AIFP's Plan A risks 600x faster AI explosion if deal fails
Massive compute buildup could turn a year-long intelligence explosion into a single day.
Tom Davidson's LessWrong post outlines a critical flaw in AIFP's Plan A for safely developing superintelligence. The plan envisions a US-China deal to pause software progress from ~2030 while continuing to scale compute hardware. This creates a massive 'compute overhang'—by 2033 total compute increases ~100x, by 2040 another ~100x (10,000x total). Since each 10x compute speeds up an intelligence explosion by ~5x, a deal breakdown in 2033 would make the explosion 25x faster; by 2040 it would be 600x faster. Additionally, software techniques accumulate publicly, so defectors could instantly implement banned algorithms. The result: a fast, uncontrollable race to superintelligence, with many projects capable of triggering it within weeks or even days.
AIFP's solution is Mutually Assured Compute Destruction (MACD): US data centers in Mongolia require Chinese 'continue' messages, and vice versa with China's centers in Canada. If either side detects a violation, they destroy the other's compute. However, Davidson identifies three challenges: detection must be nearly instantaneous (days may be too late), defectors could exfiltrate weights and algorithms before destruction, and multiple projects may independently attempt explosions. The dry tinder problem fundamentally questions whether Plan A's cure is worse than the disease—especially if the default trajectory already carried lower extinction risk.
- Compute overhang: 100x by 2033, 10,000x by 2040, accelerating explosion 25x–600x if deal breaks.
- Software overhang: public algorithm research lets defectors instantly implement dangerous techniques.
- MACD detection must be near-instantaneous—days could allow complete intelligence explosion before destruction.
Why It Matters
AI safety plans that stockpile compute might create a faster, more dangerous extinction risk than the problem they solve.