Comment by ianm218

Well there's a world simulation model and then the driving model.

You can imagine improving i.e. a specialized math model (problem in, theorem out) with a normal LLM that knows lots of problems and theorems generally.