Shawn Wang: preference

Transcript context

…So, um, typically they're referred to as a reward model and that's a, you know, that's like a term from reinforcement learning. And for that, you can start off with something very simple, which is, do you think the user is going to respond to it? That's a simple one. So you can, you can train, you can take 50 million messages and, and look at all the sorts of messages users reply to, which ones they don't. And then you can train this, this reward model to evaluate completions. And so it knows like, okay, if you say this, the user is not going to respond. So don't bother sending it to the user. If you say this, the user is definitely going to engage with it. So send them, send them that. There's an interesting parallel between MLAs and MLAs. I think we use at the top, spreading out to different experts and then at the bottom with rejection sampling, choosing from different paths. I totally agree. That's the stuff that is the future of AI. I think that's the exciting stuff. And there's a parallel between that. Why was AlphaGo able to be superhuman? Right. It's this ability to generate many different paths. Tree search. And tree search. Exactly. So I think if you want to talk about what would intelligence look like, it looks much more like tree search. Combining the generative nature of these LLMs with a really good tree search. And that's what opening I've done with O1 and O3.…

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