Wednesday, November 18, 2026

More On The Background To AI Math Training - And Answers To The AI Sample Problems

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.
In my Sept. 25th post, I noted the statement of the American Mathematical Society concerning this claim and the consequences. First, the AMS statement:

"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.)

Moving on to other AI advances:

  • 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.
I will be doing a post on the Riemann hypothesis in the near future and AI's role.

  • 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).
Next year I plan to do a two-part examination of the 4D Yang-Mills Existence problem and more on the AI contribution.

  • 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.  


Back now to the answers to the sample problems from my Nov. 12 post:

Grade school soln.   Take temp. 20    Subtract 80   =  - 60
Then add  30 =  - 60  +  (30)=  - 30        (Ans.C )   

SAT level soln.: First convert 26 mins. to hrs.  26/60 = 13/30 hr.

Then average speed of horse =   11 mi/ (13/30) hr. =  11 mi/ 0.433 hr.
 = 25.4 mph

H.S. competition soln.

Each complete circuit (spin) = 3600  
 
She spins a total of 2,25o degrees. Then total no. of spins:      

=  22500  /3600  

= 6.25 or 6 and 1/4
 
Since she's initially facing N the extra quarter turn at the end puts her direction East. (E)

No comments: