AI’s Science Push: From Math to the Rest
▼ Summary
– Neil Thompson presented research at MIT indicating that AI models are rapidly upending theoretical scientific fields beyond mathematics.
– The study found that top AI models increased their ability to solve theoretical problems in physics and economics from 2% to 35% between mid-2024 and mid-2026.
– Researchers tested OpenAI models on recent research papers, ensuring the AI solved problems independently without access to training data or the internet.
– OpenAI recently claimed to have solved the Navier-Stokes equation, sparking controversy and accusations of academic displacement among mathematicians.
– Some mathematicians are calling for limits on AI usage, arguing that the problem-solving process is as valuable as the final answer.
The Rapid Expansion of AI’s Scientific Reach
Artificial intelligence is poised to disrupt theoretical disciplines beyond mathematics, according to recent research presented by Neil Thompson, a scholar at MIT who specializes in computing and AI progress. In a concise presentation, Thompson argued that the current upheaval in math is merely the precursor to significant shifts in other rigorous scientific fields. His analysis indicates that areas such as probability theory, statistics, economic theory, high-energy physics, gravitation, cosmology, and quantum mechanics are on the brink of transformation due to the accelerating capabilities of frontier models.
Data from Thompson’s study reveals a dramatic leap in performance between mid-2024 and mid-2026. During this two-year window, top-tier AI systems improved their success rate on theoretical problems within these domains from approximately 2 percent to 35 percent. This trajectory suggests that the technology is not plateauing but rather gaining momentum. As Thompson stated:
“Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI’s ability to do them.”
Methodology and Experimental Constraints
Thompson conducted this investigation alongside Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist. Their work aimed to understand how computer science advancements influence societal innovation and broader scientific inquiry. To test the limits of current AI, the team selected problems explicitly stated in research papers published online between December 2025 and May 2026. They evaluated several OpenAI models, including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5.
A critical component of the experimental design was ensuring that the models were genuinely solving problems rather than retrieving memorized answers. The researchers chose papers that had been published after the training cutoff dates for GPT-5.5 and earlier iterations. Consequently, the models could not rely on pre-existing training data for these specific solutions. These papers served as ground truth to verify the models’ independent problem-solving skills. Furthermore, the team disabled internet access during testing to prevent the AI from searching for external solutions, thereby isolating the internal reasoning capabilities of the algorithms.
While the study focused only on fields where AI could derive answers without running physical simulations or experiments, the findings imply that the disruption seen in mathematics will likely extend to other theory-heavy scientific sectors.
The Crisis in Mathematical Research
The context for this research follows a period of intense controversy in the mathematical community. Last month, OpenAI announced that an unreleased model had solved the Navier-Stokes equations, a long-standing grand challenge concerning fluid motion dynamics. This announcement sparked significant debate, particularly when a mathematician at New York University, who had been collaborating with an academic associated with Anthropic, accused OpenAI of overtaking their progress prematurely.
Since then, OpenAI and other entities have released hundreds of additional solutions to historic mathematical problems. This surge has triggered a crisis within the field, leading some mathematicians to call for restrictions on AI usage. Critics argue that the educational and intellectual value lies in the process of solving problems, not just in obtaining the final answer. The rapid automation of these tasks threatens to undermine the fundamental practices of discovery and verification that define scientific rigor.
(Source: Wired)