Looking Inward: Language Models Can Learn About Themselves by Introspection