Shawn Wang: uncertainty

Transcript context

…Yeah, so that figure right at the top of the constitutional AI paper is worth seeing if you don't have it immediately pop into your head where they essentially compare like constitutional AI to other RLHF that they're doing internally. And that's something that most RLHF papers don't do is like they have little dots on the lines to indicate intermediate checkpoints and be really great to see more RLHF papers showing how per epoch or per half epoch of training because most RLHF is only a few epochs, at least in the open models, like what is happening there. People release checkpoints, but that's how we should be thinking about it because the optimizer is so strong and it's like we don't know what's happening in this kind of intermediate land. I don't know if this is a relevant comparison for you, but OpenAI also recently released a weak to strong generalization paper where they actually talked about a few intermediate checkpoints for GPT-4. Any comments on the comparison between constitutional AI and weak to strong generalization? I didn't see the paper. I think I saw people criticizing it for like just being like safety washing from the fact that they're like talking about GPT-2 still, which is such a kind of like odd model to focus on. I didn't really look at the paper. I think that it's a thing with OpenAI. It's like they're sharing less than they know. So I think they probably have things that are pretty cool that they're doing internally. And I'll summarize for listeners who may not have seen the paper because it's impossible to keep up and everything. I do think that what constitutional AI and RLAIF represents is that we are starting to come to a point where it's just impossible for manual human preference data collection to scale. And the only way to scale this is to trust our AI overlords to model our human preferences. And constitutional AI was the first version of this. What the second version or what weak to strong is, is that anticipating a future of the need for super alignment, where the thing that we're trying to control is smarter than us. So you take GPT-2 and try to use GPT-4 to teach it to be smarter than itself, because this is what we're going to have to do in the future as well. When we are not, we're no longer fully in control. Are we the metaphorical GPT-2? No, we're like not even in the process anymore at the point of super intelligence.…

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