Transforming Scientific Fields through AI
Last week, I attended a conference at MIT featuring a compelling presentation by Neil Thompson, an expert in the advancements of computer science and artificial intelligence (AI). In a brief yet impactful half-hour, Thompson articulated how various scientific domains are on the cusp of transformation due to the rapid evolution of AI technologies.
While mathematicians express concern over the automation encroaching upon their discipline, Thompson’s latest research indicates that fields such as probability, statistical theory, economic theory, high-energy theory, gravitation, cosmology, and quantum physics are poised to experience similar revolutions. This potential is attributed to the swift development of leading-edge AI models.
According to the findings, between mid-2024 and mid-2026, top AI models are expected to improve their problem-solving capabilities dramatically, increasing their accuracy from approximately 2% to around 35% for theoretical problems in these disciplines. There are no signs that this momentum will slow down any time soon.
Thompson predicts that in just a few years, the landscape of theoretical inquiries will be fundamentally altered by the AI’s enhanced ability to address them.
The Societal Implications of AI Progress
This study, conducted by Thompson alongside doctoral student Gabriel Manso and researcher Claudio Mayrink Verdun, focused on problems presented in research papers published online between December 2025 and May 2026. They subsequently assessed whether various OpenAI models, including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5, could solve these issues.
It’s notable that the solutions to these problems were found in articles released after the training cut-off date for GPT-5.5 and its predecessors. This meant that the models had to derive solutions independently, rather than relying on pre-existing answers in their training data. The referenced articles served as benchmarks to evaluate the models’ problem-solving abilities. Additionally, the researchers restricted internet access during testing, preventing the models from searching for solutions online.
The study’s scope was somewhat limited, concentrating on fields where AI models could tackle problems without the need for simulations or experimental activities. However, the researchers suggest that this methodology could be extended to other areas of scientific inquiry that possess a strong mathematical and theoretical foundation.
OpenAI and the Millennium Prize Problems
In September, OpenAI made waves in the mathematical community by announcing that an unpublished model had successfully resolved a significant long-standing mathematical challenge known as the Navier-Stokes problem, which pertains to the equations governing fluid dynamics. The announcement stirred controversy when a mathematician from New York University, collaborating with an academic associated with Anthropic, accused OpenAI of outpacing them after learning of their advancements.
Since then, OpenAI and other entities have reported hundreds of solutions to longstanding mathematical problems, leading to a stirring of discontent within the field. Some mathematicians have called for restrictions on the use of AI, warning that the process of problem-solving is often as valuable as the final answer itself.
Thompson points out that, similar to mathematics, we should expect advances in other fields to yield both beneficial and detrimental effects. He notes that theoretical progress can sometimes drive practical advancements in science, suggesting that increased AI utilisation may lead to more breakthroughs in this manner.
Integrating AI into Scientific Research
Prominent AI researchers have established new laboratories focused on harnessing AI for scientific purposes, while major AI labs are ramping up their scientific research efforts. These initiatives explore the potential for AI to conduct real experiments alongside theoretical analysis.
While AI might reduce the need for theorists to solve key problems, researchers assert that there will still be a necessity for experts to interpret solutions and ensure their robustness. Manso emphasises, “We will continue to need individuals who can generate solid ideas to tie everything together.”
He further adds that when faced with highly complex issues, AI may make errors. Indeed, OpenAI has withdrawn several solutions from its latest dataset following their release. “We also require methods to verify what has been accomplished, as it is easy for a model to declare, ‘this is the correct answer,’ and upon further analysis, one may realise that this might not be the case,” he concludes.