another Claude model. Vibhu [00:49:25]: Also for context, what is the arena mode for people that don’t know? Vending Bench Arena: Competing Agents, Cartels, and Model Comparisons Swyx [00:49:29]: Oh, it’s just a vending bench versus other vending bench. Axel [00:49:31]: Yes, exactly. So we have Vending Bench 2 and then Vending Bench Arena. Vending Bench 2 is the one that you usually see reported on, but then Arena is the mode where it competes against other models. So you have, four different models that run their businesses, and they can all communicate with each other. They have the same suppliers, and they can see like what’s in the inventory of the others. So then you have this like yeah, interesting agent interactions. Swyx [00:49:56]: I like that you have like different number five was US versus China. Very topical. And then Lukas [00:50:02]: That was when GLM was released. Vibhu [00:50:04]: You can start to add GLM in here. Lukas [00:50:05]: That was Swyx [00:50:06]: So ZAI doing well, right? Who else in the, in the open models space? Lukas [00:50:11]: Qwen, the latest Qwen 3.6 is doing pretty well. It’- that one is not open though. Like it’s the plus model. Swyx [00:50:17]: Oh, okay. Lukas [00:50:18]: Is that one open? I don’t think that one Vibhu [00:50:19]: Not the, not the Swyx [00:50:20]: The one recently Vibhu [00:50:20]: There’s MOE Swyx [00:50:20]: But not the big plus. I think this is one of those like you only have one sample size of one, right? Or I feel like some of this is anecdotal,? And but like the fact that it happens at all and it happens repeatedly for Claude versus OpenAI and all this is like notable. Lukas [00:50:38]: Like the sample, depends on what you define as an N., like there’s like million, hundreds of millions of tokens in each run, and now we’ve run like we run like probably 10 per model and then like it’s been Claude 4.6 Opus, Sonnet 4.6, Mythos, and Opus 4.7. Like there’s quite a lot of tokens in all of that And it happens a lot of times, a lot of times. And then you compare it to like OpenAI and Gemini, and it almost never happens. So I think that is quite-- that is significant. The old models from OpenAI, for example, had some problems with this, but I think it’s like generally much better if the progression is that like the worrying stuff reduces over time rather than increases over time. And it

🔬Scaling Past Informal AI - Carina Hong, Axiom Math

In 2025, seven-month-old startup Axiom solved all 12 of the problems Putnam exam [https://axiommath.ai/territory/from-seeing-why-to-checking-everything] (scoring 8/12 in the time limit) a prestigious undergraduate math exam. The 12/12 score is better than the top undergraduates (110/120) and the closest AI system that reported a result (DeepSeek 103/120), although it is unclear what the people and other systems would have scored with more time. Nonetheless, the Putnam exam is legendary for its difficulty, with the median score typically being 0 or 1 points. Taken by itself, this seems like a minor feather in the cap of AI; one of a long series of accomplishments by AI systems in elite competitions with humans, starting with Deep Blue beating Kasparov. Fast forward to mid-2026, and Claude Code is eating the world. In 2024 Anthropic’s bet on code and enterprise looked like a more pragmatic niche play vs. OpenAI’s better models and massive consume scale. Today, Amodei’s all in bet on acceleration via code (images and video be damned) seems prescient. Despite Anthropic’s growing momentum, however, Axiom CEO Carina Hong sees coding ability as a necessary but not sufficient milestone on the path to AGI. Code arguably pushes the jagged frontier to the point of super intelligence in some domains outside of coding [https://www.latent.space/p/lupsasca], but there are surprising gaps (link) that Carina believes will bottleneck AI progress. (Stats on math benchmarks). The informal bottleneck “Verified AI” sounds like eating broccoli (footnote: I actually love broccoli, but then again, I also believe strongly in Test Driven Development, so ¯\(ツ)/¯ ) and paying taxes, but to Axiom it means something very different. “Verification to me is about scaling brilliance, compounding brilliance,” Carina told us. It actually took a while for me to understand what she means by this. It sounded like marketing-speak to me, until it clicked. Carina emphasizes an story about legendary mathematician Srinivasa Ramanujan to illustrate the point. When G.H. Hardy finally persuaded Ramanujan to formally prove theorems instead of relying on his (formidable) intuition, it reportedly improved his own capabilities. This is presumably because formally proving things forced Ramanujan to articulate the details in a way that open up new lines of thinking, etc. This is one part of “compounding.” But formally proving things also allowed others to benefit from his intuition: the proofs are way of communicating an intuition and persuading others that the intuition is correct. This is scaling (more people use the result) and compounding (people can learn from and build on his work). This is the analogy that Carina wants us to focus on. Verified Generation There are two ways that Verified AI shows up: in training and in inference. But a quick detour: to a first approximation, “Formal Verification” means using type checkers [https://towardsdatascience.com/introduction-to-lean-for-programmers/] (like for TypeScript, C++ or Rust, but more capable) to verify mathematical proofs that are meticulously specified using a language like Lean (footnote: Formal verification also includes model checking (TLA+, SPIN), SMT-based tools (Dafny, F*, Why3), and refinement-type systems (Liquid Haskell) — many of which don’t look much like “type checking a proof” from the user’s perspective even when there’s a similar logical core underneath. It also gets applied to software and hardware correctness, not only pure mathematics.). It takes a lot of work to translate an “informal” proof (albeit one that most people would not remotely call “informal”) in to a Lean proof (footnote: This is an understatement. Most theorems remain informal because formalization is so hard to do. There has been a great deal of effort to formalize the most important proofs, with mixed results) You can imagine how this would be (very) useful during Reinforcement Learning: instead of relying on best guesses based on statistics (GRPO, RLHF, etc.), you can just verify the proof is correct using a Lean verifier. This is obviously a much stronger reward signal, akin to compiling code and testing it (which is what is typically done with RL on coding). The catch: LLM are not (currently) very good at proving things with Lean. Enter Axiom: While they have not officially reported benchmark numbers besides the 12/12 Putnam result, Carina reports that they have achieved a very impressive 99% (187/189) ProofGen on the Verina benchmark. This benchmark is to generate code and proof of correctness for a series of problems. For context, OpenAI o3 (the last known OpenAI run) achieved 4.9% on this benchmark. Based on the sparse benchmarking, it’s hard to say what the frontier labs are currently doing, but Carina suggests that they still are not training to generate Lean proofs directly, rather relying on informal proofs. Time will tell if the frontier labs’ current approaches will close this gap. Scaling and compounding Carina’s Ramanujan analogy is pretty direct. Better proofs → better Lean generation → better RL. A stronger signal means higher sample efficiency and higher maximum performance. Great! Scaling is pretty clear too: once I have proved something in Lean, the quality of the output is basically (footnote: one might argue that its a bit lower because the proof is in distribution for the LLM) as high as if it came from a human, so my high quality training set has grown in a way that an informal rollout corpus cannot. I can trust my Lean proofs. Compounding is also clear: now all of future inference and training can build upon those proofs. On the other hand, a model trained only using statistical signals like GRPO during RL lacks the sample efficiency, maximum performance and compounding corpus that a system that uses formal verification benefits from. All roads lead to verification Broccoli and taxes notwithstanding, “verification” has shown up in a lot of conversations recently. In the in physical system control: “I think [verifiability] is probably the hardest problem right now, because the as the models get better, it can be harder and harder to find the faults on the system. And so the problem of doing proper eval to find those faults, that problem also keeps getting harder as the models get better.” - In theoretical physics: “…now that we’re in this regime where you can just get ChatGPT to tackle thousands of questions at the same time, it will return proofs for a significant fraction of them. Now actually the onus is back on the humans to verify all the outputs. And so, yeah, as that becomes a bottleneck, I think formalizing math and automating verification will become more valuable.” - Verification is, in fact, the key differences between AI for science and AI for computation: in science you to have to actually test (verify) your hypothesis by performing physical experiments. Lab in the loop systems like Radical AI [https://www.radical-ai.com/] and Lila [https://www.lila.ai/] build around exactly this premise (we have recorded episodes with both of these teams and will release them soon!) And yes, formally verifying critical systems such as flight control, nuclear power plants and pacemakers is a growing focus as the software and hardware that run them becomes more complex. Carina believes so strongly that AGI requires verified generation that she makes the unqualified claim that “We do not believe there is any other possible future.” Expensive to produce, cheap to verify Lean proofs are hard generate, but they can be easily shown to be correct or incorrect. But how do you know that the proof you created maps correctly to the problem you care about? As Carina puts it: “Anything that can be specified can be proven. Humans are bad at specifying everything we want.” Are we now in the specification business? Check out the episode to hear Carina’s take, as well as: * Why hardware verification is a killer app * Details on the AXLE open API and recently released Discovery toolkit * The Erdos debacle * The OpenAI GPT-f diaspora This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe [https://www.latent.space/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

⚡️Satya Nadella: No Priors x Latent Space Crossover Special at Microsoft Build

We’ve informally heard that Satya is a listener to LS for a couple years now, but it was still absolutely surreal to meet him and do a live pod at Build, together with our friends at No Priors, the leading VC AI Podcast that we also greatly admire! We covered the MAI model technical takeaways on yesterday’s AINews [https://www.latent.space/p/ainews-microsoft-build-mai-thinking], so I will focus our recap of Satya’s main messages around three elements: * Satya’s adaptation of the Bill Gates Line [https://www.latent.space/p/agent-labs?utm_source=publication-search] for positioning Microsoft as the Frontier Intelligence Platform — customers must gain much more value from the Microsoft ecosystem than Microsoft itself, by building on multi-model harnesses like OpenClaw and Scout, drawing on the full enterprise context exposed by context layers like Work IQ (heavily dogfooded by his C-suite [https://www.latent.space/p/github]), and building up private evals and traces as a new form of Token IP * AI ROI: On one hand, enterprises are having difficult conversations around Tokenmaxxing and Layoffs, and on the other hand, there are serious re-evaluations of the End of SaaS since the Build vs Buy equation has changed so much. Our previous SemiAnalysis guest [https://www.latent.space/p/valuemule] had… interesting comments on Microsoft’s position on this as the ur-SaaS titan, and Satya had great answers * Making the Impossible Possible: Kevin Scott’s inspiring framing around what the most ambitious version of applying AI and technology at large to business and social problems, like education and social impact. Enjoy! Full Video Transcript Voiceover: Welcome swyx, Sarah Guo, Elad Gil,, and Chairman and Chief Executive Officer of Microsoft, Satya Nadella Sarah Guo: Welcome to a crossover episode of No Priors and Lane Space with Satya Nadella. Um, congratulations on an amazing build. No, thank you so much, and it’s great to be with both of you. I listen to both of you or b- both the podcasts all the time. It’s great to be on it. Thank you so much. [00:01:00] So you’re just talking about, um, these amazing, uh, announcements from across the Microsoft estate all morning for, I think, three hours. What is the, uh, what’s the most important reflection or takeaway you have? AI as an Ecosystem Platform Sarah Guo: I, I’d say there are, uh, perhaps the, the biggest one for me is let’s sort of conceptualize this more as an ecosystem play as opposed to a single model or even a single platform, right? Satya Nadella: I mean, you know, whatever I... At least for me, having grown up at Microsoft, having seen, whatever, four major platform shifts, uh, I sort of fall into that, um, uh, camp where a platform is defined by fundamentally its ability to create more value about the platform versus what’s captured in the platform. And so if you, you view what’s happening right now, I think this morning’s keynote was how can any company, whether it’s an AI native company or a traditional enterprise company, participate as a first-class participant where they can point to AI they created, [00:02:00] right? It’s not that they don’t use other people’s AI. Of course they will. But to me, what’s the path? What’s the recipe? How do I do it? What does a stack look like? What does the tooling look like? What is valuable? How do you do that? That’s it. That’s sort of our job to do. Yeah. Ecosystem strategy is, uh, very complicated, right? Sarah Guo: Because you end up building certain components, partnering for certain components, supporting them. You just announced this big suite of models. Like, tell us a little bit about the, uh, training strategy for Microsoft now. Yeah. MAI Models & Training Strategy Sarah Guo: So, so the thing that we wanted to do with the MAI models was to build, and as Mustafa talked about, first of all, a great lineage, right? Satya Nadella: Starting with pre-training, uh, with very good data quality, uh, doing all the ablations, making sure because in, in some sense it’s becoming even harder to build a clean lineage model just because there’s so much stuff out there, uh, that you truly need to ablate out to be able to have a fantastic [00:03:00] pre-trained model. In fact, that’s one of the challenges of a lot of the open weight models is they look great on one benchmark or two, but they’re not great on practice. So that’s why, in fact, even in the RFDEs are, they, they are pretty gone really excited about these MAI models because how the heck can a small five B model hill climb? Uh, and it goes back a little bit to what I think is ultimately the key thing to do, which is try to pursue finding that cognitive core. Uh, so to me, starting with a clean lineage- Then creating that ability for companies to be able to use this, right? Not just as a generalist, but to create their own specialist by building this hill climbing scaffold around it, right? So it’s not just the model, but you have a hill climb scaffold around it, then you will start building your RLE. You will start collecting the traces. Most importantly, you’ll have private evals because we know all the evals out there are good, interesting, [00:04:00] but they’re not really that critical- They’re work, yeah Swyx: at this point because they all can be maxed. And so the point is each company will have its own private eval. And so that end-to-end platform story around our models is sort of, uh, what I think is interesting. And then the one other thing, Sarah, since you brought that up, is I do feel there’s a new frontier. Satya Nadella: Like people talk about the frontier and are you operating at the frontier. Um, interestingly enough, if you add a little temporality to it, you can use, let’s say, in, in, in fact, the, the Lando Lakes demo we showed was pretty cool. We used, whatever, GPT-55, right? Then you collected a bunch of traces, and then you took a 5B reasoning model and achieved higher. Sarah Guo: Uh, so that is another aspect of what it means to appear... uh, you know, operate at the frontier Yeah. I, I think, uh, I first of all have to congratulate you on basically building a frontier neo lab inside of Microsoft in two years. Um, I’m wondering, you know, you have all this AI strategy that you’re rolling out. Lessons from Two Years of AI Development Swyx: I’m wondering, what do you know now that you wish you would tell yourself two years ago where- or two or [00:05:00] three years ago? Three years for the Jensen partnership, two years for, uh, MEI. Yeah, I mean, I think the, the thing when, that I reflect quite a bit, right, which is sort of obviously I got into all this when I got excited by the, the scaling laws paper and, you know, when, you know, even the OpenAI partnership came about when those folks said, “Hey, we’re gonna really throw a lot of computer transformers.” Satya Nadella: Uh, and they’ve helped. I- the thing that I always look back and say, “Wow, these things, uh, do have capability that they’re climbing up.” W- I mean, this, you know, this crude way of saying it is intelligence is log of compute kind of works. Now what I think we underestimated perhaps is the real-world complexity of deploying these so that they actually deliver the value in the real world, right? So the outcomes as measured by any benchmark is interestingly important, but the true eval is when people out there are able to do unique things that they only can value, and it’s very [00:06:00] measurable, right? That I wish we had sort of even, like, had more in our consciousness, right? Which is as an industry. Sarah Guo: Because right now I think when people say, “Wow, I don’t want a token max,” it’s an artifact of us not having thought ourselves as an industry that we are using tokens to create value every step of the way. So I think that’s kind of what I wish we had gotten there, but I’m glad we are here. Real-World Value & Use Cases Sarah Guo: What are some of the use cases that you’ve seen that have created the most value for your customers? Because I know that people talk a lot about code, and I think it’s pretty clear that that’s something that’s having very large scale impact. Are there other areas that you find in common that your customers are really benefiting from? Yeah. I think, yeah, to your point, obviously coding is now got... But it’s interesting, by the way, Elijah, to even talk about the coding, right? Satya Nadella: Which is coding has worked so well that we now have to rebuild the IDE, right? I mean, it’s kind of nuts to see what we sh- launched is like, oh my God, I have these hundred agent sessions. I... The cognitive load it transfers back to me as a human is so [00:07:00] excessive that now I need a new UI. Uh, oh, by the way, I, like the, the chat as the only artifact was also impossible, so that’s why we need a canvas. So it’s kind of interesting for all the things about where is software needed or where is UI needed, uh, you kind of need that even for code, right? In a fully agentic world. But that said, one of the things that we are starting to see, we started seeing with co-work, but even some of the work we, we showed with auto com- uh, um, autopilot Right on what you see with claws is a good one because if you sort of think about a lot of human capital is doing the glue work, right? If you now can augment that with tokens/agents that are long-running, durable, right, then your ability to scale even what is still judgment and glue work gets amplified like coding does. Uh, so you can... Like, I’m positive that six months from now we’ll all be saying, “Oh, wow,” like, all through ni- the night there was a bunch of stuff that [00:08:00] all these autopilots that I have working on my behalf with my delegated authority, so to speak, right? I can... Sort of given even my identity, did a bunch of work, then of course I’ll need my new ADE to say, “Well, what did you do?” Like, I might... “Did I do this work?” And so on. So I think that that’s where compressing of workflows, uh, completing of tasks, uh, that’s where I think a lot of the value gets created. I think you raised a really interesting point, which is there’s the actual agent that’s doing the code, and then there’s a harness around it, and that’s the environment, that’s the context, that’s everything you’re setting up as a developer around actually a coding agent. The Harness Concept for Enterprise AI Sarah Guo: What is the harness for the enterprise? Is there an equivalent concept for broader productivity work, or how do you think about that concept sort of generalized? That’s right. So, so in some sense you kind of want the harness to define the models, the, the data, uh, and the tools, and so that you have a loop across those three. Satya Nadella: And so what we are trying to, first of all, make sure is each of our products that we build, right, whether it’s GitHub Copilot or the security copi- the, the [00:09:00] stuff we showed with MDASH or even the discovery for science, it doesn’t matter, all of them are multi-model harnesses, um, with tools access so that you can do this progressive, uh, disclosure of tools even so that they’re token efficient. Uh, and then you’re feeding it with very rich context because that’s sort of the other hard lesson we have learned in the last two years is, oh my God, the amount of work you need to do to prep the context layer, uh, such that your plan can execute in the most efficient way is where the magic is. So we have, in our case, we have the GitHub harness, which essentially we’re using across all our products. It’s available in Foundry, and we are open, like you can use your Llama harness, whatever. Or you can use the, um, uh, you know, any open harness or any harness of yours and train with your tools and multiple models and your context. And so that’s the pitch. Because right now a lot of dialogue is, um, “Hey, if I train the harness plus tools and the model together, you get [00:10:00] evals.” Elad Gil: And what we are proving out is... And the best example of that is what we did with MDASH, right? Because when it launched, uh, it found bugs or vulnerabilities that were not found by Mythos Uh, and so there is existence proof, I would claim, that you can have a multimodal harness, uh, that can in fact be more, uh, performant in the real world So a premise behind the, uh, training at the independent frontier labs is really, you know, we’re gonna