dexinwin September 27, 2026
Artificial intelligence has taken another step into the territory of frontier scientific research.
Anthropic has announced that its Claude model, operating with minimal human supervision for several days, completed anine-loop calculation of a six-particle scattering amplitude in planar N=4 supersymmetric Yang-Mills theory.
The result surpassed the previous publicly known eight-loop record achieved by human theoretical physicists. The calculation was subsequently checked by physicists familiar with the problem.
From Eight Loops to Nine
The term “loop” refers to a level of quantum correction in quantum field theory calculations.
As physicists move to higher loop orders, the mathematical complexity of the calculation generally increases dramatically. More interactions and corrections have to be taken into account, resulting in increasingly large symbolic expressions and increasingly demanding computational requirements.
Claude was asked to tackle a specific frontier problem: the six-particle scattering amplitude at nine loops in planar N=4 supersymmetric Yang-Mills theory.
The theory is not intended as a complete description of the real universe. Instead, it is a highly symmetric theoretical model that physicists use as a testing ground for mathematical techniques in quantum field theory.
Before Claude's result, researchers had pushed the same calculation to eight loops.
Anthropic's experiment extended that result to nine loops.
One Prompt, Then Days of Autonomous Work
One of the most striking aspects of the experiment was the limited amount of human intervention.
Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude Science, a research-oriented environment built around Claude, to tackle the problem.
The initial instruction was relatively simple: calculate the six-particle amplitude in planar N=4 supersymmetric Yang-Mills theory at nine loops.
Rather than continuously guiding the model through each step, the researchers instructed Claude to continue working and provide progress updates every few hours.
The model then worked for several days, using Python and symbolic mathematics tools including SymPy.
It also applied a technique known as thebootstrap method, which allows researchers to constrain the possible form of a complicated physical quantity by imposing known theoretical requirements.
Anthropic said Claude ultimately obtained the result through two different computational approaches, with the results agreeing.
Why Is a Nine-Loop Calculation So Difficult?
Scattering amplitudes are mathematical expressions used by particle physicists to calculate the probabilities of interactions between particles.
More accurate calculations can help researchers compare theoretical predictions with experimental measurements from facilities such as the Large Hadron Collider.
The challenge is computational complexity.
At higher loop orders, calculations become dramatically more complicated. Many scattering amplitudes can only be calculated to relatively low loop orders, while only a small number of special theoretical models allow researchers to push much further.
Anthropic's research account noted that many practical amplitude calculations are performed at two loops or below, making a nine-loop result in a highly sophisticated theoretical model a significant computational challenge.
The “Yang” in Yang-Mills
The breakthrough also highlights a connection to one of the most important theories in modern particle physics.
The “Yang” inYang-Mills theoryrefers to Chinese-American physicist Yang Chen-Ning, better known as Chen-Ning Yang or Yang Zhenning.
In 1954, Yang and Robert Mills developed the Yang-Mills framework, which became a foundational part of modern gauge theory.
Yang-Mills theories are closely connected to the theoretical descriptions of several fundamental interactions, including the strong and weak nuclear forces and electromagnetism.
The particular theory used in Claude's calculation — planar N=4 supersymmetric Yang-Mills theory — is a highly specialized extension used primarily as a mathematical laboratory.
Claude did not discover a new law of physics. Instead, it pushed a difficult calculation within an established theoretical framework to a higher level than the previous human record.
Human Physicists Still Played a Critical Role
The result also illustrates why AI-generated scientific calculations still require human verification.
Lance Dixon, a theoretical physicist at SLAC and Stanford who was involved in the earlier eight-loop work, independently checked Claude's nine-loop result.
That verification is essential. Producing a mathematical expression is not the same as establishing that the expression is correct.
Researchers need to examine the calculation, test theoretical constraints and compare the result with independent methods.
Anthropic said the nine-loop result was checked and that the calculation was consistent with independent work.
At around the same time, a research team led by Chinese Academy of Sciences physicist Song He also made progress on the nine-loop amplitude, using GPT-6 to assist with part of the constraint calculations.
The parallel efforts suggest that AI systems are increasingly becoming part of advanced theoretical physics workflows.
A Result Achieved Without Massive Computing Budgets
Another notable aspect of the experiment was its reported computational cost.
Anthropic said the work did not require millions of dollars in computing resources.
One of the approaches used roughly 96 CPUs for about a week, while the overall Claude usage for each approach was estimated at around one to two thousand dollars.
That does not mean that anyone could simply spend a few thousand dollars and reproduce the result.
The calculation depended on decades of theoretical work, sophisticated mathematical techniques, existing software and carefully designed research infrastructure.
What Claude demonstrated was the ability to combine those existing resources and techniques, organize the computational workflow and carry out a large calculation with relatively little ongoing human supervision.
From Research Assistant to Research Operator
The broader significance may lie in the changing role of AI in scientific research.
Until recently, AI systems were primarily used to assist scientists by searching literature, writing code, summarizing papers or helping with calculations.
The Claude experiment demonstrates a different model.
Researchers can provide a scientific objective and allow an AI system to spend hours or days decomposing the problem, writing code, running calculations, examining intermediate results and continuing until it reaches a result.
Anthropic's account emphasizes that the achievement was less about Claude independently inventing new physics and more about its ability to execute a complicated computational recipe while coordinating software and computing resources.
That distinction matters.
Has AI Begun Discovering New Physics?
The answer, based on this experiment, is not yet.
Claude completed an extremely difficult calculation within an existing theoretical framework. It did not discover a new fundamental particle or propose a new physical law.
The next major milestone would be different: an AI system independently developing a new physical principle or mathematical method that humans had not previously considered, and then producing predictions that could be tested.
Computing farther into an existing framework and discovering genuinely new scientific ideas are two different capabilities.
For now, the nine-loop calculation is evidence of the former.
A New Model for Scientific Work
Moving from eight loops to nine may sound like a small numerical step, but the computational complexity behind that extra loop is substantial.
The experiment raises a broader question about the future of scientific research.
Could research increasingly follow a model in which humans define important questions, AI systems perform large-scale computational exploration, and human scientists verify, interpret and build upon the results?
There is no definitive answer yet.
But one point is already clear: in this particular theoretical-physics problem, Claude has demonstrated that an AI system can take on a calculation that previously represented the frontier of human computational work and push it one step further.
For artificial intelligence, it is a new computational milestone.
For physics, it is another indication that AI is moving deeper into the workflow of cutting-edge scientific research.
微信扫一扫打赏
支付宝扫一扫打赏