Sunday, November 22, 2026
English edition

Development

What happens to mathematics when AI becomes an oracle?

November 11, 2026 Development Source: Ars Technica

What happens to mathematics when AI becomes an oracle?

Share this article

“The mathematician in each of us will always want to know more,” Strogatz and Townsend wrote. “The human being must insist that knowledge serves something larger than itself. That might be the work of our century: letting our power grow without losing our soul. Big Math keeps getting bigger. So must we.” Ars caught up with Strogatz to learn more. Ars Technica: Let’s start by defining “big math,” because it’s such a broad umbrella term. Steven Strogatz: Operationally, big math is “big” relative to human cognitive capacity. So it can be surprisingly small and yet big. We also wanted to expand it to things that people would think of as computational engineering—in the building of the Burj Khalifa, that’d be civil engineering or design architecture. It’s there in medical imaging. It’s in getting us to the Moon. The lunar module was controlled by a tiny computer that only had to think about 20 or 30 variables at a time to steer the LEM down to the Moon’s surface. But it’s not something a person could do in their head. Big Math is not simply big in the ordinary sense of big numbers. There’s an undercurrent in the book about the importance of linearity and sparsity. When problems are linear, they can be scaled up to immense size, and they stay tractable. If they’re non-linear, you don’t have any hope of dealing with them with pure math. You can do them with computers. But for the people who want rigorous theorems, we’re in terra incognita when we deal with anything that’s non-linear and big. Small math can be very hard. Historically, we had small math problems with one or two variables, arithmetic, algebra that you learn in high school. Even to this day, a lot of pure math is relatively small. It doesn’t concern itself with enormous numbers of variables simultaneously because it’s totally intractable. Cryptography can be small, but it’s so complicated with its non-linearities and other algebraic difficulties. So big is one way to be difficult, but there are a lot of other ways. Nowadays, trillions of parameters are typical in the biggest of the large language models. But infinity is not big from this point. Infinity, which you would think of would be the biggest of big math, is not because calculus fits perfectly nicely in a human head. Infinity turns out to be very tractable in a way that a billion is not. There was an Alabama voting rights case where Roberts actually voted with the plaintiffs, who said that there was real harm being done to Black voters in Alabama with one of Alabama’s maps. The state was claiming there’s no way we could get a reasonable map to have two districts with Black majority voters. Moon Duchin showed that was wrong. You could generate thousands of maps that would do that. So it just wasn’t true what the state had claimed. After the book was copyedited and ready to go to press, a Supreme Court decision was made that, in fact, you can partisan gerrymander as much as you want. It doesn’t violate the Voting Rights Act. We wrote a more optimistic story. Ars: How has math changed over time from a philosophical standpoint? Strogatz: I think a crisp way to encapsulate it is the role of the computer. For a long time, a computer was a servant to mathematicians and engineers and scientists. It was no more philosophically threatening than a microscope or a telescope. It could extend our senses. It’s a chisel, it’s an anvil, it’s a tool. It helped us do calculations faster; things that would be too hard for people to do at scale you could do on a machine. We had mechanical calculators. They became electronic computers in the 1950s with the invention of the transistor and got much faster to give us today’s supercomputers. We program them; they do exactly what we tell them. So they’re servants. The first philosophically big change is in the 1950s, when computers start to be used as more than servants; they become scouts. You could send the computer out into unknown, unexplored territory in the mathematical landscape. The clearest example was work by Enrico Fermi and his colleagues (Pasta, Ulam, and Tsingou) that we wrote about in the chapter on simulation. Mary Tsingou was the one who did the calculations in what was traditionally called the Fermi-Pasta-Ulam numerical experiment. What distinguished it from earlier uses of computers is that Fermi and colleagues had no idea what the answer was going to be. It’s one thing to use a computer to calculate the power released by an atomic bomb blast or to calculate an artillery shell’s trajectory. We didn’t know the exact numerical values, but we knew what was happening. With this experiment, they were trying to simulate a simplified model of how atoms could, in the jargon, formalize as chemical bonds jiggle. You want to see the decay to entropy. How do systems degenerate to greater and greater disorder if they’re just left to their own devices? Fermi thought that the non-linear aspect of chemical bonds—meaning if you’d stretch a chemical bond by twice as much, it doesn’t pull back exactly twice as hard, so it’s not strictly proportional—might account for the so-called second law of thermodynamics, the decay to entropy. Pencil and paper math was not up to the job because you couldn’t calculate these non-linear interactions far enough to see the decay into entropy. So they ran it in the computer, and a shocking thing happened: It didn’t decay. It showed a strange echo. You started all the energy in one of the modes of vibration of this system, and then it started to degenerate. But after a