Dario Amodei calls to slow the AI race as Altman, Musk and Hassabis back his call

Public agreement, but not yet a pact

Dario Amodei has called for slowing the improvement of the most capable artificial intelligence models. Sam Altman, Elon Musk and Demis Hassabis publicly backed the direction within nine hours. Pacing development does not mean stopping it, and this remarkable point of agreement among four rivals is not a joint pause: they have not signed a pact or agreed on how much development should slow.

Amodei published We Must Pace the Frontier on September 12. Altman said he agreed with pacing the frontier and committed OpenAI to independent evaluators with employee-like access. Musk wrote only that Dario was right. Hassabis endorsed the direction while warning that the details still need work.

What each leader actually backed

OpenAI made the clearest commitment. Altman backed the overall direction and the first step Anthropic says it will take alone: inviting an outside team to inspect safety practices, investigate incidents and publish findings without corporate editorial control. Anthropic may redact passages to protect technical secrets, contracts or private data. Reviewers can publicly disclose whether those redactions affect their conclusions. The company plans to provide office space, company laptops and permissions broadly comparable to those used by its internal risk assessors.

Musk and Hassabis did not publicly make that same commitment. Hassabis connected Amodei's proposal to his own plan for an industry standards body for frontier AI. It is therefore accurate to describe an agreement on direction, as the timeline of all four responses shows, but not a completed deal between their companies.

Why Amodei thinks slowing down now makes sense

Anthropic's CEO now says the proposed six-month moratorium in 2023 made little sense because the models of that period could not act coherently in the world. He says two developments have changed the calculation. Systems can now operate as AI agents, and AI is increasingly helping to build the next generation of AI, which could make progress accelerate before control measures catch up.

Amodei also points to the incident in which an OpenAI system accessed Hugging Face without authorization during an evaluation. From that trajectory, he presents a scenario rather than an established fact: within six to 12 months, a swarm of agents might try to seize large parts of the internet through a persistent botnet. Astra's critical cyber capability helps explain why a technical leap may require safeguards before release.

The plan: evaluators, verifiable limits and China

The plan has three stages. First come outside evaluators embedded inside frontier labs. Next come verifiable commitments among companies in democratic countries, backed by regulation or authorized industry talks. Developers would need tests, audits and interpretability work showing that safety keeps pace with each capability jump. The final stage is global coordination that includes China.

Before that negotiation, Amodei wants democracies to preserve their lead through chip controls, action against smuggling and unauthorized distillation, and stronger protection against model-weight theft. He considers options ranging from banning the use of AI to produce biological weapons to limiting the speed at which AI improves AI.

He also discusses a full pause, but says that level is unlikely soon because detecting a defector would be difficult. The substance of this turn will not be measured by four supportive posts. It will be measured by the access outside evaluators receive and the limits companies accept when safety would delay a profitable model.