Practitioner slams US AI safety bill: 10^25 FLOP threshold would freeze innovation, fund lawyers
sanjaykalra · x · 2026-09-24
Enterprise AI practitioner Sanjay Kalra publishes a long critique of a proposed US AI "safety" bill, calling it a panic statute rather than real regulation.
Key points:
- He concedes the industry needs guardrails (hallucinations, unauthorized agent actions, security incidents) but argues the bill's definitions — "exceeds human cognitive performance across most domains" — are written for prosecutors, not engineers.
- The 10^25 FLOP threshold captures nearly every frontier model enterprises rely on; the "precursor characteristics" list sweeps in dual-use research like drug discovery, with no carve-outs for medicine or materials science.
- An immediate pause on training or fine-tuning "advanced" models until a new cabinet department writes rules means useful work stops first.
- Consequences: legal uncertainty freezes budgets, drives talent overseas, and a unilateral US freeze won't stop China or other programs from continuing to train.
- He argues export controls and allied coordination are the only instruments that have ever constrained technologies of this profile.
More from Safety
- Proposal: a frontier AI mutual insurer to enforce safety standards with skin in the game — Miles_Brundage · 2026-09-24
- Jade Leung steps down as UK PM's AI Adviser, becomes Vice-Chair of AI Security Institute — ersatzben · 2026-09-24
- UK appoints Dr Jade Leung as Vice-Chair of the AI Security Institute — Tom_Westgarth15 · 2026-09-24
- AI Is Compressing Cyberattack Timelines in Healthcare, and Incident Response Plans Aren't Ready — moniquejmorrow · 2026-09-24
- Are Conversation Logs Enough to Debug an AI Agent That Took a Wrong Action? — Informal-Dust4499 · 2026-09-24
- Lawfare Proposal: Frontier AI Labs Should Build a Mutual Insurer to Price Catastrophic Risk — Manderljung · 2026-09-24