Does a LLM have a utility function? — LessWrong