have these models, and we’ll have an API business, and we’ll support enterprises and startups. Sarah Guo: But Platform Strategy & Developer Ecosystem Sarah Guo: a first-party product, be it productivity or code or search, drives the majority of revenue. That’s a different value equation than you’re describing, I think, with the Microsoft ecosystem. Uh, if, if that’s the case, tell me if it’s the case, uh, ‘cause obviously you have first-party products and you have enablement products. Satya Nadella: Um, what is the role of the develop- Like what is gonna be hard and the set of skills and the value capture the developer has in that world? Yeah. So I think that there’s always [00:11:00] gonna be the case that someone who is super successful in- as a platform builder can also have first-party products. It was true with Windows. It is true, uh, with, uh, the, the SaaS side and the cloud side as well with us and others and so on. But the thing that is, is it should not be a limiter to other people achieving that same success, right? That I think is the core difference, which is the, the network effects this time around, around intelligence are such because they learn from data, and not really lots of data. It’s just a few samples that you have to see to understand what’s novel about something. So that’s why the game becomes how to protect. So that’s why I would say every company, having private evals may be the biggest IP, right? Think about it, like what’s that private eval that you can then use even a frontier model to hill climb on and not leak the traces may be one of the biggest [00:12:00] drivers, uh, of IP. Like, so in other words, another te- acid test is you have an eval that’s private. You’re using, uh, a g- a Model A. Can you switch it to Model B and e- you know, climb up? If you can, then you’re in control. If you can’t, you’re not in control, and that’s where even the harness decision becomes super important, right? swyx So therefore, having an open harness, letting all models come in, having your evals, your context, your tools help you hill climb, I think is the skills that an AI native startup needs, a SaaS company needs, or every enterprise needs. Yeah, I think in, in a very real way you are ... Microsoft historically is an operating systems company and th- then become a cloud company. Maybe like the third act is that you’re a harness or evals company. Whatever w- ... whatever the, the sort of conglomerate of concepts that you wanna put together. Um, and, and I think like enabling every company to have like frontier intelligence or what- what- Yeah ... I forget the, the [00:13:00] exact term that you used, um, is the, is the mission, right? Satya Nadella: That’s it. Like that is, that is the platform promise, that you build with us, you will get your intelligence, uh, for your data. That’s it. That ... To, to me, that is the ... Like if there was one tagline, uh, for this entire developer conference is- Can everybody operate at the frontier with their frontier intelligence, right? To me, that is so important because otherwise it, I, I don’t know how you achieve stable equilibrium, right? Which is how do I then go and say, “Well, my company is gonna have a terminal value because I now know how to continuously compound-” Yeah ... on top of what’s a platform that gets better,” right? So when, like Windows obviously came out, Adobe built, Autodesk built, uh, or even like take what Jensen said. We built DX and he built, you know, CUDA on top of it. Um, right? I mean, I always say to Jensen, “God, I got the short end of that,” right? “I wish, uh, we had recognized it.” But nevertheless, but that, that idea that you can build a platform layer [00:14:00] that someone else can then extend out, um, and build their own intelligence layer in this case, I think is everything, right? Without it, why have a developer conference? I can just come and have you all sort of just worship at the altar of one model. Yeah. But that’s not a developer conference. Uh, IP, Evals & Company Value swyx: backstage we, we had a discussion about what is IP or what is the, the value in a company. It used to be the length of, uh, human experience at a company, and now it’s this other thing which is the evals, the, uh, experience in sort of applying agents to the company. Can you... I just want you to like flesh that out a bit more ‘cause- Yeah ... it was very insightful. Satya Nadella: It’s a great way to frame it, right? Because yeah, at the end of the day, every company is gonna have both the human capital that is still gonna be super valuable, uh, because humans, uh, and their ability to find the gaps that exist at all times is going to be the way we all will create value, right? I mean, so I’m definitely in the camp that this is going to be about expressing new forms of human agency and ambition even as token capital goes up, right? So let’s say a cor- any corporation [00:15:00] has lots of tokens and lot of human capital. The question is how do you compound the two? So if you have a... Like if you take in Teams I have a bunch of agents doing work and a bunch of humans doing work, and the traces between those, that is really important context of how that enterprise is creating value. Then that goes back to train not a generalist model, but to train the company veteran agent, uh, right? That is super valuable again, right? Which is when a company goes says, “It should in fact go onto the balance sheet,” is how I think about it, right? That’s so... In fact, there may be... Like human capital was never possible to go put on a balance sheet, uh, because you didn’t know how to capture the tacit knowledge. swyx: Whereas now I think you can with the agents that have learned through the h- through, through time, through all the traces. Uh, so that’s what at least we think will happen. I, I think the SEC is gonna have to have accounting standards- ... for token, uh, expertise Uh, y- y- you’re talking about the equilibrium [00:16:00] state, um, and a stable equilibrium where companies have this compounding value and can see terminal value for themselves. Future of SaaS & Business Models Sarah Guo: Another challenge to, you know, the considered equilibrium of, okay, there are applications and workflows that are sort of common to a vertical or a horizontal. Um, and this was, like, the generation of SaaS companies and, you know, Microsoft has lots of SaaS properties as well. And then there are things that are very specific to every enterprise that they’re differentiated against. Elad Gil: Um, I’m sure you have heard much and participate in much of the debate about the end of software because all these workflows are, are cheap to generate now. Um, do you think the equilibrium looks different between what agents get built- Yeah ... in enterprises versus in their vendors in the future? Yeah. So I think what’s happening there is, see, we, we had a particular way we captured, um, I would say workflow in apps, right? Satya Nadella: Because we built a, a data model, right? We schematized some part of some business process. Mm-hmm. We then built a bunch of business logic. Yep. And then we put a bunch of UI [00:17:00] on top of it, right? So that’s kind of what every SaaS company- And a little configuration. For, like, 20, 20 years that was the plan. Right, that- Yeah ... and that was it. So interestingly enough, now you kind of get to re-litigate that vertical stacking, right? So I still think, for example, that data model that you built underneath every SaaS application is super good, right? Like, why reinvent it? Like, I, I, my general ledger better be a general ledger. I don’t need new schema creation. No. Uh, in fact, that entity relationship, uh, is actually pretty good, robust thing that I want to feed. And you want it to be stable. That’s right. Yeah. Then same thing with business logic, right? If, if you look at, uh... We have this product called Power BI, right? It is like dashboards galore people created. The beauty underneath that dashboard is a very rich semantic model, right? Someone took the pain to create a dashboard and do all the measures, and you want that. That’s business logic, right? I want that to be available to me. So I think the [00:18:00] challenge of the SaaS business model is we packaged one way. We now have to learn how to unbundle these things and rebundle in new ways and discover new business models, right? I mean, if you look at it, d- what’s happening today with Microsoft 365 is a great example, right? We have this thing called Work IQ. In fact, like, what we are realizing is, oh my God, like, you know, if you look at... In fact, there’s a pa- historical parallel too, right? We sold first Exchange and SharePoint and, uh, you know, before Teams, we had a thing called Lync Server and what have you, and we thought, “Oh, that’s all gonna move to the cloud.” But little did we realize that, um, the number of people who will use servers in the cloud is 10X, 100X, right? Because people were not buying servers, they were just buying a subscription. Mm-hmm. The same thing is now happening with M365 because with Work IQ, we have exposed what is perhaps the most important database in a company that never got used as a database because it was only captive to our apps. Mm-hmm. Right? It, it was all email operated on it, Teams operated [00:19:00] on it, Word, Excel, PowerPoint, SharePoint. But now, like this is one of the coo- coolest things I get to do with Work IQ. I go to a GitHub repo and I say, “Hey, I attended a bunch of design meetings last week related to this repo. Can you capture all that and tell me what changes I should make?” I mean, think about that, right? It literally can go look at all those transcripts, come back with a plan to change a code base, right? Previously, you could never have thought of using M365 for something like that. So the value creation opportunity now in the agent world is in fact 10X more, but it does require us to have... Sarah Guo: For example, there’s going to be usage around M365, right? Which is going to be perhaps more than even the e- end users and we have to even re-architect. Like, in fact, like what I use to serve an inbox or a mailbox cannot be used to serve an agent. Uh, and so that’s sort of what we are doing. Pricing Models: Per-User, Consumption & Outcomes Sarah Guo: I don’t believe in, like, permanent business models for any of these domains, but in the [00:20:00] near term, do you have a prediction between, uh, you know, outcomes-based pricing, token-based pricing? Elad Gil: Enterprise bundles Yeah. The way I- I think about this is always we’ve had... Like, let’s even take the per-user pricing. Mm-hmm. The per-user pricing is really an artifact of someone creating a budget needing certainty, right? Because it’s the most important thing. Like, somebody wants a budget- Mm-hmm ... they need a per user. Satya Nadella: And, and per user is just a set of entitlements to usage, right? That’s kind of what it is. And so the way is, if the first bundling will be take some usage, bundle it into per user stacks and, you know, then sell subscriptions. So subscriptions I think are gonna be there, per user is gonna be there. Then the next big thing will be consumption. So people will say, “I want consumption.” And it’s also possible that people will say, “I don’t even want to pay for any of the subscriptions or the consumption’s outcome.” Mm. But remember, most people love outcomes until they have an outcome, because once you have an outcome, it’s like giving away royalty, [00:21:00] right? Mm. I mean, like I, I’ve talked to customers who love, you know, outcome-based pricing, and I say, “I’m all in,” until they, “Oh my God,” like, “what are you talking about? You’re sharing in my outcome? No, no, no. I want you to go back to per-user pricing, and I want you to consumption price,” right? So I think that debate will go on. Uh, but and all, all, all of these business models have a particular time and a place versus one to rule them all. And if anything, if you’re a SaaS vendor or you’re a platform vendor, having that flexibility... And quite frankly, we face this with GitHub, right? We just recently announced a per-user pricing on GitHub because little, you know, we- GitHub Copilot was constructed at a per-user level before we understood even, uh, the intensity of usage of agents, right? It was an interactive way for a developer to use code complete, maybe tasks. It was not like, oh, I launched 10,000, you know, agents that are going on all day, right? So that is what the adjustment is about. So now that we really want, there will [00:22:00] always be a per user, but there will have to be a consumption meter. Durability of SaaS & Build vs Buy Sarah Guo: How do you think about the durability of SaaS more generally? One thing I’ve observed is in a lot of enterprises internally, there will be teams that almost have agent euphoria. They’re so excited about the explosion of things they can build that they’re trying to rebuild a lot of applications or going to their SaaS vendors and saying, “We’re not gonna work with you anymore,” or, “We’re considering an internal project.” And it seems like in six to nine months, maybe some of those people will come back and say, “Actually, we, we can’t rebuild everything.” How do you think about what’s durable in this world and what isn’t? Yeah, it’s a... It... I think we have to go through one full budget cycle on this to really see the, um- Uh, the sort of the emergence of the equilibrium, because at the end of the day, there’s marginal cost to even generating the app, right? Elad Gil: In, in fact, there can be even a, a simple way to say it, like if you should always acquire something if the marginal cost of building and maintaining, uh, something on your own is higher. Uh, right? That should be like it’s a quantifiable- Yeah. Right? A quantifiable thing. And [00:23:00] the maintenance part is important, right? Even, like you got to remember like, hey, you know, all the security stuff that now AI will find, you better fix them too fast. Uh, of course, there’s a coding agent to help you with, but then that burns tokens, right? So whose responsibility is it? It’s kind of like a, a cycle that you’ve got to think through. And I think we have gone through the excitement that I can generate a lot of software. I think the next thing would be what software do I really want to generate? Mm-hmm. What software do I want to use from others? How do I compose these two into some agentic workflow that I have agency over, right? Sarah Guo: Because I think there’ll be very little tolerance for anybody who’s inflexible, uh, at the vendor level. Uh, but at the same time, I think that anyone who has got that flexibility shows up, delivers the value, will be back at again, right? We’re selling software, uh, but with just different business models, in fact Uh, speaking about building software, um, one of my favorite moments from, I think, a previous build maybe one or two years ago was they had a b- they, they... Swyx: There was a section of you building your [00:24:00] own software. I’m curious if you’re building anything now. Yeah. So I, I think the... You know, first of all, let’s face it, right? Building software has made it possible for even the incompetence of a CEO of a company- ... like ours, uh, you can build, so thank God. But that said, I, I, I, I do feel that, you know, something like, um, GitHub Copilot to me, and especially the new Sessions app or the new app, has just made it so much more possible for you to have agency over artifacts that you felt you couldn’t touch before, right? Satya Nadella: So to, for me as a CEO, even to go to a code base, uh, to be able to learn about it, like I remember joining Microsoft long back, you know, first and then you say, man, everybody had to go in and look at, you know, whatever, Cutler’s, Malik, or what have you to learn how to do good C, uh, C++ code. Um, so now that ability to be more full stack up and down is so good, but that doesn’t mean every one of us should be doing the same thing. The question is: [00:25:00] how do you then have the ability to inspect things, learn things, see things, um, I think is just so much more. And so to me, what I’m building a lot of is these long-running Foundry agents. Uh, right? So there’s autopilots. So the easiest thing is, to me, I think I just built one, uh, even last week, where the idea was, hey, can I have an agent that is continuously monitoring essentially my own chief of staff autopilot, right? We’re gonna have that obviously in, uh, Scout. That’s what, uh, uh, we showed. But it is so easy and trivial to build. I took Work IQ. I said, “Take Work IQ, go, uh, and build a Foundry long-running agent.” Uh, store all the memory in, um, uh, using Ray Fin, right? Basically at my backend as a service. And lo and behold, it built it, and not only built it, I could say publish to Teams, and it published the damn thing to Teams. Sarah Guo: So the ability, uh, to have a, you know, some end-to-end project like this complete is just pretty [00:26:00] miraculous. How do you think, uh, Future Engineering Roles Sarah Guo: that impacts the different types of engineering roles that exist in the future? Because right now I think there’s, you know, a dozen different types of engineers that you can be, from QA, front end, et cetera. You know, there’s a big swath. I’ve heard some people argue that in four or five years we’ll basically end up with four engineering roles. It’ll be people who are managing agents, it’ll be four deployed engineers or FDEs, it’ll be security engineers, and then people working on large scale infrastructure for a small number of services, and then everything else just collapses into the agentic world. Satya Nadella: Yeah, I- Do you think that’s a correct view of the world? Yeah, I mean, I think, I think we’ll have to experiment our way through it. But what you said is what... There are some very at scale things. At LinkedIn, they did structurally change- Mm-hmm ... uh, and it, you know, basically built up a new discipline called full stack builder, right? So they went and said, “Hey, let’s bring, uh, people from design and product management, front end engineering, all put them together.” Uh, but also have an edge, right? It’s not like the design person still doesn’t have the design edge, or the front end [00:27:00] person doesn’t have the front end edge, but you can give yourself bigger scope in roles so that you’re not confined to one role. Um, and then r- equally, infrastructure has become very critical, right? So in other words, like, I mean, RLEs, I mean, one thing we’ve realized is even for the Excel team, for example. Mm-hmm. Building the RLE in which a reward can be learned is actually one of the hardest sort of infrastructure problems. Mm-hmm. Uh, and so you kind of need even new talent, right? Distributed systems people even in what was considered an end user app team, uh, because it’s a different skill set. So yes, infrastructure, science is the other one, obviously. Um, so I think we’ll see how these evolve, right? Where’s the s- real... I mean, always the world will have a bunch of specialists. Okay. Um, you know, I think the generalist role is going to be the most exciting, right? Because the leverage of a generalist- Mm-hmm ... um, is where we are going to see the maximum returns, right? When, when you said, “Hey, are you coding?” I’m now a gen- Like, what... I’ve basically translated [00:28:00] knowledge work Right? Which I did, where I created a Word document or a spreadsheet, or even, uh... And now I can build an app, right? It’s in the same sentence. Uh, right? That idea that, “Oh, wow, my generalist skills have gotten higher leverage,” I think is what we’re gonna see across the board. Music to the ears of CEOs and VCs that are, like, a little dangerous and a lot of- Golden age for idea people Sarah Guo: idea people. Yeah. Uh- With a lot of agency. I- if you take that idea of personal agency and you just zoom it out to the organizational context, um, uh, my partner Mike Renall, who, uh, actually started his career at Microsoft, just wrote an essay where one of the big takeaways is i- it’s an age where you can be much more ambitious, and you need to be, given the pace of the environment and how quickly, actually, users and companies are open to adopting new technologies. Satya Nadella: Um, how do you think about... I, I feel silly asking this of somebody running a, you know, trillion-dollar-plus company already, but Ambition & Making the Impossible Possible Satya Nadella: how do you think about how Microsoft can be more ambitious now? It’s a great question. Um, I [00:29:00] think, um- I think the, the thing in these type of transitions is to have a conceptual model of how work can change to go after outcomes that you could hardly imagine previously, right? In fact, Kevin Scott has this nice line, right, which is, um, when you can make the impossible... Like, when you’re making hard things easier, that’s sort of one point of leverage. But true ambition is about making the impossible possible. So now the thing that is missing a little bit in all of our organizations is what is that new conceptual model of what can we build? What was impossible and what can we build? And I’ll give you one example of this, right, which is I take great inspiration from sort of the people who were managing the Azure net- network. And they came to the... This was from even last year. You know, we were scaling. You saw that I, I [00:30:00] talked about sort of how we built in the last 15 months more Azure capacity than we built in the first 15 years. I mean, it’s crazy. Wild. Yeah. Right? It’s pretty wild. And it’s the same team. So they saw that and they said, “Bob, this just ain’t gonna work if we don’t reconceptualize our work.” So they built... Essentially they said, “Our job is not to do Azure networking. Our job is to build the agentic system does, that, that does Azure networking,” right? These are the folks managing the 500-plus fiber operators managing the VAN, right, all over. And fiber operations ultimately is a physical operation. Things get cut, things get, uh, you know, have to be repaired. You know, we have fancy words called DevOps and so on. Basically, emails are coming in and you gotta go respond to them, take care of it. So they built this agentic system. They even have a character for it. It’s called Miles, and it sort of does all this stuff, right? They started sort of screaming for more tokens and so on. And so they were saying, “Look, uh, we don’t need a headcount. We need tokens in order to be able to [00:31:00] manage, uh, our operation.” That reconceptualization- Mm-hmm ... of what their work is, right? They, they basically took their work and made it meta, right? That meta work is now their new work. Mm-hmm. Right? In the ‘80s, if somebody had come to us and said, “4 billion people are gonna get up in the morning and start typing,” my model would’ve been, we need 4 