Is the middle option asking LLM to generate a classic ML model? Or generate tons of them and pick the best?