Anthropic CEO Dario Amodei Proposes Three-Step Plan to Pace AI Development

Dario Amodei proposed a three-step plan titled 'Pacing the Frontier' to maintain a balanced pace of AI development, ensure safety, and address geopolitical dilemmas. The plan does not aim to halt model training or technological progress, but rather to secure sufficient time for alignment and safety measures, with third-party evaluators verifying these efforts.

As Step 1, Anthropic committed to unilaterally introducing an 'Embedded Evaluators' system, granting third-party evaluation teams access rights similar to internal employees to verify compliance with safety practices. This approach has precedents similar to regulatory supervisors in the banking industry.

Step 2 is 'Democratic Coordination,' where frontier AI companies in democratic nations collaborate to establish common safety standards and limits on unchecked AI progress. This may require government support or antitrust exemptions and could include checkpoint mechanisms requiring alignment certification based on model capabilities.

Step 3 involves attempting 'Global Coordination' between democratic and authoritarian nations. Amodei analyzed that banning the use of biological weapons (Level 1) is highly feasible, mandating risk testing before model release (Level 2) faces verification challenges, limiting the pace of recursive self-improvement (RSI) (Level 3) is possible but difficult like SALT treaties, and comprehensive pacing or halting (Level 4) is unlikely due to incentives to evade surveillance.