The article in brief
● AI has found an unexpected result in theoretical physics for the first time.
● AI can now solve complex problems that researchers have been working on for years.
● Researchers who don’t use AI are already at risk of falling behind this year, believes physicist Alex Lupsasca.
– We are experiencing a turning point for science. In 2026, AI will transform research in the same way that AI transformed programming and software development in 2025, says Alex Lupsasca from the Department of Physics and Astronomy at Vanderbilt University in Nashville, USA.
With the help of AI, he and his colleagues were able to show that so-called gluons, sticky particles that hold the protons and neutrons together in the atomic nucleus, can influence each other under conditions under which physicists generally believe this is not possible.
– As far as I know, this is the first time a problem of this caliber has been solved by an AI.
Some of the researchers behind the study had been trying for a year to prove mathematically that the influence was possible, without success. It wasn’t until they fed the question to the company Open AI’s latest AI model that they provided proof in the form of a simple formula.
– We checked the formula by hand. “A large part of the article is our proof that the formula is correct,” says Alex Lupsasca.
The study is available in the Arxiv database and was presented at the AAAS scientific conference in Phoenix, USA.
A year ago Would Alex Lupsasca have described himself as an AI skeptic?
– I thought Chat GPT was a great tool for proofreading emails and grant applications. But it will never be useful for what I do: very, very advanced mathematics, he says.
When he tried to describe a magnetic field at a so-called pulsar, a rotating neutron star, he came up with a long and complicated equation.
– The problem is that you never know whether it is possible to simplify such an equation. Sometimes it works, sometimes it doesn’t. And it takes a lot of time.
Alex Lupsasca asked Help on the Mathoverflow website, where mathematicians help each other with difficult problems. As an experiment, he also sent the question to a friend who had access to Open AI’s latest model called o3.
He got a solution, both from the AI and from a mathematician on the website.
– The funny thing was that the AI thought for eleven minutes before giving me the answer. “I’ve never seen an AI think for so long,” says Alex Lupsasca.
The AI answer wasn’t quite right. However, it turned out that to be successful he had to use a result that had been published in 1957 in a small mathematical journal in Norway.
“It was the first time I realized that AI might even be able to do what a theoretical physicist does,” says Alex Lupsasca.
The decisive turning point for Alex Lupsasca came last June. He managed to show why a moon orbiting a black hole cannot produce tidal effects in the same way as moons around planets, and he published the result on Arxiv.
– It took me many weeks to get to this, but it’s also something I’ve thought about throughout my career. So you could say it took me ten years to get there.
According to Alex Lupsasca, there are at most ten people in the world who could solve the problem.
– That’s why I’m very proud of this result, he says.
Facts:Archive
Arxiv (usually spelled arXiv) is an open database where researchers share research results in areas such as physics, mathematics, astronomy, computer science and statistics before submitting them to a scientific journal. The studies there have therefore not yet been officially reviewed by other researchers.
When Open AI released a new version of Chat GPT, he gave her the problem. The AI pondered for 18 minutes. Then it gave him the same equation he came up with himself.
– You can imagine how I felt. I have dedicated my entire life to becoming the best in this particular field. I am an elite mathematician. And in less than half an hour the machine did what I’ve tried to do throughout my career.
Alex Lupsasca contacted Open AI to ask them to check if the AI had access to his article. However, it had just been released and was not included in the program’s training data. It had figured everything out on its own.
– Since then, I’ve seen others go through the same thing. It typically takes three days to calm down and is similar to the five stages of grief, acceptance and denial, says Alex Lupsasca.
Eventually he began to see the possibilities himself. Since October he has been part of Open AI’s new science department.
AI development is now underway So fast that it will completely change what it means to be a physicist, says Alex Lupsasca, just as computers and computing programs have done over the last 50 years.
– A theoretical physicist or mathematician who now uses AI for calculations is an early pioneer. But I expect that by the end of the year, anyone not using AI will be left behind.
Facts:AAAS
The American Association for the Advancement of Science, AAAS, is the world’s largest association for scientists.
The organization organizes a conference every February. The first meeting took place in 1848.
This year’s conference was held in Phoenix, Arizona and the theme was “Science at Scale.”
AAAS also publishes the journal Science.
Read more:
Five questions about Open AI – has Chat GPT become smarter than a human?
Here are the 10 biggest scientific breakthroughs of the year
Maria Gunther: No turning back for American research
