Why not branch prediction with LLM?