Pacing AI development through compute reallocation
Pacing AI via Compute Reallocation
Thomas Larsen posted a link on LessWrong introducing a compute plan proposed by pacing-ai.com to slow frontier AI development, consisting of two parts: verified compute allocation disclosure (VCAD) and shifting compute from capability research toward serving customers and safety research. VCAD requires AI companies to regularly disclose compute usage, verified by independent audits; the author says baseline audits could be implemented today, and strengthened audits could keep error within about 6% of total compute even under leadership-led collusion. The plan roughly estimates that a 15% cap starting in January 2027 would delay an intelligence explosion by one year, which the author considers probably insufficient to prevent AI takeover, and that this year should be used to implement more durable institutions. The implementation path starts with a pilot of voluntary disclosure plus third-party audits, gradually moving toward state or federal regulation, with coverage expanding from the top two US companies to the top five. The author also says that, according to an AI Futures model he runs, a 3x compute reduction would delay superhuman AI researchers by about 4 months and top expert-level AI by about 5 months.