Proposed 'speed limit' on AI progress could delay expert-surpassing AI by 1-2 years
ronbodkin · x · 2026-11-19
Ron Bodkin weighs in on AI governance pacing mechanisms, backing a speed limit on measured model progress and proposing a new one: testing human comprehension of automated AI R&D, so only AI contributions people can validate get used — serving both as a safety-case requirement and a pace cap.
His preliminary modeling suggests holding measured progress at the current rate would delay AI that outperforms top experts (TED-AI) by 1-2 years, depending heavily on humans' ability to judge AI ideas' real validity. He agrees with Anton Leicht's warning that the overhang if the limit is lifted would be large.
Related event: Capping AI capability growth could delay expert-level AI by 1-2 years(3 posts)→
More from Safety
- General Catalyst CEO Taneja: AI race is a prisoner's dilemma that someone must break first — htaneja · 2026-11-19
- 脑洞:生成自带 AI 伪影的字体,让截图一律被判为 AI 生成图 — justalexoki · 2026-11-19
- NGINX2Shell: pre-auth RCE found in NGINX 1.31.8 mainline, disclosed with OpenAI's help — Eric_Wallace_ · 2026-11-19
- White House mulls disclosure rule for Chinese open-weight models; founder says distillation sidesteps it — bindureddy · 2026-11-19
- Federal audit: 4 major US AI models produced info that could help plan mass-casualty attacks — Polymarket · 2026-11-19
- VCs argue AI should be superaligned to end users only — and say legislation should mandate it — StewartalsopIII · 2026-11-19