Shawn Wang: belief

Transcript context

…Yeah. I would say like it does remind me of FlashAttention a little bit in a sense that like kind of an equivalent thing to the thing it's replacing and it's just faster, cheaper, just better. It's a very different optimization tool. There's essentially the thing in my mind that I can't get past is the difference between the control you get in training a reward model and then training a policy because essentially everything you want your reward model to do might not be everything that you train the policy to do in the RLHF step where you have like the two different prompt distributions. But with DPO, you're doing both at once. So you don't control that. And we don't know if you have fancy engineering abstractions and like test your reward model to do different things if that separation is really important. And I think that's where this like benefit at the absolute biggest scale and most investment could come from. But DPO is one update. Like it is one model. You can't separate that. So like that's the thing to know. Probably doesn't matter for most people, but it is very different. And like I was asking somebody who was on some of those earlier OpenAI papers that's not OpenAI anymore. And they were like, I wish we had thought of that. So like it is a really cool idea. And like that's the type of thing that academia still can do and can do really well and hopefully continues to do.…

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