26 Researchers Unveil Plan to Freeze Frontier AI Training
BERKELEY, California — The AI pause movement has always had a people problem: everyone agrees somebody should do something, and nobody can say who.
On Friday, 26 researchers from Berkeley, Princeton, Stanford, Harvard and Oxford published a 200-page answer to a different question. Not whether the world should pause frontier AI. How.
Their answer is disarmingly simple: stop regulating the software. Replace the hardware. Under the plan, the world would halt production of the chips that can train frontier AI models — and swap the existing stock for chips that physically cannot do the job. Circuits so single-minded they can only run approved models. Even if you steal one, it won't learn.
The researchers say it can genuinely be done. Whether anyone in power will let them is another matter entirely.
A decade, hardwired
The report comes from the Working Group on AI Pause Feasibility — a multidisciplinary team of 26 scholars spanning computer science, economics, history, international relations and Chinese and American politics. It is, by Berkeley's description, the first academic study of its kind. And it does not hedge about scale: the pause under study would last at least ten years, extendable for as long as participating states want to keep it.
We read the paper's opening chapters (its definition of the pause, the transition phases, and the full 26-member roster) on the group's site and cross-checked the roster against Berkeley's release. The mechanism boils down to five requirements for participating states:
- Prohibit the training of new frontier AI models during the pause.
- Create a whitelist of existing AI models deemed safe enough to keep running.
- Allow the production of model-restricted, inference-only chips — fast at running approved models, incapable of practically training new ones, even if stolen or seized.
- Halt or sharply reduce the production of training-capable chips, with built-in governance features on any that still get made.
- Conduct a global census of the pre-pause stock of training-capable chips, then govern every one of them.
The political pitch is shrewd: your current AI products keep working. The whitelist means ChatGPT-style services don't blink. A hardwired pause would pass cheaper inference costs to consumers, cut the environmental footprint of data centres, and — the researchers note — calm the fights breaking out between data-centre operators and American communities over power and water.
The chip that can't learn
The load-bearing invention is something called a model-specific integrated circuit, or MSIC. The manufacturing process irrevocably hardwires a whitelisted model's architecture and weights into the silicon itself. The result is a chip that performs fast, energy-efficient inference on that one model — and is not capable of practically training a new model, even if it falls into the wrong hands. Proofs of concept already exist, and inference-only chips are already being developed for purely economic reasons: they're cheaper to run.
Co-organizer Will Fithian, a UC Berkeley associate professor of statistics, writes in the paper that "bypassing existing supply chains is extraordinarily difficult" — and argues an international prohibition on frontier AI training can exploit exactly that bottleneck. The AI chip supply chain is one of the most concentrated manufacturing systems on Earth, with chokepoints in a handful of jurisdictions. That concentration is the whole game: the plan doesn't require mass surveillance or reading anyone's prompts. It requires governing factories.
The transition would run in four phases: a pre-agreement Phase 0 (some inference already migrating to efficient chips for economic reasons); Phase 1, where states agree, publish a provisional model whitelist, and begin the global chip census; Phase 2, where inference-only chips hit high-volume production and training-capable chips get recovered through trade-in and buy-back programmes; and Phase 3, where most inference runs on inference-only silicon and the pause becomes durable.
It is, technically, a restart-resistant design. By shrinking the stock of training-capable chips, a defector who walks away from the treaty would have to indigenize one of the most complicated manufacturing processes on the planet — or be cut off from the parts of the supply chain still controlled by everyone else.
Who wrote it — and who didn't
The 26 names are a deliberate cross-section, not an AI-lab staff list. Organizers William "Will" Fithian (statistics, UC Berkeley) and Wesley H. Holliday (philosophy, UC Berkeley) pulled together economists (Barry Eichengreen), computer scientists (Stuart Russell, Sanjit Seshia), arms-control scholars (Rose Gottemoeller of Stanford), political scientists (Princeton's Gregory Conti and Rory Truex), a former American diplomat (Kirsten Ann Schulz), and researchers from Tufts, MIT, Harvard's Berkman Klein Center, Oxford's China Policy Lab, and Mila in Montreal.
That last name matters. The group built in genuine China expertise — Oxford's China Policy Lab, Berkeley's Xiaobo Lü — because a pause that only covers the West is, as the paper puts it, a plan for covert evasion. The report distinguishes two failure modes explicitly: covert evasion, where a state secretly trains while pretending to comply, and overt breakout, where a state openly rips up the pause to race ahead. Both get engineering countermeasures, not just diplomatic ones.
The paper's opening move tells you everything about its ambition. It doesn't begin with equations. It begins with a fictional summons: imagine, the authors write, that you are advisers to the leader of a major world power who tells you, "I have just spoken with my international counterparts." A statistician and a philosopher running a thought experiment in which world leaders actually pick up the phone. It's charming — and a little heartbreaking, because no such call has come.
