Alessio Fanelli: belief

Transcript context

…I've just, I've just, you know, been connecting the dots. I mean, this was the key thing that a bunch of labs were working on, which is like, can you create a reward signal? Can you teach yourself based on a reward signal? Whether you're, if you're trying to learn coding or math, if you could have one model say, uh, be a grading system that says like you have successfully solved this programming assessment and then one model, like be the generative system. That's like, here are a bunch of programming assessments. You could train on that. It's basically whenever you could create a reward signal for some task, you could just generate a bunch of tasks for yourself. See that like, oh, on two of these thousand, you did well. And then you just train on that data. It's basically like, I mean, creating your own data for yourself and like, you know, all the labs working on that opening, I built the most impressive product doing that. And it's just very, it's very easy now to see how that could like scale to just solving, like, like solving programming or solving mathematics, which sounds crazy, but everything about our world right now is crazy. Um, and so I think if you remove that whole, like, oh, that's impossible, and you just think really clearly about like, what's now possible with like what, what they've done with O1, it's easy to see how that scales. How do you think about older GPT models then? Should people still work on them? You know, if like, obviously they just had the new Haiku, like, is it even worth spending time, like making these models better versus just, you know, Sam talked about O2 at that day. So obviously they're, they're spending a lot of time in it, but then you have maybe. The GPU poor, which are still working on making Lama good. Uh, and then you have the follower labs that do not have an O1 like model out yet. Yeah. This kind of gets into like, uh, what will the ecosystem of, of models be like in the future? And is there room is, is everything just gonna be O1 like models? I think, well, I mean, there's definitely a question of like inference speed and if certain things like O1 takes a long time, because that's the thing. Well, I mean, O1 is, is two things. It's like one it's it's use it's bootstrapping itself. It's teaching itself. And so the base model is smarter. But then it also has this like inference time compute where it could like spend like many minutes or many hours thinking. And so even the base model, which is also fast, it doesn't have to take minutes. It could take is, is better, smarter. I believe all models will be trained with this paradigm. Like you'll want to train on the best data, but there will be many different size models from different, very many different like companies, I believe. Yeah. Because like, I don't, yeah, I mean, it's hard, hard to predict, but I don't think opening eye is going to dominate like every possible LLM for every possible. Use case. I think for a lot of things, like you just want the fastest model and that might not involve O1 methods at all.…

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