In a recent talk at MIT, Neil Thompson, a scholar of computing progress, presented evidence that theoretical science is on the verge of a radical transformation. As mathematicians grapple with automation, Thompson’s latest study suggests that economics, quantum physics, and cosmology are the next fields to be disrupted by the capabilities of frontier AI models.
The Performance Leap
The study, conducted by Thompson alongside Gabriel Manso and Claudio Mayrink Verdun, tracked a staggering evolution. Between mid-2024 and mid-2026, leading AI models improved their success rate in solving theoretical problems from a mere 2% to approximately 35%. The researchers tested models including GPT-5.5 and o3 using problems from research papers published between December 2025 and May 2026. This ensured the solutions were not part of the training data, forcing the models to work them out independently without internet access.
Crisis in Math and the Need for Verification
This progress arrives amid significant tension. Last month, OpenAI rocked the scientific community by claiming a solution to the Navier-Stokes equations, a long-standing challenge in fluid dynamics. The announcement sparked controversy, including allegations of "gazumping" from NYU academics and warnings from mathematicians that the process of discovery is as valuable as the answer itself.
Despite AI’s growing prowess, the MIT researchers emphasize that human expertise remains indispensable. OpenAI recently had to withdraw several solutions due to inaccuracies, proving that experts are still required to verify outputs and provide the conceptual "connective tissue" for complex problems.
