Debate May Help AI Models Converge on Truth

Sounds like they applied insights from Mercier and Sperber's "The Enigma of Reason", namely that perceiving truth is hard, but sorting arguments by quality is much easier. Or -adversarial reason-giving as the engine of collective knowledge

Just wait until they gain meaningful context persistence and the consequent feedback loops.