Alexandr Wang says no one knows how to solve AI alignment, proposes scalable oversight

Meta Chief AI Officer Alexandr Wang says nobody yet knows exactly how to solve AI alignment, which he calls one of the most open scientific questions in AI. As reported by Rohan Paul on X, Wang proposes scalable oversight, in which a separate set of AIs monitors more capable models and must improve alongside them.

Wang says Meta's Muse already uses a version of this approach, with a sentinel agent that checks the main agent's actions.