AI 2027 team's AI 2040 'Plan A' slowdown proposal sparks fierce debate

On September 13, the AI 2027 team (Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, Daniel Kokotajlo) released 'AI 2040: Plan A,' a positive vision that delays superintelligence development to 2040, with two international governance schemes (Plan A and Plan S) responding to their earlier forecasts of AI takeover or irreversible power concentration. The report immediately triggered fierce debate over whether coordinated slowdown is feasible.

Confirmed

• Plan A's core mechanisms include continuous compute monitoring, guaranteed ability to destroy compute in emergencies, and research transparency requirements ensuring compute is used for agreed purposes (per DavidSKrueger).
• The transparency clause aims to sustain a slowdown equilibrium: compute is concentrated in the US and China, algorithmic insights naturally diffuse and arm competitors, so the proposal expects no one to pursue 100x-efficiency architectures and instead work only on directions that pay off at high compute scale (per anpaure).
• Gordic Aleksa enumerated Plan A's key assumptions, such as US–China cooperation, and raised preliminary critiques.
• Two CAIS researchers proposed 'Whole lab inspections'—mutual US–China auditor exchanges to verify AI pauses—countering the objection that international slowdowns cannot be verified (per DavidSKrueger).
• Commentator Kate Willett sharply criticized the report's end-state vision as having a 'machine god' flavor (via zetalyrae).

Unconfirmed

• ueaj's dual critique—too slow, since coordinated slowdown is needed far earlier than the timeline; too radical, since 60 billion H100-equivalent units by 2034 or sustained 50% GDP growth is unimaginable—remains personal opinion without author response.
• Alternative proposals remain contested: ueaj suggests negotiation representatives or strict hierarchies replacing full transparency, with all training information uploaded to a top layer that is transparent only among top actors; anpaure endorses the logic while calling full research transparency absurd, likening it to open-sourcing nuclear secrets.

Why it matters

• This is the AI 2027 team's first systematic move from forecasting to policy, putting concrete governance tools on the table while exposing the gap between 'everyone agrees on slowing down' and 'designing a workable slowdown mechanism,' providing a target for future international AI governance debates.

2026-09-13 ~ 2026-09-14 · 13 related posts

• CAIS researchers: US-China mutual 'whole lab inspections' can verify an AI slowdown — DavidSKrueger · 2026-09-13
• [source] AI Futures' AI 2040 report: monitored compute with emergency destroy as Plan A — DavidSKrueger · 2026-09-13
• AI2040's Plan A blasted as both far too slow and dangerously fast on compute scale — inerati · 2026-09-13
• AI researchers debate slowdown: full research transparency likened to open-sourcing nuclear secrets — anpaure · 2026-09-13
• [source] Gordic Aleksa Dissects Kokotajlo's 'AI 2040' Plan A: A Chain of Low-Probability Assumptions — gordic_aleksa · 2026-09-13
• [source] AI 2040: Plan A — Delaying Superintelligence via Total Research Transparency — gordic_aleksa · 2026-09-13
• Alternative to full AI transparency: per-country hierarchies with top-level coordination — _ueaj · 2026-09-13
• Debate: is research transparency the key to a slowdown equilibrium, or self-defeating? — anpaure · 2026-09-13
• Debaters question AI2040's transparency logic as compute concentrates in US and China — anpaure · 2026-09-13
• US and China own most compute; algorithmic proliferation arms your rival automatically — anpaure · 2026-09-13
• AI 2027 team unveils 'AI 2040: Plan A' positive vision, critics slam it as a 'Machine God' endgame — zetalyrae · 2026-09-14

2 near-duplicate retellings: _ueaj · anpaure