Shawn Wang: uncertainty

Transcript context

…Yeah, yeah, yeah. But eventually those will work. Like, so it, it's always a, like, like even we put out vibe. We also are quite, are quite like upfront with like, if the more people use it, there's a lifetime. It's like a car right. After you drive, run, run a certain mouse, it, it is time to shelf it. Right? Yeah. So I, I don't think there's like a, actually like a. Like a good solution. In general, I'm also like a bit I think this is like important for the community to think about, right? But like, is it like a fundamental limitation that any benchmark that goes out? Like, also there's also one thing is that in the past people used to like withhold test set, right? Like squat or something. They used to withhold test set. But then, like, after a while, I think people also realize that like, when you withhold, like MMMU, no, like when you withhold, it's like so much extra work for like the community to like eval on this that they just don't do that, right? It's either your. Dataset becomes, your benchmark becomes unpopular. I think it's also incentive things, right? So if you, let's say you are, you want to run like a contest, right? And then your goal as an academic is to get as much citations as possible on this benchmark paper, right? Like, then you, or like this, this, you want to be as famous as possible. You will not want to withhold the test set, because if you withhold the test set, and then people have, like, there was once, like, I mean, like many years ago, There were even some benchmarks where you had to, like, package your model and send it to them to run. And, like, these benchmarks never ever, like, took off. Like, took off. Just because, like, so at the end of the day, right, it's, like, It's the root problem, like, incentives. Like, it's the, also, the benchmark, the benchmarking problem is also, like, an incentive problem, right? So, like, it's also, like, like, people want to show their model is the best. And then the game masters want to gain as much clout as possible. And I think, also, LMCs also get caught into some, I don't have a, I don't have a take on this, but, like, there's, like, people who also feel that, They are also optimizing for hype, right? Their own cloud, right? So there's all this, I think it's a lot of interesting, like I don't know what field this will be, but I don't know, like, I think there's a lot of papers to be written, right? I mean, about how these incentives like rewards and incentives, like, kind of it might not be soft, so, I don't know. I would say SweetBench is probably the one that's kind of broken out this year as like now a thing that everyone wants to compete on as if you're a coding agent. I don't know if you have a view on it, but it's just, like, it should be known to be hard. You should be able to make progress on it quickly. That makes you popular and cited a lot. Yeah, yeah, yeah, yeah, yeah. Multi modality versus omni modality. So this is a little bit of commentary on GPT 4. 0 and Chameleon. I don't know if you saw the Chameleon paper from Meta. Briefly saw it yeah, I'm not, I didn't really take a look at…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.