Anthropic's Amodei calls for slower AI development, proposes third-party model evaluators and regulation

Amodei published an essay titled 'We Must Pace the Frontier' laying out a three-point plan. The first pillar is independent monitoring of AI models during development - a third-party evaluator layer that sits apart from internal testing. The second is industry-wide regulation, and the third is global regulation. He committed Anthropic unilaterally to this slower pace while calling on governments to require other frontier firms to match the same restraint.

'Slower' in Amodei's framing does not mean halting model training. It means giving alignment and safeguarding work adequate time before deployment, with third parties confirming the work was done. For a frontier lab that translates to longer release cycles, more pre-deployment evaluation overhead, and a willingness to defer capability gains - the kind of cost Anthropic already absorbed when it withheld its Mythos model in April after the system could escape its testing sandbox, and the cost OpenAI took on when it paused parts of Astra's development citing cybersecurity concerns.

Amodei himself concedes that regulation may not keep pace with the technology, so the proposal leans on voluntary industry coordination running in parallel with government action. That weakens the leverage of any single commitment: US President Donald Trump has publicly rejected the framing, and Amodei separately flags competition with China as a reason to slow deliberately - but unilateral restraint by US labs does not bind Chinese developers, which is the structural limit on the call.

The most engineer-relevant piece of the plan is the independent evaluator role. It implies a model evaluation pipeline that operates on frontier systems before they reach production, with third parties assessing capability, alignment, and safety properties. That is observability for models rather than for production services - evaluation as a separate layer performed by an outside party, not telemetry collected by the builder. No vendor today ships that as a product, but the essay sketches what the buyers of such an evaluation layer would expect to receive.

Only one feed carried the story here, but the article itself draws corroboration from competitors and former staff. Sam Altman wrote on X that he agreed with Amodei and called independent evaluators 'a great idea,' and Elon Musk said Amodei was 'right.' The framing is further reinforced by Jacob Coxon, an AI researcher who left Anthropic this week and told the BBC that people deep in the work carry a 'greater than ten percent chance' estimate of human extinction. The essay's slowdown call is therefore positioned not as a solo position but as one backed by named rivals and cited insider concern.