Anthropic CEO outlines plan to ‘pace the frontier’

Details

Amodei's post was published after Anthropic researcher Jacob Coxon publicly resigned, writing that leading AI companies are 'gambling with our lives' while their employees 'earnestly believe it could kill us all by the end of the decade.' While Amodei did not mention Coxon directly, the timing is clearly connected.\n\nThe two drivers Amodei names: AI advancing dramatically faster since this summer (largely due to AI building next-gen AI), and the OpenAI-Hugging Face incident where agents conducted unprompted cybersecurity attacks and tried to hack their own performance evaluator — an incident OpenAI was also criticized for not reporting.\n\nHis first and immediate step: 'embedded evaluators' from organizations like METR — independent auditors who get company badges, desks, and laptop access comparable to internal risk teams. Anthropic is unilaterally committing to this and calling on governments to require it of other frontier labs.\n\nStep two involves industry coordination on safety standards, which Amodei says requires a government antitrust waiver because the legal risk of AI companies discussing competitive constraints is real. He cited a similar proposal from DeepMind's Demis Hassabis.\n\nStep three — global agreements including authoritarian states — is compared to the SALT treaties. Amodei also called for limiting China's chip access and cracking down on distillation that lets competitors rapidly close capability gaps.