Summary
OpenAI announced that a group of AI agents, using a next-generation model described as significantly more capable than the recently released GPT-6 Astra, has produced a proof for the Navier-Stokes Millennium Prize problem — a fluid-dynamics equation unsolved for 90 years and carrying a $1 million bounty. Matthew Berman explains what the Navier-Stokes equations are, why they matter practically (aircraft design, weather prediction, chip cooling), and what it means that an AI system has now cracked one of mathematics’ hardest open problems.
But the more immediately charged story involves two mathematicians — Tristan Buckmaster (an Anthropic employee working on a personal project) and Levent Alpaji — who spent a year on the same problem using OpenAI’s Codex tool, reached a breakthrough in mid-August, and then watched OpenAI announce its own solution days after learning of their progress. Buckmaster publicly accused OpenAI of using the direction of their private Codex drafts to prompt its model toward the solution, and of rushing publication to control attribution credit. OpenAI categorically denied accessing private user data while simultaneously acknowledging it “cannot rule out” that de-identified training signals from their usage helped improve the model.
Berman uses the episode to discuss recursive self-improvement, the blurring line between user data and model training, and the practical implication for anyone building on top of frontier AI platforms: assume your work may eventually inform the capabilities of the model you’re competing with.
📺 Source: Matthew Berman · Published September 09, 2026
🏷️ Format: News Analysis
