Shawn Wang: preference

Transcript context

…So I think we are ignoring the fact that you have a giant shortcut, which is you can behavioral clone everything humans already know. And that's what we solved with LLMs. We've solved behavioral cloning, everything that humans already know. Right. So like today, maybe LLMs is like behavioral cloning every word that gets written on the internet in the future, the multimodal models are becoming more of a thing where behavioral cloning the visual world. But really, what we're just going to have is like a universal byte model, right? Where tokens of data that have high signal come in, and then all of those patterns are like learned by the model. And then you can regurgitate any combination now. Right. So text into voice out, like image into other image out or video out or whatever, like these like mappings, right? Like all just going to be learned by this universal behavioral cloner. And so I'm glad we figured that out. And I think now we're back to the era of how do we combine this with all of the lessons we learned during the RL period. That's what's going to drive progress. I'm still going to pressure you for a few more early opening stories before we turn to the ADET stuff. On your personal site, which I love, because it's really nice, like personal, you know, story context around like your history. I need to update it. It's so old. Yeah, it's so out of date. But you mentioned GPT-2. Did you overlap with GPT-1? I think you did, right? I actually don't quite remember. I think I was joining right around- Right around then?…

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