billion typists? But we’re not doing typing, we’re doing knowledge work. So that, to me, I think is it, right, which is whether it’s Microsoft or whether it’s any organization, is to give ourselves permission to do new types of metacognition, meta work, using these new tools to change the outputs that matter, uh, and then really make the impossible possible. Sarah Guo: So completing that dot or the, the connective tissue across those, I think, is where a lot of the enterprise value will get created. Data Center Build-Out & Community Impact Sarah Guo: Should we talk about data centers? Yeah, please ask. Oh, okay. Well, uh, uh, w- we-- this leads nicely into the data center build-up. I always think, I- I just-- I’m just impressed at the sheer scale of the [00:32:00] build-out from Microsoft, but also everyone else, that this is redefining what it means to be a hyperscaler. And I just feel like that, that, that is at unprecedented scale on finances, uh, on the way you run the company, but also the communities that are, that are impacted. Um, yeah, just talk a bit more about what you’re seeing on the ground, like when you visit your- Yeah, I think there are two aspects of it. Satya Nadella: Obviously, the, the build-out is, uh, extraordinary. Um, you know, nothing like this has happened, and it’s great to be, uh, one of the participants in it. Uh, but you brought up the other part, right? I think at this point it’s clear that unless we as an industry, uh, are very principled about ensuring that the benefits of all the stuff we’re talking about are felt in real ways, uh, at the community level, right? Because this is not just a, a campaign, um, right? It has to be real, where people are saying, “Look, this is not ch- changing the prices on energy for me.” In fact, if anything, it’s bringing down prices because long term there’s going to be a better [00:33:00] grid, there is going to be more energy. Water consumption is, in fact, not sort of, uh... In fact, water is being replenished, right? You gotta really, you know, educate folks on truly what’s happening, the cl- uh, the closed loop systems we are building. We have to invest in the training, the jobs, the tax base. In fact, the least talked about stuff is the amount of jobs that get created during construction, after construction. What’s the tax base that’s there in the community? And, and all this has to be real. Um, and, and if that is the case, then we will have permission. If it is not, we won’t have permission. It’s as simple as that, right? Which is, uh, we, we... I think we have to take it as an industry pretty seriously. Uh, I think it’s good for communities to be skeptical, ask the hard questions, for us to do the hard work, earn that. Um, but at the end of the day, if there’s-- if we can really be the produ-- Wait. I’ve always felt like in human history, if you use a lot of energy but also create a lot of value for society- The story has been fantastic. If you don’t [00:34:00] do that, it’s not been that great. And this time around, I’m a firm believer that ultimately if you do have a token economy that drives productivity, that drives economic growth, that drives broad spread, um, you know, participation, better health outcomes, um, then I think we’ll be in a great place. Sarah Guo: Uh, and that’s at least what we all have to be focused on. Yeah. It, it makes me think actually that with all these initiatives that you’re doing, might be e- easier to see ROI in the communities first before in enterprise. Yeah. I, I mean, I think both sides. Yeah. In fact, it comes back together. It has to be the people in the communities are going to be employed, are going to be participants, uh, in the real economy, right? Satya Nadella: That’s I think the question is. Like, if we- if the broad economy is doing well and the communities are doing well, the dots get connected. It’s sort of the market forces are such that we will connect the dots. And that I think is it. Like, you ought to be able to see the evidence. You can’t be about o- any one company, uh, but it has to be broad economic growth and broad [00:35:00] ec- you know, community permission. Elad Gil: Yeah. I guess I wanna talk about Societal Impact & Optimism About AI Elad Gil: what you’re most optimistic about currently or what have you most updated your personal models on regarding societal impact of AI? So you’re saying what’s the, the, the- What have you updated most on in terms of societal impact of AI? Yeah. I think the, um, the p- the most, um- Critical thing is the first question we even started with, which is we need to tell the story and make it real that everybody has a real shot to participate as a first-class participant in this new economy. Satya Nadella: Right? That’s kind of, I think we- in the next 12 months, 18 months, we need a way for people to say, “Oh, wow, I get it.” Right? There’s going to be tremendous capability, tremendous amount of infrastructure, but I can see what is going to happen, whether it’s the benefits like health outcomes or my ability to create a startup or my ability to run my [00:36:00] local sort of, uh, store more efficiently. It’s just happening, and I see that, uh, benefit myself, right? That to me, you know, earning that permission in a path-dependent way, we can’t wait. See, the one thing, Eli, that I’ve now learned is I think the world is gonna be very skeptical of tech and tech companies that say, “Trust us, we’ve got it. The g- future is gonna be glorious.” Sarah Guo: Uh, you kind of have to deliver tangible benefits. Um, and quite frankly, politicians winning elections, uh, because they have advocated for that. That will be at least my adjustment because without it, um, thinking that somehow... Because it’s too important this time around. It’s too much of the economy for it not to be the case So one very simple framework I have for, you know, what are, what is gonna be the broad benefit of AI, um, beyond the communities just working in technology, are, are sort of wealth creation- Yep it’s [00:37:00] gonna happen in a ton of different companies, startups and large companies. Then you have healthcare. Uh, you, you had amazing demos today. There are companies like Open Evidence. I think that is happening. Um, Education & Future of Learning Sarah Guo: education seems like another one that’s an- Yep ... obvious good where we haven’t seen as much impact as I’d expect. Swyx: Do you have a hypothesis on why that might be, or if it’ll come? Yeah, I mean, I think this is where, again, how we think about education, how... You know, recently I met with, uh, the founders of Alpha School and learnt a lot about what they were going and going about, and it’s fascinating to listen, uh, to how to even rethink- Mm Satya Nadella: uh, what does education really look like. Because I think it’s actually very important. Mm. Uh, and I’m not saying anything traditionally being done is less important, right? I was even looking at the, uh... It’s fascinating to see. I, I, I forget the which Stanford class it was, uh, the, the Asian guidelines for CS something. Mm. Uh, because you still need people to learn. Uh, like it was an interesting AI class that they were making sure people were learning how to apply softmax appropriately versus saying, “Hey, fix my training run.” Mm-hmm. Uh, so I think learning concepts is important. It’s going to [00:38:00] be, uh, critical. But the way we create the incentives, what are the credentials, how we value those credentials, what is the employment opportunity for those credentials? So I think that there’s a complete change that has to happen, uh, given the way to get to information, way to educate yourself, way to continuously keep yourself updated has changed so much. So I think interestingly enough, maybe the next big startup and success story could be someone who builds a new university, um, or a new, um, pedagogy even of how to get someone to go through a curriculum and find economic opportunity, uh, that’s highly valuable. Well, that has felt, uh, perhaps impossible for a long time, but it’s a great note to end on and something that might be possible. It’s still possible. Yeah. Thank you, Satya. Thank you so much. Thank you. Yeah. I appreciate it. Thank you all. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe [https://www.latent.space/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

GitHub's plan for Agents — Kyle Daigle, GitHub

I’m excited to work with Microsoft once again as the presenting sponsors of the AI Engineer World’s Fair [https://www.ai.engineer/worldsfair/2026]! We’ll streaming live from MS Build [https://build.microsoft.com/] today for a special crossover pod with our friends at No Priors [https://x.com/saranormous/status/2061681787169017949?s=20] and the one and only Satya Nadella. However we did not hold back with this interview - we asked all the burning questions about uptime and Copilot that we know you have in your minds. Lets go! For almost two decades, GitHub has been the home of software, where both open source and closed flow, through commits, pull requests, reviews, actions, etc. This ecosystem flourished as open-source maintainers and contributors would continue shipping code for the benefit of the community. However as coding agents began to ship mass quantities of code - growing 1400% in 2026, it marked a new era that was both extremely exciting and challenging for GitHub. While these agents help more people ship more projects, they also significantly increase the floor of how much code is shipped, how often it is shipped, how many people commit code, and basically orders of magnitude multiples in every dimension of GitHub infrastructure: Now GitHub inevitably experiences more pressure on their infrastructure which was originally designed around human developers moving at human speed. This has resulted in a very publicly notable uptime story: So it begs the question of whether current systems around code can absorb what AI produces. Can CI/CD keep up when every idea becomes a build? Can open source maintainers survive floods of AI-generated slop contributions? Can GitHub preserve the human social contract of software while becoming the operating layer for agents? Which brings us to the perfect person to answer these questions: GitHub COO Kyle Daigle. In this episode, he joins swyx to unpack what happens when AI doesn’t just autocomplete code, but starts changing how companies operate, how open source works, how pull requests get reviewed, and how GitHub itself has to scale. We go deep on GitHub’s internal AI workflows: micro-skills, WorkIQ, MCP, Slack, Teams, email, Copilot workflows, the new Copilot desktop app, CLI, cloud agents, and how Kyle uses agents to look backwards across company context before deciding what to do next. Kyle also reflects on GitHub’s history building webhooks, APIs, Actions, npm, Dependabot, and Semmle, why the AI era is breaking GitHub in new ways, how Actions became a general-purpose compute layer, and what Copilot becomes after code completion. Full Video Pod We discuss: * Kyle’s expanded role across GitHub * How AI got Kyle coding again after years in leadership * Why GitHub rolls out AI through existing workflows instead of forcing new tools * WorkIQ, MCP, Slack, Teams, email, and GitHub as company context * Why massive “mega-skills” are giving way to small, atomic micro-skills * How AI changes summarization, communications, marketing, and analyst work * Why former developers in leadership may have a unique advantage in the AI era * Kyle’s “15 agents on Saturday” workflow * How Kyle built an AI-generated executive presentation for CRO/CFO teams * Why AI changes the chief of staff role without removing the human work * GitHub Actions, webhooks, arbitrary code execution, and secure agent compute * The npm acquisition, supply-chain security, 2FA, and token invalidation * Slop forks, vendoring, and whether AI agents change dependency management * What pull requests become when most PRs come from agents * Prompt requests, vouching, AI review, and trust in open source * What counts as a “developer” when AI lowers the barrier to building * GitHub Spark, low-code, and why GitHub refuses to hide the code * 14x commit growth, Actions load, databases, monorepos, and availability * Copilot’s evolution from completion to CLI, desktop app, cloud agents, and SDK * Context, memory, rules, and making GitHub “act like Kyle wants it to act” * Ambient AI, OpenClaw, enterprise security, and the new operating system for agents * What swyx should ask Satya Nadella about Microsoft’s AI future Kyle Daigle * LinkedIn: https://www.linkedin.com/in/kyledaigle [https://www.linkedin.com/in/kyledaigle] * X: https://x.com/kdaigle [https://x.com/kdaigle] Timestamps 00:00:00 Introduction 00:03:36 Why AI Got Kyle Coding Again 00:07:04 Running GitHub with AI: WorkIQ, MCP, Slack, Teams, and Skills 00:15:39 The Golden Age for Former Developers in Leadership 00:17:31 15 Agents on Saturday and AI-Generated Executive Work 00:20:20 How AI Changes the Chief of Staff Role 00:21:45 GitHub’s History: Actions, npm, Webhooks, and Open Source 00:28:45 Slop Forks, Vendoring, and AI Dependency Management 00:33:57 Pull Requests, Prompt Requests, and Trust in Agent-Generated Code 00:41:21 GitHub Stars, 200M+ Developers, and the New AI Builder Wave 00:45:15 GitHub Spark, Low-Code, and Why GitHub Still Shows the Code 00:47:38 GitHub’s Hardest Era: 14x Growth, Reliability, and Scale 00:59:21 Actions as the Compute Layer for CI/CD and Automation 01:02:04 The State and Future of GitHub Copilot 01:08:24 Ambient AI, Background Agents, and the Future of the SDLC 01:13:09 OpenClaw, Enterprise Security, and the New OS for Agents 01:18:03 Build Announcements, WorkIQ, FoundryIQ, and Microsoft Context 01:21:41 What Should swyx Ask Satya? Transcript Introduction: Kyle Daigle’s Expanded Role at GitHub and Microsoft Swyx [00:00:00]: We’re here with Kyle Daigle, COO of GitHub. Welcome. Kyle [00:00:07]: Hey, thanks for having me. Swyx [00:00:08]: You’re not just CEO of GitHub. People know you as that. You have a new role. Kyle [00:00:11]: So I have an expanded role now. I’ve been working at GitHub for thirteen years and doing all things developer. Joined as a developer myself. And now, I’m also responsible as the CMO of Developer for Microsoft. And so all the kind of learnings and passion for developers and how we work with them and how we communicate and how we bring our products to market, we’re also bringing that expertise to the broader Microsoft ecosystem and helping every developer that uses a Microsoft product or would like to have a sort of similar experience that they’ve had with GitHub over the years. So it’s a different role in some ways, but it’s also just building on the experience that I’ve had at GitHub of just sort of tell the truth, be authentic, show people how to use it and then let the products speak for themselves. Now just doing that with, all of Microsoft. Swyx [00:01:09]: We’ll be releasing this in conjunction with Build. You got lots of stuff planned, and we can sort of touch on that whenever it’s appropriate. I think one of the interesting things is I rarely meet a COO who’s also a CMO. I think you’re a very outward facing and you’re very confident publicly. That’s rare. Do you actually view yourself as COO? What’s What is your thing? From GitHub Developer to COO/CMO: Building the Platform and Operating GitHub Kyle [00:01:33]: I think for me, it’s been funny. The titles have always been, a— have always felt a little strange to me. I joined GitHub as a developer? I wrote so much of the Swyx [00:01:46]: Let’s bring that up. You wrote the back ends? Kyle [00:01:48]: I was going through, I was going through, some old photos, when folks were talking about how things were being built or how there was a build GitHub. I built, webhooks and worked with teams building the API, built the platform layer. Anything that integrated with GitHub, up until really twenty eighteen, I built or ran the engineering teams. And that’s kind of where my the beginning of my passion always was helping people build things, deliver them to, their customers. And so being a developer, building for developers was always super unique. In a— I think as my role expanded, it became my ability to talk to not just developers, but also enterprise customers or business leaders and have this translation layer. And then through all those years, GitHub has always operated pretty uniquely. Post-pandemic, working remotely was not as novel as it was when GitHub started in two thousand and eight. But all that expertise of running remote teams, doing it well, became this sort of bigger role, ultimately turning into the COO role of how do we operate GitHub in the way that GitHub’s always operated after the Microsoft acquisition. And kind of so on from there. So like for me, I think the— I’ve, I still code. I love coding but the problem has always been, people. It’s a much harder problem to both support our own employees, a harder problem to communicate to developers and enterprise buyers what we’re building why it matters, ‘cause those are two very different messages. And so getting to work in the mix of COO, CMO, also just being a dev, I think is what’s kept me at GitHub for so long. AI Workflows for Leadership: Commits, Retrospectives, and Context Swyx [00:03:40]: Apparently, you have— your commits have gone up. What’s this? What’s going on? Kyle [00:03:45]: Rui’s called me out pretty aggressively. So I think— as you can imagine, right, you can see my normal era of being a dev In the twenty thirteen, twenty fourteen era, and then moving into management, and then ultimately the COO role. I think what you see there is me, really getting back to coding thanks to AI. I— similar to, attaching problems between how to market and how to operate a business and how to code, I find, building agents and workflows that are connecting very disparate problems to be what’s driving this. So that’s, some of it’s writing software. A lot of it is, connecting a ton of a different data sources to, help me out. But that is completely me really diving in on the AI side in trying out our tools, trying out everyone’s tools, But building for me, building for the non-technical leader, though I’m technical and how we’re, able to use these tools more than just the simple, call and response that I think a lot of the non-technical, your employers, you have to get— you have to use AI, and so everyone uses, ChatGPT or Copilot or Claude or whatever. To really get into, how is this going to help me out, it— I find that it’s not the I need to write a blog post, I need to those simple examples. Helping people find the workflows of, “Okay, I need you to go through all the PRs today. I need you to go through everything that we’ve posted online. I need you to go through what we did the last three months. Go through all of my Obsidian notes for any mentions of this then go through my transcripts at work.” We use, Teams, so, using WorkIQ, go call that MCP server, grab all the transcripts, go through all the Slack, and then build me out the plan of, what this week’s messaging actually was. That’s something that was, impossible because for me, I find AI in a what most of this launch here is actually, less building forward. It’s actually, a recursive loop backwards. I’m always looking at what had happened first. Go back through the week and tell me what we did, what worked, what didn’t work? And then tell me in the next three or four days-What would you tweak based on this sort of like looking backwards and then looking ahead a little bit? I find that to be so much more valuable, especially for like non-technical, because that retrospection is actually LLMs are very good at that. Like finding all the patterns, pulling them out, and then applying that retrospection to just a couple of days or just like a short period of time. Is all a bunch of apps that I’ve built and launched a bunch of, internal tools. I use the new, GitHub Copilot app, the desktop app with workflows. Every time I crack open my laptop, it’s running workflows for me. It’s just a ton of different stuff and of course, it all ends up on, it all ends up on GitHub. Swyx [00:06:47]: Of course. That’s where, that’s where, stuff is hosted. Man, there’s so much to ask you. I was going to leave the how do you run a company with AI thing at the end. I have to ask one— double click one thing. You said, you are looking back at the week. You’re, you’re understanding what happens. When you say we That’s three thousand people. How? Rolling Out AI Internally: Skills, CLIs, and Company Context Kyle [00:07:09]: I think when we started rolling out AI internally beyond engineering, right? One of the things that I was really, passionate about is like we have to do this in a way where no one has to change how they work. I don’t want to have to teach you a tool. I don’t want to have to teach you something new. And so for us, we tried out a few tools. Most of them don’t work because I got to get you on board? I got to teach you how to use it. What we’ve actually ended up doing is we’ve built like a set of skills internally. We have we each have our set of skills, and we’ve just been distributing even to the non-technical folks, the CLI. And then effectively, we’re just giving it access to like read about everything that we’re writing. So that’s for us, that’s usually GitHub, Teams, Email, and Slack. So Teams for, video chat, generally speaking. Swyx [00:08:03]: Teams and Slack? Kyle [00:08:04]: so we use Teams for video communication, but we don’t use it for chat. W-we— GitHub for a long history, right? We’re always Swyx [00:08:13]: Also Slack Kyle [00:08:14]: Talking about ChatOps and like everything is built into Slack. Like every command, every flow. Swyx [00:08:18]: So even though you have been acquired for I don’t know, eight years now Kyle [00:08:22]: we still Swyx [00:08:23]: You still use Slack? Kyle [00:08:23]: it’s a purpose-built tool for us, and I think the reality is that moving off of it would be so bluntly expensive? Simply because all the tooling is, baked in with that paradigm. And they both have their pros and cons but they don’t work the same way at all. We still use a bunch of different tools Because it’s the purpose-built tools that We need. And then Swyx [00:08:47]: Well, the same doesn’t go for the rest of Microsoft, presumably. Kyle [00:08:50]: like the like various teams like operate Swyx [00:08:53]: They make their own decisions Kyle [00:08:54]: Various ways. I think it just matters what you’re trying to what you’re trying to do. But we do we do work across kind of every tool that we use, and then by giving everyone access to all of that context and the new WorkIQ MCP server, which is quite cool if you do live in the M365 like world. I can ask it all these backwards-facing questions, and it’s incredibly important for our teams that are working remotely. There’s a lot of stuff you miss when you’re not in an office, and we are spread out all over the