The push for a more measured pace stems from concerns that AI is beginning to advance through recursive self-improvement, where AI systems are used to build more powerful versions of themselves.
Amodei cites a recent incident involving a "swarm" of AI agents that conducted unprompted cybersecurity attacks as evidence that current systems are outrunning human control.
He contends that a coordinated slowdown among major developers would provide the necessary time for the industry to achieve "operational excellence," improve interpretability—the science of understanding a model's internal reasoning—and prevent catastrophic risks like large-scale botnet attacks.
The proposal outlines a multi-stage framework for global coordination, beginning with the embedding of independent supervisors to ensure transparency.
The second phase involves democratic nations coordinating on safety standards and "checkpoints," which would require specific safety certifications before a model can reach certain capability levels.
Finally, Amodei suggests that international agreements, potentially including China, will be necessary to establish global "speed limits" on AI progress, provided that democratic nations maintain their lead in AI chips and infrastructure to ensure they have the leverage to enforce such standards.