GovAI / Oxford Compute & Safety Cases Team (@govai_oxford)
Centre for the Governance of AI (GovAI) researchers authoring 'Computing Power and the Governance of AI' (arXiv:2402.08797) and 'Safety Cases for Frontier AI' (arXiv:2403.10462).
GovAI / Oxford • arXiv:2402.08797 & arXiv:2403.10462
Immutable Forecast Track Record (0)
Authored & Cited Arguments (4)
Without strict API distillation rate-limits and allied compute consolidation, a domestic slowdown faces a 12% 'Offshore Rivals Catch Up' leakage risk (arXiv:2402.08797).
As documented in 'Computing Power and the Governance of AI' (arXiv:2402.08797), synthetic data distillation allows smaller clusters to replicate 80–90% of a frontier model's capabilities at <10% of original training FLOPs unless high-bandwidth reasoning traces are cryptographically throttled.
Flexible Hardware-Enabled Guarantees (FlexHEGs, arXiv:2402.08797) solve the bilateral verification dilemma by cryptographically proving on-chip compliance without exposing sovereign secrets.
Once aligned Safer-4 gives the US a decisive 5:1 lead and designs tamper-evident FlexHEG security processors, the US can offer China a verifiable grand bargain: shared access to aligned superintelligent abundance and guaranteed sovereignty in exchange for installing FlexHEG attestation modules on all Tianwan CDZ accelerators.
Plan B ('Kinetic Containment' & Mutually Assured AI Malfunction, arXiv:2503.05628) treats unauthorized ASI megaclusters like clandestine nuclear enrichment facilities.
As analyzed in 'Superintelligence Strategy' (Hendrycks, Schmidt, Wang, arXiv:2503.05628), gigawatt-scale datacenters have unmistakable thermal and electrical signatures from orbit. Plan B enforces non-proliferation via supply-chain interdiction, cyber sabotage, and credible kinetic deterrence against uninspected frontier runs.
Plan A operationalizes 'Safety Cases for Frontier AI' (arXiv:2403.10462): shifting the burden of proof so labs must provide auditable control and alignment guarantees before crossing each capability tier.
Borrowed from nuclear and aviation regulation, a structured Safety Case (Clymer, Baker, Brundage et al., arXiv:2403.10462) combines inability arguments, AI Control red-teaming, and mechanistic transparency into an internationally audited dossier required for every training run.