Pierre de Fermat

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI uses Fermat as the lens through which to understand AI’s characteristic failure mode: the production of confident claims that look like knowledge and are not yet knowledge. The machines we have built are conjecture engines. They see patterns no human would catch, propose connections that turn out to be real, and produce claims at extraordinary speed that sound authoritative—and arrive without the chain of reasons that would make them trustworthy. Fermat is the original author of that gap, the man who most vividly dramatized the difference between a claim and a proof, between a pattern seen across five cases and a theorem that holds for all of them.

The cycle reads Fermat against the present on three fronts. First, his margin note is the model for what AI researchers call hallucination: the fluent fabrication, the confident argument whose form is present and whose force is absent. Second, his false conjecture about Fermat numbers—that all numbers of the form two raised to a power of two plus one are prime, a claim checked in five cases and spectacularly wrong in the sixth—is the exact failure mode of generalization from training data: a pattern confirmed in a million instances that fails in the million-and-first. Third, his secretiveness about his methods is the historical rehearsal of the opacity problem in modern AI: the machine gives us the answer and withholds the reasons, exactly as Fermat gave his contemporaries the results and kept the methods to himself.

The cycle’s deepest use of Fermat is his demonstration that the cost of verification is not a tax on knowledge but its price. The claim took a moment to write. The knowledge took three hundred and fifty-eight years. The ratio between the cost of asserting and the cost of justifying is one of the most eloquent figures in the history of thought—and it is precisely the ratio that machines have inverted in our perception. By making the production of claims nearly free, they create the illusion that validation has become cheap, when in fact the validation is exactly as expensive as it ever was. The orange pill, in this reading, includes a willingness to pay the verification cost that Fermat left his successors to pay.

Origin

Pierre de Fermat was born in 1601 in Beaumont-de-Lomagne in southwest France. He trained as a lawyer, bought his way into a position as a councillor at the parlement of Toulouse—one of France’s regional high courts—and spent his working days on civil and criminal cases. A contemporary report on his judicial work complained that he was distracted and confused. His mind, plainly, was elsewhere. Mathematics was his obsession, pursued in stolen hours, and he was nonetheless one of the two or three greatest mathematicians of his century. He published almost nothing in his lifetime, preferring to communicate through letters carried across Europe via the monk Marin Mersenne, who functioned as a human switchboard for the scientific world of the day. He announced results, frequently without proof, sometimes as challenges flung at the mathematicians of France, England, and Holland: solve this, if you can.

Three contributions define his legacy and his relevance to AI. In number theory, his results—including the famous Little Theorem and the scattered conjectures of his margin notes—founded a field that would eventually produce the cryptography on which all digital security depends. In probability, his 1654 correspondence with Pascal over a gambling puzzle founded modern probability theory: the mathematics of reasoning under uncertainty that is the native language of every machine learning system alive today. In optics, his principle of least time—that light takes the path requiring the least travel time between two points—was the first rigorous demonstration that nature optimizes, the principle on which gradient descent and all of modern machine learning now runs.

Fermat died in 1665 in Castres. His son Samuel published his marginal notes in a 1670 edition of Diophantus, and so the unproven claims entered the world as facts without foundations. The note about the general theorem—that the equation x to the n plus y to the n equals z to the n has no solution in positive whole numbers for n greater than two—became known as Fermat’s Last Theorem, last because it was the final assertion to remain unproven. Andrew Wiles settled it in 1994, using elliptic curves and modular forms that Fermat could not possibly have possessed. Almost no one believes Fermat had a correct proof of the general case.

Key Ideas

The gap between conjecture and proof. A conjecture is a claim believed but not demonstrated. A proof is a chain of reasoning that compels assent—that converts belief into knowledge. Fermat lived in the space between them more flagrantly than any mathematician of comparable stature. The whole subsequent history of number theory can be read as the labor of supplying the proofs his intuition outran. The distinction is the fault line running through everything AI currently is and is not: the machines are superb at the conjectural gesture and helpless at the demonstrative one.

The confident error. Fermat conjectured that all Fermat numbers—numbers of the form two raised to a power of two, plus one—were prime. He checked five cases and became convinced. He was wrong. Euler proved it in 1732, finding that the very next number in the sequence factors easily. Fermat did exactly what a machine learning system does: he saw a pattern in the cases available to him and extrapolated it into a law. The cases were real; the pattern was real; the extrapolation was wrong, because the counterexample lay just past the edge of what he had checked. Five primes did not make a theorem. A million confirming examples do not make a guarantee. This is the precise failure mode of generalization from training data.

The co-founding of probability. In the summer of 1654, Fermat’s correspondence with Pascal over the Problem of Points founded modern probability theory. Fermat’s method was to enumerate all possible continuations of an interrupted game and count the fraction favoring each player. The insight that the fair division depends not on the history of the game but on its possible futures was radical. The mathematics of counting the futures is the direct ancestor of every probabilistic system that now powers AI: the language model predicting the next word, the risk system predicting the default, the recommender predicting the click.

The principle of least time. Fermat’s 1662 discovery that light takes the path requiring the least travel time between two points was the first rigorous demonstration of an optimization principle in nature. From this single assumption he derived the correct law of refraction. The modern understanding generalizes it to the principle of stationary action, which describes an astonishing range of physical phenomena as optimization over a quantity. This is the exact logic of machine learning: neural networks train by minimizing a loss function, gradient descent finds the configuration that makes a quantity as small as possible. AI training is, in a direct and unbroken line, Fermat’s light ray scaled to a billion dimensions.

The black box legacy. Fermat withheld his methods while announcing his results, driving his contemporaries to genuine frustration. He gave them the destination and hid the road. Mathematics, to its enormous credit, refused to accept his results on his authority—it insisted on the proofs, and took those proofs, when finally supplied, as the only thing that turned his confident claims into knowledge. A world that accepts the outputs of opaque AI systems as the basis for consequential decisions is doing what mathematics refused to do for Fermat: taking the answer without the reason, trusting the result without the demonstration. Whether we hold the machine to the standard that mathematics held Fermat—reasons, not just results—is one of the defining choices of the age.