SemiAnalysis Weekly
00:16:44 · Jordan Nanos, Andrew Megalaa
[00:16:44] S0: good, but Opus 4.6 [00:16:45] S2: was also really good. And like those two models were just like... Yeah, like I think, you know, give it a few more iterations. Like there's... There'll be no need for me to look at a lot of this code. And it was like a very clear trend. And I had to like reconcile that myself because I really like coding at the time. And that was like my thing. Like, was just like good at coding and systems. But yeah, I just had to like figure out like, oh, actually, you know, I have to take the higher level things now. Like, this is no longer a job that like I'll always be doing, like writing code by hand. Though I still think, like, you know, human plus AI is still better than AI alone just because, like, there's, like, a little... This piece of entropy that you can provide that the models don't have themselves. And so, you know, still valuable.
00:17:28 · Max Kan
[00:17:28] S1: Yeah. But funny enough, the the reason why I wanted Andrew to talk about that is I think Brian had a very similar moment for material science when he made the task for a benchmark because he was like, holy shit. You know? If if any human material scientist could do what, like, Astra just did, they're probably, you know, a top... Like, they're definitely better than the media material scientist already. Like, I'm like, I just watched Astra do this. And so I think we're probably going to sort of have these similar moments in lots of other... In domains like chip design and material science and sort of all the other GDPX domains in the imminent future. And that's something that you also hope you'll be able to see sort of in the actual benchmark results as we're able to evaluate, you know, Opus 5.5, Astra 6.1, 6.2, or whatever.
00:18:22 · Andrew Megalaa, Jordan Nanos
[00:18:22] S2: I think we got, like, a really good tweet. Someone asked about whether, like, you know, GDPval had human baselines for how long tasks take, you know, to like put things in And this was really funny because like I responded to this and I was like, actually, like it doesn't even make sense to have human baselines anymore. Like in the material science tasks, like Astra solved a subtask of a task better than like a published state of the art reference. Like when the models are like essentially past expert level, how do you even put a human baseline to that? And And some of these chip design tasks are just so long horizon and would take teams at companies to even do... And those, there's no human baseline there. How do you even put a number on it? [00:19:07] S0: It's a lot of money. Whatever the hourly equivalent is in human work hours is a lot of money. And I just threw up on screen the the cost per task for the material science works where the biggest ones, like was at 40 or $50 per task, and many of the models are solving these tasks, some of them in $10, which is I think, a pretty unbelievable statement. Maybe you can talk about reading some of the rollouts in more detail. What were some of the pieces of model behavior that surprised you the most?
00:19:07 · Jordan Nanos, Andrew Megalaa
[00:19:07] S0: It's a lot of money. Whatever the hourly equivalent is in human work hours is a lot of money. And I just threw up on screen the the cost per task for the material science works where the biggest ones, like was at 40 or $50 per task, and many of the models are solving these tasks, some of them in $10, which is I think, a pretty unbelievable statement. Maybe you can talk about reading some of the rollouts in more detail. What were some of the pieces of model behavior that surprised you the most? [00:19:45] S2: Yeah, I mean, the models find very creative ways to solve the tasks. And this is again, going back to the whole unfair verifier thing. People don't expect the way that models solve tasks because models are not human. They just solve things in very weird ways. Like, you know, maybe use module B instead of module A, for example. But sometimes just like have completely different taste in like how things are done. And sometimes it's correct and sometimes it's not. And that's like why we have to like go through this transcript. And it's, like, hard to even go through them sometimes because, you know, we're not human experts. But luckily, you can benefit a lot from chatting with AI about certain rollouts to, like, figure out whether this stuff is, like, correct or not. The the model's actually so smart, like, on their own that with the... With with human probing, you can figure out whether something is, like, fair or not fair. Yeah. It's a very interesting experience for sure.
00:20:44 · Jordan Nanos
[00:20:44] S0: Maybe we can talk about the results. I'm getting... I'm realizing that I haven't even shared the, like, flagship leaderboard screenshot that's at the top of the article yet. Maybe you guys can talk about, like, what you learned about the models themselves. I think particularly we've been going back and forth and trying to explain the difference between the flagship leaderboard across all domains and then compare it to the next one where you've got the leaderboard where you filter out biology, cybersecurity, and alignment where safety classifiers are causing models that have better performance in the weights somewhere to not actually demonstrate it on the tasks in the benchmark?
00:21:28 · Max Kan, Andrew Megalaa
[00:21:28] S1: Yeah. Yeah. I think that's a really good point worth emphasizing because we kinda went back and forth between which of those screenshots was the one that should have been the headline result. The one where you include all the domains versus the one where you exclude cyber, biology, and alignment because those were the three domains where sometimes safety classifiers kicked in and therefore sort of sandbag the score for some models. Like, if you look at Astra on cyber, for example, the score is, like, 0% because it just refused to do any the cyber task. And, ultimately, we decided that the one including all the domains is better. But I think it actually is sort of fair to penalize overly strict safety classifiers to some degree. I mean, I know you've experienced this yourself, Jordan, sort of just doing, you know, your day to day work at SemiAnalysis, then you're you're hit with a cyber safety classifier from, Opus Ration. You're like, well, fuck. I guess I gotta use, like, GLM 5.3 now or whatever. [00:22:32] S2: No. Jordan Jordan was... Jordan had a very interesting insight. He actually figured out that Muse Wonder is, like, a top tier cyber model independently. And you were right. Like, you were totally right. Like, Muse Wonder and Muse Wonderful were just top tier for, like, what you can get out of the box for a cyber model. [00:22:50] S1: Yeah. Yeah. Yeah. That... Which was, again, a very surprising result [00:22:53] S2: Yeah.