OpenAI's Math Results Move Expert Judgment Into the Prompt: 722 Papers, an 88-Hour Navier-Stokes Proof
shashib · x · 2026-11-03
OpenAI's recent math results mark a shift where expert judgment moves into the prompt: in September its agents produced a Navier-Stokes Millennium Prize proof using 10,000 agents over 88 hours, formalized in Lean; on October 6 the company released 722 AI-written manuscripts on hundreds of open problems, 162 with main results formalized in Lean. The Clay Mathematics Institute says the problem has "apparently been settled," with expert review ongoing.
Rutgers math professor Alex Kontorovich notes mathematicians working with AI focus on two jobs: stating the problem and checking the result. The author extends this split to enterprise teams writing instructions for agents — budget people for both. An arXiv 2025 estimate adds context: a 1,000-step argument at 99% per-step accuracy has just a 0.004% survival odds, underscoring why formal verification matters.
More from AGI Musings
- Hot take: robotics value will accrue to users of robots, not robot makers — ZeYanjie · 2026-11-03
- Sovereign AI is creating the sharpest haves-vs-have-nots split in 30 years, says investor — arieljalali · 2026-11-03
- DHH: programmers may be worse at AI coding than non-coders — '100%' — 2C_ornot2C · 2026-11-03
- Hypothesis: high-progress places tend to have lots of cults, from Bauhaus to AI safety — Afinetheorem · 2026-11-03
- Researchers propose adding Readability to ARR review criteria amid LLM-generated paper flood — TuhinChakr · 2026-11-03
- Viral post: the next generation will never know 40-hour weeks, student loans, or aging — rand_longevity · 2026-11-03