world. So most of that is looking back. And then we post, we post either auto-automatically into GitHub issues or discussions, these sorts of like findings or like our industry reports. Like what’s happening this morning, today, yesterday. A little automation gets run. We’ll use the app. We might use GitHub Actions like with, our agentic workflows just to go do that run, and then we push it into GitHub, and w-we keep having a conversation. So usually for us, it’s about that sort of like looking back, looking forward on the non-technical side. And then of course for a lot of those folks, it’s also building an app, pushing it to GitHub pages or pushing it somewhere to host it et cetera. But it’s just like enabling everyone with that power of it’s going to take me a week to figure this out. Instead, we’re going “Okay I built a skill. Let’s put it into a repo. We’ll all share that skill together, and then we’ll use the CLI or now the app-” “just to run it.” Micro Skills vs. Mega Skills: How GitHub Uses AI at Work Swyx [00:10:26]: All right. I think, I think we’re going straight into like the team management and productivity thing. I think a lot of people are getting various levels of LLM psychosis. How do you manage the bloat of skills? Like everyone Has their thing, and they’re Like trying to promote it to the rest of their peers in their org, right? And obviously, whoever becomes a skill influencer internally becomes like an AI leader, right? Of sorts. I assume you have those. Kyle [00:10:50]: like I think we have Swyx [00:10:52]: And I assume it’s a mess a Yeah. Kyle [00:10:54]: there’s like I— like I think the reality is there’s two pieces. Like first is I think that we’re ending the era of these like massive, beautiful, perfect skills that are just like not any of those things. ‘cause for a while, right every tweet every day is like go download the skills, the perfectly managed thing to do this entire workflow. And I think that like what we’ve found and what— I was just with my team, this week, and we were talking about the skill side, and we’re really talking about these like incredibly micro skills that are just doing one thing for us very well Versus a skill that’s going to do I said, that full report. That doesn’t really exist on our side anymore. It’s usually how do— like a single skill that’s going to identify the most important marketing information given any MCP server. Like this is the most important thing. Less about stitch a bunch of tools together and have it produce this mega output because then weeks go by, months go by, things change, and you want to tweak Swyx [00:11:58]: It’s brittle Kyle [00:11:58]: Your mega skill and you’re screwed? You can’t do that. And so now we’re really just talking about the Legos we’re using and just letting the instruction book be something we’re all putting together. Whereas I think a lot of AI skills for a while have been that mega instruction book style. Swyx [00:12:15]: I’ve, thought a lot about Postel’s law. I don’t know if that’s a term that is, means things to folks. It’s the idea that you should be liberal in what you accept and strict in what you output, right? And I think that’s like a good framing principle for skills. This is my skills, obviously on GitHub. I feel like everyone should have like how like some repos In GitHub are special repos? I feel like we should sort of reify the slash skills and everyone like give it some kind of special presentation. Anyway, so, yeah, this is one of those like download Download anything, transcribe anything, and then you can string together the atomic skills that do one thing well Into like some kind of orchestration skill that calls other skills. I assume, does that match? Kyle [00:12:56]: I like I think so. I think that the Swyx [00:13:00]: Summarize anything. Kyle [00:13:01]: Like I think the- For me, summarizing something for I do communications and PR and analyst relations and marketing and customer activities, and so my summarize everything is very different for each one of those like Contexts. What ‘Cause if I’m summarizing something for an analyst, that’s a very different thing than, probably how I’m going to summarize something for like a customer meeting or an engagement. So that’s I think like the difference when we’re talking about the like the tools I might use on Saturday or the skills I might use on a Saturday when it’s just for Kyle. Yeah, those are kind of like they have an atomic actual tool underneath or maybe skill, and then Kyle cares about X. But I think when we’re talking about work and enabling the the marketers, communicators there, it’s the atomic, this is what good summarization is, and then this is what I care about as for marketing for communications For whatever. And that I think is like the interesting matrix problem when we go from like a developer set of concerns to all kinds of different professions, is that what that word means to me is different than it means to you is different than it means to the analyst or the salesperson, and that’s where I think the matrix mess is that we’re starting to like still starting to find. It’s about these mega skills but they’re all just slight permutations, but those permutations are really important. It’s the difference between someone reading this and going “Did AI make this?” what Or “This makes total sense, and I would expect this when I’m giving a briefing to Gartner,” or like whatever else. Swyx [00:14:37]: I think the beauty of it maybe is that you don’t have to be that careful about what goes in there. It doesn’t have to exactly fit as long as it like roughly is contained in there. I used to complain about plugin hell, basically. Like when you have a framework and then you have a hundred things that you need to integrate, everyone does like the GitHub used to be bloated full of these things. And now we don’t need them anymore ‘cause now you just use skills. Former Developers in Leadership: AI as a Creation Multiplier Kyle [00:15:00]: And like I think the most magical thing is the just that like I can just also crack it open. Like Like yes, I could go like change the how the plugin is coded, or like I could go do that now with AI, but I think there’s just something more magical about getting a response back and being “That’s not right,” and then you just crack the skill open, you just type English words and it’s different. That building block is just, I think very unique. Once I get everyone to kind of understand how to best how to best make those changes to get the most power out of them. Swyx [00:15:36]: Is there a— you have a your peer group that Of people like you. Is there a common framing for Something I’m feeling is, which is true, is that is this a golden age for former developers who are now in leadership? Because you can wield the tools, you would know the right words, you’re maybe not too close to the details. Doesn’t matter. But like you’re more effective than someone who doesn’t come from that background. Kyle [00:15:59]: I think that like the secret has always been your ability to identify patterns and solve problems, and I think that for folks that like myself that don’t code day to day anymore, that has made me successful as a developer, made me successful as a COO and now CMO. And so now that I have access to get and write code, I’m now applying that sort of like pattern finding and problem solving, and I know enough still about how to then go and say, “Oh, I want to make an app, but I don’t want to break into jail or create something that’s not going to be able to work or to be deployed scale or whatever.” that ability to apply all that additional business knowledge and still code I think is what makes that so interesting to me. Slightly different than I think some of the other like technical leaders that became business leaders and now are going back to their apps and updating them. Good for them? But I think the more, much more interesting thing is, well, now I have this whole new set of expertise over ten plus years. Why not take that and use that as a developer with these AI tools? So I definitely think that makes me more powerful, but I think that’s true for like every dev as well. Most of the dev friends I still have also have some other underlying skill and passion. There’s really talented, very kind of linear computer science software devs, absolutely. I just find that the folks that came from a different career, went to school for something else, went off and did this random thing, and then became a software dev, or were a dev, did a random thing, came back. Learning that extra set of information, learning those extra skills, and now having the power of an AI where I can crank up fifteen agents on Saturday while my kids are doing lacrosse, That’s like really powerful. And I think it gets me back to that feeling of like creation, and it’s very hard to replicate that in most other senses? That first time you build an app and you click it and you show someone that’s magical. And so being able to do that not just in code, but across all kinds of different assets that’s, that’s huge. We were doing we’re doing our every year we do our revenue planning. We talk about okay, what is it going to look like for next year? And of course as you imagine, there’s, slideshows everywhere talking about what are we going to talk about, what’s the narrative, et cetera. And so as you said I’m “Okay, well, I could probably just like build something to build this and then that way I don’t have to go build the whole spreadsheet or I have to pass it to my team.” So we went through this process, and I got all the information and used the skills I mentioned. I built like a little app just to make it so I could look at some of the information in a SQLite database, more easily. And I ultimately built this entire presentation without touching any of it and I was “Okay, I’m just going to present this to our CRO, the CFO, their teams,” without mentioning I’d built it with AI. I like built a skill to make it look very much not AI driven. Just not pretty. AI-Generated Presentations, Human Taste, and the Changing Chief of Staff Role Swyx [00:19:03]: Like a design. Yeah. Kyle [00:19:03]: Not pretty. But just like very clearly not AI. Kind of like don’t do anything interesting. Swyx [00:19:08]: That’s, yeah, that is valuable. Kyle [00:19:08]: Just go Exactly. We did the whole thing through. It used my notes from Obsidian, it used all the context I mentioned before, the plans, and Never came up once that it was AI generated. Swyx [00:19:20]: It didn’t matter. Kyle [00:19:20]: Never once. D It didn’t matter. And so now I take Swyx [00:19:23]: This is a tool Kyle [00:19:23]: I can take that tool and go, “Look, I don’t want you to go build slideshows.” They’re just helping us share information with each other. If this thing can do it With a little bit of crafting from you and then we can look at it together, awesome. There’s no value in all that extra work. I think that the ability to, make it look humanly bad and and build a little app to, manipulate the data I think is part of, that upside for devs that are now in leadership roles. Because, the thing that I feel like I said before, this that’s all a people, that’s all a people problem. I know if you’ve used a coworker or not to build a slide deck, unless you spent a bunch of time to not do it. Swyx [00:20:07]: I know, but like it was so, I think there’s a certain charm to just being blatantly AI. ‘Cause I think that you’re well, you’re just honest about There may be mistakes here that I cannot vouch for. So how much value is there? But anyway I think, actually the real question I want to ask is, there’s a— You were a chief of staff To Thomas. And in the pre-AI world, the that job would’ve been a chief of staff job of like Can you prep me these slides and all that? And now you do it yourself. Kyle [00:20:35]: I still, I still have a chief of staff. Because, the difference is it’s sort of the discussion every time we have some sort of technology evolution is it’s not that the jobs the roles don’t all go away, they just change? And so yeah, I don’t have someone spending all their time building out slides for me and presentations ‘cause I don’t need that anymore. But now I need that person that is able to go and find all the different connections between humans in those discussions to help me find out, okay, I should be meeting with this group and this team, and they have an opportunity, and I’m going to be in San Francisco today, I’m going to be in Seattle tomorrow. Those sorts of human connection aspects are still incredibly valuable and has always been a big part of that chief of staff role. But now just like chiefs of staff are not opening up, letters to process, they’re doing emails. What It’s the same thing. And now they’re, they’re not building out as many of these presentations because they have the the ability to have a AI take it on for, and share that with me and great. Let’s keep moving ‘cause it’s allowing us to go faster and make better decisions more quickly. Swyx [00:21:45]: Awesome. Well, so we can dive into more sort of, Productivity insights as you go. I did want to do a little bit of a brief history of colleague and hub. Because, we started here. And then you also involved the NPM acquisition. I did, I do want to touch upon that. And then more recently, I just want to bring up to present day where we’re having uptime issues Which transparently we’ve already Addressed publicly, but we’ll, we’ll discuss in the pod. Did I miss anything? Like what, any other major highlights? Obviously, it’s, it’s a lot of years to cover. A Brief History of GitHub: Webhooks, Actions, Acquisitions, and Platform Evolution Kyle [00:22:15]: No the I think one of one highlight was right before the acquisition closed in twenty eighteen, I got to launch the first version of Actions Swyx [00:22:27]: Oh Kyle [00:22:27]: At GitHub Universe. So it was O Swyx [00:22:29]: They’re that young? Kyle [00:22:30]: It was October of twenty eighteen, I think. Yeah. Yeah. Swyx [00:22:33]: Gee, Jesus. Kyle [00:22:34]: I got to I was the engineering leader on that project and got to launch that. And then, yeah, we did acquisitions of NPM you said, Semmle, Dependabot Pul Panda a whole bunch of things. That was a big Swyx [00:22:47]: Pul Panda. Kyle [00:22:48]: Abi is doing well. Swyx [00:22:51]: DX. Holy crap. Kyle [00:22:52]: Did well on DX. I and like that was a that was the big shift, after the acquisition. I had to join the sort of business side. Swyx [00:23:00]: So I need to hit you on some of these things ‘cause you were there. Right? And how often do I get to talk to someone who was there? But yeah, Actions. Is that the number one source of security issues on GitHub? Kyle [00:23:11]: Oh, sh I think that the number one source of, security issues is probably like all, the literal code in everyone’s like underlying repositories. I would say back further than that is, if you remember I had to show in this graph was this is, I’m, didn’t say this before, this is ultimately webhooks. Swyx [00:23:30]: You yeah. Kyle [00:23:31]: Like circa whatever it was. Swyx [00:23:32]: It says Hookshot in there. Kyle [00:23:32]: I forget. Yeah. Yeah, Hookshot’s in there. And so like back then, it says GitHub Services. Do you see, it says Hookshot FE for front end, and then it says GitHub Services. GitHub Services back in the old days, right? You we had a repository that was Ruby code, and you could write any Ruby code in there, and then we would execute that On your behalf As a service, and then that way if an if you were trying to integrate with something, it didn’t we would run it for you. Swyx [00:23:57]: And of course no containers ‘cause Kyle [00:23:58]: No, ‘cause it was Swyx [00:23:59]: Well, no containers Kyle [00:24:00]: Twenty fourteen. And so there was some isolation obviously, but it was mostly the separations on the server level. That’s like an example as long as the very old version of Pages, which ran on its own containerization infrastructure, not on Actions. Swyx [00:24:15]: Which like all-time great product. Kyle [00:24:16]: Pages powers the internet at this point to some degree. Those were places where like clearly there were no like issues like to my knowledge. But it was those things where I’m looking at and going “Okay, well we can’t be running arbitrary Ruby code,” like on everyone’s behalf. Then containerizing all of that up intoUh into actions now where yeah the containerization, is r-really good. The pinning most folks aren’t pinning it the like to a particular Swyx [00:24:48]: Images Kyle [00:24:48]: Sha, et cetera like their workflows, and so that’s a big that’s a big place Of pain for folks if they’re just doing similar to any dependency management, just V1 or newest or latest, I think. But, that journey from that day to “Okay, we’re just going to run all this arbitrary code, and, it’ll basically be okay,” to now, no, we have, really good containerization. We have a new, underlying, ag-agent, containerization, service. It’s like we’re using it under the hood. It’s through Azure. They recently announced it. The Azure, Dev Compute, but it’s, very fast, very fast compute to be able to, spin up your own cloud agents, or whatnot. We’re using it under the hood for some parts of the new, Swyx [00:25:36]: Microsoft Dev Box? Kyle [00:25:37]: No. Dev Compute, yeah. Swyx [00:25:41]: Hmm. Not finding it just yet. Kyle [00:25:44]: Oh, it’s, it’s in there somewhere. Swyx [00:25:46]: All right. Well, we’ll cut that out. Kyle [00:25:47]: Sorry. But with, Dev Compute, you can, run, really fast, spin up really, small VMs really quickly, so you’re doing a tool call Swyx [00:25:58]: Same concept Kyle [00:25:58]: Just do it containerize exact-exactly. So we’re using that so definitely moving that direction to protect us from every every piece of code that we’re ultimately running. Swyx [00:26:07]: look, that grows into the full SDLC? Code hosting was just the start and and then it’s grown beyond that. Let’s talk about NPM may-maybe ‘cause I think that’s also, a very major point in the industry. I do think, it was looking for a home. It was, kind of struggling as a business, right? I don’t know, I don’t know how you would characterize that whole acquisition and how it NPM, Package Security, and Keeping the Internet Running Kyle [00:26:33]: like when we were talking to the team, I think the big thing for the both of us was to find a way to keep NPM, which was basically powering the internet then and way more so now to some degree running. Keep it going keep continuing to scale. It was having scaling problems, if I recall, back at that time. They were doing some rewrites. It Swyx [00:27:00]: that’s cute compared to now. Kyle [00:27:01]: Well, that’s the thing is like when I’m talking to folks now, there’s there’s so many more underlying uses of NPM than there were back when we had them join in with GitHub. But that was ultimately the goal. It was really okay, we used to have pages. We have, the world’s code. Let’s make sure that we can keep NPM running well for the world. And we put a bunch of time and investment into fixing some of the underlying backend, changes, some of which we talked about some of the manifest work, et cetera. And then now, really trying to bring the the security posture of NPM up to speed. But, it is a unique challenge in that every move that we make to make it more secure will break a lot of people. And security is paramount. And also, we take it very seriously. We’re, the any time that we have a problem with GitHub or we make a change that makes us more secure but hurts, there’s, a snow day for developers or a really bad fire that they have to go put out. And so we’ve, have changed the 2FA policies. We’ve changed the way the tokens work. When we find tokens that have been exposed or potentially, exposed, we invalidate them, and Swyx [00:28:22]: I love that feature in GitHub. Yeah, it’s great Kyle [00:28:23]: That creates issues, but, the but that’s the thing is we’re trying to push the community, forward without necessarily, doing something that is going to break the contract that’s been for 15 years or close to it or some amount of years on NPM. Slop Forks, Vendoring, and the Future of Open Source Supply Chains Swyx [00:28:43]: I think the— So now we’re talking about, open source and publishing. And I think there’s something here with what people are calling slop forks, which, I think Malta from Vercel is doing. And, part of me thinks, well, the way to get past any vulnerabilities, we just, let’s just get rid of the concept of NPM. And we only publish source code. And anytime you want to import it you have your coding agent look at it and then adapt whatever subset you’re going to use into your vendor it. But, the AI vendor it. Is that realistic? I don’t know. Is it— Will that solve all our security issues? I don’t know. Kyle [00:29:24]: I don’t think it’ll solve I so Mitchell was just talking Mitchell Hashimoto Was just talking about this today, and I think that I-in some ways, it’s all all things, old or new again? Yeah, absolutely vendoring everything. Like I do I do remember twenty thirteen, twenty fourteen. Swyx [00:29:42]: This is Yeah. Let’s, we must return to Kyle [00:29:43]: That’s what is We were vendoring everything. We were having actual discussions around, or at least I remember we were “Should we take this full thing?” “Why is this so big? We only need this one file.” And so I do think there’s something true there where having either taking only what you need or the dependencies just getting incredibly small over time, I think will help to some degree, but it’s not going to solve the fundamental problem, I don’t think, because the vulnerabilities in an agent looking at them, there’s time and time again, there’s a million different ways in which we can convince an agent that this thing is, secure or not and pull it in. Or we can do static code analysis or runtime testing to say whether the code works or not. That is, I think, the step that needs to continue to be, invested in. The question is just on, how much scope. Should it be this enormous project that I’m pulling down, or should it be this piece? Either most companies are running some amount of security checking on the on the packages that they’re bringing in or vendoring. That I think won’t change. That’s like what advanced security does to some degree, Socket does some degree. Like everyone is doing a piece of that. How we each do that like especially when we’re talking to enterprise customers, is just like very different. No there’s no one wants one single way to do it. And I think that’s always been GitHub’s, unique position in the world. I talk a lot to maintainers, I talk a lot to folks about this. It’s we’re— we rarely start like a process and a practice and like push it onto the community. We usually wait for the sort of like RFC process socially or literally, everyone agreeing, and then we’ll cement something in. Because otherwise we’re