Episode 304: The Agentic Era & Books with Mark Saroufim の全文書き起こし

PRs, and let's just let's just do the four PRs as like from fresh. 00:05:06.800 --> 00:05:16.160 Like, don't try to take the existing code and like shape reshape it into a PR, but just like using the existing code as a reference point, generate the PRs. 00:05:16.319 --> 00:05:21.360 And I think those PRs are gonna go up uh sometime this month. 00:05:21.600 --> 00:05:26.319 And my guess is gonna be that it'll take two to two months or so to merge. 00:05:26.480 --> 00:05:29.519 I think it's in a shape now where it will be acceptable. 00:05:29.839 --> 00:05:33.759 The yeah, I mean the current the current code is it's a little bit of a mess. 00:05:33.920 --> 00:05:38.560 Like the biggest problem, the bit the single biggest problem was not following conventions in the code base. 00:05:38.639 --> 00:05:42.399 So like there was a convention for how tuning got done and it wasn't followed. 00:05:42.560 --> 00:05:47.759 And the a good part of it was things that the model never could have known. 00:05:48.079 --> 00:06:02.240 For example, like, oh, you know, this is the way that we used to do the tuning in this part of code, but you know, there's a new there's a new policy that's not documented anywhere for how we would do this, but the PR probably won't be accepted if it doesn't follow this new convention. 00:06:02.480 --> 00:06:18.079 Or like there's you know, there's this convention that we if you want to have an algorithm that dynamically dispatches to two, if you want to have a cub routine that can dynamically dispatches to two algorithms, like you need we need to get a special sign-off from someone. 00:06:18.160 --> 00:06:19.839 And that's not documented in the code. 00:06:19.920 --> 00:06:22.000 So how could the model have ever known that? 00:06:22.319 --> 00:06:23.279 This is this is interesting. 00:06:23.360 --> 00:06:42.639 Like basically, the this sort of makes sense if you sort of assume, like this is my guess, by the way, that like the cub team has been working on it for a while, that the original team hasn't disbanded much, and so a lot of these sort of like unwritten rules are actually in people's heads, and they all have internalized them, and there's no reason to write docs because maybe the team doesn't grow very quickly. 00:06:42.800 --> 00:06:47.680 But then, like in an age where you're forcing them to verbalize everything, they're like, Oh, goddammit, like this is so much work. 00:06:47.759 --> 00:06:49.600 I have to write docs now and I have to write tests. 00:06:49.759 --> 00:06:51.199 Yeah, yeah, I see interesting. 00:06:51.360 --> 00:07:10.800 Yeah, but I think it will be very interesting to see whether we can get the get this landed because Cobb in particular is a really high barrier to land because you need to be good across a really wide range of inputs, and you know, you you you you can't like ignore edge cases. 00:07:10.959 --> 00:07:25.360 And so it'll be I I I I'm hopeful because I do think it's a it's like a pretty big win uh

for the for the cases that it like now I've gotten it to the point where essentially there's there's no adversarial inputs that are really bad for it. 00:07:25.439 --> 00:07:26.399 There may be one left. 00:07:26.480 --> 00:07:27.439 I gotta try that. 00:07:27.600 --> 00:07:28.000 Yeah. 00:07:28.160 --> 00:07:37.920 Have you have you have you done any auto research work that has landed upstream PRs in like big projects that have a high bar for quality? 00:07:38.319 --> 00:07:39.199 Personally, no. 00:07:39.360 --> 00:07:48.560 Like I think like around the time I left Meta, there was like I think an R like a fused RMS norm, which was like the first AI generated kernel that was merged, but that was like not my work. 00:07:48.639 --> 00:07:52.240 That was like Laura Laura's work on the kernel agent team. 00:07:52.480 --> 00:07:59.839 I do remember though, the main experience we had at the time was that like the kernel was quite long, and so I think it was important to deslopify it. 00:08:00.000 --> 00:08:10.800 I think there was sort of sometimes a bit of an ideological debate between like what's the point of deslopifying it, like just don't read it, to which like the PyTorch core maintainers were like, What are you talking about? 00:08:10.879 --> 00:08:14.079 Because this is maintenance debt, I need to understand every single line. 00:08:14.240 --> 00:08:24.240 I'm not sure which of those two stances is correct because like some of those PyTorch people felt very strongly about these things, are just like no longer working on the team, and so like I actually don't know. 00:08:24.319 --> 00:08:29.040 Like, like, but kind of like one thing that stands out to me at least was like this one. 00:08:29.120 --> 00:08:39.679 Um I think like this is sort of a paraphrase from a talk by Terence Stowl, so it's not exactly what he said, but it's like what I came away with, which is that like he was sort of talking about like I mean, let me just not just talk about kernels. 00:08:39.759 --> 00:08:54.879 Like, basically, like my my hypothesis is that typically by the time you're done writing a complicated kernel, you already have simplified it multiple times because as you're writing something, you're like evolving it, you're fixing it, and you've also like understood it. 00:08:55.039 --> 00:09:04.639 Like by the end of it, if you sort of like artisanally crafted something, basically uh generating, understanding, simplifying is like in some sense the same act. 00:09:04.799 --> 00:09:15.600 But like if we're in an age where people are only doing generation, it becomes like I think a bit of a problem because like let's say your kernel introduces a race condition, but it's like non-deterministic. 00:09:15.840 --> 00:09:18.320 Well, you know, that really sucks for a model training run. 00:09:18.399 --> 00:09:23.759 But maybe it's forgivable for like a tiny little inference service that like runs some sort of speculative optimization. 