OpenAI's NS math result reportedly took $15M of compute; efficiency questioned
rbhar90 · x · 2026-09-14
Robert Bhargava, processing OpenAI's claimed NS mathematics result, notes one stat: OpenAI threw roughly $15M of compute at the problem.
His open questions:
- The breakthrough seems to have already been made by human mathematicians; in another universe, could $15M funding human mathematicians have solved it instead?
- The hard part is unknown denominators — how many problems did OpenAI try that failed? Citing his DeepRetro paper's 8% success rate for reaction step prediction, he back-of-envelope estimates $150M in failed millennium problem attempts.
Conclusion: still very impressive, but potentially dramatically less compute-efficient than funding human mathematicians.
More from AGI Musings
- Dario Amodei's new essay calls to pace the frontier; Anthropic opens systems to third-party evaluators — ZeroStateReflex · 2026-09-14
- RL luminary Szepesvári: researchers actually want AI models that generate insight — CsabaSzepesvari · 2026-09-14
- Pedro Domingos: swap "AI" for "interpolation" and doomer screeds fall apart — pmddomingos · 2026-09-14
- Commentary: frontier AI leaders don't trust each other, so they ask government to chaperone — bilawalsidhu · 2026-09-14
- Cognitive Scientist's Essay: Rethinking Human-AI Relations After Cyberattacks and Navier-Stokes — DistractedDendrite · 2026-09-14
- TCS researcher reflects on problem-solving vs theory-building in the AI era — jasondeanlee · 2026-09-14