Last week, a presentation at the Massachusetts Institute of Technology (MIT) delivered a stark and compelling forecast: artificial intelligence is on the cusp of fundamentally reshaping numerous scientific disciplines. Neil Thompson, a researcher specializing in computing and artificial intelligence progress, articulated this vision during a concise, yet impactful, half-hour talk that has since sent ripples of anticipation and apprehension through the scientific community. His latest research indicates that fields such as probability, statistics theory, economic theory, high-energy physics, gravitation and cosmology, and quantum physics are likely to experience significant upheaval in the coming years, mirroring the seismic shifts already being observed in mathematics.
The AI Frontier in Theoretical Science
Thompson’s study, conducted in collaboration with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, focused on the evolving capabilities of advanced AI models in tackling complex theoretical problems. The research team meticulously analyzed problems presented in scientific papers published online between December 2025 and May 2026. They then rigorously tested the performance of several leading OpenAI models, including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5, on these theoretical challenges.
A critical aspect of the study’s methodology was its commitment to assessing the AI models’ genuine problem-solving prowess, unadulterated by readily available solutions. The researchers specifically selected papers whose publication dates fell after the known training cutoffs for GPT-5.5 and its predecessors. This crucial detail ensured that the models were not simply retrieving pre-existing answers from their training data. Instead, they were compelled to derive solutions independently, offering a more authentic measure of their inferential and analytical capabilities. To further isolate the models’ inherent abilities, internet access was deliberately restricted, preventing any external recourse to online solutions.
The findings are striking. Between mid-2024 and mid-2026, the study reveals a dramatic increase in the AI models’ ability to correctly solve theoretical problems across these scientific domains. The success rate surged from a mere 2 percent to an impressive 35 percent. Crucially, Thompson and his team observed no indication that this trajectory of rapid advancement is decelerating.
"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," Thompson stated during his presentation. This prediction underscores the accelerating pace at which AI is moving from a tool for analysis to a generative force in scientific discovery.
A Precedent Set in Mathematics
The implications of Thompson’s research are amplified by the recent, highly publicized events within the field of mathematics. In September, OpenAI announced that an unreleased AI model had successfully solved the Navier-Stokes equations, a long-standing grand challenge in mathematics that governs the behavior of fluid motion. This breakthrough, while celebrated by some, also ignited significant controversy. A mathematician at New York University (NYU) accused OpenAI of "gazumping" them and their collaborators after learning of their own parallel progress on the problem, highlighting the intense competition and the potential for AI-driven discoveries to disrupt established academic timelines and collaborations.
Since this initial event, OpenAI and other entities have reportedly presented hundreds of additional solutions to previously intractable mathematical problems. This proliferation of AI-generated solutions has precipitated what many in the mathematical community are calling a "crisis." Some prominent mathematicians have publicly called for limitations on AI’s role in problem-solving, arguing that the process of mathematical discovery and rigorous proof is as valuable, if not more so, than the final answer itself. This sentiment stems from a concern that over-reliance on AI could diminish the development of critical thinking and intuitive reasoning skills among human mathematicians.
Expanding the AI Horizon: Beyond Pure Theory
Thompson’s study intentionally focused on scientific fields where AI models could deduce solutions without the necessity of running complex simulations or conducting physical experiments. This constraint was deliberate, aiming to isolate and quantify the AI’s capacity for abstract reasoning and theoretical problem-solving. However, the researchers’ conclusions suggest that the "shock waves" currently reverberating through mathematics are poised to extend to other disciplines that are heavily reliant on theoretical frameworks and abstract thinking.
The implications for scientific progress are multifaceted, encompassing both potential benefits and significant challenges. Thompson acknowledged that theoretical breakthroughs often serve as the bedrock for practical advancements in science. Consequently, an accelerated rate of theoretical discovery driven by AI could translate into a faster pace of real-world innovation.
The scientific community is already witnessing a surge in initiatives dedicated to harnessing AI for scientific discovery. Prominent AI researchers have launched new ventures specifically focused on developing AI tools for scientific research, and major AI laboratories are significantly increasing their investments in this area. These efforts are exploring not only AI’s theoretical capabilities but also its potential to conduct and analyze real-world experiments, moving beyond pure computation to encompass the full scientific inquiry loop.
The Evolving Role of the Human Expert
While AI may reduce the necessity for human theorists to solve key problems from scratch, the researchers emphasize that human expertise will remain indispensable. Gabriel Manso, a co-author of the study, stressed the continued need for human insight. "We are still going to need people to generate good ideas to connect the dots," Manso stated. This highlights the role of human intuition in framing research questions, interpreting AI-generated results, and integrating disparate pieces of information into a cohesive scientific narrative.
Furthermore, the complexity of the problems AI tackles means that errors, even in advanced models, are a distinct possibility. OpenAI itself has acknowledged this by withdrawing several solutions from its latest dataset post-publication, underscoring the need for rigorous validation. "We also need ways to verify what was done, because it’s very easy for a model to tell you ‘this is the right answer,’ and when you look at the entire solution, you start to see maybe it is not," Manso elaborated. This points to the critical ongoing role of human oversight in ensuring the accuracy and soundness of AI-derived scientific conclusions.
A Glimpse into the Future of Scientific Discovery
The rapid evolution of AI capabilities, as evidenced by Thompson’s research and the recent developments in mathematics, signals a transformative era for scientific inquiry. The ability of frontier AI models to independently solve complex theoretical problems suggests a future where the pace of discovery could accelerate dramatically. However, this acceleration brings with it profound questions about the nature of scientific expertise, the validation of AI-generated knowledge, and the ethical considerations surrounding AI’s role in pushing the boundaries of human understanding.
As AI models become more sophisticated, their impact will likely extend beyond theoretical disciplines. The current focus on fields where simulations or experiments are not immediately required is a strategic starting point, but the long-term implications suggest that AI could eventually contribute to, or even drive, discoveries across all branches of science. This necessitates a proactive and adaptive approach from researchers, institutions, and policymakers to navigate this evolving landscape responsibly and harness the full potential of AI for the betterment of scientific knowledge and human progress. The conversation initiated by Thompson’s presentation at MIT is not merely academic; it is a critical dialogue that will shape the future of science for decades to come.
