ChatGPT cracks decades-old gluon amplitude puzzle, marking AI's first major theoretical physics win

ChatGPT spits out surprising insight in particle physics

An unlikely new contributor has entered the ranks of theoretical physics: ChatGPT. For decades, physicists believed a particular obscure interaction involving particles called gluons could never happen. Now, ChatGPT has revealed that the process can in fact occur, albeit somewhere deep inside the murky guts of protons and neutrons. Researchers announced the possibility last week at the annual meeting of AAAS, which publishes Science. "The ideas are not revolutionary," says Zvi Bern, a particle theorist at the Mani L. Bhaumik Institute for Theoretical Physics at the University of California, Los Angeles. "But what is revolutionary is that a machine can do this." Gluons are the massless quantum particles that convey the strong nuclear force, which binds particles called quarks into protons and neutrons, as well as protons and neutrons into atomic nuclei and gluons to each other. Because the strong force is so strong, the theory of it is nearly intractable mathematically. Theorists can describe each specific interaction of gluons -- two gluons ricocheting off each other to spawn a third, for example -- with a mathematical expression known as a scattering amplitude that, roughly speaking, gives the probability of the interaction occurring. But, even for the simplest interactions, those scattering amplitudes can be dauntingly complex and nearly impossible to evaluate. ChatGPT helped crack one of these puzzles. Gluons spin a bit like tops, and one can spin either in the direction it is traveling, like a football thrown by a right-handed quarterback, or it can spin the opposite direction, like a football thrown by a left-handed quarterback. In the latter case, the gluon is said to have negative helicity. For decades, physicists thought that in the simplest collisions of any number of gluons, at least two of all the particles had to have negative helicity. If only one had negative helicity, the scattering amplitude had to be zero. About a year ago, however, three theorists spotted a potential loophole: A lone gluon of negative helicity could interact with others of positive helicity if all the particles were moving in roughly the same direction. Now, they had to prove it. Andrew Strominger, a theoretical physicist at Harvard University and co-author of the study, says he and his colleagues initially thought they could confirm their hunch in a few weeks. But the calculations ended up being cumbersome and time consuming. After months of working by hand, Alfredo Guevara, a high energy physicist at the Institute for Advanced Study, finally discovered a pattern in the team's scattering amplitude formulas. The equation that described four gluons looked like the one for five gluons, and so on. The team hoped to generalize the formula and show that the interaction could happen for any number of gluons, but the resulting expression was dozens of terms long and essentially unworkable. The group suspected an elegant and clean formula was hiding in this quagmire: One had been discovered in the 1980s for a similar type of gluon interaction. But even after a year of work, the researchers couldn't simplify what they had. Around the same time, Alex Lupsasca, a theoretical physicist at Vanderbilt University, joined the newly launched OpenAI for Science team and was tasked with improving ChatGPT's science abilities. He connected with Strominger, his graduate adviser, and discovered that this gluon problem would be the perfect test subject. "I thought, it's probably not going to work, but we'll find out why not" and adjust the artificial intelligence (AI) model accordingly, Lupsasca says. After some first attempts to probe the model, the theorists asked OpenAI's latest and most advanced public model, ChatGPT-5.2 Pro, to simplify the expression for four gluons, which it did in about 20 minutes. Then they asked to do the same for five gluons, then six. GPT-5.2 Pro managed to reduce a sum of 32 terms to a product of only a few, all on one line of text. Finally, the group asked for a guess of the generalization of the formula for any number of particles. This time, it replied within a minute or two, providing what it called an "obvious" generalized formula. Worried the answer might be a hallucination, the researchers checked the formula and couldn't find anything wrong. "All of a sudden, I felt like my machine turned from a machine into a live being," Strominger says. Next, the group took the generalized formula from GPT-5.2 Pro and fed it into an internal OpenAI model that's under development, which the researchers privately call "SuperChat," prompting it for a proof. After 12 hours of processing, the model spat out a robust proof that passed human checks. Within hours of its posting to the preprint server arXiv on 12 February, the team's paper was trending on social media. And when Lupsasca presented the team's results during a 13 February session at the meeting, physicists in attendance were shocked. "What the OpenAI agent was able to do is impressive," says Aida El-Khadra, a particle theorist at the University of Illinois Urbana-Champaign. The paper's authors are optimistic about AI's future place in the sciences. "This is a change of paradigm in the way we do physics," Guevara says. He thinks AI will be able to do for physics what it has recently done for programming: become so good that human users can rely on it for day-to-day tasks without extensive scrutiny. Lupsasca, who claims he was an AI skeptic only a year ago, says, "I think there is some kind of threshold that is being passed." The wider physics community has received the news with cautious optimism. Bern and El-Khadra expect assistance will help with routine tasks, find errors in research, accelerate the paper writing process, and bridge information across fields. Both cite concerns about the use of AI in academic physics, such as researchers' unacknowledged use of AI in their work or the possibility that large language models will automate many of the research tasks traditionally used to train graduate students. But neither is worried that ChatGPT is coming to take their jobs. "None of this feels to me like scientists will be replaced," El-Khadra says. Lupsasca hopes researchers can use AI to solve the biggest problem in theoretical physics: reconciling quantum mechanics and gravity. He plans to extend the team's approach to gravitons, hypothetical quantum particles that convey the force of gravity, and to find a way to mathematically describe a special version of quantum gravity that remains well-behaved even at high energies, perhaps "by the end of the year."

