What happens to mathematics when AI becomes an oracle?

Applied mathematicians Steven Strogatz and Alex Townsend have released a book, Big Math: The Hidden Codes That Run Our World, examining how computation has reshaped mathematics and what happens as problems grow past what any person can hold in mind. Strogatz, long known for his 1998 paper with Duncan Watts on small-world networks and for popular books including Sync and Infinite Powers, discussed the project in an interview. The collaboration grew out of Townsend's course on the top 20 algorithms of the 20th century; the concept shifted toward big matrix calculations and finally into the current book.
Strogatz's working definition is relative rather than absolute: big math is big compared with human cognitive capacity, so it can be surprisingly modest in raw number and still qualify. The book's examples stretch across civil and design engineering on the Burj Khalifa, medical imaging, and the Apollo lunar module, whose small onboard computer tracked only 20 or 30 variables at a time to steer the lander down, yet handled something no person could do mentally.
Two structural properties carry much of the argument. Linear problems scale to immense size and stay tractable; nonlinear ones resist pure-mathematical treatment, even though computers can grind through them. Sparsity matters for the same reason. The result, in Strogatz's account, is that once problems are both big and nonlinear, anyone seeking rigorous theorems is in uncharted territory. He also draws a counterintuitive distinction between size and difficulty: modern large language models routinely carry trillions of parameters, while infinity itself is tractable because calculus fits comfortably in a human head. Cryptography can involve tiny numbers and still be brutally hard because of its nonlinear structure.
One extended example is redistricting. Legal standards call for districts that are contiguous, reasonably compact, respectful of natural boundaries, and roughly equal in population, but compactness has no crisp mathematical definition. Moon Duchin and colleagues generated large ensembles of maps that satisfy the legal criteria but are otherwise random, producing a baseline against which an actual map can be compared. If a map's outcome falls far outside that ensemble, it is atypical, though Strogatz is careful to say this does not by itself prove malicious intent. He recalled an Alabama voting rights case in which the state argued no reasonable map could yield two Black-majority districts and Duchin's work showed otherwise, generating thousands of maps that could. He also notes that after the manuscript was copyedited, a Supreme Court decision held that partisan gerrymandering does not violate the Voting Rights Act, undercutting the more optimistic framing the authors had written.
Strogatz frames the philosophical shift in three stages. The computer began as a servant, no more threatening than a microscope, extending human calculation at scale. In the 1950s it became a scout: the Fermi-Pasta-Ulam-Tsingou numerical experiment, with Mary Tsingou doing the calculations, probed a simplified model of vibrating chemical bonds expecting to watch energy decay toward entropy, and instead saw most of the energy return to its original mode, a recurrence that helped open soliton theory. The lesson was that in nonlinear territory, intuition fails and the machine reports back from places people cannot walk.
The current stage, he argues, is the oracle: systems that produce results we do not understand and cannot explain themselves. Google Translate works impressively without having yielded new principles of linguistics, and AlphaFold, honored with a Nobel Prize in chemistry, solved far more protein structures without discovering new physics or chemistry. Strogatz expects this opacity to deepen. The remedy he sketches borrows from medicine, where drugs reach approval through controlled trials without full mechanistic understanding: even an opaque oracle can be checked empirically by how often it is right.
The cost is epistemic. Pure mathematics has long been prized as the one discipline offering absolute proof, with Euclid's results still standing after more than two thousand years, unlike anything in the science or medicine of that era. Treating math more like science or engineering means surrendering that certainty. Strogatz also points to the culture of productive struggle and the aha moment of sudden pattern recognition, and suggests younger mathematicians feel the loss most acutely. He raises Stephen Wolfram's idea of computational irreducibility, the claim that many systems admit no compressed description and can only be run, while noting Wolfram is demonstrably right about many simulations but may or may not be right about nature.
The interview also touches on recent anxiety in the field, including OpenAI's release last month of 722 AI-generated math papers, which the discussion refers to as the mathocalypse. The book's closing sentiment, as the authors put it, is that the mathematician in each of us wants to know more, while the human being must insist knowledge serves something larger.
Why it matters: Teams building on machine learning already work with systems whose outputs outrun their explanations, and this framing suggests evaluation practice borrowed from medicine, repeated validation rather than full mechanistic understanding, may become the norm (an inference from the discussion, not a stated recommendation). For product and research groups, the practical question is less whether a model's underlying math is comprehensible than whether its reliability can be measured well enough to trust. Mathematics itself may increasingly resemble an empirical discipline, a shift that affects how results are verified, taught and defended.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Mathematician Steven Strogatz discusses his new book, Big Math: The Hidden Codes That Run Our World .