Inference-only chip proposal would bar RL rollouts for models in training
wfithian · x · 2026-10-31
Clarifying a compute-governance debate between David Duvenaud and teortaxesTex, wfithian explains the design of "inference-only chips": they can efficiently run inference only on approved models, not on arbitrary weights — inference for general weights is defined as a training operation, so RL rollouts for models being trained wouldn't be allowed. Duvenaud prefers "slowdown" over "pause" and admits workability is uncertain, but argues the proposal pushes out the workable-vs-effective Pareto front.
Related event: Debate: Can More Inference Compute Substitute for Training Compute?(2 posts)→
More from Safety
- Wired podcast: a contractor's Claude agent got into a Pentagon system — Wired AI · 2026-10-31
- Tyler Alterman on the hacking panic: more Y2K than COVID, use 1Password and update software — jd_pressman · 2026-10-31
- Anthropic's VS Code extension on Cursor backdoored for nearly 12 hours, up to 5,000 devs exposed — Ars Technica AI · 2026-10-31
- Security researcher: vibe coding will force software to consolidate around mature products — jd_pressman · 2026-10-31
- Every Used Codex and Agents to Survive the Railway Credential Breach — every · 2026-10-31
- Researchers debate whether more inference compute can substitute for training compute — teortaxesTex · 2026-10-31