AI safety via debate

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

Source: OpenAI News — Published May 3, 2018 at 07:00 UTC — Category: Models