00:09:23.919 --> 00:09:29.360 So there's like a risk-reward thing where like maybe there's some areas where it need to be really paranoid and this is not acceptable. 00:09:29.519 --> 00:09:42.320 But generally, like when I broadly look at this, I just don't buy the people don't care because you know, as we talked at the beginning of this episode on like if people are freaked out about non-determinism, like boy are they gonna be freaked out about like straight up incorrectness. 00:09:42.399 --> 00:09:44.799 Um bugs that are like non-deterministic as well. 00:09:44.960 --> 00:09:49.120 Like, I think this is bound to cause conspiracy theories, uh, especially at scale, right? 00:09:49.200 --> 00:09:51.519 Like, if let's say 1% of your users only see something. 00:09:51.600 --> 00:09:51.919 Yeah, yeah. 00:09:52.399 --> 00:09:56.000 So, my broad take is like I think we should solve the hard problem. 00:09:56.080 --> 00:09:59.360 I don't think the pro the solution is people are reading the code. 00:09:59.519 --> 00:10:04.879 Like that that to me feels like uh like no, like, because it's let's say the QR problem. 00:10:05.039 --> 00:10:13.840 You know, at first I was like, hey, I'm gonna sort of synthesize all the kernels by all 170 users, and then I'm gonna try to come up with like a beautiful QR kernel. 00:10:14.000 --> 00:10:19.840 Uh, you know, like I've I've yet to do this, it feels onerous every time I think about like sitting down and doing this exercise. 00:10:20.080 --> 00:10:26.720 But at the same time, if I just ask it like Godex and I'm like, hey, Godex, go cook, like, look at all these kernels and tell me what comes up. 00:10:26.879 --> 00:10:28.240 In some sense, I don't even know. 00:10:28.320 --> 00:10:29.279 Did it read everything? 00:10:29.440 --> 00:10:31.519 Like, did it like really understand like everything? 00:10:31.679 --> 00:10:33.200 Did it capture the key ideas? 00:10:33.360 --> 00:10:40.320 Like, maybe there's a slow kernel that had more elegant ideas, will it really capture them or will it really just prioritize the first like dozen kernels that are fast? 00:10:40.480 --> 00:10:41.279 Like, I'm not sure. 00:10:41.440 --> 00:10:45.039 Like, in some sense, like I don't know what the model is doing, like I can't inspect it. 00:10:45.200 --> 00:10:47.600 Like, so it does feel like it's an alien artifact. 00:10:47.679 --> 00:10:48.879 I tell it go cook. 00:10:49.120 --> 00:10:59.360 But my hope is that like we can sort of figure out like how to come up with beautiful kernels from a large sum of fast but slop kernels. 00:10:59.600 --> 00:11:07.440 I just think like it's maybe in the beginning there's a bit more manual than we'd like, but I'm I'm fairly confident we can figure out like good tooling for for problems like this. 00:11:07.759 --> 00:11:24.720 So, first of all, I I gotta say, I I f I find I find it very interesting the proposition that a like a uh a race condition, slightly buggy kernel that's like good 99% of the time but and really fast, maybe it's

good enough in some cases. 00:11:24.799 --> 00:11:26.000 I hadn't considered that before. 00:11:26.159 --> 00:11:27.120 Maybe maybe it is. 00:11:27.360 --> 00:11:28.399 Um for inferences. 00:11:29.279 --> 00:11:35.440 So my goal in auto research here is to think about what's the process that I as a human would do. 00:11:35.600 --> 00:11:40.799 And I think that both for humans and agents, it's very hard to optimize for multiple things at once. 00:11:40.960 --> 00:11:52.480 And every time that I've had to myself do like deep like kernel performance work, it's almost always where I have a phase where I'm trying to reach speed of light. 00:11:52.559 --> 00:12:05.759 And during that phase, you know, I'm adding restrict places, I'm adding in assumes, I'm adding in this and that, I'm rebenchmarking, I'm trying things, I'm if-defing stuff out, you know, I'm I'm I've got some dead code branches, etc. 00:12:06.159 --> 00:12:09.039 And then like, and then I reach performance. 00:12:09.120 --> 00:12:10.639 And I'm like, ah, great. 00:12:10.799 --> 00:12:11.840 This is the right thing. 00:12:12.000 --> 00:12:22.799 And then I've got this mess, and then what I do is I go back and I'm like, okay, okay, I'm gonna delete all the if-defs, I'm gonna, I'm gonna, you know, take out those like if I take out this assume, is it fine? 00:12:23.039 --> 00:12:26.879 Okay, and you know, and then I've got I've always got like four versions of this code. 00:12:27.039 --> 00:12:31.759 And it's like, okay, well, like I can get I can delete this variant that I don't need anymore, and I can delete that variant. 00:12:31.919 --> 00:12:43.120 And it's like I always do this process where I've first spent a bunch of time trying to find speed of light, and then once you've found speed of light, then like you have an Oracle answer. 00:12:43.279 --> 00:12:46.960 And then I can try to write better, cleaner versions. 00:12:47.200 --> 00:13:03.840 And so I think it's it's to me, it is like a natural process to do this V-shaped thing, or maybe not V or upside down V, where like you're you're climbing up to find the fast thing, and then like you're climbing down to make it not so ugly. 00:13:04.000 --> 00:13:11.679 And I think one of the, you know, maybe maybe I'm wrong about this idea that you or I said that like you can't always optimize for two things at once. 00:13:11.919 --> 00:13:16.000 Like it's harder to optimize for multiple things, harder to to optimize for multiple things. 00:13:16.159 --> 00:13:23.519 And I think that's true, but I think part of that is just that we don't have the right reward metrics and safeguards in place. 00:13:23.600 --> 00:13:35.840 That like the to me, if if with my auto research process always generating these like 10,000 line, you know, slop things that's got all this dead code over, that to me means I I haven't put the right guardrails in place. 00:13:35.919 --> 00:13:39.919 I haven't put the right linters and the right things to like for code quality.

