AI Governance 'Plan A' Sparks Debate on Superintelligence Risks and Coordination

Recently, multiple AI safety experts discussed a governance proposal known as 'Plan A' (or 'AI 2040: Plan A'). Aimed at addressing the risks of superintelligence (ASI) seizing control or irreversibly concentrating power as predicted in scenarios like 'AI 2027,' the proposal advocates for delaying the intelligence explosion through great power coordination and a 'verified slowdown,' sparking widespread attention on the future of AI governance.

Core Mechanisms and Interventions

The primary goal of 'Plan A' is to buy time, delaying an ASI breakout originally forecasted for the late 2030s to the year 2040. To achieve this, experts like @JeffLadish note that the plan suggests embedding monitoring infrastructure in all major data centers. This would allow major powers like the US and China to verify each other's activities and forge arms-control-style agreements. Furthermore, @DKokotajlo emphasizes the importance of 'Total Research Transparency' in the era of superintelligence, viewing enhanced transparency as a key intervention to prevent excessive power concentration as AI deeply integrates into the economy and military.

Reactions and Route Controversies

The proposal has been praised by figures such as @JeffLadish as one of the most systematic and comprehensive responses to superintelligence to date. @eli_lifland compared it against alternatives like the full-speed 'Plan C/D' or the nationalization-like 'Plan B,' arguing that a 'verified slowdown' is the least bad option currently available. However, the plan also faces significant skepticism. @sebkrier pointed out that the proposal lacks concrete implementation paths for regulatory cooperation and compute tracking, warning against assuming that a highly cooperative and stable regulatory order will naturally emerge. Additionally, the 'dry tinder' controversy highlighted by @AdrienLE suggests that a slowdown could accumulate massive excesses in compute and algorithms, which could ironically and significantly accelerate an intelligence explosion if unleashed. @JeffLadish also acknowledged that as technology progresses, the deployment of secret data centers and untracked chips will become easier, posing a formidable challenge to monitoring efforts.

2026-07-10 ~ 2026-07-11 · 16 related posts

Full story(5 episodes)→