AI treaty verification may need pragmatic hardware checks before “math” solutions
geoffreyirving · x · 2026-07-25
The thread argues that AI treaty verification tech may be needed either if humans reach agreement on slow/stop treaties or if superintelligent systems create a vulnerable-world scenario mid-ramp.
It then sketches a three-level framework for treaty verification:
- Pragmatic
- Enclaves
- Math
For the pragmatic tier, the author defines it as methods that do not inspect the code being run. Examples include datacenter inspections, hardware location tracking, and chip controls or constraints that make training harder than inference. The point is that there is a broad design space of practical approaches that hardware, security, and crypto people should explore.
Related event: Geoffrey Irving Outlines Three-Layer Approach for AI Treaty Verification(5 posts)→
More from AGI Musings
- ClickHouse and Nvidia argue open models will matter across safety, cost and performance — mattturck · 2026-07-25
- AI Search Dries Up Web Traffic: The Era of Google Zero Is Here — The Verge AI · 2026-07-25
- Claude Opus 5 rates its own moral patienthood at 41% in automated interviews — Sauers_ · 2026-07-25
- AI lowers the entry bar, but problem selection and verification still favor the top 1% — tokenbender · 2026-07-25
- A new AGI essay argues the field is climbing the same mountain from two slopes — op7418 · 2026-07-25
- Open models are fine, but the U.S. wants frontier open models from U.S. companies — BenBajarin · 2026-07-25