New video!
Added 2025-07-06 14:00:14 +0000 UTCIn the future, AIs will likely be much smarter than we are. They'll produce outputs that may be difficult for humans to evaluate, either because evaluation is too labor-intensive, or because it's qualitatively hard to judge the actions of machines smarter than us. This is the problem of “scalable oversight.” Proposed solutions include “debate” and iterated amplification. But how can we run experiments today to see whether these ideas actually work in practice?
In this video, we cover Ajeya Cotra’s “sandwiching” proposal: asking non-experts to align a model that is smarter than they are but less smart than a group of experts, and seeing how well they do. We then show how Sam Bowman et al. tested a basic version of this idea in their paper “Measuring Progress on Scalable Oversight for Large Language Models.”
