I was sitting in the back of a lecture hall at MIT last week, half expecting another carefully hedged presentation about the future of artificial intelligence. But Neil Thompson, a researcher who studies progress in computing and AI, did something more unsettling. In the span of about thirty minutes, he convinced me that several scientific fields are about to be turned upside down, and not in some distant, hypothetical way. It is happening now, and it is happening faster than most of us are prepared to accept. Thompson’s talk was built around a striking new study, conducted with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, that looked at whether the most advanced AI models can solve real theoretical problems in fields that most people would assume are too abstract, too creative, or too hard for machines. The study focused on probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. These are not the kinds of problems with straightforward formulas or easy answers; they are deep, messy, structural puzzles that define the frontiers of human knowledge. And yet, between mid-2024 and mid-2026, the best AI models went from correctly solving about 2 percent of these problems to correctly solving about 35 percent of them. The jump is staggering)Skip. No signs of slowing either. As Thompson put it, within a few years, he expects a lot of theory questions to be totally transformed in terms of AI’s ability to do them. I walked out of that lecture with my head spinning, trying to imagine what a world like that looks like for the people whose life’s work is to ask those questions.
It would have been easy to discount the study as hype, but the methodology was careful and genuinely rigorous. The researchers did not just hand the models a stack of textbook exercises or ask them to recall famous results. Instead, they took problems stated in research papers posted online between December 2025 and May 2026, and then tested a series of OpenAI models—o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—on those problems. The crucial part is that the solutions to these problems appeared in papers that were published after the training cutoff for the models being tested. In other words, the AI could not have memorized the answers from its training data. The papers served as ground truth, a way of checking whether the models had actually solved the problems on their own, and the researchers also blocked internet access so the models could not go looking for solutions online. This is what separates this study from so many other breathless AI claims. It was not a test of retrieval or pattern matching; it was a test of genuine problem-solving, at least in the sense that the models had to synthesize ideas, reason through constraints, and arrive at answers they had never seen before. There were limits to the study, of course. The team focused on fields and problems in which AI models could reason about theories without needing to run simulations or physical experiments. That means the study actually understates the broader disruption, because it only looked at the areas where it is even possible to give a text-based problem to an AI and expect a text-based answer. But the very existence of such a test, and the rapid improvement it revealed, suggests that the shock waves currently being felt in mathematics are only the beginning.
To understand why this is such a big deal, it helps to think about what these problems actually look like. A theoretical problem in quantum physics or cosmology is not a homework question. It is often a knot that a researcher has been pulling at for years, an open problem that requires not just technical skill but also deep intuition, creativity, and the ability to recognize hidden connections between seemingly unrelated ideas. For a long time, these kinds of problems seemed safe from automation. They are so far removed from the pattern-recognition tasks that AI excels at, so dependent on human judgment and experience, that many people assumed the machines would eventually write code, handle legal documents, and maybe compose music, but not solve the deepest theoretical puzzles in physics and economics. The study shatters that assumption. It shows a trajectory from near total failure in 2024 to a situation in which the best models can solve a third of these problems by mid-2026. If that curve continues, even at a slightly slower pace, it will not be long before AI is solving the majority of theory problems that researchers are currently publishing. And this is not just about speed. It is about the kind of science that becomes possible. Imagine a young physicist with a brilliant idea about gravity, but not enough technical machinery to prove it. In the old world, that idea might remain a vague intuition. In the new world, an AI could work out the proof, or at least outline the steps, and the physicist could focus on refining the idea and interpreting what it means. That is a thrilling possibility. But it is also terrifying, because it raises urgent questions about what happens to the people, institutions, and practices that have long been built around the slow, human process of solving problems.
