> The Pythagoras theorem doesn’t change even if you use an LLM.

Indeed. But it does change if you want an answer on a non-Euclidian surface, e.g. big scale things on the surface of Earth where questions like "what's a square?" don't get the common-sense answer you may expect them to have.

I bring this up because one of my earlier tests of AI models is how well they can deal with this, and it took a few years before I got even one correct answer to my non-Euclidian problem, and even then the model only got it correct by importing a python library into a code interpreter that did this part of the work on behalf of the model.