Why can’t powerful LLMs learn multiplication?