Scholars publish 210-page paper on feasibility of a hardwired pause of frontier AI training

Berkeley statistician William Fithian and Cambridge AI safety researcher David Krueger, together with 25 interdisciplinary co-authors, have released a 210-page working paper, "The Feasibility of a Hardwired Pause of Frontier AI Training," systematically arguing for the feasibility of a global "hard pause" of frontier AI training, and exploring policy options such as halting production of training chips and freezing frontier AI for a decade.

Confirmed

• The paper takes compute as its lever: AI chips and fabs are highly specialized and their locations are largely known, so cutting off frontier training compute is technically feasible—the authors argue this is not "torrent-banning" fantasy; as long as no one has the "means" to continue frontier training, the race can be frozen.
• On design, the paper offers one concrete implementation of a "hardwired pause" while leaving multiple adjustable "knobs" for policymakers to facilitate international agreements; per Krueger, the scheme captures most of the benefits of "removing compute" while being easier for parties reluctant to lose access to existing models to accept.
• Fithian's post analyzes feasibility and mechanism design from economic, international-relations, historical and game-theoretic angles, suggesting the relevant processes could take roughly a decade based on historical trends, and that banning frontier training means efficiency gains at the frontier slow sharply; the overall conclusion is that a pause is a realistic policy option if decision-makers are willing to pursue it.
• Krueger notes in a long thread that a hardwired pause could severely suppress secret AI projects: even if one party secretly pulled ahead, they would still have to advance the entire frontier while competing with many near-peer models, making "taking over the world" extremely hard.
• Krueger has publicly offered a $1000 bounty for a rebuttal that convinces him the plan has major flaws or that a better alternative exists.

Not yet confirmed

• Krueger admits he remains nervous that the paper's argument isn't "watertight" and is still weighing it; his personal view is whether the plan can reduce societal-scale risk to acceptable levels depends on how advanced AI already is at the time of the pause, and capabilities are rising fast.
• One core point of contention is "scaffolding singularity": even with model weights frozen, agent scaffolding around models and their usage patterns could keep evolving with algorithmic progress, and the risk is greater if AI can automate that R&D.

Why it matters

• This is a rare effort to move "pausing frontier AI" from slogan to a 210-page interdisciplinary feasibility analysis covering technical, economic and international-coordination dimensions, giving the policy debate a concrete handle.
• Krueger revealed he largely stopped research a year ago to focus on his venture EvitableOrg; he made an exception for this paper and offered a public bounty for rebuttals, showing the authors want to be examined in the open.

2026-10-10 ~ 2026-10-10 · 15 related posts

• [source] Berkeley-led group proposes "hardwired pause": stop training chips to freeze frontier AI for a decade — wfithian · 2026-10-10
• Researchers argue a pause on frontier AI training is a practical policy option — wfithian · 2026-10-10
• Cambridge's David Krueger Proposes Freezing the Compute Supply Chain to Pause Frontier AI — DavidSKrueger · 2026-10-10
• Krueger's pause paper leaves many policy knobs for an international agreement — DavidSKrueger · 2026-10-10
• Krueger: whether the pause works hinges on how advanced AI is at pause time — DavidSKrueger · 2026-10-10
• Krueger warns of a 'scaffolding singularity' even if model weights are frozen — DavidSKrueger · 2026-10-10
• [source] Cambridge's David Krueger outlines a hardwired AI pause to curb secret frontier projects — DavidSKrueger · 2026-10-10
• Krueger: a hardwired pause would strongly disincentivize secret AI projects — DavidSKrueger · 2026-10-10
• Krueger returned from a year off research to work on the AI pause paper — DavidSKrueger · 2026-10-10
• [source] Cambridge's David Krueger offers $1,000 for flaws in his AI safety plan — DavidSKrueger · 2026-10-10
• New paper: Can a hardwired global pause of frontier AI training actually work? — wfithian · 2026-10-10

4 near-duplicate retellings: KatjaGrace · wfithian · wfithian · wfithian