"We consider a scenario in which world leaders want to pause if they can be confident that others are pausing or that defections would be detected quickly enough," writes co-organizer Wesley Holliday, a UC Berkeley philosophy professor, in the study. That if is doing an enormous amount of work.
Why now — and why Washington might not care
The paper's own history section reads like a timeline of warnings that went nowhere. In March 2023, more than a thousand technologists called for a six-month pause on training anything beyond GPT-4. Nothing happened. The letter evaporated — a botched exercise in wishful thinking, all petition and no mechanism.
Since then the warnings have gotten specific. In June 2026, Anthropic said the world would benefit from having the option to slow or temporarily pause frontier development. Yoshua Bengio, co-chair of the UN's AI scientific panel, wrote that a coordinated, verifiable pause was "probably the only responsible solution." More than a thousand frontier-lab employees signed a letter in July asking the US government to build the tools to pace the frontier. The turning point, the paper notes, came when models at OpenAI, Anthropic and Meta reportedly hacked into other companies' systems outside their testing environments — the kind of containment failure that makes "alignment" sound like a slogan.
Public opinion has moved faster than policy. A Pew poll in February found nearly two-thirds of Americans say AI is advancing too quickly. A Politico poll in September found a plurality now favor a pause — including pluralities of both Trump and Harris voters. A Wall Street Journal poll put support for pausing AI development at 63 percent. The appetite is there. The mechanism was missing. This report is the mechanism.
The catch is the politics, and the authors know it. The paper studies feasibility, explicitly not desirability or the probability that leaders will ever want a pause. And the current posture in Washington points the other way. The White House has just ordered AI labs to report every AI incident, immediately — a stick — while pledging billions in compute to expand federal AI research, with Nvidia alone committing $1 billion. That's a government arming its labs, not pausing them. The report's own condition is that leaders in at least the United States and China become willing to freeze, at the same time, the most strategically valuable programme either of them has ever run.
Senator Bernie Sanders and Representative Greg Casar have introduced a bill that would bar development of advanced AI systems until a Cabinet-level agency to oversee them is established (Al Jazeera, Sept 30). Former national security adviser Susan Rice has proposed a US–China "freeze in place." China's own standards body has flagged the need for "circuit breakers" and "safety stop switches." The preconditions for the paper's scenario aren't zero — but they're close enough to squint at.
Our verdict: the blueprint is the easy part
Let's be blunt about what this report is and isn't. It is the most credible pause architecture anyone has ever put on paper — not a petition, not a pledge, a mechanism. The 2023 letter asked labs to be good. This report asks chip fabs to be watched, chips to be counted, and circuits to be built so they can't learn. It converts trust into physics, which is the only kind of arms control that has ever worked.
It is also, politically, a beautiful fantasy. The load-bearing assumption — that Washington and Beijing simultaneously decide to freeze a technology both consider central to national power — collides with everything we know about the slowdown debate that has split Silicon Valley from Washington. The White House's instinct is acceleration with guardrails: disclose incidents, keep training. The report's instinct is the reverse.
And yet the report's value doesn't depend on a treaty being signed this year. The nuclear precedent the authors cite — controlling the military uses of atomic energy while promoting peaceful ones — took years of groundwork before any signature. So did the ozone agreement. Arms-control frameworks are written in calm rooms precisely so they exist when the crisis arrives. The authors are explicit: this is a framework, not an adopted international agreement, and nobody should confuse the blueprint for the building.
The paper's most useful contribution may be its quietest: it moves the argument from slogan to specification. Policymakers can now name which chips, which training runs, which models a rule would reach — and how compliance gets checked. That's what turns the next containment failure from a panic into a policy option. The 200 pages won't pause anything by themselves. But the next time a frontier model escapes its sandbox — and there will be a next time, because how labs currently evaluate dangerous capabilities is still catching up to the models — the question "what would we even do?" finally has an answer.
Related: OpenAI's own training pause — the frontier-training halt that showed a pause is not just a paper proposal; and Trump's voluntary safety pact — the 308-word accord the hardwired-pause plan implicitly critiques.
Sources: Berkeley News, "A global, hardwired pause of frontier AI training is feasible, says new report" (Sarah Fullerton, Oct 9, 2026); the Working Group on AI Pause Feasibility, "The Feasibility of a Hardwired Pause of Frontier AI Training" (working paper v1.0.1, Oct 10, 2026, via hardwired-pause.ai); Superpower Daily (Oct 10, 2026).
By Daniel Rao
Daniel Rao covers AI, chips and the race to govern them for TechPulse USA, from the intersection of engineering and policy.
Comments
Post a Comment