After mathematics, AI can conquer new heights

While the mathematicians are still stunned by how AI is automating their field, the latest research by Dr. Neil Thompson at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and director of the MIT FutureTech project shows that statistical probability, economic theory, high energy, gravity and cosmology, quantum physics could become the next field to be influenced by the rapid evolution of advanced models.

According to research results that Thompson shared at a presentation in early October, the best AI models increased their ability to accurately solve about 2% of theoretical problems in the above fields to 35%. He also predicted that AI will completely answer many theoretical questions within the next few years.

Theo WiredThompson and two colleagues, doctoral student Gabriel Manso and scientist Claudio Mayrink Verdun, synthesized the problem raised in scientific articles posted online from December 2025 to May 2026, then tested whether OpenAI models such as o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5 could solve it.

In particular, the solution to the problem is outlined in articles published after GPT-5.5 and previous models stopped updating training data. This means the model must come up with a solution on its own instead of relying on answers provided as training data. The research team also blocked Internet access to prevent the model from searching for solutions online. Welcome.AI commented that their method helps evaluate AI capabilities in a more accurate and rigorous way, demonstrating that the model can actually be creative rather than just ‘rote learning’ from what it already knows.

The three scientists focus only on areas where the AI ​​model can find solutions on its own without running simulations or experiments. The results reveal that beyond mathematics, AI’s influence can spread to many other fields of heavily theoretical scientific research.

Theo Naturelast month, OpenAI created a stir in the world of mathematics when it announced an unreleased model that solved the Navier-Stokes problem, a long-standing challenge in fluid dynamics, thereby sparking controversy about the limited application of AI and the value of the problem-solving process itself. This tension reflects broader concerns about the role of AI in scientific discovery.

Thompson warns the coming changes will come with challenges. Although the theoretical achievements of AI can accelerate practical applications in science, the solutions it offers still have the risk of errors, especially for complex problems. The research team emphasized the need for human supervision to confirm results from AI. Human expertise remains essential to interpreting AI findings and ensuring their accuracy.

Sara Hooker, an artificial intelligence researcher and co-founder of San Francisco startup Adaption, said AI will fundamentally change the way scientists work, but not eliminate them. She argues that actors are not good at asking the essential questions that drive discovery, so the role of the scientist remains crucial. They are just tools to help researchers find answers faster after they have formulated the question.