AI 2027 team's Plan A proposes slowing AI via China deal and compute monitoring
Daniel Kokotajlo and Ryan Greenblatt offer a controversial positive vision for AI safety.
Deep Dive
The team behind the AI 2027 report has released Plan A, a proposal to slow AI development through international cooperation. Key elements include a deal with China, global monitoring of compute hardware, and mutually assured compute destruction (MAIM) to enforce limits. They predict ASI by 2040 and vast economic growth. Zvi introduces the plan without endorsing it, noting that Timothy Lee questioned the plausibility of near-term superintelligence.
Key Points
- Plan A proposes US-China deal to monitor compute and use MAIM (mutually assured compute destruction) to enforce slowdowns.
- The same team accurately predicted many AI developments in their 2025 report AI 2027; now they forecast ASI by 2040.
- Critics like Timothy Lee argue the plan's assumptions about superintelligence are implausible, creating an 'epistemic chasm'.
Why It Matters
If enacted, Plan A would reshape global AI governance—but its feasibility hinges on unproven cooperation and disputed timelines.