AI Regulation Needs an Asilomar-NPT Playbook, Not a 'China Will Win' Test
krishnan · x · 2026-09-14
The author argues every AI regulation debate in DC ends with "But China will win" — the wrong test for the wrong problem. Regulation is a bet on where the bigger error sits: over-regulation freezes a technology compressing decades of productivity into years; under-regulation surfaces failure modes in production with no rollback. By 2026 both risks are live.
Multipolar reality: 151,448 derivative models built on Qwen alone by mid-2026 (180–210 repos/day); Switzerland's Apertus, UAE's Falcon, France's Mistral, India's Sarvam 105B, and Korea's K-EXAONE all ship weights that answer to neither Washington nor Beijing.
Historical precedents: the 1975 Asilomar recombinant DNA pause (practitioners write technical rules) and the 1970 NPT (contain what you can count).
A two-layer design: an Asilomar layer of published safety frameworks, independent pre-release evaluation, and incident reporting (harmonizing California's SB 53 and the EU AI Act); and an NPT-style multilateral pact on a short list of capabilities — autonomous cyber offense and bio uplift — verified at the compute chokepoint, since frontier training runs need tens of thousands of chips in one visible building.
More from AGI Musings
- TSMC paces the frontier by accident whenever it underestimates chip demand — dan_s_becker · 2026-09-14
- Why humans may still beat frontier LLMs at open-ended AI research — flowersslop · 2026-09-14
- Google AI x Econ team finds field evidence that prior expertise drives learning from AI-assisted work — soumitrashukla9 · 2026-09-14
- Debate: don't ban open source — ensure aligned models out-compute misaligned ones — basedjensen · 2026-09-14
- Consumer agents: open questions on agentic commerce, incentives and multi-agent patterns — illscience · 2026-09-14
- Hot take: disqualify anyone who called GPT-2/3 risky from AI safety debates — StewartalsopIII · 2026-09-14