> LLMs are permanently trapped in the vector space they are trained on.

A lot of the time people state the kind of fundamental limitations of LLMs very confidently when it feels like it is too early for people to really know. Like we are already well past the point where where LLMs are just pre trains on the internet with some RLHF for chatbot… Most of the effort is spent on elaborate reinforcement learning.

Is it unconceivable that future generations of LLMs could be RL’d to use einsteins visual method for theories [1] with the right tooling and geometry representations? Or just something random like that.

[1]. https://www.visualscribing.com/blog/2019-11-11-einstein-on-v...