The Croissant Principle: A Theory of AI Generalization — LessWrong