AI Futures Project proposes slowing frontier AI through compute reallocation
Pacing AI via Compute Reallocation
Thomas Larsen of the AI Futures Project released a frontier AI slowdown proposal, centered on two mechanisms: first, Verified Compute Allocation Disclosure (VCAD), under which AI companies regularly disclose how compute is used and submit to independent audits; second, reallocating compute from capability R&D to serving customers and safety research. The author roughly estimates that if a 15% cap on capability R&D compute were implemented starting January 2027, it could delay an intelligence explosion by one year, but this is very likely insufficient to prevent AI takeover, and that year should be used to build more durable policy mechanisms. The proposal designs workload records, capability and non-capability classification, and a regular disclosure process, paired with two tiers of third-party audits, basic and enhanced; the author estimates that, provided auditors have sufficient internal information, even in the face of management-led collusion, enhanced audits can keep error to about 6% of total compute. The text also lists seven types of cheating attacks, and recommends starting with voluntary disclosure, gradually transitioning to state or federal regulation, and moving toward international regulation in the long term.