while, 98 percent of the energy came back into the original mode. That was the echo that they called a recurrence phenomenon. It led us to a whole new branch of math that we call soliton theory. It was an early indication that once you walk into the nonlinear jungle, human intuition breaks down. We couldn’t walk there, but we sent our scout, this computer probe, into the wilderness of the mind, and it was the only thing capable of doing the calculations and reporting back to us what it found. So is it that we’ve artificially restricted ourself to systems that are transparent enough that we can develop science that our insight illuminates? I think Wolfram would claim that most of nature is computationally irreducible in the sense that there is no simple insightful theory into what’s going on. The best you can do is just run a simulation and watch what happens, similar to just letting nature unfold and watch what happens in reality. In the jargon people use nowadays, you shouldn’t expect a compressed description, a simplified model that will account for what you’re seeing. We have seen those. Newton could explain the motion of the planets from a few simple principles and that worked, but that’s, I think in Wolfram’s take, artificially simple. Most of science is not going to end up being like that. We don’t know if Wolfram is really right about that. Wolfram, I think, is demonstrably right that many computer simulations are computationally irreducible, but nature might not be. Ars: If you can’t see into the models, and they can’t explain what they’re doing, how do you know they’re right? Strogatz: Let’s stipulate that we don’t understand them, and that they are not at the present able to explain themselves to us. Let’s put a little asterisk next to that part, though, because it could be that they’ll become very good teachers. Or maybe we don’t understand Google Translate or AlphaFold right now, but we will in 10 years. I do believe that there will be systems that are truly opaque because of human limits, that we just won’t be fast enough or smart enough or have enough memory. AIs will show us something that we just won’t be able to understand even with their help. What happens then? Then I think math starts to be more like science or medicine, or engineering. At the FDA, for a new drug to be approved, it has to be shown to be safe by doing a lot of controlled studies and randomized trials. There’s a whole methodology for opacity that people in medicine have worked out. But the doctors do not require knowledge of the mechanism of action. You don’t have to know how aspirin works. You don’t have to know how electroconvulsive therapy works for people with intractable depression. Even if we don’t understand, we can still live with these technologies or drugs by these methods. You don’t have to have first principles understanding. If the AI is an oracle, you can still check it by seeing how often it is right. Ars: That is the opposite of what AI does. I guess I worry that it will do the same thing to math. There is a beauty to math, an aesthetic element. If all you’re focused on is getting the answers and ignore the journey, those are qualitative losses. Mathematicians have individual voices, too. Strogatz: Totally. There are people whose style is very visible in the questions they ask, in the approach they take to them, in the way they explain themselves. I have cried on camera for several interviews over the past few months because there are lots of things maybe being lost. But I want to put an underline on the fact that we don’t really know. For example, what if AI gets really good at teaching us? I find when I interact with it, that it can be a very effective explainer. And why do we have to feel sad? Maybe we can understand math better than ever with the help of this machine. We may not get to make the greatest discoveries because the machine will probably be making them, but we can still make little discoveries. I never made a great discovery. I’m not at the level of a Fields medalist, but I can still enjoy math and so can people way below me. There are recreational mathematicians who take pleasure in discovery, whatever they figure out on their own. Just like when a person solves a puzzle, it’s not like a jigsaw puzzle or Sudoku is important. You just have the pleasure of thinking. There are people who love the workout part of it. There are people who like the competition. We shouldn’t pretend. There are people who are thrilled by the idea of being the first to make a big discovery. That’s animating for some folks, and it will be hard for them. Strogatz: So far the AIs are not good at asking questions. I think they could be trained to ask good questions. A graduate student is often not good at asking questions when they start because in college you’re not taught to ask questions, you’re taught to give answers. The hardest thing in training a new PhD or postdoc is helping them figure out how to ask questions. I think we could do that with AIs, too. There is a feeling of loss, but I don’t think we should shortchange the feeling of excitement. There’s so many practical, important benefits that could happen in energy technology and climate and medicine, theoretical physics, pure math. I mean, we’re making astonishing discoveries through this machine. So for a lot of people, it’s a very joyful time. It’s very confusing. That’s the only honest assessment. We end the book with Faust. Usually people hear “Faustian bargain” as, “You have sold your soul to the devil, now you’re going to pay.” Goethe’s version is more ambiguous: that if we keep striving to build something bigger than ourselves, maybe we can be redeemed. But it’s still a very dangerous game that we have chosen to play.