Theoretical physics has long rewarded a specific kind of individual genius — the lone mathematician who sees what no one else can. That model is being disrupted. A new analysis from The Economist details how AI systems are working through some of the hardest open problems in the field, from string theory's vast landscape of possible vacua to the multi-loop Feynman diagrams that would consume years of human effort to compute by hand. Physicists aren't calling it automation. They're calling it a new kind of collaborator.
Google DeepMind is driving most of what's actually happening. AlphaProof has demonstrated formal mathematical reasoning. FunSearch applies program synthesis to mathematical discovery. AlphaTensor didn't just replicate known algorithms for matrix multiplication — it found better ones. Researchers working with these systems describe something more surprising than pattern-matching on prior results: genuine contributions to problems that have been open for years.
Large language models from Anthropic, OpenAI, and Google are doing different but related work in the hands of working physicists. They draft proof sketches, check dimensional analysis, translate between mathematical formalisms, and propose ansätze — informed guesses about the shape a solution might take — for equations that have resisted closed-form solutions for decades.
The range of physics being touched is striking. In gravitational wave science, AI is speeding up amplitude calculations relevant to LIGO and the planned LISA observatory. In condensed matter, neural network methods are tackling the many-body Schrödinger equation. Most dramatically, DeepMind's collaboration with EPFL on plasma control in tokamak fusion reactors has moved beyond pure engineering into territory with real implications for fundamental physics. MIT and CERN are both integrating these tools into workflows that were previously entirely human-led.
The Economist is careful about what this means. AI may be compressing into years what would otherwise have taken a generation of physicists decades to achieve. That creates genuine discomfort alongside the excitement: when the reasoning behind a discovery is buried inside billions of model parameters, does the scientific community actually understand it? And if AI drives down the cost of certain classes of theoretical work, what happens to physics funding and academic careers over the next decade? The physicists who are actually using these systems, for now, mostly describe exhilaration. The frontier, they say, is moving faster than it ever has.