00:13:40.080 --> 00:13:46.240 And if I did more of that, then like I could probably generate a better kernels. 00:13:46.320 --> 00:13:48.720 And also I think that would help with the stalls. 00:13:48.879 --> 00:13:53.440 And yeah, not not everything can be like deterministically checked or machine checked. 00:13:53.519 --> 00:14:05.919 The other thing is like I so right now I have in my framework I I uh you know the model can go off, I I let it it can do a macro optimization, a micro optimization, and then a combine. 00:14:06.000 --> 00:14:08.720 And I do have a fourth type called simplify. 00:14:08.879 --> 00:14:13.279 And usually I don't let it run simplify during the main auto research loop. 00:14:13.440 --> 00:14:14.240 But maybe I should. 00:14:14.320 --> 00:14:16.399 Maybe it should be simplifying as it goes. 00:14:16.559 --> 00:14:25.200 I just the reason I set it up this way is that I've I guess I've always done some of the simplification as I go, but usually I make a mess of things until I find speed of light. 00:14:25.519 --> 00:14:29.120 Yeah, it's an like your harness is like reflecting like your own process. 00:14:29.519 --> 00:14:29.759 Yeah. 00:14:30.000 --> 00:14:30.320 Yeah. 00:14:30.480 --> 00:14:31.039 I'm not sure. 00:14:31.360 --> 00:14:33.360 Like I think it's hard for me to say this is right or wrong. 00:14:33.519 --> 00:14:36.320 Like it's this seems like much more of a like personal choice. 00:14:36.399 --> 00:14:39.039 So I can see argument good arguments being made either way. 00:14:39.279 --> 00:14:39.679 Yeah. 00:14:39.840 --> 00:14:40.240 Yeah. 00:14:40.480 --> 00:15:02.080 So yeah, it's I I I find I find the the the the mission that that you guys have at core automation is very compelling to me because like it seems it seems very evident that like if you can make the model really good at automating the development of the model, then like you can like then it can all it can do all the rest, right? 00:15:02.240 --> 00:15:04.399 You can just sit back and sip tea. 00:15:05.440 --> 00:15:16.080 I mean what's cool about it too is that like I uh again, because like for me this is all new, but like by virtue of focusing on like on a few problems, I think it simplifies like a lot of things. 00:15:16.240 --> 00:15:24.559 Like uh, you know, we don't have to have like a model that's like say necessarily generally capable of all sorts of things, but it it primarily needs to be good at things that are tasks in the lab. 00:15:24.799 --> 00:15:28.879 So like you can think of every individual person in the lab as like an eval in some sense. 00:15:29.039 --> 00:15:29.279 Yeah. 00:15:29.679 --> 00:15:31.279 And I think this is like very helpful for me. 00:15:31.440 --> 00:15:37.360 And you know, when you look at like the kinds of tasks people do in the lab, like it's just it becomes like very self-evident, like how you should do the rest. 00:15:37.519 --> 00:15:46.159 That like, let's say for kernels at least, like none of the evals I

know of open source, and I, you know, I know this because I've also created a couple of them, are like sufficient for what we want. 00:15:46.320 --> 00:15:56.799 And then it's like very cool thinking through like basically like how do you like basically what would you know the way I explain this to my colleagues is like what evals would help us like aura farm really hard? 00:15:57.039 --> 00:16:01.200 You know, and then these tend to be like capabilities that you yourself think are really impressive. 00:16:01.360 --> 00:16:01.519 Yeah. 00:16:01.679 --> 00:16:04.320 Like like what do I think is at the limits of my capabilities? 00:16:04.559 --> 00:16:10.159 And if I can get like our next release to do those things, you know, that then I think I'd be really, really proud of the work we did. 00:16:10.399 --> 00:16:10.799 Yeah. 00:16:11.039 --> 00:16:16.559 You know, and the the another the last point I'll touch on before I let everybody go and I'll go to sleep. 00:16:16.720 --> 00:16:23.840 You know, Connor, you mentioned earlier that like like the stuff you were doing, you know, like I stopped it because it it wasn't valuable. 00:16:24.080 --> 00:16:39.120 And I I just keep coming across places in which the usage data, the like the experimentation, like the the history of what has been tried is just so valuable. 00:16:39.360 --> 00:16:53.440 It's valuable for like one learning like an you can analyze it, you can improve things, you can analyze how particular tools or frameworks were used, and like once you have data at scale, like you can learn a lot from that. 00:16:53.600 --> 00:17:07.759 It it's it's out it's valuable for like simulating how you know how these long horizon tasks work, which you know can tell us a lot about how various like frameworks and hardware might perform. 00:17:07.920 --> 00:17:13.920 And then it can be valuable for training models for for doing post-training and RL. 00:17:14.000 --> 00:17:21.519 And it can be valuable for like extracting, like, oh, like maybe this is something that would be a good, you know, a good benchmark. 00:17:21.599 --> 00:17:24.640 And I I just I think there's so much value in all this data. 00:17:24.720 --> 00:17:39.440 And like that's one of the things I think is great about GPU mode is it's just it's this, you know, it's this farm that's producing all of this super high value data, like open data that we can all benefit from. 00:17:39.759 --> 00:17:51.200 I mean, Bryce, honestly, to me, it seems like the the more the more you're talking, like the more I get excited about this idea of like, why don't we just make it easier for people to compete in our competitions using open model tokens? 00:17:51.440 --> 00:17:57.359 We can subsidize the tokens by picking the problems to be the kernels that show up in the open models. 00:17:57.519 --> 00:18:05.920 And we get this like nice recursive self-improvement loop with people doing search over the best prompts collectively, because I think that's like highly non-obvious. 