25-Scholar Paper Proposes "Hardwired Pause": Phase Out Training Chips, Keep Inference Only
A new working paper, "The Feasibility of a Hardwired Pause of Frontier AI Training," co-authored by 25 interdisciplinary researchers, proposes achieving a hard global pause on frontier AI training through a "Hardwired Pause": gradually phasing out AI chips capable of training and deploying only inference-only chips, making new large-scale training physically difficult. David Duvenaud shared and recommended the paper.
Confirmed
- The paper "The Feasibility of a Hardwired Pause of Frontier AI Training," written by 25 interdisciplinary scholars, explores the feasibility of a global pause on frontier AI training
- wfithian (Will Fithian) proposed a path to making an AI pause "sustainable": bake the restriction into hardware by phasing out training chips and switching to inference-only chips; he collaborated with Mithuna, creator of the science channel Looking Glass Universe, on an accompanying video
- Roko Mijic relayed the same idea, arguing it would give an AI pause durability
- In X discussions around the compute governance policy paper, Duvenaud, wfithian, alextmallen and others clarified that the policy's goal is not a AI "slowdown" but a durable pause on frontier training
Why it matters
- The proposal tackles the core challenge of AI governance—how a pause can be enforceable and sustainable: traditional policies are easily circumvented by compute diffusion or regulatory arbitrage, whereas physical hardware-level limits are far harder to evade
- The joint push by multiple scholars and science communicators suggests the "inference-only chips" approach is moving from academic concept to public debate, potentially shaping future arguments over compute export controls and chip design
2026-10-31 ~ 2026-11-02 · 5 related posts
Primary sources
- Proposing a 'Hardwired Pause': swap training chips for inference-only chips — wfithian ·
- Paper with 25 coauthors argues a hardwired AI training pause is feasible via chip bans — birchlse ·
- Researchers debate feasibility of compute-governance policy for durable pause on frontier AI training — wfithian ·
- [source] Researchers debate feasibility of compute-governance policy for durable pause on frontier AI training — wfithian · 2026-10-31
- [source] Paper with 25 coauthors argues a hardwired AI training pause is feasible via chip bans — birchlse · 2026-10-31
- Proposal: make an AI pause durable by phasing out training chips for inference-only hardware — wfithian · 2026-11-02