The problem with LLMs is that they're unreliable so they only work for problems with tight feedback loop or for problems where the result can be easily verified.