SuperDataScience Podcast Episode 1038: AI's Contentious Math Breakthroughs and the Verification Bottleneck
Host Jon Krohn unpacks the remarkable and contentious run of mathematical results produced by AI recently, focusing on OpenAI's claim to have cracked part of a Millennium Prize problem. He details three major caveats regarding the scope of the proof, credit disputes, and shaky verification, concluding that for AI builders, when generation becomes cheap, verification becomes the bottleneck.
SuperDataScience Podcast Episode 1038 delves into the recent controversy surrounding AI's achievements in mathematics. Host Jon Krohn analyzes OpenAI's claim that approximately 10,000 autonomous agents running on an unreleased internal model found a blow-up singularity in the 3D Navier-Stokes equations. This relates to one of the Clay Mathematics Institute's seven Millennium Prize problems, previously deemed inaccessible to computers by Fields medallists. According to The Economist, the effort took 88 hours, involved 2.7 million inter-agent messages, cost an estimated $6.5M in compute, and included 17 hours of Lean formalization using GPT-6 Astra.
However, Krohn highlights three significant caveats. First, OpenAI resolved only two of the four official variants (the forced ones C and D), leaving the unforced variants A and B—which many mathematicians consider the heart of the problem—unsolved; OpenAI is not claiming the prize. Second, there is a messy credit dispute involving Tristan Buckmaster of New York University and Levent Alpöge of Anthropic, who were also using AI to hunt for blowups in related fluid equations. OpenAI acknowledged it cannot rule out that their drafts ended up in its training data. Third, verification remains shaky: after dumping 722 manuscripts to GitHub on October 6th, OpenAI withdrew three due to a sign error and conceded that about half were released without full verification.
The episode also surveys the backlash from the mathematical community. A joint declaration by 25 Fields medallists, endorsed by over 8,000 researchers, warns that the goals of AI companies are severely misaligned with those of mathematicians, arguing that understanding, not just solutions, is the point. Despite criticism, some scholars like Ying Shieh at Kennesaw State University view the release as a mine of new lemmas and proof techniques. Krohn concludes with a lesson for AI builders: when models can generate hundreds of plausible proofs or thousands of lines of code in hours, generation stops being the bottleneck. Verification becomes the critical constraint, and only results paired with machine-checkable evidence and human understanding will hold up.