AI 2027 Team Unveils Plan A: Global Coalition to Delay Superintelligence Until Safety Is Assured
sjgadler · x · 2026-09-10
The AI 2027 team (Thomas Larsen, Daniel Kokotajlo, Ryan Greenblatt, Eli Lifland, et al.) has published Plan A at ai-2040.com, a positive vision to delay superintelligence via an international coalition until its safety is highly confident.
Core principles: slow down whenever needed for safety confidence; near-total research transparency; broad diffusion of frontier AI across many companies and countries; self-destruct switches on datacenters for reversibility.
Key steps: audit the compute supply chain to bound "dark compute"; temporarily pause AI training while building better verification tech and secure R&D datacenters; resume R&D under safety-case-based regulation and total transparency.
Immediate actions: keep a small gap between internal and publicly deployed models, publish model specs and whistleblower channels, and build government AI capacity.
Related event: AI 2027 team unveils optimistic AI 2040 'Plan A'(3 posts)→
More from AGI Musings
- Economist Backs Tough Sanctions for Academic AI Writing: 'Acceptable Amount Is 0%' — paulnovosad · 2026-10-27
- Academics Debate: Is a <1% False Positive Rate Enough for '100% AI' Writing Labels? — paulnovosad · 2026-10-27
- Jevons paradox in software: cheaper AI means more demand for dev teams, not less — teodorio · 2026-10-27
- Doing all work with AI is a culture shift, not a workflow — gabriel1 · 2026-10-27
- Does meritocracy survive when models crack open math problems? — msalbergo · 2026-10-27
- tszzl: The weakest anti-tech arguments rely on limitations of the technology — tszzl · 2026-10-27