OpenAI's AI Agents Solve the Navier-Stokes Millennium Prize Problem

Matthew Bermango watch the original →

OpenAI claims their next-generation model agents solved the 90-year-old Navier-Stokes fluid dynamics problem in 88 hours, sparking controversy over data usage and credit with independent researchers.

The Breakthrough

OpenAI announced that a group of AI agents, powered by an unreleased next-generation model, produced a proof for the Navier-Stokes Millennium Prize problem, a foundational challenge in fluid dynamics that has remained unsolved for 90 years.

Technical Execution

The solution was achieved through a massive, automated agentic process that spanned 88 hours of compute time. The workflow involved the following parameters:

  • The agents generated 4.9 million messages during the research process.
  • The total output reached 300 billion tokens.
  • The model utilized a recursive self-improvement loop, where the AI was tasked with accelerating its own research and experimental design.
  • The process began on September 1 and concluded on September 5.

The Credit Controversy

The announcement triggered a dispute involving mathematicians Tristan Buckmaster and Levent Alpoge, who were working on the same problem using OpenAI's tools. Buckmaster alleged that OpenAI may have accessed their private drafts stored within the platform to guide their own model's approach. OpenAI has categorically denied accessing specific user data, though they acknowledged that de-identified data derived from product usage may have informed model training. Sebastian Bubeck, representing OpenAI, stated that the company attempted to coordinate a joint release and denied requesting the removal of any author from the work.

Platform Risk and Future Implications

This event highlights the significant platform risk for businesses building on closed-source models. Because frontier labs acknowledge that user data may be used to improve future model iterations, proprietary expertise shared within a chat interface can be effectively absorbed by the model provider. This creates a scenario where the platform can potentially compete with or outperform the very users who provided the training data.

  • #ai
  • #math
  • #platform-risk

summary by google/gemini-3.1-flash-lite. probably wrong about something. check the source.