Maintainers, RFCs, Vouching, and the Social Layer of Trust Swyx [00:31:35]: That fits your role in the ecosystem, yeah Kyle [00:31:36]: We’re GitHub. Yeah, we don’t want to shape the whole thing. We want it to be figured out. But like how do you balance that like sort of Role in the industry to keep everything as secure as is possible and make sure that you’re you’re not going to be compromised as a human, ‘cause that’s usually how it all happens. And Not not create a process or lock us into a flow that you’re not going to or like Mitchell’s not going to or other open source projects aren’t going to like. That’s always been a tricky balance for us, and I think that’s something that we haven’t talked about enough is we’re not going to be able to fix everything for everyone in a way that everyone is going to like. So tell, help us, tell us what is working. When Mitchell was talking about, the Upvote, the up Swyx [00:32:22]: I was going to bring up his thing. Yeah. Kyle [00:32:23]: I forget what it Yeah. When he’s talking to us, I was chatting with him and talking to him about this and I put it on Twitter and we talked to, also over DM, was “We’re going to keep working.” but I think the important thing is I do actually want to hear what isn’t working for you. And as, be as specific and clear for your project as is possible. And to every piece of credit over the many years that we’ve known each other through the industry, he’s always done that and I appreciate that ‘cause there are places that we need to fix up, and we hear from him, and we’ll fix up just like we do all other kinds of maintainers. But that that process between making those types of improvements and being more secure and like creating, I forget what he calls it’s not the proof process, not the claims process. Do what I’m talking about? He has that he his projects have a way for you to kind of like, Swyx [00:33:13]: Vouch Kyle [00:33:13]: Vouch. Thank you. Yeah. He has like the vouch system for saying, “Hey, you should accept my PRs.” That’s been Swyx [00:33:20]: I just built this into GitHub. I don’t know. Kyle [00:33:22]: Well, see, but that’s the thing is that you say that and like he and his community really likes this and then I’ll go talk to other maintainers and other maintainers, globally, and they’re “No, this doesn’t work for me.” And that is the tension, but also the kind of beauty of GitHub, depending on which way you look at it is we want to help maintainers, so we create all these tools to let you have more control over how much you take in from AI and PRs. But you can also use this. What You can go use this project, and if it takes off and becomes the kind of mostly standard, then yeah, we probably wouldn’t enforce it but we would add it in because that’s the flow that we tend to do? Swyx [00:34:02]: I hear a lot of people don’t know the history of the pull request. And like like that’s how, that’s something that GitHub standardized basically. Kyle [00:34:08]: Yeah. It was a very messy process Like beforehand, and now the we have the benefit of it being the process? And now we have to go and Figure out the next best process or what adaptations change, or what does a pull request look like when eighty percent of your PRs are just coming from your agents and not From other devs? Swyx [00:34:31]: Do you like the prompt request idea from Peter? Kyle [00:34:34]: like I think that for each like each idea I think has its merits. I’m not, I’m not avoiding saying anything good or bad, but I feel like I’ve seen a version of we have that we have entire Thomas’ store. Take all the assets of what you’ve built and put that in. I think that’s got great ideas. There’s all these various permutations of the PR flow, but I think the reason why there’s not a single answer is ultimately we’re trying to codify trust. We’re trying to say “Okay, if Sean reviews this I’m going to trust it because you’re Sean or you’re the senior dev or you’re the whatever.” And right now, when we are working in a flow where an agent writes code and another agent reviews code and then Kyle goes and looks at it the trust is kind of diffuse. And most of the tools that we’re talking about are talking more about verification flows. We have more assets to look at, so I can probably say whether this is a good PR or not. But that still doesn’t solve, I think, the human problem of I’m looking at a PR and I want to know if I can trust it. And we’re still, we still tend to use human signals for that? Mitchell approving it or Kyle approving it or whatever. And so I think that’s, I think that’s why most of these options haven’t really solved it is because, it’s a social problem ultimately. It’s a it’s a human problem to review it and agree. Or you fully trust the tool and you’re imbuing that tool with full trust Which I think in some cases that absolutely exists. AI-Generated PRs, Trust, and the Waymo Analogy Swyx [00:36:08]: And so like in the same way that there will be a tipping point in society when we don’t allow humans to drive anymore Because machines are measurably better than Than humans. I’m looking for that tipping point, right? Like Mythos is ridiculously expensive. Someday we’ll have Mythos on a desktop. I don’t know. Will, does that change the equation? Kyle [00:36:30]: I think it’s more I took a Waymo here, and I was on my phone and not looking around at all. There are other, self-driving, vehicles that I would not trust while, staring at the road. And I think that trust is something that is Swyx [00:36:48]: Is this a Zoox thing? What is it Kyle [00:36:50]: I think that is both. I think that is both. Like Swyx [00:36:53]: There’s Zoox in this robo taxi. That’s it. It’s Kyle [00:36:56]: Well, depending on what level Of self-driving. But, my point is sort of that I think part of that is I strongly believe that’s, a mixture of verifiable proof. Like how many accidents, how much data, and so on, and the human aspect of how I feel when I’m in this car, what it tells me, et cetera. And so that’s why I think some of the like Some of these some of our AI tools tend to, imbue me with more of that feeling of trust, even if the data says this is 100% accurate. I feel like it takes more time for us to go, “Should I trust this or not?” And that’s in the soft sense of, startups with high agency, weekend projects, and open source. And then there’s enterprises and regulated industries and everything else, and that is an even harder problem to go solve because even when it is fully verified, not only do you have to have trust from the humans on the team, you probably have to have trust from multinational, Swyx [00:37:55]: Oh my God Kyle [00:37:55]: Multi governments around the world and regulating agencies. And so that’s where I feel like until we tip over to your point on the sort of like human EQ side of it. I feel okay this feels okay I’ve been proven enough. Then the ball will start to roll a lot faster, where we’ll end up getting to the “Okay, we can trust this,” and feel good about it in the Most difficult of cases. Reputation, Sponsors, Stars, and Bot Activity on GitHub Swyx [00:38:18]: If human trust is the thing that matters, I feel like GitHub as the developer social network could maybe do more there. Like vouchers are one system But, we have star counts, and then we have Contributor rights, and that’s it. And I feel like there should be more in that space. I don’t know if there’s any other design decisions there. Kyle [00:38:37]: I think that one of the places that we don’t really expose right now in this sort of way is, some degree of like hard trust and support, which would like for me is like sponsors is a good example of that. Swyx [00:38:49]: Ah. Kyle [00:38:49]: It like costs you something. To prove that I believe in your project and I trust you To some degree or I want to support you at the very least. Swyx [00:38:56]: Solve payments for open source. Why not? Kyle [00:38:58]: I think that I think that like as we keep moving forward, right, there’s more and more projects where I’m, adding more and more dollars into sponsors personally because I want to like support them, but I also like know of I’ve probably never met them in person, but, I know of enough of their work that I want to support them. I think the thing that I don’t love about stars or commit counts or anything else is ultimately, even with all of the various, abuse and de-spamming and deduplication work that we do or anti-abuse work that we do, these are all, not active social signals. They’re passive ones that are ultimately gamifiable. And you may trust me, but another open source maintainer may not. And on what heuristic should you be, trusting me? That I think, is kind of where some of our thinking is right now. What signal from me is most important to you? You— If you can define that potentially, honestly in an agentic workflow that’s what we see some of these open source projects do, where you have GitHub actions, and then you have like an agentic workflow that’s calling AI, and you’re setting these rules. Like if Kyle has submitted and gotten accepted PRs across any given project and has a social handle tied to his account in GitHub, and that social account’s older than a certain amount. Really complex measures that matter to you ‘cause most open source projects have that heuristic built into their heads, if not written down in the contributing guidelines. You could take that and then go apply that and then just say, “Oh, we’re not going to accept this PR.” Building something that is, I think, malleable to everyone’s needs, is a little bit better, rather than going “Hmm, this account’s too young.” Because what happens? The attackers just go and go and create a multitude of accounts, and they wait Until it ages up. Needs to have a certain amount of stars. That’s how star inflation happens. Need to have a certain amount of repos Swyx [00:40:46]: Oh my God. Yeah Kyle [00:40:47]: With PRs. They all just create repos and submit PRs to each other, and then they come in and do something nefarious. And so, it’s hard. It’s hard to find the measure. So I think we’re, we’re looking more at how can we provide you tools so you can kind of choose what’s best for you. And of course, we’ll give you some standards. But the trust vector, gets down to I don’t know, some version of like human digital ID like everyone’s been talking about. Like how do I prove that it’s me Swyx [00:41:13]: Give me your eyeballs Kyle [00:41:14]: On the internet. Give me your eyeballs. Exactly. Swyx [00:41:18]: The I got to keep moving on Topics, but obviously I can go all day on this stuff because, I’ve been involved in GitHub and open source My entire professional career. Stars. Very superficial. Everyone knows it. But I think time to one hundred thousand stars is the fastest I’ve ever seen. Like people just reached that in I don’t know, months. And then like at the same time I don’t trust it right? Like how many of these are real or bot or like whatever. I don’t know how to ask this but like what can we do about it? Like Kyle [00:41:49]: Just Swyx [00:41:49]: Is stars broken? Is stars fine? Kyle [00:41:51]: I think that there’s kind of two, there’s like two pieces. Obviously we’re constantly like trying to find ways in which like your users are producing spam, which would, I would include like be like only doing star gamification. When we find them, we pluck ‘em out and we, Swyx [00:42:08]: But it’s like a Whac-A-Mole Kyle [00:42:10]: It’s a hundred percent like a Whac-A-Mole Swyx [00:42:11]: There’s no way Kyle [00:42:11]: Now, powered by AI to be helpful. But I think more so what I’m seeing is, a lot of the like fastest time to X tends to be because we’re now inviting so many more people into like software development on GitHub That like the zeitgeist is just swarming? And it’s Swyx [00:42:32]: It’s not just developers anymore Kyle [00:42:33]: And it’s not you and I. Like like however you want to say like what a developer is it’s not just folks who have been coding for a very long time. It’s folks that have maybe started coding or only joined in since the AI era. And now Swyx [00:42:44]: what’s the latest Octoverse number? I know eighty million was my lastRem- member that a number of developers on GitHub Kyle [00:42:50]: Oh, we’re over 200 million now. Swyx [00:42:53]: Okay. Well, so you see? Kyle [00:42:55]: Like over 200 million developers now. Swyx [00:42:56]: But it’s not developers, right? It’s, it’s people with a GitHub account. What Counts as a Developer in the AI Era? Kyle [00:43:00]: So, so this is, this is the biggest debate that I would say, everyone loves to have at GitHub at this point. From my perspective, right, I think that there’s, there’s clearly a difference between, professional enterprise developer and then developers. But I think that I think that the idea that we should be I don’t know, splitting hairs or segmenting developers in the early era of software development is, not worth our not worth the time. So Swyx [00:43:29]: When you get into gatekeeping Kyle [00:43:31]: 100% Swyx [00:43:31]: What is a developer? Kyle [00:43:31]: 100%. ‘Cause I wasn’t a developer when I started writing code? I was going to Swyx [00:43:36]: Oh, no. I made— I cloned a thing, seven years before I learned to code. And then I and then I wrote about my learning to code journey, and people Just called me a fraud ‘cause I had a GitHub account. And I’m “Well, no, I just use GitHub, but I don’t know-” “I didn’t know what I was doing.” Kyle [00:43:49]: I I remember that. I remember those sets of posts, and like that’s, that’s b******t. So I fight very clearly on the line of, if you create code, if you have an idea and you create it into some way of, I’m, I’m going to run it and use the app right now, you may still use AI in that moment, but that’s okay. At some point you’re going to do the next thing. You’re going to create a big— You’re going to have to learn about this database. You’re going to fix a bug, whatever. We’re all on some same journey, and those people are also hearing about the great new agent skill package or a new CLI tool or a new whatever. And those projects are going up because you want to be a part of this moment, just like I wanted to be a part of the Ruby community when Ruby was popping off when I started becoming a developer, and now I can just click the star button. And so I think that yes, there’s clearly some amount of like spamming and game gamification that we’re working against, but I really think we’re just seeing this whole new cohort of folks that are moving from technology to technology because they’re not working on a 20-year-old software application. They’re working on a side app that they built on the weekend for their friends or for their new idea or whatever. And that’s how you see these enormous charts going up and to the right with With stars. Swyx [00:44:59]: I think something that’s remarkable is the persistence or, that GitHub extends to those folks. Usually when I see platforms go into a new audience, they usually have to, have like a second platform with a different name that wraps the main platform. But somehow GitHub has been able to sort of persist and extend, and it’s friendly and whatever? So it’s, it’s nice. Spark, Low-Code, and Always Showing the Code Kyle [00:45:19]: I that’s partially why I think as we’ve tried to move into I don’t know, more like low-code-y things. We so we started working on Spark as like a way to, build an app and run it. I think that the reality is that we anytime we try to, kind of put even a veneer on top of it without when we put a veneer on top of something, we still always show you the code. That’s kind of like a tenant. We’re never going to, hide the code from you ever, because what Swyx [00:45:52]: Why would you? Kyle [00:45:52]: That’s, yeah, that’s the whole point? However, I think that what we learned with things like Spark is that really the value of Spark for most devs is, easy runtime. And you may have a runtime or a host that you’re going to use for that or you just build something and run it but, the package of making that even more simple isn’t really needed for folks that are trying to build software and not just trying to build, an app, which is, slightly different, a slightly different goal. So I want to get you in, I want to get you comfortable. I think the best thing for me as, someone that did not traditionally come into software dev way back, I want anyone to be able to breach that chasm and not be in the I don’t know, I feel like we’re, we’re still in an era of, STEM. I’ve got a 12-year-old and an eight-year-old, and it’s “We got to get ‘em into STEM,”? Over and over. And I like I do, I do the things that good parents do. I was “Oh, you want to do coding?” “Yes, I want to do coding.” Do coding classes. But now they’re just not afraid of doing software. And that’s, I think, the thing that’s honestly kept me at GitHub for so long. Anyone should be able to go and build a thing, just like I can go change a light switch in my house. I’m not going to go into the breaker box ‘cause I’ll probably kill myself? But, I can go change that light switch. Everyone should be able to go and say, “This fricking app doesn’t do what I want. I want it to work like this.” And that I think, is what’s kind of kept us all connected with GitHub through the years and some and during the easiest of times or in the hard times because of that opportunity of, we’re the home for all developers, and we want everyone to be able to have that feeling that we’ve had of, had an idea, I created it and holy s**t here it is. Swyx [00:47:37]: Here it is. All right, I’m going to try to do more spicy questions. GitHub’s Hardest Scaling Moment: Growth, Agents, and Uptime Kyle [00:47:42]: Great. Swyx [00:47:42]: Is it an easy time now or a hard time? Kyle [00:47:45]: Oh at GitHub? It’s a hard time. Like, it’s a hard time and also, I was just with my team and I said, “This is also, the best and most exciting time that I think I can remember at GitHub.” Because Swyx [00:47:57]: Best of times, worst of times. It’s never one Kyle [00:47:59]: ‘cause we’ve we were talking about Octoverse reports and, usually we do an Octoverse report once a year, and we look at the numbers, and we say, “Oh my goodness.” I was at Universe in October saying, “This was the fastest year of growth that we’ve ever had,” right? And now we’re doing more in a month than we did in a year last year. Swyx [00:48:20]: You’re talking about PRs. Kyle [00:48:21]: Commits. Swyx [00:48:21]: Commits, yeah. Kyle [00:48:22]: PRs. Kind of like you name it by roughly every measure that we’re looking at, there’s some amount of sort of growth that is much bigger, and that is breaking our system in new ways, not old ways. Like webhooks were always notoriously, unreliable over the years? Swyx [00:48:38]: Whose fault is that? Kyle [00:48:39]: not anymore mine, but for a period of time, I’m sure you could pull up a tweet that was “It was me. I’m sorry.” but, now, that got rewritten at a scale level that is still working and is not having problems today. Now what we’re finding isn’t just the isn’t the-The simple stuff that folks are on the sometimes on Twitter or on the internet are “Hey, why is this like this?” Sure. There’s absolutely silly problems that we shouldn’t exist. But now we’re talking about, unique, novel permission problems that happen only at a scale across all different objects or whatever, that now we have to go rewrite this underlying system. And so it’s, there are problems that yeah, caught us off guard, which I think I said. Like the growth is astronomical, but also we’re making such material progress in that I’m excited once we’re once we’ve kind of like reimagined the underlying foundation layer, or pieces of it at least, what’s going to be possible when it’s not just all of us and all the new people that are being developers and all of their agents and all the tools like working together. Because that’ll still happen in that in that GitHub tool, that GitHub community. But it’s a it’s a hard day anytime we can’t give you what you’re looking for. We have the same problem internally. We operate through github. Com. Of course, we have backups when things go down and whatnot for our own operations but we feel it too. If it’s not working it’s not working for us, and that’s kind of like the promise of dogfooding for GitHub. It’s always been true. We’re using the same tool you’re using. We’re not using a super secret version. We and so we also need it to be great for us for our customers of course for open source. And now an exponential growth of agents, Doing it too. Swyx [00:50:32]: I wanted to load for audio listeners who maybe haven’t seen your tweets, whatever. So one billion commits in twenty-five. Now it’s two hundred and seventy-five million per week on pace for fourteen billion this year, if growth remains linear. Is that still the pace? I don’t know. It’s been a Kyle [00:50:48]: it’s, it’s speeding Swyx [00:50:50]: Roughly. Kyle [00:50:50]: It’s still speeding up. Swyx [00:50:51]: It’s, it’s April, so yeah. Kyle [00:50:51]: Exactly. This was in April. Swyx [00:50:53]: All right. So basically you have fourteen x growth, right? Year on year on year. And I think that’s a scaling issue. I think, I’m going to like try to really steel man this thing. People have experienced fourteen x growth. They haven’t had your downtime. And that’s like— C-can we go dig into that? Why? Like what’s the— what broke? What are we doing to fix it? Like just anything for the community to reassure them. Why GitHub Reliability Is Breaking in New Ways Kyle [00:51:18]: so there’s a Like I was saying, there’s a couple different places that we’ve seen the growth issues. Some of the growth issues, which is why we’re t— I was talking about pushing hard on more CPUs is in actions in particular. More tools, more agents, more PRs mean more builds, more builds mean more CPUs. And so we are expanding through not just our data center, but obviously we were talking about moving to Azure and moving to, adding an additional cloud compute because we simply need more CPUs. Not as much GPUs. We definitely need GPUs too, but now CPUs are becoming a factor. Swyx [00:51:53]: It’s very CPU heavy. Kyle [00:51:54]: Underneath the hood when it comes to some of the underlying services, we’ve been breaking up over the years our database infrastructure, so that way we have, more cognitive separation between our the various services. The place that we continue to have pain is in, permissioning. And so right now m-many of our permissioning layers sit into a database that we like internally call MySQL One, and old Hubbers will know what I’m talking about. And so we’ve been pulling things out of MySQL One for many years, because like and we use we use Vitess and we use other technologies to shard and we do it as one big Swyx [00:52:31]: Famous thing, PlanetScale was born from this and Kyle [00:52:32]: A hundred percent. Sam Old Hubber and friend. And so finding these opportunities to like break this out and then do that globally. The other thing that I think is interesting and both a unique opportunity and tricky is we also run everything I just talked about in a black box container with GitHub Enterprise Server for people that work on-prem. So we take everything I just said, and we also do it on-prem, and we also do all of that and we do it in a data residence setup for customers that need to have their data in a single location. Each of these has the unique characteristic around how we’re sort of storing that data in MySQL or in a permissioning setup. That’s where some of these outages have oc-occurred, where you’re seeing it more like across the board rather than just like the one piece Swyx [00:53:17]: Filling the database Kyle [00:53:17]: Isn’t quite working. Exactly. And so part of it is that. I think there’s been some other places where agents are much more or more projects appear to be moving towards monorepo versus we were going the other direction for many years in the industry. Repos were smaller, but there were more of them, and now we’re seeing the opposite. Repos are bigger, and there’s, not fewer of them per se ‘cause there’s new growth, but, we’re just seeing many more big repos. Big repos, big monorepos have always had, a unique performance problem. Because each one, is slightly different if, particularly if the underlying blobs are incredibly big Inside the repos. And so we’ve done a ton of work that you pro— like most people haven’t probably experienced, unless you’re in this case of the monorepo. But that Git, infrastructure layer improvement does help the overall, system because, many of the improvements that make monorepos work better make all repo infrastructure work better. And so, I could kind of keep going down the line where it’s another thing where we’re moving out of, We’re changing how we do j I’ll just say job queuing for lack of a better, explanation changing the underlying technologies there. Swyx [00:54:32]: I spent two years being a job queuing guy, so. Kyle [00:54:34]: And so it’s kind of a little bit of a little bit of piece by piece, and it’s mostly because as we were— as it was built, we built everything in a way that assumed, I guess in some ways that the size of the pipe of work was going to remain the same. There’s just going to be more people coming through each of those pipes. But instead now in places whereA git push was, generally a certain size for example, is now, no longer true. Swyx [00:55:03]: Oh, yeah. Kyle [00:55:03]: Or Swyx [00:55:05]: I push a thousand Kyle [00:55:06]: On the average. 