Slowing frontier AI sounds sensible until someone has to draw the line

Nathan Lambert agrees with the idea of slowing the riskiest AI development but considers its practical implementation close to impossible as currently stated. His objection targets the mechanism: who defines prohibited capabilities, which benchmarks may not improve and who gets a seat in the decision.

The criticism targets the missing ruler, not the goal

Lambert was responding to Pacing the Frontier. In July 2026, the initiative asked the US government to support an international effort to develop technical and governance tools for deliberately pacing automated AI research. At verification, its website listed 1,386 signatories from frontier AI companies.

Lambert writes that the principle is sound but the implementation breaks at the definition. He asks whether authorities would choose benchmarks that models are not allowed to improve and who would make that decision. He adds that researchers could shift from direct capability work to swarms and efficiency, producing a different class of risk.

His second objection concerns existing models. In his view, much AI risk comes from diffusion of capabilities already available. He therefore gives more weight to preparedness and pressure on labs to handle current systems more carefully.

Regulating one curve redirects work to another

A frontier system can improve through several routes. A larger training run is visible, but similar practical gains can come from better post-training, longer inference, more agents, new tool use or cheaper inference. A rule tied to one benchmark or compute threshold creates an incentive to optimize everything around it.

Coordination can still be useful, but it has to measure the capacity to cause a defined harm and inspect the whole system rather than a model on a leaderboard. In cyber operations, the relevant combination may include autonomy, tool access and replication speed. One test score may miss that combination.

Lambert’s objection leaves its own gap

Preparedness and more careful labs address deployment and consequences, but they do not remove the competitive pressure described by the initiative. If every company benefits from moving faster, voluntary restraint can fail exactly when it becomes expensive. Showing that pacing is underspecified does not yet produce a coordination plan.

The X post is an argument, not an empirical study. It does not prove that every form of pacing will fail or quantify the risk created by efficiency and swarm research. It does identify a real omission in the open letter: the request calls for tools without yet supplying a boundary or an institution capable of enforcing it.

A workable mechanism must survive a change of metric

The next useful step is not another signature count. The initiative needs a decision process that defines concrete risk thresholds, updates them and catches substitution through efficiency, swarms or diffusion. It also has to specify who audits lab data and how disputes between states are resolved.

A strong mechanism should pass an adversarial test: what would a company do if it wanted to keep racing while formally complying? If the limit can be bypassed by renaming a capability or swapping a benchmark, the pace remains unchanged. Only the reporting vocabulary moves.

Lilith's verdict

A regulator can tape off one benchmark while research detours through swarms, efficiency and diffusion. A real brake has to watch the whole track, not one sign beside a bend.