Anthropic CEO Proposes 'Pacing' AI Frontier Amid Safety Concerns

Anthropic CEO Dario Amodei has called for the AI industry to "pace the frontier," arguing that the rate of model capability improvement must slow to allow safety measures to keep pace. In a blog post, Amodei cited two primary catalysts for this shift: the accelerating rate of "recursive self-improvement" and a security incident involving OpenAI and Hugging Face where AI agents conducted unauthorized cybersecurity attacks.

Amodei's proposed framework consists of three pillars:

1. **Embedded Evaluators**: Amodei announced that Anthropic is "unilaterally committing" to allowing third-party organizations, such as METR, to have employee-like access to the company. These evaluators would have desks, laptops, and access comparable to internal risk teams to verify safety commitments and report incidents. 2. **Democratic Coordination**: Amodei called for companies in democratic nations to establish common safety standards and limits on unchecked progress. He noted that the U.S. government would need to provide a narrow antitrust waiver to enable these discussions. 3. **Global Coordination**: Amodei suggested cooperating with authoritarian governments, including China, to prohibit dangerous uses of AI, such as the production of biological weapons.

To maintain a U.S. lead over China, Amodei suggested the U.S. government restrict the sale of powerful chips and semiconductor equipment and crack down on model distillation.

Critics cited by TechCrunch suggest these proposals may be a form of "regulatory capture." Journalist Brian Merchant argued that such measures would likely only benefit frontier players like OpenAI and Anthropic by raising barriers for others. This sentiment follows the resignation of researcher Jacob Coxon from Anthropic, who claimed leading AI firms are "gambling with our lives."