Berkeley-led group proposes "hardwired pause": stop training chips to freeze frontier AI for a decade
wfithian · x · 2026-10-10
A working paper by UC Berkeley statistician William Fithian and 25 interdisciplinary coauthors examines whether a global pause on frontier AI training is feasible, amid a WSJ poll showing 63% of voters favor pausing AI development.
- The "hardwired pause" plan (designed to last at least a decade): prohibit frontier training while allowing inference, and stop making training-capable chips, gradually replacing them with inference-only chips (specialized for a restricted set of models, hardwired to be impractical for training).
- Threat models: covert evasion (a state secretly trains a strategically decisive model) and overt breakout (a state defects and races to ASI). Short-term risk is low: exponential hardware buildout barely keeps up with scaling, and most chips for next year's training runs don't exist yet.
- Long term: better algorithms could undermine verification; the authors estimate training would need 100x efficiency gains to keep the plan verifiable—roughly a decade at historical rates. Afterwards, training could resume or the pause be extended.
The paper also covers economic implications and what IR, history, and game theory say about pause design, concluding a pause is a practical policy option if leaders choose to pursue it.
More from AGI Musings
- Stanford-led team unveils Swap-seq: prime editing deletion screens to map enhancers, silencers, topological elements — anshulkundaje · 2026-10-10
- Scientists brief EU's Virkkunen on agent swarm risks: alignment lagging capabilities — S_OhEigeartaigh · 2026-10-10
- Mathematician John Urschel solved a decades-old problem with GPT Astra, staying optimistic on AI research — soumitrashukla9 · 2026-10-10
- 'Association for Human Mathematics' bans AI via three vows, mocked as self-defeating — chrisfirst · 2026-10-10
- Liv Boeree: AI probably isn't sentient, but torturing it is still bad for you — HarperSCarroll · 2026-10-10
- Data centers could become SpaceX's biggest revenue stream, with $1T revenue on the horizon — Dr_Singularity · 2026-10-10