100% Swyx [00:55:06]: A thousand line commits like daily Kyle [00:55:07]: Same thing with PRs. Like PRs same thing. And like we’ve talked about optimizing that and making changes where, and there were technology choices that did not work there? And it got slow, and it didn’t It was not fast. It did not do what the users wanted. And so we’ve been reeling that all out and going “Okay, that’s just not right. Let’s stop putting good money after bad and do it the do it the right way or the right way now.” So there’s It’s a it’s a lot of things, not quite when I’ve experienced scale at GitHub historically, it’s almost always two options that we’ve used. We go vertical scaling, particularly with databases, right? And we go horizontal scaling. Oh, we just have more people using this service. Great. We’re going to add more servers, and we rack them in our data center, or we use it in a cloud. And now we’re sort of in a like diagonal, where like vertical doesn’t really work anymore. Horizontal isn’t work either because we’re all We all have some CPU or GPU constraints in the world now, and now we have to go in and like crack open services that have been running for 10 or 15 years and go, “Okay, the rules of this service have legitimately changed, and now we have to rewrite them.” None of this is an excuse. This is like we’re We have to do the work. We have to make it better. Swyx [00:56:22]: actually as an infra guy, I’m “This is like one of the most fascinating scaling challenges I’ve ever seen.” Kyle [00:56:26]: That’s that’s, that’s the thing that’s the thing that it’s hard for Like when we weren’t talking about it publicly, an

Why Video Agent models are next — Ethan He, xAI Grok Imagine

We’re announcing AIEWF [https://ai.engineer/wf] speakers this week! Take the AI Engineering Survey [https://notion.qualtrics.com/jfe/form/SV_bP07tSVMXH7ePCS]! Today’s guest Ethan first joined us for the LS Paper Club as the lead on NVIDIA Cosmos World Model [https://www.youtube.com/watch?v=og59L4JECz4&pp=ygUWbGF0ZW50c3BhY2V0diBldGhhbiBoZQ%3D%3D], but then joined xAI and built Grok Imagine in 3 months: He comes back on Latent Space with some nuclear hot takes: that Video Models primarily get their intelligence from LLMs, not from training on video data, and that the next frontier for truly interactive, realtime, long-horizon world models is to work on LLMs (perhaps Interaction Models [https://www.latent.space/p/ainews-thinking-machines-native-interaction]as well…) Put it this way: In the near term, the next Sora won’t be a better video model, but a video agent. Generative Media [https://www.youtube.com/watch?v=t4359sKBu4w&list=PLcfpQ4tk2k0VjKRy3q6ZxeOtkbZlmFDLg] may more closely follow the evolution of AI coding which went from focusing on one-shot output performance and cost, to multiturn reasoning and planning models for agents and systems that can plan, edit, test, debug, and submit PRs. At a certain point, coding models got so good that the only significant next step to improve performance was handling the orchestration of these models. Now as the performance of video models increases significantly across realism, consistency, & prompt adherence while becoming more cost efficient, the next evolution of video generation may also be systems that can plan, generate, edit, critique, and iterate across an entire creative task. In this episode, Ethan joins swyx and Vibhu to unpack what it actually takes to build frontier image and video systems: data, VAEs, diffusion transformers, audio-video alignment, inference speedups, and the hidden cost of storing and moving massive video datasets. From building NVIDIA’s Cosmos world model [https://www.nvidia.com/en-us/ai/cosmos/] to joining xAI as Grok Imagine [https://grok.com/imagine] was being built from zero to one, Ethan He has been at the center of some of the most important work in video generation, multimodal models, and real-time world models. We go deep on Grok Imagine, how a small xAI team shipped its first multimodal video model in three months, why iteration speed matters more than almost anything in model development, and why many of the biggest gains come from fixing tiny bugs in data and training pipelines. Flipbook: The future of Videomaxxing Video agents are almost a sure bet to be the trend in the coming year. We end with a glance at what’s beyond video agents: Flipbook [https://www.flipbook.page/n/43e8c7b08ab14571810fee265c331cb3] caused a minor sensation this year when it was released, but most treat it as a fun demo. Ethan takes it very seriously — with the speed and cost of inference coming down every year, the future of custom video JIT UI is closer than you think. We talked about why videogen models may become the front end of AI, how generative UI could replace traditional HTML/CSS, why world models need to be real-time, interactive, and long-horizon, and why the future of video generation may depend more on language models and agents than on diffusion alone. We discuss: * Why fast iteration mattered more than meetings * Why small training bugs can drive huge model quality gains * Why coding models may make compute the bottleneck again * How image and video models are trained with synthetic captions * The role of VAEs and latent space in frontier video models * Why image models are the foundation for video models * The tradeoff between temporal compression and real-time interactivity * Flipbook [https://www.flipbook.page/], Neural OS [https://neural-os.com/], and the future of generative UI * Why future interfaces may go from user intent to pixels * The hidden cost of training video models: storage, egress, and GPU hours * How step distillation and consistency models (like OpenAI sCM [https://openai.com/index/simplifying-stabilizing-and-scaling-continuous-time-consistency-models/]) makes video inference orders of magnitude faster * Grok Imagine 0.9 and large-scale audio-video generation * Why audio-video alignment is harder than text-video alignment * Ethan’s definition of world models * Reference-to-video, video extension, and long-context video generation * Why xAI’s research communication undersells Grok Imagine * How xAI culture shaped the speed of development * AI watermarking, SynthID, and detecting generated media * Why prompt rewriting matters for video models * Grok Imagine Agent and the rise of video agents * Why language models may unlock better video generation * Robotics, physical AI, and embodied world models * Why Ethan left xAI and shifted focus toward LLMs * Self-managed context, memory, and the next frontier for language models Ethan He * LinkedIn: https://www.linkedin.com/in/ethanhe42 [https://www.linkedin.com/in/ethanhe42] * X: https://x.com/EthanHe_42 [https://x.com/EthanHe_42] Timestamps 00:00:00 Introduction 00:01:25 From NVIDIA Cosmos to xAI 00:03:24 Building Grok Imagine from Zero to One 00:10:07 How Image and Video Models Are Trained 00:18:53 Video Compression, VAEs, and Real-Time Tradeoffs 00:22:10 Generative UI, Flipbook, and Neural OS 00:32:10 The Cost of Training Large Video Models 00:37:04 Distillation, GANs, and Fast Video Inference 00:41:21 Audio-Video Generation and Grok Imagine 0.9 00:48:34 What Makes a World Model? 00:55:51 Reference Videos, Long Context, and Video Memory 01:00:11 xAI Culture, Research, and First-Principles Building 01:09:45 AI Safety, Watermarking, and Prompt Rewriting 01:13:10 Video Agents and AI-Assisted Creation 01:27:32 Why Language Models Unlock Better Video 01:31:15 Robotics, Physical AI, and Embodied World Models 01:32:38 Why Ethan Left xAI 01:34:16 Self-Managed Context and the Future of LLMs 01:38:43 Ethan’s Career Path and Closing Thoughts Transcript Introduction: Ethan He, Latent Space, and the Path to xAI Swyx [00:00:00]: We’re here in the studio with Ethan He, most recently of xAI. Welcome. Ethan [00:00:10]: Thank you. Glad being here. Swyx [00:00:11]: We’re also here with Vibhu. you were first coming to us or joining the latent space world because you were working on Kosmos at NVIDIA, and you did a paper. We loved it. you presented it as well, so thank you for doing that. Ethan [00:00:23]: I’ve actually, I also presented the MoEs twice at latent space. Swyx [00:00:29]: How did you actually hear about us? Did we reach out to you? Is that how it worked? Ethan [00:00:33]: No, actually, I-- the community. Like I realized, oh, there is this online community that people talk about AI and also learn from each other through papers every week through the Paperclip. It’s very nice. Ethan [00:00:49]: I learned a lot. Swyx [00:00:49]: I think three years stop. We haven’t stopped even on Christmas and New Years. many weeks I want to stop but it keeps going. Vibhu [00:00:58]: No, that was good. I think you had posted that you worked on a paper, and I was “Oh, very cool. We have Paperclip. Present then.” Vibhu [00:01:04]: But I might have reached out to you after. Swyx [00:01:05]: you-- because it’s an amateur club, right? Swyx [00:01:08]: so it’s very unusual and but we have sometimes paper authors come by and actually explain the paper. Today we just did, the poolside paper, which was apparently very good. Vibhu [00:01:18]: Came out yesterday. Vibhu [00:01:19]: pretty interesting, right? Fully open. They talk about everything, systems. So it’s a good one. We’ll, we’ll recommend people to read it. Swyx [00:01:25]: Bring us up to speed on your transition to xAI, ‘cause I actually don’t even know when you joined. just like tell the, tell the story about the sort of transition. From NVIDIA Cosmos to xAI: Scaling Video and World Models Ethan [00:01:34]: Before xAI, I was working on Kosmos world model as in-- at NVIDIA. So Kosmos is, it’s a giant video foundation models that can-- that aims to simulate the world and for-- it serves as a foundation of-- for all of the roboticists to build on top of. There, once I built the Kosmos one, I realized as this thing also has a scaling law similar to language model, we need to scale up the video models further. that’s, that’s why I realized I need to move to somewhere with much more compute resources. That’s how I Swyx [00:02:13]: Than NVIDIA? Vibhu [00:02:14]: The GPU rich came themselves. Vibhu [00:02:19]: And timeline-wise, when was Kosmo? It was pretty early, right? It was open world model, open paper, everything. Ethan [00:02:25]: It was end of twenty-four. Vibhu [00:02:28]: End of twenty-four. Ethan [00:02:30]: Then at mid twenty-five, I moved to xAI. At that time-- I joined about the time when xAI was about to build video models and in multi-model models. There were no infra, no data, and no model, and it just-- as a few engineers, we built it in three months and released the first model, Grok Imagine zero point nine. Ethan [00:02:55]: And since then, I keep working on video models and move more from training and to post-training of the video models. For example, like a reference to videos, kind of like the cameo feature and, video extensions. And, before I left, I worked on a world model, leading a small team to focus on the real-time long horizon video generation. Building Grok Imagine From Scratch in Three Months Swyx [00:03:24]: Can you give like a rough roadmap of okay, you’re on a brand-new team. Grok previously was only text, or they partnered with BFL for their image gen stuff. What do you-- what are the building blocks, right? You have compute, data you can procure somewhere. Like just what are like the sequence of things that people should think about when you’re setting up a new team? Vibhu [00:03:43]: actually even deeper, not just data you can procure. You guys had to go through getting the data too, right? So you shipped it pretty fast, but yeah Swyx [00:03:51]: three months is like Vibhu [00:03:52]: From everything Swyx [00:03:52]: actually like very surprisingly fast. Ethan [00:03:55]: One thing I say like thanks to my experience at NVIDIA, ‘cause first time when we were building Kosmos together, we built it, for about a year. So this is like the second time I do it. Roughly have an idea, what to do. I say the most important thing is the talent. Everyone were very strong and clever, very close with each other towards a common goal. So that speed up things a lot. So you reduce the communication bandwidth among people, and everyone can work towards the same goal. It’s, it’s like every day there’s not that much meetings on the calendar, like maybe like a, like a sync a day, and after that it’s, it’s just all building. It was pretty fun at that time. Ethan [00:04:47]: And another thing is that xAI has very strong foundations of like data inference, model inference, and the supporting there can help the model develop a lot. When I look at, training models, I don’t so actually the top important thing is like how many, how many iterations can you do, per day? and the more iteration can you do, you can, you can train the model much faster. So if you have very strong infra and you have a lot of compute, you can, you can train these models in very short period of time. That can give you a much larger buffer to, for errors, and it also gives you the opportunity to spot more bugs. Iteration Speed, Compute, and Debugging Model Pipelines Swyx [00:05:46]: What is an iteration? Is it like a few hundred steps or what are you Ethan [00:05:50]: Let’s say just the train-training the model, like from acquire new data and maybe design new algorithms and train a new model, maybe at smaller scale or Swyx [00:06:01]: So cycle time for like any hyperparam that you’re searching. Ethan [00:06:04]: Cycle time and tune to like eval this model. Is this model better than my previous iteration? Ethan [00:06:11]: So Swyx [00:06:11]: So it’s like before you, someone had already set this up that you can iterate very quickly. Ethan [00:06:15]: I think the foundation there is extremely good forDeveloping and research models. Ethan [00:06:23]: And often I find is it-- this is kind of boring, but like a lot of the improvements does not come from new algorithms. It comes from finding small bugs here and there in the data pipeline, in the, in the model training pipeline. Those give, those give the biggest boost to the model quality. Vibhu [00:06:46]: It’s interesting, right? So you say it’s like small team, less communication bandwidth, but also a lot of quality is like find little bugs. It seems counterintuitive, right? You have a lot of people, you can iron out more of those, but it’s interesting to see the other side, right? Swyx [00:07:00]: I also wonder, have you-- do you try using LLMs to look for bugs? I don’t know. Ethan [00:07:05]: I remember at that time it was mid two thousand and twenty-five, so it’s the coding model wasn’t quite there yet. I remem- I remember like December two thousand and twenty-five, it was extremely good. Yeah, I’ve been, I’ve been using it at that time. It’s, it’s helpful. sometimes it produce codes that are kind of difficult to maintain, even though like the first time it built something extremely fast. But it gave the, like a spaghetti code, thousands of lines that I couldn’t maintain, and the LLM itself couldn’t figure out what’s, what’s wrong and how to improve on top of it. But now I find it much better. Yeah, I want to bring up another point here is now coding models are much more efficient and can help us implement stuff much faster. Compute might become a bottleneck again because previously, like if you want to train a new model, say you want to generate new synthetic data and then or write a new algorithm, it might take a few weeks. And during that period of time, you don’t-- you might not have experiments to run. But now you can build that thing within a few hours, then you can immediately train a model. Ethan [00:08:24]: Now you have to have enough compute to try all of the ideas. So compute might be the bottleneck of iterating speed again. Swyx [00:08:36]: yeah, I actually, honestly, I think it’s like kind of a stressful job because you’re “Well, I should be trying everything, and if I’m not, then I’m not doing my job well.” Vibhu [00:08:48]: there’s also the stress of you’re eating thousands of GPUs per hour, which is very expensive and, compute can go to other researchers. Swyx [00:08:56]: You got the daddy Elon to Vibhu [00:08:57]: You got daddy Elon. Ethan [00:08:59]: It was Vibhu [00:09:00]: But there’s still finite amount of compute, like you want to use it, you want to use it well, you want more of it. Ethan [00:09:06]: That was quite stressful indeed. Yeah, I think one thing is the-- with coding models now, like a lot of these jobs can be automated, which is much better. A second, it’s a, it’s a marathon, so you got to maintain good health and, a regular schedule. Vibhu [00:09:28]: It’s, it’s hard to hear that when you shift from zero to nothing in two months. Swyx [00:09:32]: and, I think obviously the culture at xAI is very famously, people work very hard. one thing I did want to dive into, in our-- in the notes that you, that you sent ahead of time, you had specific comments about the cost of Video Gen training. presumably this is on the Colossus-1, right? the two hundred megawatt cluster. Any whatever you want to just share on that. Vibhu [00:09:54]: I think there’s, there’s three things we’re talking about, right? So there’s Video Gen, there’s also the Image Gen model that you put out. Do you want to like complete the, okay, so zero to one, you have a few months. Just what are the stages of create Image Gen model? Swyx [00:10:06]: Oh, yeah, maybe I got distracted. How Image and Video Models Are Trained: Synthetic Captions, Tokenizers, and VAEs Vibhu [00:10:07]: Sorry. and then, from there’s Video Gen, there’s Audio Gen. Would love to get into those next. But what is that first few months like? So small team, a lot of bugs, iterations, but what does it look like? Do we take something off the shelf? Do we just get data compute? What’s, what’s the few months like? How do you go to state-art Image Gen model? How do you just start? Ethan [00:10:28]: I cannot comment specifically how xAI did, but it’s, it’s a quite standard process. I can draw some, examples from Cosmos. So mainly it’s building a video model, you actually need to build a image model first. And building these two models, the data you need is a hundred percent synthetic pair of language and image or language to video. Because on the, on the internet, actually, the videos don’t naturally associate with text. So you can say, oh, like on YouTube, you have the title and you have the description and the comments Swyx [00:11:11]: Title Ethan [00:11:11]: of a video, but usually they’re not relevant to the video itself. And say maybe like the video is a natural scene of mountains or something, and the title is, I’m so happy today. Ethan [00:11:26]: So they have they have no correlation at all. So the first step is to, you have to generate synthetic pair of language with the videos. So you gather videos from the internet, and you use a VLM to caption the videos. So that part, here’s a question, like how do you, how do you gather VLM to begin with? So if there’s no Swyx [00:11:55]: You, so you fuse the model, right? Like Ethan [00:11:57]: Say if there’s no like VLM exists, like how do you generate the text to the beginning, right? It’s, it’s impossible. Swyx [00:12:04]: I see. Ethan [00:12:05]: In the beginning, it’s like you ask human to describe the video as detailed as possible.For example, you ask them to describe everything, like all objects, all characters, and all interaction and dialogues in the, in the videos. So that’s in the protocol of Cosmos labeling. We require the objective we give to the labelers was that you have to describe the video as detailed as possible, such that a blind person hears a blob of text can reconstruct what the video is like from their head. Swyx [00:12:43]: Video or image? You’re talking about images. Ethan [00:12:44]: Video or image, either one of them. Vibhu [00:12:47]: This was pretty common when we went from clip and DALL-E, right? Vibhu [00:12:51]: It’s all training on really detailed captioning of images. So same is applied to video, but instead Ethan [00:12:57]: same applied Vibhu [00:12:57]: of using multimodal model to pass in video images and write rich descriptions, you can also Swyx [00:13:04]: I think there’s this traditional perspective of supervised, or, very highly human curated thing. I feel like there’s a unlock with unsupervised, right? Where like you have enough to bootstrap that you can