Safety researchers propose $10M compute pre-release tests for destabilizing AI capabilities

Jsevillamol · x · 2026-09-20

Greg Burnham proposes frontier labs spend $10M of compute in pre-release testing on easily verifiable "destabilizing" innovations like polynomial factorization. Jsevillamol agrees, noting inference scaling favors defense only if compute holders actually commit a significant budget to it, and argues testing shouldn't stop at verifiable domains: try overcoming governments, designing lethal pathogens, and other hard-to-verify risks.

Original post →

More from Safety

Safety channel →