LLMs are literally next token prediction engines and nothing else.

Diffusion for text is not even an academic toy at this point and will likely never be a real thing.