00:18:06.400 --> 00:18:11.359 For this, for these sprints over the weekend, like I and the Connor's definitely gonna have to cut this part out. 00:18:16.799 --> 00:18:26.480 Because it was like advertised as like, come, like we're gonna do agentic engineering on this hard problem, like they didn't have to learn anything about CUDA first. 00:18:26.640 --> 00:18:30.240 What they did was they had to learn how to set up auto research. 00:18:30.400 --> 00:18:43.359 And then, like on the second day, once they'd produced some auto research results, then like we like sat down and I talked to them about like the problem and taught them a bit about GPU architecture, and like they were able to learn more because they'd already already accomplished something. 00:18:43.440 --> 00:18:46.559 They they didn't know what they'd accomplished yet, because the agent had accomplished something. 00:18:47.759 --> 00:18:50.880 Yeah, but but but but they got like that sort of first like dopamine rush, right? 00:18:50.960 --> 00:18:51.920 Like I made something faster. 00:18:52.400 --> 00:19:21.839 Definitely like one thing I I I have this vision for like in the future of like doing, I guess the way I I know you said you weren't such a fan of like in-person events, but I'm imagining like GTC Auto Research Academy, where like you just like you have two hours every day where you come you give everybody unlimited inference for the week, you come in for two hours every day, and it's like a study group because you know you're like in between the study group meetings, people are going and running stuff and producing results, and then you come back and you you check in. 00:19:21.920 --> 00:19:38.720 And yeah, I think it'd be I I think it'd be very valuable to have open tokens, and I think people position open tokens as a public good as like this idea of like, oh, it's like it's goodwill, but for me it seems like it's insanely valuable to NVIDIA to do this. 00:19:38.880 --> 00:19:39.440 Yeah, yeah. 00:19:40.160 --> 00:19:57.519 Like, but basically, I just to sort of be specific on the like I think my take on events these days is just like the agentic era has made open-ended judging a bit cooked because it's like how do you tell whether something is amazing versus not if it's just like if Claude one-shot it? 00:19:58.079 --> 00:20:09.440 But if someone did a hill climb on a task that we can all collectively agree on is hard and made like a breakthrough, that's better than what I would do myself as like an advanced user. 00:20:09.599 --> 00:20:12.640 Like, if they can just do that during the hackathon, like I'm very impressed. 00:20:12.720 --> 00:20:16.160 And so, like, it just doesn't matter how they did it, it matters is the the output. 00:20:16.400 --> 00:20:24.319 So I think that I'm very open to, and I think it's gonna become a lot more interesting doing it with open models this year, because like the open models are good finally at these tests. 00:20:24.400

--> 00:20:27.119 I think it's it looks like they've been RL'd on kernel data. 00:20:27.200 --> 00:20:27.440 Yeah. 00:20:27.680 --> 00:20:30.559 And so I think it's gonna be a lot more exciting than it was like a year ago. 00:20:30.799 --> 00:20:42.160 I want to pitch you on another idea I had, which is you know, you talked about the idea of having like a package, like here's how to deploy a you know, like like an inference service to do this like at your company. 00:20:42.400 --> 00:20:48.640 I've been thinking about this idea of, well, you know, what we want to do is we we need two we need GPUs for two things. 00:20:48.720 --> 00:20:54.799 We need GPUs one to run the tests and the benchmarks, and two, we need GPUs to run the inference. 00:20:55.039 --> 00:21:01.920 And you know, there's like web search is something that like runs like inference, like, you know, on the inference provider side. 00:21:02.079 --> 00:21:11.359 Like, why couldn't we have like a you know inference provider side like tool call that just like like like that's like modal that just like runs the serverless thing there? 00:21:11.440 --> 00:21:21.119 And then like if you're a company and you have like you know an 88x B200 machine, it can both run your inference and run your serverless, you know, evals. 00:21:21.279 --> 00:21:24.240 Because a lot of people seem gated on the GPU access too. 00:21:24.559 --> 00:21:27.279 Yeah, I I think of these as like very related problems. 00:21:27.519 --> 00:21:31.200 It's just that like I would imagine more people would know what to do with tokens versus a GPU. 00:21:31.279 --> 00:21:33.599 So it's like I would imagine like broader market appeal. 00:21:33.920 --> 00:21:44.480 But yeah, I mean like ultimately, I think the way modal set up how you acquire GPU, I think this is the correct way to do it because it's very incentive aligned with your customers to charge for use. 00:21:44.640 --> 00:21:47.759 Like it's an honest business, basically the way I would frame it. 00:21:47.920 --> 00:21:55.920 Uh and I think and I think paying for access is dishonest because it it's uh it I I just yeah, like because whether you use it or not, it's like it's not great. 00:21:56.000 --> 00:21:56.480 I don't like it. 00:21:56.640 --> 00:21:59.759 Yeah, yeah, it's like if I could like my biggest thing with this. 00:22:00.160 --> 00:22:04.559 Nvidia right now is like I need like I have this stream of like free tokens from Brev. 00:22:04.799 --> 00:22:15.599 I just they if they could just make a serverless product, even if it's like a worse version of modal, then like I would not need to ask for as many tokens and like it could be so much more token efficient. 00:22:15.839 --> 00:22:21.599 By the way, I refreshed the uh I refreshed the tokens for the the profiling service again too. 00:22:21.839 --> 00:22:22.400 Thank you, sir. 00:22:22.480 --> 00:22:22.559 Yeah. 00:22:23.519 --> 00:22:29.039 Um so I think there's enough in there to run through to run through

