Shawn Wang: uncertainty

Transcript context

…Yeah, I mean, the outputs of LLMs, are they synthetic data? They probably are, but I think it depends on the goal you have. If your goal is you're creating synthetic data with the goal of trying to distill GPT-4's superiority into another model, I guess you can call it synthetic data, but it also feels like disingenuous because your goal is I need to copy the behavior of GPT-4 and- It's also not just behavior, but data set. So I've often thought of this as data set washing. Like you need one model at the top of the chain, you know, unnamed French company that has that, you know, makes a model that has all the data in it that we don't know where it's from, but it's open source, hey, and then we distill from that and it's great. To be fair, they also use larger models as judges for preference ranking, right? So that is, I think, a very, very accepted use of synthetic. Correct. I think it's a very interesting time where we don't really have good social models of what is acceptable depending on how many bits of information you use from someone else, right? It's like, okay, you use one bit. Is that okay? Yeah, let's accept it to be okay. Okay, what about if you use 20 bits? Is that okay? I don't know. What if you use 200 bits? I don't think we as society have ever been in this conundrum where we have to be like, where is the boundary of copyright or where is the boundary of socially accepted understanding of copying someone else? We haven't been tested this mathematically before,…

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