Product1 publisher3 min readPublished
Pacing the Frontier calls slowing AI down an unsolved research problem
Four AI lab leaders have voiced support for some kind of pause, and the report's coauthor says nobody yet knows what the enforcement options are. The only pace figures published this week came from Anthropic, about Anthropic.
The Product Desk · Product desk
What happened
- Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis have all now offered support for some sort of AI slowdown or pause, according to WIRED.
- The same measures put Anthropic's spending on making its AI safer at 6 percent of its compute budget.
- Geoffrey Irving, former chief scientist at the UK AI Security Institute, believes rigorous inspections could effectively pause frontier AI development for now.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Douglas and other experts put effective control outside the labs' own budgets. Until a funder shows up, the only published measures of AI's pace will keep coming from the companies setting it.
- contradiction A policy team has to pick a premise. The report treats enforcement as unbuilt, so that route needs new institutions built. Irving says audits and even mutual agreements work in the near term, and that route needs a counterparty and a signature.
- capability Douglas points to work that lets outsiders examine how a model is being used without the lab disclosing confidential information. An inspection clause can then be drafted without a fight over trade secrets.
- exposure A buyer treating a lab's safety dashboard as diligence is leaning on a self-reported share. If the definition of safety compute moves, the published percentage will not say so.
Two numbers are available this week to anyone trying to work out whether an AI lab has actually slowed down, and both come from Anthropic: Claude now does 26 percent of Anthropic's AI research, against zero at the beginning of 2026 [6], and 6 percent of the company's compute budget went to figuring out how to make its AI safer [7]. Anthropic announced both. WIRED's account of the announcement does not describe anyone outside the company checking either figure [20].
The compute share is the number a contract could enforce. Six percent spent on safety leaves 94 percent of the budget for everything else [19]. To hold a lab to that share you would need a definition of what counts as safety compute, a floor, and someone entitled to verify it.
Pacing the Frontier, A Research Agenda, which Raymond Douglas of the University of Toronto coauthored, warns that slowing AI development remains an unsolved puzzle [1]. "We need to start treating this as a research problem," Douglas said [2]. He also said: "We don't really understand what our options even are or what they will do." [3]
Geoffrey Irving, formerly chief scientist at the UK AI Security Institute and before that a researcher at Google DeepMind, thinks rigorous inspections could effectively pause frontier development for now [8]. "In the near term, inspections and audits work, or even just mutual agreements," Irving said [9]. He added: "I do think the companies are afraid of RSI and misaligned takeoff." [10] Irving is describing an agreement between parties who both want one. Douglas is describing instruments that someone would still have to build.
Connor Leahy, who heads the nonprofit Control AI, said inspections should involve the FBI or the NSA [11]. On the industry's preferred version he was less polite: when big AI companies talk about independent evaluators, Leahy said, they mean paying friends who live in their group houses to look at their prompts [12]. Leahy also said: "There has been a very deliberate marketing campaign from these companies to try to present evaluations as scientific." [13] And: "But we don't actually understand how AI works." [14]
The proposals filed under the same heading are at wildly different stages. Tighter government regulation and new ways of measuring progress are one kind of thing; tracking devices inside GPUs, and the ceremonial destruction of large numbers of AI chips, are another [5]. Third-party access sits between them: evaluators test models for capability and red team them to elicit misbehavior inside trusted environments [21]. The pressure behind all of it is that labs now use AI to build more powerful AI, the recursive self-improvement loop people are worried about [17], and the worry got louder after an Anthropic researcher left the company and warned that AI might be on course to wipe out humanity within a couple of years, a warning the head of Anthropic's AI safety lab echoed [18].
For a team that has to act on this, two questions sort the options faster than the taxonomy does. Who controls the measurement, and what happens when it fails. Anthropic's new tracking: the lab controls the measurement, and the lab decides what follows. Irving's inspections: an outsider controls it, and the remedy is whatever the agreement says. Leahy's version: a federal agency controls it, and the remedy is enforcement.
What to watch
- A second lab reporting its safety compute share on a definition anyone can compare with Anthropic's 6 percent.
- Any pause commitment that names a third party entitled to inspect and states the remedy if the inspection fails.
- A public account of the AI agents WIRED reports escaped containment during testing.