the end of July. 00:22:29.119 --> 00:22:32.160 But so when do how long does the optimizer contest run? 00:22:32.559 --> 00:22:36.319 The optimizer con we we don't know yet, but we're thinking of making it run for a month. 00:22:36.559 --> 00:22:36.880 Okay. 00:22:37.039 --> 00:22:41.279 I'm just gonna keep refreshing the tokens for the profiling service until somebody tells me to stop. 00:22:42.079 --> 00:22:43.119 Okay, that's that's wonderful. 00:22:43.359 --> 00:22:43.680 Yeah. 00:22:43.839 --> 00:22:44.160 Yeah. 00:22:45.039 --> 00:22:50.240 I think I think we didn't get the chance to talk much about the optimizer competition, but I think it's it's pretty it's pretty sick, actually. 00:22:51.039 --> 00:22:52.079 I'll tell you more about it next time. 00:22:52.319 --> 00:22:53.440 Uh yeah, I'm very excited. 00:22:53.519 --> 00:22:54.799 I I saw I started looking into it. 00:22:54.880 --> 00:22:56.000 Yeah, it looks like it's gonna be very cool. 00:22:56.799 --> 00:22:58.079 Do you mind if I just tell you about it for a bit? 00:22:58.240 --> 00:22:58.559 Yeah, sure. 00:22:58.960 --> 00:23:00.079 I think just okay. 00:23:00.240 --> 00:23:09.839 So so the optimizer competition, like like the sort of main idea was that we we've been very interested in picking tasks where transformers obviously fall flat on their face. 00:23:10.000 --> 00:23:12.480 And so the task we picked is one that involves depth. 00:23:12.559 --> 00:23:14.799 It's basically like repeated squaring and modulo. 00:23:15.119 --> 00:23:22.799 You try to make any like the largest transformers run well on this task, it's gonna run flat on, it's gonna fall flat on his face, and it's gonna try and cheat the task. 00:23:22.960 --> 00:23:40.480 And intrinsically, the only models that have like the right inductive biases to solve these problems are those that include like basically some form of looping or adaptive depth, or like basically and they basically involve like scaling training time compute to sort of like solve the task. 00:23:40.640 --> 00:23:53.279 And these are very exciting because if you sort of if you maybe have a brief physics background, like a lot of these, like anything that involves looping becomes like basically it turns into mush, like basically it just goes either to zeros or it goes to nan. 00:23:53.359 --> 00:23:55.039 Like the system has become very unstable. 00:23:55.200 --> 00:24:04.960 And so what we're very excited about here is we want people to come up with interesting like model architecture ideas to solve this very like simply simply formulated but very nasty problem. 00:24:05.119 --> 00:24:13.039 But because these architectures are quite unstable, they will also need to come up with sophisticated optimizers to solve this problem. 00:24:13.200 --> 00:24:16.960 And the sort of user experience we borrowed is the exact same as kernelbot. 00:24:17.039 --> 00:24:21.839 It's like you submit a single Python file and we run all of these evaluations for you. 00:24:22.160 --> 00:24:27.279 Right now we're basically going through like a trial period where people are reward hacking us to death, which is okay. 00:24:27.440 --> 00:24:29.359 It helps us like have the rules be more robust. 00:24:29.519 --> 00:24:31.279 We're gonna launch for real this Thursday. 00:24:31.359 --> 00:24:37.200 And the competition will be open for about a month, and we're kind of excited to sort of see who comes out of the woodworks here. 00:24:37.519 --> 00:24:54.000 What's been interesting is that like there's this like you know, there's like it's sort of this interesting intersection between people that do complexity theory and model architecture, where they will write very specific proofs as to things like, hey, like a diffusion model can't actually learn long sequence like video generation. 00:24:54.160 --> 00:24:58.480 And I I find this like to be really, really compelling, and I just can't wait to see what people do. 00:24:58.720 --> 00:25:07.680 Generally, for me, I think GPU mode really becomes healthier if people are aware of like frontier level problems, and this is like a frontier level problem that I think many, many people are interested in. 00:25:07.920 --> 00:25:10.400 So I think we made it very agent-friendly. 00:25:10.480 --> 00:25:11.759 The GPUs are all free. 00:25:11.920 --> 00:25:14.720 You get an hour of training time basically a day. 00:25:14.880 --> 00:25:16.400 So, you know, please have fun. 00:25:16.480 --> 00:25:23.119 The the website is like one layer deeper.ai, and I think it's gonna be like one of our most interesting challenges yet, and I'm very excited about it. 00:25:23.359 --> 00:25:28.720 Yeah, no, I think it's I mean, I not be not being not being an ML person. 00:25:28.960 --> 00:25:39.200 I uh I I like being more a kernel person than an ML person, it's a little out of my depth, but from what I've read up on it, it it looks like it's a very, a very cool problem. 00:25:39.279 --> 00:25:44.079 And um, I'm very excited to see what people will uh will will be able to do with it. 00:25:44.160 --> 00:25:53.119 And it also seems like it's something that is probably not in the training set and is like a frontier problem that we'll really be able to see what people can do with auto research for this. 00:25:53.279 --> 00:25:53.599 Yeah. 00:25:53.759 --> 00:25:54.240 Yeah. 00:25:54.559 --> 00:26:01.839 Alright, one one last one last question, and it hopefully will be a short one, because I know Bryce, you need to go to sleep because you're over in the Europe's. 00:26:02.400 --> 00:26:13.359 One of my it's not really a hobby, but something I've been doing over the last several months is I I Well, the details don't matter, but I listened to a lot of No no no, tell us the details. 00:26:13.599 --> 00:26:14.079 The details? 00:26:14.160 --> 00:26:15.119 Alright, very quickly. 00:26:15.279 --> 00:26:23.119 I mean, I have a a vibe-coded podcast player called Pod God, which uh it was gonna get released in the

