Using LLMs for universal computing would be crazy inefficient. On top of that they aren't deterministic (at the moment) so building systems on top of that would be shaky