Nathan Lambert: evaluation

Transcript context

…No, you have some awesome slides. So for people listening on podcast only, we're going to have the slides on our show notes, and then we're going to have a YouTube version where we run through everything together. Sounds good. Yeah. I think to start skipping a lot of the, like, what is a language model stuff, everyone knows that at this point. I think the quote from the Llama 2 paper is a great kind of tidbit on RLHF becoming like a real deal. There was some uncertainty earlier in the year about whether or not RLHF was really going to be important. I think it was not that surprising that it is. I mean, with recent models still using it, the signs were there, but the Llama 2 paper essentially reads like a bunch of NLP researchers that were skeptical and surprised. So the quote from the paper was, meanwhile, reinforcement learning known for its instability seemed a somewhat shadowy field for those in the NLP research community. However, reinforcement learning proved highly effective, particularly given its cost and time effectiveness. So you don't really know exactly what the costs and time that Meta is looking at, because they have a huge team and a pretty good amount of money here to release these Llama models. This is just the kind of thing that we're seeing now. I think any major company that wasn't doing RLHF is now realizing they have to have a team around this. At the same time, we don't have a lot of that in the open and research communities at the same scale. I think seeing that converge would be great, but it's still very early days. And the other thing on the slide is some of Anthropic's work, but everyone knows Anthropic is kind of the masters of this, and they have some of their own techniques that we're going to talk about later on, but that's kind of where we start. 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?…

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