Alek Westover and Nathan Sheffield of Redwood Research want regulators to cap the compute AI companies put into making models more capable, and they have written the rulebook for counting it. Their post on the Redwood Research blog argues that a cap is only as good as its definition of what counts. Their answer is blunt: every workload is a capabilities activity unless it earns an exemption.

The authors say they want to pace the frontier, whether through legislation or voluntary commitments. A limit on capabilities compute only works if companies cannot relabel their work to dodge it. They set three tests. Restricting the measured compute must substantially slow AI progress, the line must hold up under adversarial pressure, and reasonable experts must largely agree on how an activity is classified.

The simplest definition passes all three, as the authors admit: treat everything a company does as capabilities work. They argue for a narrower one anyway, for two reasons. It leaves labs compute to improve safety, and it lets outside parties use models, which spreads power and information about what they can do. If the narrow version proves hard to adjudicate, the authors say the “narrow safety” and “narrow product development” exemptions could be cut.

Only four exemptions exist, in this order: safety work (narrowly drawn), serving customers, product work on narrow models, and compute that sits idle. Each applies only if the work does not improve general model capabilities. Idle compute means a whole GPU doing nothing for a non-negligible stretch of time, so an underused core does not qualify. Narrow product work covers things like integrating AI into Google Docs or glasses, protein-folding models, and recommendation algorithms.

The safety exemption is where the authors spend the most effort. It covers six items: evaluating model alignment, testing and improving company security, safety research that is not too dual-use, monitoring, running misuse classifiers, and incident investigation. The research whitelist is short. It names interpretability, chain-of-thought informativeness, model persona work, model organisms research, and theoretical alignment work without near-term capability uses. Labs must publish a project plan at the start and a detailed public report at the end for that research to qualify.

The boundary cases show how much the proposal leans on caution. Running monitors counts as non-capabilities work, but training them does not. The same split applies to misuse classifiers. Scalable oversight, dangerous-capability evaluations, and fixing hackable RL environments all stay in the capabilities column. The authors give a test for the gray zone: if a company tweaks a training recipe, checks the result against a cyber evaluation, and keeps the best tweak, that evaluation is capabilities work whatever it is called.

Serving customers is exempt with three exceptions. Labs that outsource capabilities tasks to customers, train on customer data, or serve customers doing central capabilities work such as building RL environments must count that inference. For zero-data-retention workloads, the company must either refuse the customer or presume the compute counts.

The authors say false positives are cheaper than false negatives, and that compute-light activities hardly matter. They also say third-party auditors will be needed to check how the definition is read, and that it will need adjusting if companies game it.

The post closes with a call to action. Anthropic has put out its own figure for the safety slice of its AI R&D and has urged third parties to measure and report that number. The authors note that Anthropic did not explain how it defined safety research and has not yet had auditors verify the figure. They urge developers to let auditors report the capabilities and non-capabilities split using a definition like theirs. For any lab that already publishes a safety share, the practical question is whether it would accept a stricter default.

Alek Westover and Nathan Sheffield, “How should we define ‘capabilities activities’ for the purpose of pacing?”, Redwood Research blog (publication date not stated in the source).