Would you be a better RLHF labeler than GPT-4? — LessWrong