Can a chatbot be a co-author? AI helps crack a long-stalled gluon amplitude proof

Like many scientists, theoretical physicist Andrew Strominger was unimpressed with early attempts at probing ChatGPT, receiving clever-sounding answers that didn't stand up to scrutiny. So he was skeptical when a talented former graduate student paused a promising academic career to take a job with OpenAI. Strominger told him physics needed him more than Silicon Valley. Still, Strominger, the Gwill E. York Professor of Physics, was intrigued enough by AI that he agreed when the former student, Alex Lupsasca, Ph.D., invited him to visit OpenAI last month to pose a thorny problem to the firm's powerful in-house version of ChatGPT. Strominger came away with much more than he expected -- and the field of theoretical physics appears to have gained a little something too. "Incredible," Strominger put it, acknowledging that AI quickly reasoned through a problem he wasn't sure he could solve himself without unlimited time. Strominger had carefully chosen a problem that had eluded concerted collaborative efforts at solving but was understood well enough to clearly post to AI. Neither scientist expected a breakthrough. Even Lupsasca, who has served as a research scientist at OpenAI since last fall, imagined the problem would probably trip up the AI, giving them an opportunity to provide feedback and help improve the large language model's reasoning around complex theoretical physics. Instead, the internal ChatGPT -- what Strominger dubbed "Super Chat" -- eventually solved the thorny problem in its entirety. AI as a fifth collaborator Four physicists -- Strominger, Lupsasca, Cambridge University's David Skinner, and the Institute for Advanced Study's Alfredo Guevara (who had worked with Strominger as a junior fellow in the Society of Fellows from 2020-2024) -- worked with ChatGPT as a powerful fifth collaborator. ChatGPT-5.2 pro broke the logjam, proposing an answer, and Super Chat proved it was correct after 12 hours of running. The group then spent a week breaking down the solution, checking the calculations by hand, and turning it into a paper, the result of which ("Single-minus gluon tree amplitudes are nonzero") was published as a preprint on the arXiv server. While the specific findings will interest what Strominger called "the cognoscenti in some sub-field of theoretical physics," the broader takeaway for those without a physics Ph.D. requires no fluency with gluon amplitudes. "It's the first significant discovery in theoretical physics that is done by an AI," said Lupsasca, who is also a former junior fellow. "Maybe we'd have figured out a clever trick the next day," Strominger said about the efforts of the team of physicists. "Maybe we'd have never gotten it." The scientists collaborated using both the publicly available ChatGPT-5.2 pro and the in-house Super Chat, which can "think" through complex problems for 12 hours at a time. Strominger found the experience exhilarating. "There was a moment when I felt like I was working with a creative person," he said. "Not just a machine that was crunching through stuff. You know, that's all psychological, but it felt that way." The physics puzzle behind the breakthrough Strominger typically works on three or four problems at a time, progressing incrementally but steadily. In this case, though, he had stalled while attempting to prove a conjecture about gluons, the particles that mediate the strong force binding the nucleus of an atom together. Amplitudes are the complex quantities used in quantum mechanics to provide probabilities for the outcomes when atomic particles interact. Physicists sometimes presumed a certain kind of gluon amplitude could not exist. Strominger, Skinner, and Guevara thought otherwise. Guevara worked out an exceedingly complex expression of these amplitudes, but they could not finesse it into something simple. They even tried feeding it into ChatGPT last spring, without success. "It just fumbled," Strominger said. "The latest model is a whole new ballgame." How Lupsasca became 'AI-pilled' Enter Lupsasca, who had recently transformed from skeptic to proselyte. A year ago, having tried only the free version of ChatGPT, he considered it useful mainly for proofreading grant proposals. Then he got stuck trying to find a solution to a differential equation describing magnetic fields around pulsars. "Usually in this game, there's always some trick that you have to pull out of a hat," a "special identity," or formula, to unlock the answer, he said. A friend with a subscription to the pro version of ChatGPT-3 suggested he feed it an experiment. In 11 minutes, it solved the problem using a special identity published in an obscure Norwegian mathematical journal in the 1950s. (Still, it made a very "human" mistake, resolving the hard part but adding a typo to the answer.) Last June, Lupsasca published a paper after deriving new black-hole symmetries, what he called "one of my coolest calculations." He was feeling good -- "I can count on my hands the number of people in the world that could have done that" -- until he tested out the new ChatGPT-5.2 pro upon its release in August. In less than 30 minutes, AI crunched through calculations that had taken him considerable time and brainpower. "That's when I became "AI-pilled," as people say," Lupsasca said, determined to join the vanguard of what he called "the most significant change in theoretical physics in my lifetime." OpenAI for Science takes shape He reached out to OpenAI, which had prioritized the model's ability to transform coding but hadn't taught it to work on complex physics problems. ChatGPT simply learned while absorbing oceans of information. So the company accelerated the launch of OpenAI for Science, a program to hire specialized faculty to reinforce ChatGPT's reasoning, starting with math and theoretical physics. In October, they made Lupsasca their first hire; he took leave from the faculty at Vanderbilt and made his mentor Strominger his first invitation to test ChatGPT. What this means for human scientists Since Strominger's return to campus, people have asked whether AI might render him obsolete. "Call it vanity. I think I'm irreplaceable," he mused. "I think it's just the opposite. I think it will empower us to do more, but we have to retool. Good scientists have to retool all the time."