AI safety via debate

【RSS 原文 · 未经 AI 筛选或翻译】We’re proposing an AI safety technique which trains agents to debate topics with one another, using a human to judge who wins.