Nathan Lambert: prediction

Transcript context

…Can we do just a one-second RL version? So you come from a robotics background, which RL used to be, or maybe still is, state-of-the-art. And then now you're seeing a lot of LLM plus RL, so you have the gym fans, Eureka, you have MPU, which we had on the podcast when they started with RL. Now they're doing RL plus LLMs. Yeah. Any thoughts there on how we got here? Maybe how the pendulum will keep swinging? I really think RL is about a framing of viewing the world through trial and error learning and feedback, and really just one that's focused on thinking about decision-making and inputs in the world and how inputs have reactions. And in that, a lot of people come from a lot of different backgrounds, whether it's physics, electrical engineering, mechanical engineering. There are obviously computer scientists, but compared to other fields of CS, I do think it's a much more diverse background of people. My background was in electrical engineering and doing robotics and things like that. It really just changes the worldview. I think that reinforcement learning as it was back then, so to say, is really different. You're looking at these toy problems and the numbers are totally different, and everyone went kind of zero to one at scaling these things up, but people like Jim Phan and other people that were... You saw this transition in the decision transformer and papers and when people are trying to use transformers to do decision-making for things like offline RL, and I think that was kind of like the early days. But then once language models were so proven, it's like everyone is using this tool for their research. I think in the long run, it will still settle out, or RL will still be a field that people work on just because of these kind of fundamental things that I talked about. It's just viewing the whole problem formulation different than predicting text, and so there needs to be that separation. And the view of RL in language models is pretty contrived already, so it's not like we're doing real RL. I think the last slide that I have here is a way to make RLHF more like what people would think of with RL, so actually running things over time, but a weird lineage of tools that happen to get us to where we are, so that's why the name takes up so much space, but it could have gone a lot of different ways. Cool. We made it one slide before going on a tangent.…

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