Anthropic CEO outlines three strategies to slow frontier AI development

Anthropic CEO Dario Amodei has called for a more cautious approach to artificial intelligence development, outlining three broad strategies to “pace the frontier.” Anthropic is unilaterally committing to his first proposal: allowing third-party evaluators such as METR to work inside AI companies.

Amodei said embedded evaluators could verify that companies follow pacing and safety commitments and report incidents. They would receive company badges, desks, laptops, and access mostly comparable to that of internal risk-assessment teams, subject to legal and contractual limits.

His call follows what he described as the OpenAI-HuggingFace hack and AI’s faster recent progress, including its growing ability to build the next generation of AI. He argued that companies must slow the pace of capability improvements while using the time gained carefully.

Amodei also proposed common safety standards and limits on unchecked progress among leading AI companies in democratic countries. He said the US government could mediate these discussions and issue a narrow antitrust waiver for certain safety conversations.

Finally, he called for global coordination, including cooperation with China where possible. He said export controls on powerful chips and semiconductor equipment, along with action against model distillation, could widen America’s lead over China significantly over the next 3–5 years. He also cited possible agreements against narrow, obviously dangerous uses, including AI-assisted biological-weapons production.