AI proofs now span hundreds of thousands of Lean lines — verification becomes a second compute meter
demian_ai · x · 2026-10-14
- demianai notes AI-written math proofs now run to hundreds of thousands of lines of Lean, the mechanical proof-checking language. No human will read them; only the checker knows they're right — so trust shifts to whoever wrote the checker.
- He argues this "verification tax" is coming for everything agents touch: OpenAI says it's spending more and more compute on secondary monitoring, a second system watching the working agent. Every deployed agent quietly runs a second meter.
- His framing: file verification under inference demand that scales with the first customer, not safety cost. When quoting agent token growth, ask how much is the work vs. checking the work. The more autonomous the agent, the bigger the watcher.
More from AGI Musings
- Judea Pearl maps Western civilization's "narratives" to assumptions in causal models — yudapearl · 2026-10-14
- Researcher: The world economy is being retooled for imminent human obsolescence — zetalyrae · 2026-10-14
- Viral post captures young professionals' despair: 'I'll never get to do that because Dario said so' — zetalyrae · 2026-10-14
- Researcher Reflects: 'Amazing' AI Discoveries Forgotten Within Weeks — ftm_guney · 2026-10-14
- Zuckerberg calls out AI 'pacing' rhetoric, Domingos voices support — pmddomingos · 2026-10-14
- Founder: Bangalore feels 20 years behind SFO in AI adoption and ambition — prasanna_says · 2026-10-14