Mathematicians Are The Canary Down The AI Coal Mine

Main Ideas

1. Why the Navier–Stokes Equations Matter

The episode begins with OpenAI’s claim that a team of AI agents had proved a famous problem about the Navier–Stokes equations. The Clay Mathematics Institute attached a $1 million prize to the problem in 2000. The hosts explain why these equations matter before examining what the AI claim does—and does not—show.

The equations describe how fluids move, with applications ranging from blood flow and weather to aircraft and rockets. Their foundations include:

• The continuum approximation: Treat a fluid as a smooth substance rather than as separate molecules. This usually works when fluid conditions change over distances much larger than the gaps between molecules, but can fail in very thin air or at tiny scales, such as near spacecraft re-entry conditions or in capillaries.
• Conservation of mass: Fluid cannot simply appear or disappear.
• Newton’s second law: The equations relate a fluid’s acceleration to the forces acting on it.
• Pressure, viscosity, and external forces: These account for how fluid parcels push or drag on one another, as well as other forces acting on the fluid.

The discussion traces these ideas from Cauchy’s general description of fluid stresses to the more usable Navier–Stokes equations. A key simplification is assuming a Newtonian fluid, whose viscosity stays constant rather than changing with applied force, as it does in some non-Newtonian fluids.

2. What Makes the Equations Difficult

The equations include a nonlinear term: a fluid’s movement affects how its velocity changes elsewhere in the flow. The hosts compare its potential feedback to a microphone picking up sound from a nearby speaker and amplifying it repeatedly. Viscosity can act like a brake on this effect.

The famous mathematical question is whether, starting from smooth initial conditions, the equations always continue to have smooth solutions—or whether the flow can develop a singularity, with quantities such as velocity growing without limit in finite time. In fluid-mechanics terms, this is often described as the flow “blowing up.”

The episode illustrates how mathematicians study simpler special cases:

• Stokes flow: With very high viscosity and slow movement, viscosity dominates. Hannah Fry demonstrates this with food coloring in corn syrup: slowly stirring and then reversing the motion can bring the colors close to their original arrangement.
• Euler’s equations: Removing viscosity produces a simpler ideal-fluid model. The hosts describe this as a useful mathematical simplification, even though a perfectly viscosity-free fluid is an idealization.
• Boundary layers: Near a solid surface, fluid velocity changes from almost no movement at the wall to faster flow farther away. The hosts connect this to dust on fan blades and to dimples on golf balls, which affect airflow around the ball.

3. What OpenAI Claimed—and the Caveat

In the recording, the hosts describe OpenAI’s use of roughly 10,000 coordinated AI agents over about 88 hours, exchanging around 2.7 million messages. The transcript mentions an estimated computing cost of $15 million. The agents reportedly found a construction in which a smooth flow develops a singularity in finite time.

The important qualification is that the result depended on a complicated external-force term, denoted (F) in the discussion. The presenters argue that, while the work may technically address the problem as formulated, it may not answer the deeper question that motivated many mathematicians: whether the fluid equations themselves, without an artificially troublesome force, can produce a blow-up.

A later update in the episode says that OpenAI had published more than 700 mathematical manuscripts in the intervening weeks, including claims related to major open problems such as the Riemann hypothesis. The hosts say mathematicians had not yet had time to check or understand all of them.

4. Why Mathematicians Are Concerned

The criticism is not presented simply as fear that AI will replace mathematicians. The hosts describe several broader concerns:

• Correctness is not the same as understanding. A proof may establish a result without revealing why it is true or what it teaches us about fluid motion or the physical world.
• Proofs may be hard to inspect. The episode mentions Lean, a system that can check formal proofs, but stresses that verification does not necessarily provide human-readable insight. Some AI proofs could take years to understand.
• Attribution and influence are unclear. Mathematicians who worked on a related Euler-equation problem accused OpenAI’s models of possibly drawing on or echoing their work. The hosts note that people and research teams may have overlapped, making it difficult to separate independent discovery from shared influence. OpenAI, as described in the episode, denied stealing ideas.
• The motives behind the work matter. The hosts question whether mathematical breakthroughs are being pursued for insight or mainly to demonstrate AI capabilities to investors and the public. They argue that mathematics is valuable not just for producing answers, but also for curiosity, explanation, and deeper understanding.

The hosts compare this to earlier cases where a claimed proof was so difficult to follow that it did not immediately advance the field. They mention Shinichi Mochizuki’s lengthy proof of the ABC conjecture as an example of a result that other mathematicians struggled to evaluate and understand.

5. A Proposed Way to Think About AI’s Capabilities

Michael Stevens lays out three stages:

1. Interpolation: Generate something within the range of patterns and material already represented in the data.
2. Extrapolation: Use existing human knowledge to push beyond what people have already achieved. The hosts suggest that mathematical AI may be operating here.
3. Abstraction: Identify a deeper pattern that explains the data and phenomena in a new way. They say this would be a more profound kind of discovery than producing a correct but opaque proof.

The central warning is that mathematics may be an early indicator—or “canary in the coal mine”—of questions society will face elsewhere: whether we want merely rapid answers, or also the insight and human understanding that make discovery meaningful.

Other Segment

The episode includes a Cancer Research UK sponsorship message. It describes progress in cancer treatment and research, including the organisation’s work on detection technologies and potential treatments, and directs viewers to its website for further information.

Speakers and Sources Featured

Speakers

• Hannah Fry — Mathematician and fluid dynamicist; explains the equations and the AI controversy.
• Michael Stevens — Co-host; asks questions and discusses the implications of AI-generated proofs.

People, Institutions, and Sources Mentioned

• OpenAI — Its AI agents and mathematical manuscripts are central to the discussion.
• Clay Mathematics Institute — Established the $1 million prize for the Navier–Stokes problem.
• Cancer Research UK — Sponsor featured in an advertisement.
• Isaac Newton, Augustin-Louis Cauchy, Claude-Louis Navier, George Gabriel Stokes, and Leonhard Euler — Historical figures associated with the mechanics and equations discussed.
• Anthropic — Mentioned as another AI lab reportedly working on the problem.
• Shinichi Mochizuki — Cited in discussion of the ABC conjecture proof.
• Lean — Formal proof-checking system mentioned in connection with verifying proofs.
• An unnamed group of mathematicians — Described as having proved a singularity result for the simpler Euler-equation case.
• Deis Saris — Mentioned by Michael Stevens in an analogy about AI deriving a major theory from historical information.