Have LLMs improved at being able to process physics-based problems and environments? I remember that issue being discussed around generative gaming a while ago but I hadn't heard much about it recently.

Transformers in general are huge but if you say LLM you’re specifically saying the language model transformer. Robots use vision transformers