If large language models have one redeeming feature for safety researchers, it’s that many of them think out loud . Ask GPT-4o or Claude 3....