Coping with dumb LLMs using classic ML