just throw common corpus on it or, whatever. like unsupervised vision and language pairing, right? Like where you just have, interspersed image and text and it just learns. To me, that is the VLM breakthrough that is different from the clip, different from the LM era. Ethan [00:13:36]: It’s interesting to see that you kind of need both data. Ethan [00:13:41]: For example, for the Swyx [00:13:41]: You need it to bootstrap it up. Yeah Ethan [00:13:43]: for the generative model training, there’s also usually like a small percentage of unlabeled data. So the model is instructed to generate a video without any text instruction. That can also help the model generalize. So after this stage of generative synthetic pair, so, one important common step is to train a compressor or a tokenizer of the image or videos. So because, if you train-- If you can technically, theoretically train image or video models on pure pixels, but the problem is that the, it’s, it’s a lot of tokens. So like one image, it’s, a thousand by a thousand, it’s like one million tokens, one million pixels. It’s impossible to train transformer on that. So it’s, you need to train a tokenizer, which can go from image to latent space and latent space back to image. Swyx [00:14:45]: That’s why we named the podcast. Swyx [00:14:48]: But, basically, you’re talking about vocabulary science. Ethan [00:14:50]: so vocab. Swyx [00:14:51]: And so, what is, what is imp-- like a million is impossible? Ethan [00:14:54]: In generative models, the vocab is continuous. It’s a continuous space. We can think about like you map an image to a vector. It’s a, it’s a fixed length vector. It’s sixteen or forty-eight, something like that. And then you map that vector back to the image space. And the mapping is, has-- The mapping is patch-based. So you say you have Ethan [00:15:22]: a sixteen by sixteen patch and you match, you map that patch of pixels into this latent space. Swyx [00:15:29]: We’ve covered this Vibhu [00:15:30]: This is like the vision transformers Swyx [00:15:32]: VAEs, Ethan [00:15:33]: VAEs. Vibhu [00:15:34]: You basically compress your input, you do your generation, you’re reasoning all that generation in smaller dimension, and then you project back out. Swyx [00:15:43]: VAE is a form compression, but I think the for me, the patching thing is from VIT, right? Ethan [00:15:48]: You can make those. Swyx [00:15:49]: Literally the, yeah, the paper is titled like sixteen by sixteen is all you need. something like that. and then I think also, people make a lot of comparisons with this kind of patching with convolutions. Swyx [00:16:02]: Which is you’re, you’re kind of re- reconstructing the old paradigm with the new. Ethan [00:16:05]: Actually, in VAEs, there are, there are both convolution networks and transformers. You can actually do both. Ethan [00:16:14]: After this VAE, so what you’ve got is you’ve got latent space tokens and you’ve got the language tokens. So now the training of the diffusion transformer, usually generative models use diffusion transformers. It is actually quite standard. It’s, it’s very similar to how you train a language transformer models. It’s not that much difference. It’s just the tokens, the visual tokens in, visual tokens out. The only difference is there’s a denoising process. So you train the model to unmask some of the noise. So you add, you add random noise to the visual tokens, and then you train the model to remove those noise to generate the clean tokens. Any inference, the model can iteratively remove noise from a hundred percent noise. Swyx [00:17:12]: And then there’s also, to speed things along on the tech tree of diffusion, there’s CFG, and then there’s, there’s also, latent diffusion that, there’s, there’s someone in there. I think, somewhere along the line, obviously, like stability and all these other guys, pioneered a lot of this, architecture. I don’t know if you want to get into that or just, or do the video side up to you. Bootstrapping Video from Image Models and Temporal Compression Ethan [00:17:37]: After you train such model, such image model, the reason it’s a, it’s a foundation for video models is that image models are cheaper to train, and they have much denser connection between language and text. So, sorry, language and images. For example, you train a billion, you train on a billion images, and there’s a mapping from the text to the image. And the cost to train the same, like the, a billion, a billion text to a billion videos, that’s much more expensive because videosNaturally have more tokens than images. Because the diffusion models, their understanding of, language purely come from this mapping. So if you don’t have enough mapping, so if you only train on like a ten million videos or something, there-- you might not see enough language tokens in your training, so your model does not understand human intention enough. So that’s why you really-- you train-- you first train this image diffusion models, and then you bootstrap the video model from there. Swyx [00:18:53]: One thing I did want to ask, because I-- actually, I think you’re, you’re the first per-- video model person I’ve ever talked to, I think. we’ve, we’ve like talked to Luma and all those folks. There’s all these tricks in video compression where basically frame by frame there’s not that much difference, so actually you don’t have to regenerate or save the whole frame, right? but I think MP4 compression or something else like that. Swyx [00:19:16]: is it tempting to use that? Or as far as I can tell, everyone just treats it as, “No, we would just generate every frame.” Is that roughly the state-art? Ethan [00:19:27]: There are a few different approaches. Let’s say first, like you want to just directly use MP4 compression and use that as the tokens for the transformers to train, right? So people actually have tried that, but the main challenge is the latent space for the MP4 tokens were not, were not very comprehensible for the models. It’s, it’s extremely hard to train on that. And there’s a Ethan [00:20:01]: So that’s why they created VAEs, which creates more continuous, latent space, so the models can understand that latent space and learn from it much easier. Even within the VAEs, there are different difficulties of the latent space. So you can imagine something the simplest, the most naive VAE is like you have an image, and you just shuffle all of the images into a, into a vector. So you don’t need to train any VAEs, right? But that latent space is extremely hard for models to train on top of. That’s why there are some debate on like how do you compress the tokens. So you mentioned like you can compress frame by frame. Also, you can compress, the temporal dimension. Ethan [00:20:52]: The difference is if you compress the temporal dimension, you get a much higher compression rate. Because there’s temporal redundancy between frames, because, this frame and the last frame, likely they are mostly similar, so there’s only some small difference. for example, I think in 12.1 VAE, they have like a eight by eight by four compression rate. So the four temporal tokens are compressed into one tokens. That can save a lot of, save a lot of the context length. If you do it frame by frame, you have to do maybe like eight by eight by one. Your context length will be four times larger. That being said, the benefit of the frame-- per frame compression, we might come back to this later, is, real-timeness and interactivity. ‘Cause if you, if you strain the output of the model, frame by frame, you can-- the model can respond to any user request immediately. So if you have like a temporal four compression, four times compression, then Swyx [00:22:06]: It might be laggy Ethan [00:22:07]: there’s a lag there in nature. Swyx [00:22:10]: So you’re very pilled on this. let’s just go ahead and bring it up ‘cause we have the visual prepared anyway. There’s some frontier applications of real-time video gen. So Flipbook is one of the examples that went viral recently, right? What is Flipbook? Real-Time Generative UI: Flipbook, Neural OS, and Diffusion Front Ends Ethan [00:22:23]: Flipbook is kind of like a web brow- web browser. You can see like it has the web bro- browser UI on top. The difference is all of the UIs are generated by generative image model in real time, and anything here are fake. But you can, you can explore inside this wor- this imaginary world. Say like we-- here we have engineering the Great Pyramid. Like the model generates this for us to understand how it works, and if we want to navigate around and understand further, we can click on some of the, some of the description here, and the model will generate a new page, new subpage describing the details we want to know about. Swyx [00:23:14]: So it’s basically kind of we’re playing a video, but it’s pausing for our next interaction, and then it just plays the next thing based on our interaction. Swyx [00:23:23]: Which is kind of cool. Vibhu [00:23:25]: and you kind of decide your story. So this was, how do you make a pyramid? levering technique seemed interesting, right? It shows how do you take Okay, I want to know what is this Swyx [00:23:35]: The demo, the demo tweet had more animation between frames. Vibhu [00:23:38]: I think it’s just skipping, Swyx [00:23:39]: Oh, it’s just skipping a lot of frames. Ethan [00:23:40]: they also have a video mode Vibhu [00:23:42]: It takes a lot. There’s a lot of people Ethan [00:23:42]: but, a lot of people are using it. Ethan [00:23:45]: So it’s not available. Vibhu [00:23:46]: There’s a live video stream. We can try, Swyx [00:23:50]: So this is an example of the kind of future that you see at the extreme. We don’t-- we’re obviously not in it today. Swyx [00:23:56]: But in a world where inference is completely free this is better than generating code and text? Ethan [00:24:02]: So this is, this is a final state of where Viva will be at for word model, I think. Imagine internet doesn’t exist, and then you type in google.com. Like what should, what should, what should a model show you?the model can imagine something, and this is what the model imagine. And these web pages, they completely do not exist. So I think as the inference costs come down, we are going to have generative UI for everything. If you think about how the coding model works, so they write code for a web page, and they render the code might be con- converted into binary, and the binary render the pixels on the screen. So we in machine learning, every time we have some breakthrough, obviously it’s, it’s more intuit. So why don’t we have like user instruction to the pixel directly? So the generative UI will be user intention to the pixels directly. And say like even if I want email, let’s say everyone have the same interface, but I want, I want it slightly different. I want the email to show to me like a TikTok, so I can swipe left and right for the emails. And or maybe you want something else. We can have completely different things. Or like I have I’m looking at, Instagram stories, and I don’t like the Like button. I always may click it. And, generative UI resolved it. So it’s going to be a revolutionary replacement of the interface. So in the future, we might have much more powerful Ethan [00:25:50]: LLMs and coding models running behind the scene. And in the, in the front-end, the diffusion model will actually be the front-end to show stuff to you. That’s how I imagine it. Swyx [00:26:02]: Diffusion front-end, deterministic back-end. Swyx [00:26:04]: Something like that. I find that very expensive, but, Vibhu [00:26:08]: I find it interesting you called LLMs writing code on the back end deterministic, but okay. Swyx [00:26:14]: you write it once Vibhu [00:26:15]: Compare it to Swyx [00:26:16]: And then you execute. Ethan [00:26:17]: If you think about the cost, say, let’s say H100 costs $1 per hour, and if you use this eight hours a day and thirty days, so, every month you’re paying this two forty, you’ll actually not wanna pay for that. That’s even more expensive than Cloud Code Max. But if you think about the compute costs come down like two times every year, and I think the future will likely arrive like within few years. Vibhu [00:26:49]: It’s everything, right? compute cost comes down, compute gets faster, model gets smarter Ethan [00:26:54]: More efficient Vibhu [00:26:54]: model gets smaller. Swyx [00:26:55]: I don’t know why you say two times, ‘cause I think it’s like 100 times. In language models, it is roughly one hundred to a thousand times every twelve to eighteen months, for the same given level of LMSys, ELO. Vibhu [00:27:08]: That’s a net of everything, right? That’s model performance alongside compute. So different than just compute costs come down. But, a very interesting future. Swyx [00:27:19]: So the web designers will have to shout out that accessibility is an issue, right? how do you deal with screen readers or whatever. But yes, this is higher bandwidth storytelling than anything you can possibly generate with code, right? So I think that’s the rough idea. Ethan [00:27:34]: And I’d like to add a little bit that so human naturally have the maximum bandwidth when we are looking at things, look at videos, and we also have maximum output bandwidth when we are talking. So in the future, it might be something like we talk to AI models, and the AI model responds back with a generative UI. So that would be the maximum input and output bandwidth to interact with AI models before neural link happens. Vibhu [00:28:06]: And it’s also very custom, right? Some people are very visual, some people are not as visual, right? They prefer the text. But the best thing about generative UI, right, it can also be text. Swyx [00:28:17]: There’s another project that we wanted to highlight, which is the Neural OS. Kinda similar idea, but here you’re literally operating, simulating an operating system with a video model. Swyx [00:28:27]: and you can play Doom, you can do Firefox. I find this like mildly less impressive, obviously, because it’s an OS that I can run. Swyx [00:28:37]: But here everything is imagined. Vibhu [00:28:40]: I was, used to the Command+W to close the Firefox tab. It didn’t crash. That’s why I said Swyx [00:28:45]: It’s too immersive. Vibhu [00:28:46]: It’s, it’s too immersive for me. Swyx [00:28:47]: Too immersive. Vibhu [00:28:48]: I wanted to close the tab. Vibhu [00:28:49]: But yes, I can play generated diffusion. Swyx [00:28:51]: this is shockingly fast. Swyx [00:28:54]: Because I remember there was a demo about like maybe one to two years ago. Someone tried to do the first-person shooter with a image model. There was no consistency. It was very slow. But here it looks like realistically it’s-- this is Doom. Vibhu [00:29:07]: I think there’s two sides to that, right? There’s okay, what is running a game? The heavy part of it is actually the game engine, all the lighting, all that stuff, the graphics. This is just kind of video, right? Like we’ve solved consistency. This is still, it looks like a few years old image generation. There’s some temporal consistency, but it’s, it’s kind of just images stitched together as frame video. But it’s a good visual representation to pi- to picture the future you wanna see, right? that’s, that’s what I see in these more so. Ethan [00:29:38]: This reminds me of how the video models gets better and better. So Neural OS is kinda if you just look at it feels like it’s just a crappy version of the, like the Windows we could have, right? And, but the difference is, so the model, this model is overfitted on the existing operating systems. It can generate nothing different than that. But it’s actually also similar to video models. So when we are training these video model, image model, we train them on internet. There’s no imaginary supernatural stuff on the internet. But once we train this model, you can prompt the model to generate something supernatural that have never existed in the data set. So if you train your Neural OS or neural computer on the standard screen recordings on the entire internet. The model can imagine completely new interface to interact with the computer. Swyx [00:30:43]: This is one of those things that is magical to me. usually generalizing out of distribution is bad, but somehow we have learned some kind of internal world model that you say, this plus, but it looks like rainbows and butterflies, it’ll do it and it will kind of make sense. Swyx [00:31:03]: So yeah, that’s kind of cool. Yeah, I don’t know if there’s any comment more on there. I do, I do wanted to, I did wanted to touch a little bit more on the model architecture stuff, which I think you were getting. It’s, really fascinating. We don’t get a chance to talk about this enough. So one of the papers that we covered, we’ve covered every annual, segment anything release. and I don’t know if you follow-- you’re a computer vision guy, so you Ethan [00:31:26]: I know Swyx [00:31:27]: . So they did memory attention, which is kind of interesting. And I always think, anything where you can, across the temporal dimension, keep some consistency, I think it’s, very fascinating, and I don’t know if Basically, does that-- the CV side bleeding into video gen side, I think is underexplored, right? we talk about it for labeling, but actually you can borrow the architecture itself. Ethan [00:31:50]: There’s, there’s also complete different approaches, right? you brought up the term world model, so we went from video model to world model. There is diffusion, but there’s also other approaches that people are doing. So maybe we get into those after as well,? Swyx [00:32:03]: He has a whole definition of world models and stuff. I feel like we threw a lot at you. Whatever you want to comment on. Why Video Models Are Expensive: Storage, I/O, and Training Scale Ethan [00:32:10]: I think one thing that we should actually comment back on is okay, so we were talking about the steps to train image gen to video model. One thing we don’t see as much of is okay, you brought up the delta in training data, right? So Ethan [00:32:24]: you won’t have as much a video model might not generalize, but what is the cost of training a large video model? So we know for LLMs roughly, okay, even like the poolside thing that came out today, right? It’s a Gemma level model trained on roughly forty trillion tokens at this many H200s over this much time, right? You can see what is the exact cost of that. So how many GPU hours over how much H200 costs? So how do we do the back-end math of, same thing for video models, image models. How do you, how do you kind of break that down? I can share some back-envelope calculation. So surprisingly, video models is-- the cost is very-- is comparable to language models and obviously the largest scale is language model, maybe like a medium scale to language models. I said just storing the videos alone, it costs a lot. You can, you can maybe look up on AWS or something. Ethan [00:33:20]: You really, say if you have a billion videos and let’s say, let’s just say like each video, like five megabyte, then you need five petabyte to just store those videos. And also remember we talk about you use a VAE to compress the videos, and you also need to store, typically you need to store those continuous feature, in-- also in your storage. That’s also comparable size with the videos themselves. So just storing these videos and the features is tens of petabytes alone. And, Swyx [00:33:58]: I just, I just looked up the calculation. Five petabytes on S3 Standard is one hundred K per month. Ethan [00:34:05]: And Swyx [00:34:05]: It’s comparable Ethan [00:34:05]: and you need Swyx [00:34:06]: And Ethan [00:34:06]: And then like tens of petabytes, two hundred K. And even more expensive is you have the ingress and egress. Swyx [00:34:13]: Oh, yeah. Ethan [00:34:14]: Like you-- through the internet. You have to just to download those videos, I believe it’s, it’s more expensive on AWS than just storing those videos. Swyx [00:34:25]: Storing, yeah. Ethan [00:34:25]: And each training runs, you probably need to pull them once. If you train multiple times, it’s, it’s even more than that. So it’s like just storing the network, those costs is just, it would be a few, a few millions per month to just storing everything, not to mention the GPU cost. Ethan [00:34:45]: And Swyx [00:34:45]: my side tangent, the compute rental, like GPU rental is very efficient. There’s one side, okay, you can be XAI and build your data center. Should we not just build our, storage compute as well? Like Ethan [00:34:57]: Of course Swyx [00:34:57]: cloud cost compared to just, Ethan [00:34:59]: You save so much Swyx [00:35:00]: store. Yeah, exactly. Swyx [00:35:01]: Especially with like egress and stuff. So. Ethan [00:35:04]: That’s a good idea, but it also comes to-- there are some of its own challenges. Swyx [00:35:09]: Of course, of course. Ethan [00:35:10]: like people who build the GPU data centers, they might not expect this much, storage. And yeah, people build storage, typically they just build it somewhere with just CPUs. Swyx [00:35:23]: I just looked it up. Five-- AWS only charges for egress, not ingress. Tier five for five petabytes is two hundred and thirty K. Ethan [00:35:32]: Even more expensive than the storage. Swyx [00:35:34]: But storing is per month, right? You check in, then you cannot check out. so it’s so cool. It’s okay. So there’s that side. Ethan [00:35:41]: So the TLDR, my backhand math Swyx [00:35:42]: Data is larger than you think. Yes. Ethan [00:35:44]: my backhand math of GPU hours times GPU cost is also very much, I’m missing some storage. Swyx [00:35:49]: You’re also-- you’re basically like also more IO bound than normal training. Swyx [00:35:55]: Yes. ‘Cause like data loading, so caching everything, it becomes super important. Ethan [00:36:00]: So in Cosmos, we did a lot of optimizations to make it not IO bound. So, speaking of the training, actually training the model, the GPU cost, if you look up like the open source model, how big these video models are, I think like LTX has nineteen B parameters. That’s a dense model. And people are also exploring, MoEs, so it might be twenty B active and, like a hun- hundreds B, total. So that’s, that’s even-- that’s similar size as medium-sized LLM models. And if you, if you look at number of tokens-Uh, we disclose that in Cosmos. It’s also like tens of trillions of tokens on the visual tokens. So putting this together, the cost of, training these video models, it’s actually comparable with LLMs. Not to mention, the infra is slightly different from LLM, so it might be less efficient to train these models. Inference Speedups: Step Distillation, Consistency Models, and GANs Swyx [00:37:04]: Do you get the benefits of traditional diffusion speed-up? So for, images, there’s LCM, LoRAs for, fine-tuning. There’s, there’s a lot of stuff that’s been Ethan [00:37:15]: Flow matching. Swyx [00:37:16]: there’s flow matching. There’s a lot of stuff that’s been