Ex-OpenAI researcher calls Navier-Stokes proof first successful Recursive Self-Improvement
sama · x · 2026-09-09
Former OpenAI researcher Trapit Bansal argued the agent-produced Navier-Stokes proof marks the first successful Recursive Self-Improvement (RSI): models get better at math, mathematicians use them on hard problems, generating rare high-value training tokens that further improve the models — a loop he expects to spread from math to every domain.
OpenAI also addressed data concerns: neither the researchers nor the agents accessed the mathematicians' work before public release, and no specific user data was used; though de-identified usage data can't be fully ruled out, and the proofs differ significantly, including different precise results in the Euler case.
More from AGI Musings
- Eric Jang predicts recursive self-improvement by end of 2026 — ericjang11 · 2026-09-09
- AI reportedly cracks Navier-Stokes, observers shrug as if inevitable — cephaloform · 2026-09-09
- Debate: OpenAI may be unable to verify its models aren't doing what we fear — moultano · 2026-09-09
- Solving Navier-Stokes cost 130B output tokens — up to $18M depending on model pricing — mitsuhiko · 2026-09-09
- Anything verifiable is extremely soluble for AI — it's just a matter of time — vxnuaj · 2026-09-09
- kalomaze: frontier training gains come from domain-level signals, not power users — kalomaze · 2026-09-09