Nathan Lambert: belief

Transcript context

…You asked people open questions. Oh, I mean, there's so much to do here. They're kind of like summarization of things that will be hinted at in the talk to this point which is like I split it up in my work between like data training and model which is essentially like how do we evaluate what's happening at the model level with RLHF. I think big labs are indexed on their own base models so they don't know like what's swapping between CloudBase or GPT-4 base how that would change any notion of preference or what you do with RLHF. I think in the open we could do that. We can swap between Lama2 and Mixedraw and kind of see like does RLHF work the same for both of those? Do they both get alpaca valve bumps when you use the same dataset in the same framework down the line? That'd be good to know if like how sensitive RLHF is. On the data we talk a lot about aggregation. On the research side there's a lot of interesting things as like does getting your data from scale or a Discord army change the quality of the data based on like professional contexts. They probably should do it internally. They should do like internal market analysis on that line. We should also mention there has been a report that a lot of these labelers use ChatGPT to do their work.…

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