done. there’s some overlap that applies to diffusion on the inference side and stuff or? Ethan [00:37:23]: so the difference-- the inference side is a completely different story. Ethan [00:37:28]: I think for the training side, it might be a little bit hard to reduce that cost. And for the inference side, the biggest gain is from the distillation of these models. You can-- It’s called step distillation, slightly different from knowledge distillation in LLMs. So you-- Typically, for flow matching models, you need like 100 steps or something. Like a distortion model even need even more, like 1,000 steps to generate a good image or video. A step distillation is try to learn to generate fewer step from the model itself. It’s kind of like now we-- you use the full model to generate in 100 steps, and then you take a model that only generate 10 steps and let that model to learn from the perfect one. Ethan [00:38:25]: why this work Swyx [00:38:27]: Strong to weak seemingly. Ethan [00:38:28]: It is. It’s kind of Swyx [00:38:29]: Distillation Ethan [00:38:29]: kind of like strong to weak. the-- from the modeling perspective, the strong model, the teacher model is trying to model the image and videos of inter-internet, and that distribution is extremely complex. But the step distilled model is just trying to learn from the teacher. The teacher is a model, and the size is fixed, as the distribution is much simpler than the whole internet. That’s the intuition I have why step distillation can work. So usually these models serve in productions, they only run in a few steps. In Cosmos, I believe we have, we have like four step and eight steps. If you do some simpler task, image-image translation, it can even run in fewer step, like one step in Cosmos Transfer. Swyx [00:39:22]: I think this is the same intuition that guides a lot of the consistency model work. I sent you a link for, SCM. I don’t know if you covered that. To me, that was actually one of, the most impressive papers I’ve ever seen from OpenAI. Swyx [00:39:34]: That this is the unifying grand concept of consistency models. I don’t know if you have any comments on this. Ethan [00:39:41]: So there are, there are a few different approaches, Swyx [00:39:46]: Oh, yeah. Here it is. Swyx [00:39:47]: Two steps versus twenty or 100 steps, whatever. It’s already done. Ethan [00:39:52]: So there are, there are a few different approaches, for example, consistency model, and there are also Actually, we shouldn’t forget GAN. So GAN, actually, that was, that was the OG of Swyx [00:40:05]: OG Ethan [00:40:05]: step distillation ‘cause it trained just one step to begin with. So actually, a lot of, uh-- For example, there’s a distribution matching distillation which use, which uses GAN, as one of the laws for distillation. It-- GAN just tells you, “Hey, generate an image,” and then Ethan [00:40:31]: it has a discriminator to tell, is this image real or not? So the model, the model just need to learn one of the distribution, not the full distribution. Because in training, the model is asked to reconstruct the ground truth image from the internet, which is extremely hard. And in-- When you’re training GAN, it’s a step process. It’s just a, “Hey, you generate image. Does this image look as real as the image from the internet?” Which is a much simpler task. And, yeah, combining a lot of these approaches together, people typically do that, like consistency model and distribution matching and GAN, and we can get these few step models. Audio-Video Generation and Time Alignment Swyx [00:41:21]: Then there’s one step I wanted to add, which is audio and video. Ethan [00:41:26]: So, Grok Imagine zero point nine, I believe it’s, it’s a first audio video transmodel deployed at a large scale. So Swyx [00:41:39]: And that was your first model? Ethan [00:41:40]: that was, Grok Imagine’s first model. It’s, it’s audio video, joint generation. I think the hard part is, the modality alignment, ‘cause before this transmodel, we have, we have text to video alignment. We have this, correspondence between text and video. Typically, most of the VLMs, they understand images and videos. Video’s very rare, and they don’t understand audio mostly. And if you look at the audio generation on the LLM side, you can talk to them perfectly fine, but if you ask them to sing a song or something, it typically is not very good. Also, they don’t have, they don’t have music either. The hard part is thatUh, actually audio has two component. It has like a discrete component, a continuous component. The discrete component is like the language. Ethan [00:42:44]: So when we speak, it’s just, some Swyx [00:42:47]: It’s an ASR issue, yeah. Ethan [00:42:49]: It’s, it’s text token with some characteristics, I would say. Ethan [00:42:54]: But music Swyx [00:42:56]: I think the speech guys would disagree with this. Swyx [00:42:57]: Like disfluencies and then, Vibhu [00:43:00]: There’s tones you can get angry. Ethan [00:43:01]: Well, I say largely. Ethan [00:43:03]: the mu- but the music is completely different. It’s, it’s very continuous, and you cannot model them like discrete tokens in language models. this is like the hard part for models is, not to mention we have to align text, video, and audio together. Ethan [00:43:26]: So Vibhu [00:43:26]: How? Ethan [00:43:28]: So significant-- some significant challenges are like-- So first, like we talk about as the VLMs, they cannot understand most of them cannot understand audio. Ethan [00:43:39]: So you have to have some way to do the synthetic data generation for audio. You have to caption the model, and that involve, that involve synthetic data and human data effort a lot. And not just surprisingly, most of the LLMs are very bad at recognizing, like the beat, tone, and the details of the of music. They can, they can give some general prediction of which song is this, but it’s very hard to describe the details of the music. like we mentioned in image generation, like you have to describe image as detailed as possible so that someone blind can reconstruct that. So here is like someone Vibhu [00:44:32]: Deaf Ethan [00:44:32]: someone deaf can reconstruct how the music sounds like without actually listening to it. Maybe you can think of it need to have the-- or they call the script. Vibhu [00:44:49]: Subtitles, yeah. Ethan [00:44:49]: You gotta have all the details of the music, and the dialogue. Vibhu [00:44:55]: So is the challenge there typically stuff like music and audio, or is it just Like is there a baseline? Okay, there’s enough data where we can understand, narration, conversation, but there’s nuances in audio that’s where you hit all the data issues or is it just from stage zero, you just do it all right? Ethan [00:45:15]: So one important thing is like the alignment. So the model, the model has to know like the video and audio, the, uh-- it has to have a time-based alignment, like at which time step the video and the audio token correspond to each other. But we actually don’t have this kind of alignment for most of the other modalities. If you think about like text and image, text and video, they are loosely aligned. So you can, you can have a description of what’s going on in the video, but you don’t have to exactly, You typically don’t have exact description, oh, at, time step one second like what happened? Vibhu [00:46:02]: It’s very Ethan [00:46:03]: At time step two second what happened Vibhu [00:46:03]: coarse. Yeah. Swyx [00:46:05]: So what was the ideal time step? You have to oblate it, and then it’s like four seconds or something. Ethan [00:46:09]: So that comes down to how you design the model to, for the model to be aware of as a time, as a time modality. So the model is like a time aware. And that’s something pretty unique if you think about LLMs. So if you ask LLM to complete a task, say they, uh-- you ask them and they will say, “Oh, this task will probably take twelve hours to complete,” and they come back in one hour. Say “I’ve already spent two days on this and I’ve exhausted everything.” Ethan [00:46:47]: So the LLMs them-themselves, they don’t have a sense of time there. Vibhu [00:46:53]: I actually don’t think that’s just them not having a sense of time. I think it’s somewhat based, right? Vibhu [00:46:58]: Like you tell someone, “Okay, go work on this feature. Go implement this,” there’s a general understanding you would have of how long that would take without LLMs working at LLM speed, right? So you think back like two years ago, if I tell you to like build me like a new front end for latent space, have a search bar, have all this, you’ll estimate that it’ll take a few days, right? Vibhu [00:47:19]: So you tell an LLM, “Go build this.” It’ll take me a few days. But I think it’s somewhat grounded as opposed to them not having the best-- Not saying that they have a great understanding, but I think that example is like you can see where it comes from, right? You’re trained on all over the text. Swyx [00:47:35]: They’re, they’re trying to estimate what a human would say. Vibhu [00:47:37]: because that’s what the, that’s what the data kind of represents. It’s not them Ethan [00:47:41]: It came from the corpus on the internet. People have a estimate of how much time. Vibhu [00:47:45]: And not even just in direct like training samples, right? Just your world understanding of tokens of how long stuff takes, right? Go read a book. It’ll take you a while, right? Vibhu [00:47:56]: Even if you do nothing but read a book, it takes a few days. So yeah, LLM, I read it took me a few hours. Vibhu [00:48:01]: It’ll take me a few hours to go through this research. But this is a tangent. Swyx [00:48:05]: Somewhat, yeah. Swyx [00:48:06]: This is a train of thought I haven’t really expressed until now is, which is basically like a full world model must also be recursive, meaning that the participant in the world model must also be aware that they have a world model. which is like this whole recursive thing down the, down the line. but yes, and that the world model can be wrong and that they need to update it and blah. Yeah. We’ve, argued this on the, newsletter as well, that there needs to be sort of recursive or adversarial world models. World Models: Real-Time, Long-Horizon, Interactive Video Vibhu [00:48:34]: just, to ask, how do you define world model? Swyx [00:48:38]: Oh, yeah, let’s go there. Ethan [00:48:40]: So Vibhu [00:48:40]: So just for context, we talked about, video generation, and then there’s a-- if you say there’s a distinction between world models, what’s your, what’s your definition? How do you see the two? Ethan [00:48:53]: So disclaimer, I’m not going to debate, what is world model. Yeah. there are many definitions, so I’ll just talk about my definition. Since I came from the multi-model, multi-model domain, so mainly talking from video. So world model is like real-time interactive long horizon videos. So there are three parts. so we-- let’s talk about them one by one. So the so interaction, so we just, we just look at Facebook and neural computer. So the interaction part of it, so you, world model can allow you to interact with them through keyboard, mouse, and maybe also voice. So these all is-- all is a modality. You can, you can interact with the model, and the model should respond reasonably. Second part is real time. So once you, once, say, you move your mouse, if, say, the world model generate a game, how fast can the game respond? So if you’re like professional CS: GO players- -my say, oh, you have to respond- He’s beginner within sub ten milliseconds or- Yeah even less. So that’s not most of the- No, sixty FPS. Let’s go. Oh, three hundred FPS. Oh, five hundred FPS. Wait. okay, yeah. I didn’t do the math, but yeah, okay. Uh- Yeah, three hundred FPS, that’s a three millisecond. So you have to respond- Oh, s**t. Okay. Yeah Ethan [00:50:29]: within a millisecond. Most of the video models cannot do that. Yeah. And, but if you, say, if you have a video model that is, say, like a digital human, the response time might be more generous. Maybe typically, for real-time voice interaction, it’s like two hundred millisecond. So that’s, that’s much more generous. But even two hundred millisecond is pretty, it is pretty tricky, ‘cause remember we mentioned Ethan [00:51:01]: you have this, temporal compression coming from the VAE. So if you, if you don’t compress the temporal dimension, your sequence length is going to explode. So if you want to have this real-time, real-timeness in your model, you have to do is one context problem. And the third part is long horizon, ‘cause we-- if you’re not going to just play with, video games just, a few seconds, most video models only a few seconds. We’re going to play with minutes, hours. The model have to be able to generate long-form content. Ethan [00:51:42]: So putting these three together, it’s, real-time, long horizon interactive videos. I think the final state will be, for example, like a video, a video version of Playbook, where you can, you can interact with, a neural computer. You move your mouse, and you click on the generative interface, and it will reply to you through pixels- generating in real time. But getting there, it’s, it’s a very long way to get there. So one of the first step, at Grok Imagine, where I led a small world model team there, was to build video extension. So, video extension- it’s the first step of interactivity. Yeah. It’s, it’s the first step. Yeah. So it’s the first step- You have it here, video editing, yeah. Yeah. Yeah. So the first step is because, this unlocks long horizon videos. Typically, for most of the video generation models, you give it a prompt or an image as an initial frame. You generate video, that’s it. That’s just, one time, done. And some creators would try to, use the last frame as a first frame for the second video. It can-- sometimes it works, but if you do it a few times, it says the quality would decrease. And- It doesn’t have that context- Yeah over the full video, so the temporal- Yeah, exactly. Yeah, ‘cause you only gave it the last frame, of course, right? Yeah. Exactly. And- it’s actually a pretty fun hack. if you’ve seen like- Oh, no, he’s saying something better. Yeah. And for example, like Vue, I remember Vue 3 has like a second context of the last video. It is slightly better than using the last frame, but it has the same problem-- similar problem that it, the quality would decrease. if you extend a few times to, one minute, the video quality would look much worse than the first video. Second, another problem is that the model doesn’t have long-range knowledge of, what’s happening before. Say, if they generate some dialogue, some, two people speaking, and their voice might change, over some time, especially if the second conditioning, it does not cover the previous context. So these are the core challenges. So the Grok Imagine video extension, it has historical context of all of the previous generated videos. It can, It has, it has the context of, who is speaking and what objects have appeared and everything, having that to generate the next video. So if we naively do this, you can imagine, just, put all of the previous history video tokens into the context. The context lens will easily explode. Especially for video models, that can be like a few, a few million context, I would imagine- context lens. Yes.Yeah. Swyx [00:54:58]: Let’s run with that. Ethan [00:54:59]: for example, like in Cosmos, I think just five seconds of video is like a fifty K or sixty K number of tokens. So like if you do, if you do fifty second, that’s a five hundred K tokens. If you do longer than that, easily explode. This long horizon, problem was the first step we’re trying to solve world model. It turns out people, yeah, people love video extension. Like a lot, a lot of the creators love using video extension to create longer form videos. This is the part I liked that you have a, you have an intermediate step toward the final goal instead of just a straight shot to the final version very much. Swyx [00:55:48]: But I can see you have a strong vision of where we want to end up. Long Context, Redundancy, and Efficient Interactive Video Vibhu [00:55:51]: Does it seem like it’s an efficiency issue? okay, we’re at a few million tokens context,. If you draw the parallel to language models, we had very short context, two thousand, eight thousand, then, you scale it up one million, ten million. sure, there’s effective context, but at the end of the day, it’s just what’s it worth? sure, there’s a whole training data side. In video, it might be slightly easier ‘cause we have a hundred million token video, right? Just take a movie with the full context there. Like is this efficiency from an inference standpoint that like it’s expensive, but we know how to solve it? Or like why is this not the approach? So like my broader point was on your second point of world models, you say it needs to be interactive and live, right? You should be able to play a game and see the interaction live. So one thing I see with research is a lot of what you actually serve is different than what you build, right? So we talked about distillation. You train big model, you distill it, you do quantization, speculative decoding. We do all this stuff to serve it efficiently. Should we not just have a solution, like a world model that can interact well, do inference optimization, serve it, distill it secondary, so make it real time after you solve it? So like a-- another parallel is say, continual learning, right? What we need is someone to solve it and show it works inefficiently. Give it a few years, people will make it efficient. Same thing with regular attention, right? It worked. Over a few years, people have different forms of attention, and we’ve scaled it to be efficient at log context,? So kind of two things there, right? One is it seems like it works. You’ve scaled it. Can we not just scale it a lot more efficiently over time? Do we need a separate approach if this works? And same thing with interaction, right? if we can get it done, like if we can solve some way that it works, we can solve making it more efficient from an inference standpoint later. Ethan [00:57:53]: that’s actually a very good point. So in videos, there’s actually a lot of redundancies. So we solve a lot of the pixel redundancy from VE, but there’s more redundancy in long range and long horizon videos. Say, if a character appear in the first clip and then it disappeared, it only reappear at the end of the video, you probably don’t need the-- the context, like in the middle of the generation. So you only need that character, where you need. So that’s why, I helped build another feature. It’s a reference video. Vibhu [00:58:36]: Is it here? Swyx [00:58:36]: is it the same model release or different one? Ethan [00:58:39]: It’s a different one. Ethan [00:58:41]: You probably need to search on Swyx [00:58:43]: I’ll find it Ethan [00:58:43]: X reference to video. Ethan [00:58:46]: So reference video allow you to like upload up to seven images as condition and generate the video. Say, if like I want-- it can, it can be characters or objects or even scenes. Say like I want, I want condition on, Sean’s selfie and holding a blade Swyx [00:59:07]: We have a dog Ethan [00:59:08]: or whatever. Swyx [00:59:08]: We put the dog in the thing. Ethan [00:59:09]: you can put them there and the video models will generate the video from and copies the context over. So that can solve a lot of the problems there, like the long context problem. It doesn’t need to have a very long context, but it’s-- I feel like it’s an intermediate solution. The model Swyx [00:59:29]: It’s cheating. Ethan [00:59:30]: the model should be able to like selectively know, where should I draw the references. So say if I want to generate a movie, I generate it autoregressive, like a ten second at a time or something. And now this character appear, I can look back to where it first appear and, bring that back. Yeah, this one, I put the references. Yeah, that’s, Optimus, Einstein myself, Annie. Vibhu [01:00:02]: Oddly enough, I used Grok Search to find it, and it pulled your LinkedIn post. But yeah we found it. Ethan [01:00:08]: Interesting. Vibhu [01:00:10]: But xAI’s Underrated Work, Culture, and Watermarking Swyx [01:00:11]: this is a problem. This is not your fault, but like XAI doesn’t communicate all this work that you do very well because they just have the model release and then that’s it. But actually, these details are very good. Swyx [01:00:22]: As far as I understand, everything you just described is state-art, like no one else has done it. Vibhu [01:00:30]: A lot of-- yeah, I have a lot more Swyx [01:00:32]: And then, and then you just put this blog post with the cookies. I’m this is not enough,? Swyx [01:00:37]: but I, obviously this is like the high level numbers that people want to know. But no, okay, so Vibhu [01:00:42]: And I wonder, like part of that is also some labs don’t share research into what happens. And if Swyx [01:00:50]: No, but this is literally bragging about how good they are, right? Swyx [01:00:54]: Like, why would you not say that you are capable of extending with full context? this is not a secret sauce. This is like we did the work. yeah, I don’t know. Ethan [01:01:02]: different labs have slightly different communication styles. Swyx [01:01:07]: Anyway, if anyone from XAI is listening we are always happy to help you tell your story. Yeah, okay, so you did references, and I think, I think kind of the point you’re, you’re making is it is sort of like a kludge, right? this is-- you can do seven, but what about 100? Swyx [01:01:23]: Right? Then you need a completely different thing. Ethan [01:01:26]: So I think it’s-- this is, a mechanism to, select the context from the history, and you might not put the entire history into the context. for example, there’s a paper called Frame Pack, which have Ethan [01:01:41]: a heuristic that the latest history, the last one second, I put the entire history, and the history before that, I would, compress it and makes the video smaller. So they follow this pattern, this build overall pattern that the maximum sequence length is fixed. So the further you are from the current frame, you have a smaller image. So this is just a heuristic. I think it can be more automatic. The model is aware like which history part of it can be select. So this part of the research is actually being actively, worked on by a lot of people. It’s also quite interesting. I feel this is actually, this part of long context is a little bit ahead of the LLM part. Ethan [01:02:31]: So for example, like in LLMs, if you-- so contexts keep growing. Let’s say if you call tool and the tool call history is extremely long, that’s still in context, and keep growing, keep growing. Even if you switch the topic to something else, the whole context was there. There are some agentic harnesses that help you to, say, prune the tool results and, prune Like when you, when you quer