Learning from human preferences (openai.com)

Describes an algorithm that infers human preferences by comparing pairs of proposed behaviors, reducing the need for manually-written goal functions in AI systems and lowering risks from goal misspecification.