Large Language Models Mostly Know What They Know