Recently, a Google DeepMind project called AlphaProof Nexus helped resolve mathematical problems that had gone unanswered since the 1950s and 1970s. These weren't the famous million-dollar problems that make headlines. They were quieter ones, the kind that working mathematicians had picked up and set aside for decades because nobody could crack them, until AI entered the picture.
Dr. Swarat Chaudhuri, a computer science professor at The University of Texas at Austin, led that project as a research scientist at DeepMind. What's notable about the breakthrough isn't that a computer processed calculations faster than a person could. It's that the AI reasoned its way through the problem much as a mathematician would: connecting ideas, testing approaches, and working through technical details until an insight emerged.
If AI can do that for a 70-year-old math problem, the question becomes: what else is within reach? Faster drug discovery. New materials. More accurate models of climate systems or financial markets. Nobody knows exactly how far this extends yet. But the trajectory is clear enough that researchers like Chaudhuri are staking their careers on it.
Why Math, of All Things
Mathematics turns out to be one of the best domains for observing this shift, for two distinct reasons. The first is straightforward: math is full of open problems, and AI is beginning to help solve some of them. That's the AlphaProof Nexus story.
The second reason is less obvious but arguably more consequential. Training an AI to perform rigorous mathematical reasoning also sharpens its reasoning more broadly, and makes it more reliable about distinguishing correct answers from incorrect ones. Chaudhuri puts it simply: "If you have an AI system that can do advanced mathematical reasoning, it's also a testament to its general reasoning ability."
So, there are really two efforts underway simultaneously. One uses AI to push into unsolved mathematics. The other uses mathematical reasoning to make AI systems themselves more trustworthy. Chaudhuri's lab works on both.
A Foot in Two Worlds
Chaudhuri directs the Trishul lab at UT Austin, where his team conducts research at the intersection of machine learning, formal methods, and software reliability. His doctoral work focused on proving that software behaves correctly, not merely testing it and hoping for the best, but formally verifying it with mathematical rigor. That background is highly relevant now, as AI systems become capable enough that their reasoning can be checked and verified rather than simply sounding plausible.
In 2025, Chaudhuri was named a Guggenheim Fellow to work on designing an AI agent that could propose new math problems and their solutions, and another one that evaluated the “interestingness” outputs of the first agent. This project emulated the curiosity-driven exploration that human mathematicians are known for and served as a catalyst for his work at DeepMind, a role independent of his position at UT. Chaudhuri sees the two as mutually reinforcing. Ideas from the frontier of industry AI research filter back to his students at UT. Problems his lab grapples with on campus, in turn, shape the questions he brings to Google. "The exchange between academia and industry accelerates progress in both directions," he says, and it isn't just a talking point. It's how his research actually operates.
Still a Human in the Loop
None of this means researchers are becoming spectators. Chaudhuri is emphatic on this point.
Someone still has to determine which questions are worth pursuing. Someone still has to supply the initial insight that launches an entirely new line of inquiry. AI hasn't assumed that role, at least not yet. "Our goal is not to replace the mathematician," he says. "But to augment their abilities. Human plus AI should be better than just AI."
That framing matters, especially as anxiety spreads among early-career researchers watching AI close in on work they’ve spent years pursuing. Chaudhuri’s view is that students must still learn to do mathematics and write code themselves, but that cutting-edge research increasingly demands fluency with these tools. The goal isn’t to become a prompt engineer. It’s to become someone who can push beyond what AI can do alone.
His own students illustrate this dynamic. A group in his lab recently used AI to make substantive progress on a decades-old open problem, not by delegating the solution and stepping back, but by collaborating with it: evaluating its ideas, redirecting its approach, and pushing it further. That's the operating model. The AI handles the technical grind and occasionally surfaces a genuinely novel idea. The human determines what matters and where to direct that effort next.
What Happens When AI Learns to Ask the Questions
Here's the part Chaudhuri finds most compelling, though it remains somewhat further out.
At present, AI largely answers questions that humans pose. Solve this proof. Verify this program. Someone still has to decide the problem is worth solving in the first place. "Theory discovery, posing new problems to solve, these are all tasks for humans, as of now," Chaudhuri says.
His lab is already developing early systems that explore new mathematical territory independently, cultivating something resembling a sense of what's interesting. It isn't difficult to imagine where this leads: an AI that identifies an anomalous pattern nobody had been looking for and determines, on its own, that it merits further investigation.
Extend that possibility beyond mathematics. An AI flagging an unusual correlation in medical data that turns out to be an early indicator of disease. Or identifying a software vulnerability that human reviewers had overlooked repeatedly. Or arriving at a mathematical insight that unlocks a problem in an entirely different field, much as abstract mathematics has quietly enabled breakthroughs in physics and cryptography for over a century.
Chaudhuri won't predict precisely when that future arrives, but he believes researchers should prepare for it now. "How do we all 10x or 100x ourselves with these tools?" he asks.
That's the wager UT Austin is making as well: equip the next generation to work alongside these systems as partners, so that when AI begins posing its own questions, there are people positioned to make the most of the answers.
The Trishul lab at UT Austin, directed by Swarat Chaudhuri, studies problems at the interface of automated reasoning, programming languages, and machine learning. Learn more about the lab's current work.