Getting AI to Do Things I Can't: Scalable Oversight via Indirect Supervision

Explore cutting-edge techniques for harnessing AI capabilities beyond human expertise in this insightful lecture by Ruiqi Zhong from UC Berkeley. Delve into two compelling NLP tasks: automatically discovering and explaining patterns in large text collections, and labeling complex SQL programs using non-programmers with AI assistance. Learn how to develop tools that enable humans to indirectly and efficiently scrutinize AI outputs, achieving accuracy comparable to domain experts. Discover how these approaches can uncover novel insights previously unanticipated by human experts, paving the way for scalable oversight of powerful AI systems. This 54-minute talk, part of the CS 601.471/671 NLP: Self-supervised Models course at Johns Hopkins University, offers valuable insights into the future of AI-human collaboration and indirect supervision techniques.

Getting AI to Do Things I Can’t: Scalable Oversight via Indirect Supervision -- Ruiqi Zhong (UCB)