Key Takeaways

  • Frontier AI models from labs like OpenAI and Anthropic excel at solving well-posed problems and checking proofs, but they lack the judgment to decide which questions matter.
  • Chess reached an analytical plateau once computers mapped out optimal move sequences, but mathematics cannot be solved because its search space expands every time a new proof is written.
  • Sabermetrics worked in baseball because play reduces to isolated, one-on-one set pieces between a pitcher and a batter. Football resists the same modeling because 22 players collide simultaneously in statistical chaos.
  • The value of human technical talent is shifting away from mechanical proof-checking toward understanding how individual answers alter the wider problem space.

Math Will Never Run Out of Moves

When computers beat world champions in chess, the game flattened. Software calculated tactical combinations faster than grandmasters, converting centuries of human intuition into look-up tables and minimax trees. The board stayed sixty-four squares, thirty-two pieces, and zero room for new rules.

John Urschel sees a clean boundary between board games and mathematical discovery. Urschel, an assistant professor of mathematics at MIT and a former NFL offensive lineman, points out that computational power cannot box in higher mathematics the way it boxed in chess. As Urschel puts it: “Definitely we don't have any consensus I'd say about how we should be using AI, what things are going to look like in a couple years, but I think we have pretty good consensus that you know math is not going to be solved.”

In chess, the universe of possible positions shrinks as pieces exit the board. In mathematics, every verified proof opens fifty unexplored corridors. AI can scan existing structures and close known loops, but it cannot saturate a domain that grows larger every time someone solves a theorem.

Football Breaks the Moneyball Model

Silicon Valley founders constantly try to import the Moneyball playbook into messy operational businesses. They assume that if statistical analysis worked for the Oakland Athletics, it can predict human systems across enterprise software, hiring, or athletics.

Urschel warns that baseball is an anomaly. “Moneyball worked in baseball because it's so such an easy problem if you think about it,” Urschel explains. “People are in set pieces. The rules are so straightforward like you know it you know someone's batting, someone's pitching, then something happens and it's really well understood. Whereas football, you've got 22 guys on a field and it's just complete chaos from like a from a data statistical point of view.”

In baseball, a pitch exists in a sterile state. You track velocity, spin rate, and exit angle. Football destroys those tidy assumptions. An offensive tackle's pass block depends on the guard's split, the defensive tackle's stunt, the safety's rotation, and turf conditions on a third-down snap. When twenty-two people move at once, hidden variables multiply faster than any model can resolve.

Verification Is Cheap, Direction Is Not

The current generation of large language models is capable at narrow execution. If you feed an established model a well-defined equation, it will walk the steps and flag errors in your logic.

Yet Urschel draws a sharp line between calculating an answer and comprehending its consequence. “These tools are very very good at solving a wellpost problem that we've been thinking about for a while,” Urschel says. “The thing that we're still trying to figure out is how generative can this be?”

Urschel stresses that understanding remains human work: “The idea is trying to look at a proof, look at a solution and understand something about what that means.” When open-access course materials make technical knowledge universally available, raw mechanics become a commodity. The scarce skill is no longer solving an assigned prompt. It is knowing which problem deserves the next five years of your life.

What to Do With This

Audit your current engineering roadmap and label each feature as either a set piece or open chaos. If a team is attempting to build generative AI features on chaotic, 22-player operational data without clean set-piece boundaries, kill the project this week and refocus them on automated verification of known outputs.