Paul Christiano's departure from government AI safety

Paul Christiano has resigned from his role as Head of Safety at the Center for AI Standards and Innovation (CAISI) and returned to the Alignment Research Center (ARC) as executive director. In a post explaining the move, Christiano describes ARC's agenda as finding mechanistic explanations for neural network behavior, which he presents as a backup plan if mainstream alignment techniques fail before broadly superhuman AI arrives. He estimates a 20-30% chance that current alignment approaches, including generalization shaping, malicious behavior prevention, and misalignment detection, break down before that point. He gives ARC roughly a 10% chance of achieving its most ambitious goals before superhuman AI makes human labor obsolete, and says success would cut AI takeover risk by a meaningful amount. ARC is hiring an automation lead and chief of staff, with a researcher round planned within months.