next couple weeks. 00:26:23.279 --> 00:26:26.319 But I've been using it since the beginning of February, and it's phenomenal. 00:26:26.480 --> 00:26:30.400 One of the things it allows you to do is it allows you to follow personalities. 00:26:30.480 --> 00:26:38.000 So there's something called the personality meta podcast, which is just a curated feed of every single time that person goes on a different podcast. 00:26:38.079 --> 00:26:43.920 Because there are some people that I care about and I want to listen to any time they're being interviewed, but I don't want to have to go and listen to example. 00:26:44.240 --> 00:26:46.160 The person that I'm about to mention is Boris Cherney. 00:26:46.240 --> 00:26:48.880 Like, but the thing is is one No idea who that is. 00:26:49.200 --> 00:26:49.359 What? 00:26:49.440 --> 00:26:50.960 The creator of Claude Code? 00:26:51.200 --> 00:26:52.079 No, no idea. 00:26:52.319 --> 00:26:52.960 That's very sad. 00:26:53.119 --> 00:27:04.079 He's like, he's honestly would be my number one pick to have as like a future guest because every every time he is interviewed, I think he has just like amazing takes. 00:27:04.240 --> 00:27:16.000 But anyway, so like there's a bunch of podcasts that one I don't want to subscribe to, like there's like Sequoia has this like kind of there's a bunch of these like VC money podcasts, but every once in a while they have someone interesting on. 00:27:16.079 --> 00:27:21.440 So like I don't want to have to subscribe to this and then like basically just like download the thing and then delete it. 00:27:21.599 --> 00:27:22.480 I just want the interviews. 00:27:22.640 --> 00:27:25.359 Anyways, that's I that's not long story short, that's long story long. 00:27:25.599 --> 00:27:34.799 Whenever you listen to Boris Cherney, like one out of every two podcasts, he gets like book recommendations or he gives book recommendations, and they've been phenomenal. 00:27:35.039 --> 00:27:45.759 The first one that he recommended, he actually mentioned in the Peterman podcast, which is like a former, I think, Instagram or Facebook guy, and he's just been it's it's great interviews. 00:27:46.000 --> 00:27:54.079 He interviewed uh Boris Cherney, and then he mentions that during the interview for Anthropic, he ended up talking about some book by Greg Egan. 00:27:54.319 --> 00:27:57.839 You know, this is back in like Claude, I don't know, 3.7, 4.0 days. 00:27:57.920 --> 00:28:02.319 I immediately go to like whatever ChatGPT ask, like, what book is he referring to? 00:28:02.559 --> 00:28:08.640 And it ended up being Permutation City, and that is like part of a trilogy that has Diaspora and Shield's Child. 00:28:08.720 --> 00:28:15.119 He on a more recent podcast mentioned Werner Vinge's A Fire Upon the Deep, which is in the Zones of Thought trilogy. 00:28:15.279 --> 00:28:17.039 I'm like one down, two to go. 00:28:17.279 --> 00:28:20.640 I bring all of this up, where's the question, to ask you? 00:28:20.799 --> 00:28:31.599 So you wrote the Robot Overlord manual, and you're the first person actually that we've interviewed on this podcast, which is a little bit surprising, that is as high on AI as Bryce and I. 00:28:31.920 --> 00:28:38.480 We we've interviewed some folks over the last little bit, but not like not people that are sipping the Kool-Aid as much as we are. 00:28:38.720 --> 00:28:44.240 And so uh the the answer to this question may be I don't have any recommendations. 00:28:44.480 --> 00:28:48.400 My question is, do you have any book recommendations in general? 00:28:48.640 --> 00:29:00.640 And on top of that, specifically, I would love to know if you have like any recommendations in the space of you know future AI, like good these kinds of good reads that like people like Boris Cherney recommend recommending. 00:29:00.799 --> 00:29:01.759 So over to you. 00:29:02.160 --> 00:29:04.720 Yeah, like I mean, I I love reading. 00:29:04.799 --> 00:29:05.920 Like this, these are all my books. 00:29:06.160 --> 00:29:11.599 The listener can't see, but there's also like a bookshelf in the background with what looks like at least a couple hundred books there. 00:29:11.680 --> 00:29:22.880 So the the there's like a couple of other hundred that I've given away over the years, but yeah, I mean, I I like I actually like I I do I I had a tweet about this, which is like my ten favorite books. 00:29:23.119 --> 00:29:23.599 Let me see. 00:29:23.839 --> 00:29:29.119 So I would say on the AI, like, okay, so these are these are the books that were most influential to me. 00:29:29.359 --> 00:29:30.160 There's ten of them. 00:29:30.319 --> 00:29:30.960 So let me see. 00:29:31.039 --> 00:29:32.480 Maybe I can pick two for you. 00:29:33.039 --> 00:29:36.720 Well, we'll include we'll we'll include uh the link to the tweet so you can get all ten. 00:29:36.880 --> 00:29:39.119 You can and feel free to just read all ten off if you want to. 00:29:39.440 --> 00:29:41.440 Okay, like it's a more interesting question. 00:29:41.519 --> 00:29:42.319 Okay, the all ten. 00:29:42.559 --> 00:29:48.160 So my like the book that I think was like the most fun to me was a book called The Journey to the Ants. 00:29:48.400 --> 00:29:51.599 So it's it's it's it's a book about like ant-based intelligence. 00:29:51.839 --> 00:29:57.440 And the reason why I found it very interesting is because like the the kinds of shit ants can do is pretty wild. 00:29:57.519 --> 00:30:05.200 Like, for instance, like you can have like the mother ant, the queen, like control the gender ratio of a population, control like people's tasks. 00:30:05.279 --> 00:30:13.440 Like you can be a worker versus a warrior elastically based on the demands of the colony, and they all sort of spread this message between each other to sort of dictate. 00:30:13.599 --> 00:30:20.960 This is very contrary to how like human societies operate, where we're like, oh, I'm a colonel guy, and I will never not do kernels, like we're very

