Key Takeaways

  • AI theorem provers from labs like OpenAI and Anthropic are claiming answers to landmark problems, including the Unique Games Conjecture, the Quasi-Riemann Hypothesis, and Navier-Stokes finite-time blowup.
  • Alex Kontorovich defines the "solved axiom" as a thought experiment: assume any math problem a human can solve with AI could eventually be solved by AI on its own.
  • Terence Tao warns that casual AI proof drops ruin the patient discovery process for graduate students, acting like plot spoilers on years of research.
  • Ankur Moitra points to the "doorman fallacy" to explain why reducing mathematicians to mere proof producers ignores their real value: framing the questions worth answering.

Automated Proofs Spoil the Training Run

When AI labs drop candidate proofs for problems that have stumped humans for decades, the initial reaction is awe. The second reaction is panic among the people who do the work.

OpenAI and Anthropic have pushed systems into territory once reserved for human specialists, targeting problems like the Unique Games Conjecture and the Quasi-Riemann Hypothesis. For senior researchers, these tools look like fast assistants. For graduate students and postdocs who spend four years wrestling with a single subproblem, a sudden machine-generated answer destroys their work overnight.

Lex Fridman highlighted Terence Tao's warning about these drops: “Much as one cannot unhear a movie spoiler or a crossword clue, one cannot explore a problem as profitably and richly once one is aware of an existing solution, and I mourn the path not taken and the opportunities lost in the frantic race to develop the technology.”

Mathematical training does not happen by reading the answer key. It happens during the two years of dead ends before the answer appears. When an automated prover skips the dead ends, it removes the friction that creates deep researchers. You get the output, but you kill the apprentice pipeline that invents future fields.

The Solved Axiom Meets the Doorman Fallacy

Alex Kontorovich pushed the tension further with a simple test. “Let's assume the solved axiom,” Kontorovich said, “which says any mathematical problem that can be solved with human interacting with AI could also have been solved by AI alone.”

If the solved axiom holds, the romantic view of human-in-the-loop math collapses. The human becomes a temporary crutch while the models scale their search steps in formal verification tools like Lean. As Ankur Moitra pointed out, this upends decades of academic tradition: “The playbook and the blueprint for success that we had for so long is now thrown out the window.”

Yet Moitra cautions against writing off human researchers too quickly. He points to the doorman fallacy. An efficiency expert looks at an apartment doorman and sees a person whose only mechanical task is opening a door. You can replace that task with an automatic sliding sensor. But the moment you install the sensor, you lose the safety checks, the package handling, the neighborhood gossip, and the human judgment that actually made the building run smoothly.

In mathematics, writing down a formal proof is the equivalent of opening the door. The real work is knowing which room is worth walking into. Machines can verify consistency across millions of steps, but they do not know what humans find interesting, beautiful, or true to physical reality.

As Steven Brunton noted, math attracts people who crave certainty: “We chose math, you know, probably because we liked rules and formalism and, you know, uncertainty is uncomfortable. I think the humility aspect is really important, though.” The discipline is learning that verification is cheap, but taste remains human.

What to Do With This

Audit your engineering or research team this week for the doorman fallacy. Identify the three hardest technical problems your junior staff are assigned to solve right now. If your immediate reflex is to plug an LLM into their workflow to hand them the final architecture, stop. Let them hit the wall for five days first, because the judgment they build while failing is the only asset your company retains when code generation drops to zero marginal cost.