00:00:00 · Qasim Wani, Ruby Justice Thelot, Ruby Justice Thelot
[00:00:00] S0: For GDPX, we observed that the average accuracy or the mean reward for all the frontier models was not more than 24%, which means that there's still a lot of room to grow for these models to improve on. Part of the reason why these models are not good at it, and we sort of talk about it in our blog post within the GDPX report, one of the big problems with models like Astra or Carbon and Opus is that they don't know what experiments to prioritize. [00:00:38] S1: We're live with Qasim Wani, who's the founder of Extrapolate, builds RL environments or co founder of Extrapolate, which builds RL environments where AI agents learn to improve themselves by doing frontier research. And it's also contributed to GDPX, which is SemiAnalysis's new eval, which seems really cool and interesting. So, Qasim, welcome to MTS. [00:00:59] S0: Thanks for having me. [00:01:01] S1: So where where should we start? We're just talking before we went on about the new meter report that says that, basically [00:01:13] S2: Since October of last [00:01:14] S0: year? [00:01:14] S1: Yeah. Since October of last year, the rate at which the Anthropic Epoch Capabilities Index benchmark score has been increasing is about double what it was previously. And they have other evidence that's nonpublic in addition. And then all this combined leads them to believe that anthropics models are providing, like, substantial r and d uplift already. So do you have any takes on this? Like, if if this this line is true, like, does this mean we're entering RSI?
00:01:46 · Qasim Wani
[00:01:46] S0: It's one of those, like, if this, then that kind of situations. I think what's interesting to zoom out a little bit is to define our site more concretely. It's something that we're talking about right before the show is different labs have a different perspective on our side and it's it's really interesting how a a close friend of mine, he he works at Google and they claim that they're also working on RSI but his team specifically does not work on model training. So it's like what even is is RSI at this point? What does RSI look like about? It's it's like it's the thing that's in vogue. The way that I would say I think about RSI is can an agent invent the next version of itself? Mhmm. You know, Anthropic in particular, their interpretation is from at least what I've heard, so internally things could be a little different, but from what I've heard, their whole interpretation of RSI is how good are models at doing pre training research, and can agents automate all of pre training. That's sort of what Karpathy allegedly was hired to do at Ant was to automate a lot of pre training research, which is absolutely one of the most critical things because everything is upstream of pre training. But there's a lot more to self improvement beyond just model training. And I think that's where we sort of get lost in the meta, which is why we created the RSI index with semi analysis for GDPX. And the biggest reason why we worked with semi analysis on the RSI capability is because if you look at pretty much every single benchmark out there today when it comes to AI research and this whole like realm of auto research, RSI.
00:03:34 · Ruby Justice Thelot
[00:03:34] S2: Yeah.
00:03:34 · Qasim Wani
[00:03:34] S0: A lot of them like PostTrainBench, Deep RSI and the Valves RSI Index, they're all like hyper focused on can agents do some groundbreaking research. And while that is like an important and a critical capability for models to be really good at, it's, as FrontierLab, you don't just tell Claude, hey, go build the next version of Claude. It's a lot more sophisticated and a more iterative by design. As researchers, we spend majority of our time looking at the data. And so it's very iterative by design. You're not just, researchers don't really have this idea of I'm going invent the next transformer. They're very like iterative by philosophy. And what we see with a lot of these benchmarks, like let's say PostTrainBench in particular, the benchmark has seven tasks and one of the task is can you take one of the open source quen models and post train it to have the best performance on AIME or GSM8K. The problem with those types of tasks is that's not what you do inside the labs. Like inside the labs you're hill climbing on 20 benchmarks at once and not a single benchmark. And that was sort of like our version of RSI. It's like let's just do what we did at the lab. So I guess a little bit of background about me. I started the RSI efforts at xAI working with Guodong Zhang, who 's one of the founders of the company. And our philosophy when we started working on this was to be able to have Grok invent the next version of Grok, it needs to be good at a multitude of It needs to be really good at model training, but it needs to be equally good at identifying what data to train on. It needs to be really good at creating the environments, doing harness design, eval design, infra. And those are like the primitives required to scale a lab. And so that's where I think a lot of the meta gets lost these days where people like, oh, RSI is just can a model train something, but it's actually much broader than that.
00:05:37 · Ruby Justice Thelot, Qasim Wani, Ruby Justice Thelot
[00:05:37] S1: Totally. Yeah. Yeah. Yeah. Mhmm. So how exactly would you define RSI? Like, what is your definition, personally? [00:05:44] S0: I would say it's really can an agent invent a new lab? [00:05:48] S1: A new lab? [00:05:48] S0: A new lab. A new lab. [00:05:50] S2: Interesting. Like, what can one agent invent a new lab? [00:05:53] S0: Or a swarm of agents. [00:05:54] S2: A swarm of agents. [00:05:55] S1: Yeah. Seems like a really high bar. Like, that seems like a lot higher than what most people describe as RSI. I mean, some people's definition is as simple as like, is the model useful to programmers or AI researchers in any way that would cause the cycle time of new model development to shorten?
00:06:17 · Qasim Wani, Ruby Justice Thelot
[00:06:17] S0: Yeah, I think that I would describe that as self improvement and not recursive self improvement. Like the nomenclature here is on the recursive piece, because the whole premise behind RSI is by virtue of models getting good at very specific capabilities, we should achieve super intelligence. And that's sort of why I really like GDPX as a benchmark because it's not just focused on RSI, it's focused on all the critical components that will yield to super intelligence. RSI just happens to be a direct north star towards measuring for superintelligence. So it is a fairly opinionated answer, but that's what you need. You need people that have an opinionated way of measuring for model progress instead of generic improvement in productivity. It's kind of like saying, what is more important, creating good PowerPoint slides or doing drug discovery? They're both important like, to to... Like, if you're a researcher who can't really talk to a lot of, nontechnical people and you have to do some, like, reporting on material science and drug discovery research, if you don't have good taste in PowerPoints, you're probably not gonna be able to, like, get your agenda out the door. And so our version of RSI is just like, can an agent or a swarm of agents invent the next version of a frontier lab? [00:07:42] S1: Well, why can't they do this already? [00:07:46] S0: They're not good at it. That's sort of how we Yeah. [00:07:49] S1: But like why? Like where where did they come up short?