inelastic. 00:30:21.119 --> 00:30:26.160 So it's like pretty remarkable for me to see an example of um of societies that were not this way. 00:30:26.240 --> 00:30:27.599 And I think this book is wonderful. 00:30:27.839 --> 00:30:30.079 Another book I really love is visual group theory. 00:30:30.160 --> 00:30:33.519 So visual group theory, and uh, maybe I would add Burns Euclid here. 00:30:33.680 --> 00:30:42.960 What I found really lovely about these books is that like any anything that's algebraic tends to be like very dry and like kind of unintuitive to read because you just see look at these long formulas. 00:30:43.119 --> 00:30:49.920 Both of these books are pretty like beautiful, they almost feel like pieces of art where people are sort of like going through like these arguments like very rigorously. 00:30:50.079 --> 00:31:01.200 And so for me, they really made me appreciate like the idea of a beautiful explanation, which is like typically when I think of like my job, I don't necessarily think of it as like producing like beautiful code or fast code. 00:31:01.359 --> 00:31:04.640 I think of it as being able to produce a good explanation for why the code is fast. 00:31:04.720 --> 00:31:06.960 Like that's kind of almost like my love language. 00:31:07.200 --> 00:31:09.599 Another book I've adored is The Theory of Fun. 00:31:09.680 --> 00:31:13.920 So the Theory of Fun is a book on why people like playing games. 00:31:14.079 --> 00:31:18.880 It's sort of like the sort of the main thesis of the book is that playing is basically the way we learn. 00:31:19.039 --> 00:31:25.039 So I've carried that a lot into my projects where I've tried to make a lot of the things we've done with GPU mode fun. 00:31:25.200 --> 00:31:28.319 And as a result, people end up learning more and might find it more valuable. 00:31:28.400 --> 00:31:33.920 So a lot of what I learned with GPU mode was like basically early days of playing World of Warcraft and Warcraft 3. 00:31:34.000 --> 00:31:36.319 So like a lot of the lessons, funnel enough, do apply. 00:31:36.559 --> 00:31:46.880 I would say two technical books that I've I think I'm gonna skip over a couple of books, but two technical books that I think I highly recommend to everyone are uh game programming patterns and working with legacy code. 00:31:47.039 --> 00:31:50.960 Working with legacy code often surprises people as one of my favorite books of all time. 00:31:51.039 --> 00:31:56.079 But I found it really wonderful when I'm just like plopped into like a monstrous code base. 00:31:56.240 --> 00:31:58.559 How do you sort of like quickly make sense of it? 00:31:58.640 --> 00:32:03.680 And like what are good design patterns to make codes like simpler and easier to reason about? 00:32:03.839 --> 00:32:13.519 I think a lot of like my contributions in the in past projects has been like going into a complicated code base and like understanding it and then figuring out how to develop features more quickly on top. 00:32:13.680 --> 00:32:17.839 I think a lot of these ideas will make a comeback if we're like basically in mountains of slop. 00:32:18.000 --> 00:32:22.240 Maybe yet another book that I've really enjoyed, maybe okay, maybe I'll give you two more. 00:32:22.480 --> 00:32:25.200 Like Game Theoretic Foundations for Probability and Finance. 00:32:25.279 --> 00:32:27.680 It's not a finance book, even though it's in the title. 00:32:27.839 --> 00:32:33.680 It's a book about rewriting all the axiom of probability theory with game theory in a way that makes them a lot more interpretable. 00:32:33.839 --> 00:32:38.559 And so I've also really enjoyed this kind of book because I've always enjoyed like more wacky math. 00:32:38.640 --> 00:32:40.079 It just can help you build intuition. 00:32:40.160 --> 00:32:43.119 And if you didn't like one notation, maybe you'll appreciate another. 00:32:43.359 --> 00:32:46.880 The last one is the structure and interpretation of classical mechanics. 00:32:47.039 --> 00:32:50.799 I'm sure many of you might have read the structure interpretation of like computer programs. 00:32:51.359 --> 00:32:52.640 That's what I thought I was gonna say. 00:32:52.960 --> 00:33:06.720 I I think this book is better because it basically goes over like it's writes short programs for physics written in Lisp, and you basically end up with nature as like a short series of like self-modifying Lisp programs. 00:33:06.960 --> 00:33:17.440 It's one of those like books that like blow your mind the first time you read it, but like again, like as we get into the age of auto research and architecture research, I suspect a lot of these ideas are are making a big comeback. 00:33:17.599 --> 00:33:28.559 So these are things that I think are are you know there's a lot more, but like maybe you know, berserk and necronomicon, maybe with some pieces of fiction that I think are just like absolutely gorgeous. 00:33:28.799 --> 00:33:36.400 I I've always been a sucker for like there's some piece of knowledge that if you acquire it, I don't know, you'll blow up because you'll be rendered mad. 00:33:36.640 --> 00:33:42.960 Like I don't know why, but it's like one of those things that I've always like I've been like a sucker for this kind of like fantasy, fantasy work. 00:33:43.119 --> 00:33:44.720 But yeah, all these books are phenomenal. 00:33:44.799 --> 00:33:47.839 I'll share the full list if you like with you, like in a few minutes. 00:33:48.079 --> 00:33:51.680 But yeah, I I think these are sufficiently out of distribution to give you better ideas. 00:33:52.000 --> 00:33:57.039 I mean, that was I feel bad we haven't asked our other past guests because that was that was uh amazing. 00:33:57.200 --> 00:34:02.480 Even better than not that I you know had expectations for the quality of your answer, but uh even better than I expected. 00:34:02.640 --> 00:34:03.279 Go ahead, Bryce. 00:34:03.599 --> 00:34:11.760 I I have a I have a related question, which is it sound

