OpenAI 提出辩论式 AI 安全训练方法

We’re proposing an AI safety technique which trains agents to debate topics with one another, using a human to judge who wins.