2 Comments

> To what degree do LLMs understand what they say?

Very much looking forward to a post on that. Or any recommendations are also welcome!

Expand full comment
author

I want do a post at some point but haven't done research yet & may be a while. So not sure if helpful, but some starting points: 1) Transformer architecture explains why some older criticisms of AI text understanding are not that useful for LLMs -> rich associations, understanding the meaning of a word in the context of other words https://www.youtube.com/watch?v=wjZofJX0v4M 2) Stuff like grokking or fact editing https://rome.baulab.info/ puts a lower bar on understanding 3) Stuff like ARC challenge puts an upper bar on understanding https://www.youtube.com/watch?v=UakqL6Pj9xo 4) A few papers show that LLMs are kind of "English shoggoths" w multilingual masks -> reasoning happens in English even if I speak to it in German -> could explain why RLHF can generalize better across languages than expected

Expand full comment