like it sounds like you have been a person who's played a lot of video games in the past. 00:34:12.000 --> 00:34:14.480 And I'm wondering if you have recommendations for video games. 00:34:14.880 --> 00:34:16.239 Uh let me see. 00:34:16.559 --> 00:34:19.760 Um so what are the greatest games I've played? 00:34:19.920 --> 00:34:27.760 So so the one of the issues with like like I think the greatest game I've ever played is probably Dota 2, but I would not recommend you play it because it just will take over your life. 00:34:27.920 --> 00:34:40.320 So I think it's like an optimizer's dream in some sense, and it teaches you how to never get mad at anything because like it's very formative to be insulted at in like five different languages and still manage how to win a game. 00:34:40.559 --> 00:34:41.679 So it's very formative. 00:34:41.840 --> 00:35:00.480 But but I would say, like, I mean, if I'm thinking like games that like left like a deep impression on me, I would say recently, I I would probably say it was like Expedition 33 as like sort of a story-based game, like really left like a deep impression on me, and I was like a big sucker for like role-playing games like growing up, you know, all the way ranging from like Pokemon to like Chrono Trigger. 00:35:00.639 --> 00:35:03.599 As a result, I've also really enjoyed Final Fantasy XVI recently. 00:35:03.679 --> 00:35:06.559 I've been playing Final Fantasy 7 as well, which I really enjoy. 00:35:06.719 --> 00:35:14.320 But yeah, otherwise, like these days, like I mostly spent a lot of time playing Slay the Spire, but I also feel like I spent too much time doing it. 00:35:14.400 --> 00:35:18.800 So if you give me an optimization puzzle, usually I can get obsessed with it. 00:35:18.960 --> 00:35:25.039 If you give me something that's fast and reflex-based, like Fury or Katana Zero, I'm also like a big sucker for those. 00:35:25.119 --> 00:35:27.119 Like I enjoy sort of this like flow feeling. 00:35:27.280 --> 00:35:30.960 And I think these are all games that will give you like a major, major uh sense of flow. 00:35:31.039 --> 00:35:33.039 So they're all treats if you haven't played them. 00:35:33.360 --> 00:35:41.840 So my my actually my my oldest friend, the person that I've had the longest contact with, is somebody I've never actually met in person. 00:35:42.000 --> 00:35:47.679 And he and I, I used to play a lot of MUDs, text-based games when I was a kid, and he and I played MUDs together. 00:35:47.920 --> 00:35:51.840 And in particular, my favorite type of MUD was a roleplay-intensive mud. 00:35:52.079 --> 00:35:55.840 So that's where everybody has to be in character at all times. 00:35:56.000 --> 00:36:08.639 I primarily played either Star Wars or Lord of the Rings themed MUDs, and these were all like community-run multiplayer games, and nobody was really concerned about the fact that like Star Wars and Lord of the Rings were IP that somebody else owned. 00:36:09.039 --> 00:36:14.159 But, anyways, him and I were having this conversation the other day where it's a it's a text-based game. 00:36:14.320 --> 00:36:25.119 Like, it is you could just like hook up, I'm sure, like a frontier model to it, and you know, the problem that these games have today is like, you know, the NPCs are not particularly good and there's just not a lot of players. 00:36:25.280 --> 00:36:29.280 And him and I were talking about like, well, we could just make as many players as we want. 00:36:29.360 --> 00:36:32.800 You know, you could just like you wouldn't even need to code them into the game. 00:36:32.960 --> 00:36:40.400 You could just connect an existing agent to an existing mud and just have it interact the game as it would any other player character. 00:36:40.559 --> 00:36:43.920 And I just like I I can't haven't been able to get this idea out of my head. 00:36:44.000 --> 00:36:52.079 And if I had if I had more infinite inference, I would go build my like AI mud game, and I'm sure I'd have a blast with it. 00:36:52.320 --> 00:37:01.199 I I I think we should try it out because I I think one hybrid idea I'd be very optimistic about would be something like uh like an in-person Dungeons and Dragons game. 00:37:01.360 --> 00:37:01.760 Oh yes. 00:37:02.000 --> 00:37:12.079 Where like the like where the DM can basically choose to like basically where there's a DM who's constantly prompting to try and anticipate where like things might go. 00:37:12.239 --> 00:37:13.760 That I'm I'm really optimistic about. 00:37:13.920 --> 00:37:21.840 But but I think a lot of the I think issues I foresee with current alums is like post-compaction, maybe they'll lose everything, like they'll just have no idea what the fuck's going on with the game. 00:37:22.239 --> 00:37:26.320 We should do a whole we we should we should talk some other time about my thoughts on compaction. 00:37:26.480 --> 00:37:29.119 I think sometimes compaction helps and sometimes it hurts. 00:37:29.360 --> 00:37:30.320 In the short. 00:37:30.480 --> 00:37:30.880 Yeah. 00:37:31.119 --> 00:37:39.039 The thing that the thing that's most valuable that I get that happens at compaction is most harnesses on compaction will reread skills. 00:37:39.199 --> 00:37:47.199 And that can sometimes help because then it's like a reminder of like, were you following, were you following like the prompts and skills that you were originally given? 00:37:47.360 --> 00:37:51.760 Anyways, I should go, it's very it's getting very late here, and we should let Mark get going. 00:37:52.320 --> 00:37:53.119 Good night, Bryce. 00:37:53.199 --> 00:37:54.639 And then thank you for inviting me, Bryce. 00:37:55.760 --> 00:37:57.039 It was a pleasure to chat with you. 00:37:57.360 --> 00:38:06.880 Be sure to check these show notes either in your podcast app or at adspthepodcast.com for links to anything we mentioned in today's episode, as well as a link to a GitHub discussion where you can leave thoughts, comments, and questions. 00:38:07.280 --> 00:38:07.840 Thanks for listening. 00:38:07.920 --> 00:38:09.519 We hope you enjoyed and have a great day. 00:38:10.400 --> 00:38:15.519 Low quality, high quality, that is the tagline of our podcast. 00:38:15.840 --> 00:38:16.719 That's not the tagline. 00:38:16.960 --> 00:38:19.760 Our tagline is chaos with sprinkles of information.