Claude Breaks World Record in Nine-Loop Scattering Amplitude Physics Calculation
nashnova research
Anthropic's Claude ran for days with near-zero human intervention and computed nine-loop scattering amplitudes, breaking the eight-loop human record held by Stanford SLAC's Lance Dixon. For the first time, AI independently completed — at minimal cost — a calculation top physicists spent years unable to crack.
What does "nine loops" mean, and why is it so hard?
Scattering amplitudes — calculations predicting what happens when particles collide — are a core tool of theoretical physics. Each additional "loop" causes the computational load to explode exponentially.
Most researchers stop at two or three loops. The most precise predictions in particle physics use only five loops. The previous human record was eight loops, held by Stanford SLAC professor Lance Dixon's team.
This means → nine loops is not "one more step." It crosses into a zone that human teams cannot realistically reach within any reasonable time or budget.
How did Claude actually do it?
Claude used the "bootstrap" method — a technique that works backward from known results to derive unknowns — developed over years by Dixon's team. It ran two independent paths in parallel, then cross-verified the answers.
The pivotal move: Claude exploited "antipodal duality" symmetry to compress the unknowns from 1.85 million down to 76,000, locking in a unique solution.
In the final result, a single kinematic slice expanded to over 30 billion terms — up from 1.67 billion at eight loops. In plain terms = the data volume swelled nearly twentyfold, and Claude's automated pipeline powered through it in one continuous run.
What did it cost, and how long did it take?
The two computation paths together cost a few thousand dollars. The raw compute for the bootstrap path alone was roughly $100 — equivalent to renting 96 CPUs for about a week.
A researcher typed one prompt, asked Claude to report progress every four to six hours, and walked away. No further intervention followed.
This means → a task that occupied a top physics team for years was replaced by one prompt, a few thousand dollars, and a few days. The gap in cost structure is the most striking part of this breakthrough.
What does the former record-holder think?
Dixon himself spent roughly two weeks checking the result and confirmed it is correct.
He wrote: "Other than me and my co-authors, Claude is probably the 'person' in the world who best understands my 2019 and 2023 papers."
He is not dismayed — Claude used methods his team developed. But he added: "When LLMs start proposing new physical principles before humans do, that will be the moment that truly touches the human soul."
Did a Chinese Academy of Sciences team also reach nine loops?
Eight days before Anthropic's announcement, a team led by He Song at the Chinese Academy of Sciences' Institute of Theoretical Physics published core skeleton data for six-particle amplitudes from two through nine loops on the academic platform Zenodo.
Dixon disclosed that the team used GPT-6 for part of the constraint calculations, but the overall framework remained human-led — a "human + AI" collaboration model, distinct from Anthropic's pure-AI path.
This reflects a shift: AI-assisted theoretical physics is moving from one-off events to a trend — two routes, two models, arriving at the same destination in nearly the same time window.
Where is the next real threshold?
The core of this breakthrough: AI completed, with near-zero supervision and minimal cost, a calculation that eluded top human teams for years.
Yet both Dixon and the Anthropic team acknowledge that Claude still relied on methods humans had already devised.
In plain terms = AI has proven itself an elite "executor," but not yet a "discoverer." The next test is whether AI can independently propose new physical principles without leaning on existing human frameworks. That would mark the real leap from tool to agent.
市场有风险，内容仅供研究参考，不构成投资建议。