OpenAI Made History. Nobody Understands It.

Summary

The video examines OpenAI’s claim that its AI system solved a major open problem about the Navier–Stokes equations—the mathematical rules used to describe how fluids such as water and air move. Its central point is that a computer-checked result may be correct while remaining difficult for people to understand, verify in context, or use to make further discoveries.

The problem: Can the equations “explode”?

• The Navier–Stokes equations describe how a fluid’s motion changes over time. They are used in areas such as weather prediction and aircraft design.
• Mathematicians have long asked whether the equations always produce sensible results, or whether a fluid’s speed could become infinite in a finite time. This hypothetical breakdown is called an explosion or singularity.
• The video explains the idea with a whirlpool: as a vortex gets narrower, it may spin faster and faster.
• The problem is associated with a million-dollar prize established by the Clay Mathematics Institute in 2000.

Two possible ways to address the problem

• Prove the equations remain well-behaved: Show that an explosion can never occur.
• Show an explosion can occur: Construct a valid example where the equations produce one.
  • In the example discussed, mathematicians may use an imagined “spoon” to stir the fluid.
  • The stirring must be smooth; abrupt or twitching movements do not count.
  • The challenge is to create a vortex that tightens and accelerates until it blows up.

The video says mathematicians in Madrid had developed a vortex construction, but their stirring method did not meet the smoothness requirement. Tristan Buckmaster and Levent Alpoge then worked on that remaining step, using AI as part of their effort.

OpenAI’s AI effort

• OpenAI reportedly used 10,000 copies, or “agents,” of a new AI system. The agents explored solutions, compared approaches, and revised their views in response to one another.
• After about 88 hours, the system produced a proposed counterexample: a smoothly stirred vortex that becomes narrower and faster without bound.
• A second AI system translated the proof into Lean, a formal proof-checking language. The video compares Lean to a calculator checking a sum: it can verify the formal steps it is given.
• The process reportedly cost millions of dollars. The video emphasizes that the AI completed in under four days a step that human researchers had worked on for about a year.

Questions about the proof and its verification

• A Lean certificate is important evidence that the formalized proof’s steps follow correctly. But it does not automatically establish that the formal code faithfully represents the proof’s English explanation.
• The video reports that three mathematicians found a place where the code and the English text did not match. They did not say this proved the result false, but one worried there could be other such discrepancies.
• It also reports that OpenAI’s account of how its AI used the Madrid team’s method was challenged by experts, who said the AI’s approach was different.
• Terence Tao and other mathematicians were described as unable to fully understand the proof. Tao’s reported concern was that, despite the formal certificate, the work had yielded little mathematical insight so far.
• The video’s analogy is a recipe translated into another language and checked by a cook: the cook can verify that the translated recipe works, but may not know whether the translation changed the original ingredients.

Disagreement over credit and conduct

• Buckmaster said that OpenAI researcher Sébastien Bubeck proposed a deal under which Buckmaster could be named in the proof but his colleagues could not, because one worked for a competitor. Buckmaster also described the exchange as threatening.
• Bubeck disputed Buckmaster’s account, denied trying to silence him, and apologized for comments about Buckmaster’s career.
• Buckmaster posted his work online shortly before OpenAI announced its result.

Broader concerns about AI-generated research

• The video says OpenAI later released 722 mathematical papers at once, with its AI addressing thousands of unsolved problems. Some papers were withdrawn the same day.
• This prompted concerns, described in the video as a “mathematical apocalypse,” about how researchers can assess a flood of AI-generated results.
• Tim Gowers is presented as worrying that young mathematicians may have less incentive to spend years mastering mathematics if AI can generate results faster than people can understand them.
• The video stresses that the central unsolved question remains: what happens without externally stirring the fluid? The result discussed concerns a carefully constructed, forced scenario, not every version of the Navier–Stokes problem.
• It also notes that the proposed vortex reaches an extremely tiny scale—described as roughly a thousand times thinner than a human hair—where the continuum equations may not directly represent real water. The video therefore says ordinary weather forecasts are not at stake.
• The larger lesson is about a growing gap between producing answers and understanding or trusting them. The same issue could arise in fields such as drug discovery, batteries, and materials science. The narrator argues that an opaque result could still be valuable if it leads to major practical benefits, but says understanding remains an important part of science.

Speakers and sources identified

• Dave Lee — narrator.
• Tristan Buckmaster — mathematician whose account of his interaction with Bubeck is presented; the video notes that Bubeck disputes it.
• Sébastien Bubeck — OpenAI researcher; his denial and apology are described.
• Terence Tao — mathematician whose concerns about understanding and mathematical insight are discussed.
• Tim Gowers — mathematician whose concerns about the effects on younger researchers are discussed.
• Levent Alpoge — Buckmaster’s collaborator, identified as working at Anthropic.
• Luis Martínez Orovio — named in the subtitles as a Madrid-based mathematician who had worked on the problem for years.
• Other sources or institutions mentioned: OpenAI; Anthropic; the Clay Mathematics Institute; Lean; Scientific American; mathematicians in Madrid.