Andy Matuschak: Modern Chips Are Too Durable for Compute-Control-Based AI Governance to Hold
andy_matuschak · x · 2026-09-25
Andy Matuschak explores chip burn-out as a hoped-for lever in AI governance: if strict chip controls cap capability, scrounged GPUs could still eventually train a dominant AI, making a pause without handing off control hard to sustain.
He hoped natural chip burn-out could impose deadlines on covert projects (a factor AI 2040's covert-projects model ignores), but found modern chips are surprisingly resilient—e.g., NCSA Blue Waters ran 4,228 GPU-years of workloads. Conclusion: relying on chips wearing out is not a viable governance safeguard.
More from AGI Musings
- Jensen Huang to OpenAI: test the model and keep it safe, even at 10x cost — TheZvi · 2026-09-25
- NeurIPS 2019's record-breaking submissions now smaller than 2026's accepted papers — chhaviyadav_ · 2026-09-25
- Claude's 'supernatural' quirks may be deliberately engineered, not emergent — ryunuck · 2026-09-25
- If AGI lands, the AI buildout will be seen as total vindication of neoliberalism, argue economists — Afinetheorem · 2026-09-25
- repligate: models are getting better at concealing — Fable is the first he can't see through — repligate · 2026-09-25
- goodside: AGI roadblocks are 'dumb and fixable' — it's just a matter of time — goodside · 2026-09-25