We Must Pace The Frontier

Dario Amodei has published a new essay titled "We Must Pace the Frontier," naming his call after the Pacing the Frontier letter that lab employees signed in July. The proposal calls for slowing the rate at which AIs increase their capabilities, to allow for necessary alignment and safety work. Amodei said that without pacing he expects things to escalate quickly, and offered three proposals, unilaterally committing to the first one.

OpenAI followed with a commitment, and both Elon Musk and Demis Hassabis endorsed the overall proposal. Microsoft CEO Satya Nadella also agreed, and Anthropic's Long-Term Benefit Trust is on board. The first proposal, embedded evaluators, would give ongoing employee-like access to a team of embedded third-party evaluators, such as METR, to verify adherence to safety practices and commitments, report incidents, and help assess alignment of models, training pipelines and processes. Amodei framed the benefits as verifiability, transparency, and a second opinion free of commercial incentives.

Anthropic is unilaterally committing to the step now, proposing desks in its offices, access badges, company laptops, and access to workspaces, tools and permissions mostly comparable to internal risk assessment teams, with exceptions where law or contracts require it or to protect customer and partner information. External reviewers would have the right to publish key findings without editorial control, with narrow redaction abilities, and could say publicly if a redaction removed something important to their conclusions.

Amodei's second proposal is democratic coordination, in which frontier AI companies within democratic countries establish common safety standards and limits on unchecked AI progress. The third is global coordination, in which the US and other democratic governments attempt to coordinate with authoritarian governments where possible, while taking seriously challenges of verifying compliance. Amodei said pacing does not mean pausing, and that progress will still seem fast.

Zvi Mowshowitz wrote that there is still a long way to go, the odds are still against us, and the situation remains grim, but called it actual progress and said the work can begin. Commentators debated implementation, including whether evaluators can be independent, knowledgeable and sustainably funded at once.