In a talk at MIT, Neil Thompson, who studies advances in computing and artificial intelligence, presented evidence that several scientific disciplines will soon be radically transformed by AI. As mathematicians experience the automation of their field, Thompson’s new research suggests that probability theory, theoretical statistics, economic theory, theoretical high-energy physics, gravitation and cosmology, and quantum physics could be next, thanks to the rapidly growing capabilities of advanced models. According to the study, between mid-2024 and mid-2026, the best AI models went from correctly solving roughly 2% of theoretical problems in these fields to around 35%. There are no signs that this progress is slowing.
How the study was conducted
Thompson conducted the study with postgraduate student Gabriel Manso and research fellow Claudio Mayrink Verdun. They took problems set out in scientific papers published online between December 2025 and May 2026 and tested whether various OpenAI models — o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5 — could solve them.
The key point was that the solutions to these problems appeared in papers published after the training data cut-off for GPT-5.5 and earlier models. This meant that the models tested had to find the solutions themselves, rather than rely on answers supplied to them as training data. These papers served as a benchmark for testing the models’ abilities. The researchers also blocked internet access so that the models could not search for solutions online.
The study’s scope was limited: the authors focused on fields where AI models could work through a problem without having to run simulations or experiments. Nevertheless, the work shows that the shock waves now being felt in mathematics are likely to spread to other areas of scientific research that rely heavily on mathematics and theory.
Context: mathematics has already felt the effects
Last month, OpenAI shook the mathematical world by announcing that an unreleased model had solved a major, long-standing mathematical problem — the Navier–Stokes equations, which describe the motion of fluids. The episode sparked further controversy after a New York University mathematician, working with another scientist affiliated with Anthropic, accused OpenAI of beating them to the announcement after learning of their progress.
OpenAI and others have since announced hundreds of new solutions to long-standing mathematical problems, triggering something of a crisis in the field. Some mathematicians have called for limits on the use of AI, warning that the process of solving problems is often just as useful as arriving at the final answer.
What this means for science
As with mathematics, Thompson says, advances in other fields should be expected to have both positive and negative consequences. He notes that theoretical breakthroughs sometimes drive practical advances in science, and suggests that wider use of AI could mean more progress happens this way.
Some prominent AI researchers have already founded new laboratories focused on applying AI to science, while major AI laboratories are expanding their scientific research. These efforts explore how AI can be given the ability to conduct real experiments alongside theoretical analysis.
AI could mean that theorists are no longer required to solve key problems, but, the researchers say, this does not mean experts will no longer be needed to make sense of the solutions and check their validity. “We’ll still need people to generate good ideas and connect the dots,” says Manso. He adds that AI can make mistakes on very difficult problems. (OpenAI did, in fact, withdraw several solutions from its latest dataset after publication.) “We also need ways to check what has been done, because it’s very easy for a model to say ‘this is the right answer’, and when you look at the whole solution, you start to see that perhaps it isn’t.”
Source: WIRED
