Nathan Lambert: belief

Transcript context

…The disconnect between more sociology work versus computer work already exists, or is it like a recent cross contamination? Because when we had Tri Dao on the podcast, he said FlashAttention came to be because at Hazy they have so much overlap between systems engineer and like deep learning engineers. Is it the same in this field? So I've gone to a couple of workshops for the populations of people who you'd want to include this like R. I think the reason why it's not really talked about is just because the RLHF techniques that people use were built in labs like OpenAI and DeepMind where there are some of these people. These places do a pretty good job of trying to get these people in the door when you compare them to like normal startups. But like they're not bringing in academics from economics, like social choice theory. There's just too much. Like the criticism of this paper that this is based on is like, oh, you're missing these things in RL or at least this decade of RL and it's like it would be literally be bigger than the Sutton and Barto book if you were to include everyone. So it's really hard to include everyone in a principled manner when you're designing this. It's just a good way to understand and improve the communication of what RLHF is and like what is a good reward model for society. It really probably comes down to what an individual wants and it'll probably motivate models to move more in that direction and just be a little bit better about the communication, which is a recurring theme and kind of my work is like I just get frustrated when people say things that don't really make sense, especially when it's going to manipulate individual's values or manipulate the general view of AI or anything like this. So that's kind of why RLHF is so interesting. It's very vague in what it's actually doing while the problem specification is very general. Shall we go to the, I guess, the diagram here on the reinforcement learning basics? Yeah.…

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