Shawn Wang: uncertainty

Transcript context

…I am not sure if it's a fundamental issue or not. I think we just don't understand transformers enough. I don't even mean transformers as an architecture. I mean the use of transformers today, like combining the tokenizer and transformers and the dynamics of training, when you show math heavy questions versus not. I don't have a good calibration of whether I know the answer or not. I, you know, there's common criticisms that are, you know, transformers will just fail at X. But then when you scale them up to sufficient scale, they actually don't fail at that X. I think there's this entire subfield where they're trying to figure out these answers called like the science of deep learning or something. So we'll get to know more. I don't know the answer. Got it. Let's touch a little bit on just Meta AI and you know, stuff that's going on there. Maybe, I don't know how deeply you're personally involved in it, but you're our first guest with Meta AI, which is really fantastic. And Llama 1 was, you know, you are such a believer in open source. Llama 1 was more or less the real breakthrough in open source AI. The most interesting thing for us covering on this, in this podcast was the death of Chinchilla, as people say. Any interesting insights there around the scaling models for open source models or smaller models or whatever that design decision was when you guys were doing it? So Llama 1 was Guillaume Lample and team. There was OPT before, which I think I'm also very proud of because we bridged the gap in understanding of how complex it is to train these models to the world. Like until then, no one really in gory detail published.…

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