The controversy that exploded last month around OpenAI’s announcement about the Navier-Stokes equations shows how fraught these questions have already become. The Navier-Stokes equations describe the behavior of fluid motion, and they are not just a technical curiosity; they are one of the most famous and challenging open problems in mathematics, with connections to weather prediction, airplane design, and even understanding turbulence. According to the announcement, an unreleased AI model had solved a long-standing grand challenge associated with these equations. The reaction was electric. Mathematicians were stunned, excited, and then deeply scandalized when an NYU mathematician, who had been collaborating with another academic affiliated with Anthropic, accused OpenAI of gazumping them after learning about their progress. The word gazumping sounds almost quaint, like a real estate dispute over a house, but in this context it was a serious accusation of scientific misconduct, a claim that OpenAI had used private knowledge about someone else’s work to jump ahead and claim credit. The episode was messy, emotional, and revealing. It showed that AI is not just a tool sitting quietly in the background; it is becoming an active participant in the scientific community, with its own competitive dynamics, rivalries, and ethical dilemmas. In the weeks since, OpenAI and other companies have announced hundreds more solutions to longstanding mathematical problems, and many mathematicians have described the feeling as a crisis. Some have started to ask whether there should be limits on the use of AI, warning that the field is being transformed so quickly that no one has time to think about what is being lost.
What is being lost, perhaps, is something more subtle than the right answer. Mathematicians have long argued that the process of solving a problem is often as valuable as the answer itself. The struggle matters. The false starts matter. The moment of insight, when a proof finally clicks into place after months of frustration, is not just a means to an end; it is a form of understanding that shapes the way a thinker sees the entire field. If an AI hands you a proof on a silver platter, without showing its work, without making mistakes, without any sense of how it thought through the problem, have you actually learned anything? This is the question that hangs over the current debate. Some researchers are calling for limits, not because they hate technology, but because they worry that the next generation of mathematicians will grow up never having experienced the slow, patient, sometimes painful journey that builds mathematical intuition. They worry that the final answer will become a commodity, while the deeper wisdom that comes from discovering an answer on your own becomes rare. At the same time, it is hard to argue that we should stop progress just to preserve the traditional training of mathematicians. We do not refuse to use calculators because they make arithmetic too easy. We do not forbid computer models because they make weather prediction less mystical. The challenge is to find a way to use AI without letting it hollow out the scientific process from the inside. That means creating new norms, new methods of verification, and new ways of teaching that treat AI as a collaborator rather than a replacement. It also means being honest about what we lose when we no longer have to struggle for every answer.
Looking ahead, it is easy to feel either desperate or euphoric, and the truth probably lies somewhere in between. Neil Thompson’s lab studies how advances in computer science affect society and innovation, and the lesson from his work is that these changes are never purely good or purely bad. On one hand, the rapid rise in AI’s ability to solve theoretical problems could democratize science in unprecedented ways. A student in a developing country, without access to elite mentors or expensive equipment, could use an AI system to explore deep questions in quantum physics or economic theory that would previously have been closed to them. That could unleash a wave of creativity and discovery that looks nothing like the research world we know today. On the other hand, the concentration of these AI systems in a handful of tech companies raises uncomfortable questions about power, control, and access. Who decides what problems the AI prioritizes? Who gets to profit from the solutions? Who is responsible when an AI makes a mistake, or when it solves a problem in a way that looks right but is subtly wrong because it was trained on flawed human assumptions? These are not just technical questions; they are social, political, and deeply human questions. Thompson’s work suggests that we have, at most, a few years before these questions become impossible to ignore. The fields he studied are already on the brink. Mathematics is in crisis. Physics and economics will likely be next peaks. The researchers who adapt, who learn to work with AI rather than against it, may find themselves liberated from the grind of calculation and able to do more creative thinking than ever. But the researchers who refuse to engage, or who cling to the idea that human thought is the only legitimate path to knowledge, may find themselves left behind. Leaving that lecture hall, I felt a strange mix of anxiety and hope. The future is coming whether we are ready or notشي. The only real question is whether we will meet it with courage, wisdom, and a willingness to rethink what it means to know something. Perhaps, in the end, the most important problem AI will help us solve is not a problem in physics or cosmology, but the problem of explaining to ourselves why we think, why we explore, and why the struggle to understand is so essential to being human.
