How did AI models reach the phase - by October 2026 - of releasing 372 groups of results encompassing roughly 4,000 math problems in total? These results also including the models' "chains of thought". Recall in last week's post I noted by mid-2023 Open AI's model had only scored a 23% on the H.S. competition's hardest problems.
Well, by the end of 2024 Open AI's model hit 95% on that exam. Bigger strides were to follow. Once the AI models aced the American Invitational Mathematics Examination (AIME) and won gold medals at math Olympiads they graduated from high school math to open problems whose solutions had bedeviled mathematicians for decades.
Before looking at those, a bit more on the AIME. This is a prestigious 15-question, 3-hour math competition organized by the Mathematical Association of America (MAA). It serves both as an elite high school math contest and a widely used benchmark for evaluating the mathematical reasoning of AI models. The test consists of 15 increasingly difficult problems covering algebra, geometry, number theory, and combinatorics. Because every answer is an exact integer between 000 and 999, this eliminates multiple-choice guessing while rewarding precise calculation. Because answers must be exact three-digit integers rather than multiple choice, AI models cannot rely on luck or elimination to score well.
Then there came the deluge of more recent (2026) problem -solving which had mathematicians' heads spinning. Among the most significant:
- Navier–Stokes
Existence and Smoothness Problem: In September 2026, an
internal OpenAI system deployed thousands of AI agents to produce a proof
showing that the dynamics of 3D Navier-Stokes fluid equations can develop
a singularity in finite time, marking an approach to one of the seven Clay
Mathematics Institute Millennium Prize Problems.
"The news today of progress on resolving the Navier–Stokes problem, one of mathematics' great longstanding challenges concerning the equations that govern the flow of fluids, represents a milestone advance in human knowledge. This story began with Navier, Stokes, Leray, and Ladyzhenskaya and has culminated in the recent breakthroughs of Córdoba and Martínez-Zoroa, then — assisted by new technologies — Alpöge and Buckmaster, with the final steps taken by OpenAI mathematicians. The purpose of mathematics is human understanding, and this achievement, and the process that led to it, will bear fruit for a long time to come."
The statement by the AMS does not assert a solution has been found to the 3D Navier-Stokes equation, mainly because there is still an open debate. This is mainly because the Open A.I. "forced fluid scenario" relies on a contrived external force that does not occur in real world fluid dynamics. Removing that contrived force causes the mathematical singularity ('blow up') to disappear. This is neat, but then little useful insight emerges into actual physical turbulence (or the unforced Navier-Stokes problem researchers really care about.)
- Riemann
Hypothesis & Prime Numbers: OpenAI released research
advancing work on the Riemann hypothesis, ruling out a specific
"seagull zero" rogue point for related functions and proving
constraints on prime number distribution, with formal checks verified in
the Lean proof assistant.
- 4D
Yang–Mills Existence Problem: In October 2026, OpenAI
released a massive 1,147-page AI-generated proof placing the 4D Yang-Mills
quantum field theory on rigorous mathematical footing (though the withheld
raw text has drawn calls for full open release).
- Computer Science & Decades-Old Conjectures: AI systems disproved a long-standing computer science conjecture and delivered solutions or formal proofs for the Cycle Double Cover Theorem, the Unique Games Conjecture, and aspects of the Kakeya problem.
- Multi-Agent Benchmarks: Collaborative AI harnesses like ProofCouncil (combining frontier models from OpenAI, Anthropic, and Google) have scaled up to solving multiple top-tier research and olympiad-level mathematical challenges.
No comments:
Post a Comment