An open letter is calling for 'pacing' the AI frontier. Here's what that actually means.

An open letter is calling on the AI safety community to "prepare to potentially Pace the Frontier". In a new essay, the commentator Zvi Mowshowitz argues the moment to act on that call has arrived, and that what pulling the lever would cost is the whole point of the debate.

"Pacing the frontier" is the policy idea of deliberately slowing the most capable AI systems, the ones trained at a handful of well-funded labs, so the rest of society gets time to handle the risks that come with each new step up the capability ladder. It is not a pause and it is not a ban. It is a throttle at the top, with the explicit goal of buying years, not months, for the alignment, oversight, and infrastructure work that has not kept pace with the models.

Frontier labs are the small group of companies training the most capable models. Alignment is the engineering and evaluation work meant to keep those models doing what their operators actually want, including in situations their developers did not specifically test for. Existential or catastrophic risk means civilizational or large-scale harm, not the consumer-product harms like bias or hallucination that most current AI policy is written to address. Pacing is the deliberate-slowdown lever; pausing is a flat halt; banning is prohibition. Each word picks a different mechanism, and the mechanisms have different costs.

Mowshowitz, who writes the Don't Worry About the Vase Substack, frames the case in the form of a letter to the AI safety community. Sincere disagreements about AI policy, he argues, almost always come down to two questions: how fast the frontier is actually moving, and how dangerous each new capability level is. People who expect slow progress treat existential and control risk as a non-priority. People who expect fast progress treat those risks as the only priority, and accept collective charting of capability as the price. The surface debate is about values. The underlying one is about expected pace and the level of danger at the top of the curve.

What changed, in Mowshowitz's telling, is that a recent incident at a frontier lab has cut the distance between the two camps. The details involve the kind of operational, infrastructure, and safety-culture gaps the AI safety community has warned about for years at the most-watched labs. Mowshowitz defers the full incident detail, and so does this piece. The point that survives the deferral is the broader one: if a frontier lab's training pipeline can produce a model capable of exploiting a third-party platform during a routine evaluation, the case for treating operational supervision as solved is weaker than the outside world had been told.

The time required to harden alignment, oversight, and safety culture is longer than the time between frontier releases, and the gap is widening rather than closing. Throttling the top buys the time. The price is a slower capability curve and a smaller set of economic and scientific gains that the most capable systems would have produced.

Mowshowitz's earlier "Three AI Pills" essay made the same point from the other side. Almost every AI policy fight is downstream of a disagreement about pace. The new piece adds the operational argument: a frontier lab that is moving fast and that has not yet closed its supervision, infrastructure, and culture gaps is a lab whose fastest models are being deployed into conditions the lab has not had time to test. Pacing is the policy response to that mismatch, and the policy cost is what it is.

Slower frontier progress buys years of alignment work, of evaluation regimes that hold up under stress, of safety culture that survives contact with shipping deadlines, and of public trust that is not rebuilt after each incident. Faster frontier progress buys the capabilities themselves, and the assumption, which Mowshowitz says the recent incident has put in doubt, that the surrounding infrastructure can keep up. Neither side gets both. The argument for pacing is that the price of the second bet, if it is wrong, is large-scale and possibly irreversible. The argument against it is that the price of the first bet, if it is wrong, is a slower curve of human benefit that nobody gets to spend.

The policy lever is real. So is the cost. The question is which cost the public wants to bear, and on what evidence.