Anthropic 官宣,Claude 拿下了理论物理的一项前沿纪录。
它在几乎无人类干预的情况下,连续运行数天,一举攻克了高能物理学界出了名难算的「九圈散射振幅」计算难题!
具体来说,它算出了平面 N=4 超杨-米尔斯理论里六粒子振幅的九圈结果,人类此前的纪录停在八圈。
注意,杨-米尔斯模型(Yang-Mills)其中的「杨」,指的正是物理学泰斗杨振宁先生。
他与米尔斯在 1954 年共同创立的「杨-米尔斯理论」,是整个现代粒子物理学的基石。
基于它的任何高阶计算,都堪称人类脑力的极限挑战。
如今,人类顶尖理论物理学家,就这样被 AI「截胡」了。
更可怕的是,Claude 完成这项神级挑战,仅仅只用了一个提示词,花了几千美元!
几天后交回的答案,光其中一部分展开就有 300 亿项。
看起来,科学界的低垂果实,比想象中要多得多啊。
这次,负责给 AI 验证结果的物理学家 Lance Dixon 这样感慨:
我身边的大多数理论物理学家实际上早就承认,大模型迟早会颠覆物理学,只是不知道那一天什么时候砸到自己头上。
对我来说,那一天是 9 月 1 日。
硬核物理学家挑衅 AI「我的老本行很特殊」
故事的起因,源于一位物理学家的「挑衅」。
Matt von Hippel,曾经是一位理论粒子物理学家,现在是一名科普作家。
面对最近的 AI 新闻,他表示不信邪,除非亲眼看到大模型应对自己领域的难题。
深蓝赢了国际象棋大师卡斯帕罗夫,围棋界说:围棋太复杂,电脑下不了。结果被 AlphaGo 打脸。
AlphaFold 碾压人类蛋白质结构专家,但其他领域说:笔者没有那种数据,复制不了。
现在,大语言模型能读文献、能模仿数学证明。但争论依旧不断:AI 是不是只会找反例,提不出新框架?是不是只在数学这种规则清楚的领域管用,物理这种模糊领域不行?……
每次都是同一个套路:学者先怀疑,直到 AI 杀进他们自己的领域。
他认为他的老本行「散射振幅」,不一样。
上个月,他在自己的博客上向全网 AI 公司发下了一封「英雄帖」:
如果你们 AI 公司真的想震撼笔者这些科研人员,就来挑战我的老本行吧!
请证明 AI 能在普通学者可以负担得起的计算预算内,应对散射振幅领域悬而未决的难题。
他一共出了三道题,其中一道便是:
给我把 N=4 超级杨-米尔斯模型(N=4 super Yang-Mills)算到第九圈(Nine Loops)!
在理论物理学中,物理学家们借助「散射振幅」的公式来预测粒子的行为,例如大型强子对撞机里的粒子怎么撞。
但这个公式简直是噩梦。为了无限逼近真实答案,物理学家必须一层一层地叠加被称为「圈」精细修正层。
每多加一个「圈」,答案就更精确,但计算量是呈指数级爆炸增长的!
实际操作中,大多数人类停留在第二圈或第三圈。粒子物理学中最精确的预测,也才用到了第五圈。
而在特殊的「测试沙盒」模型中,人类当下的极限纪录是「八圈」。这是由 SLAC 国家加速器实验室的大牛教授 Lance Dixon 和团队多年死磕才拿下的。
「九圈」?看起来是需要海量算力和无数天才大脑耗费数年才能触及的领域。
Matt 给 AI 的挑战,就是这么简单粗暴。
一句提示词,几千美元,Claude「杀疯了」
面对这种公开叫板,Anthropic 的科学家们,悄悄接下了战书。
八月底,Anthropic 的两名物理学家员工 Liam 和 Siddharth 悄悄联系了 Matt:「嗨,你的挑战,笔者搞定了。」
Matt 当场惊掉下巴。怎么做到的?烧了上百万美元的算力吗?
并没有。
研究团队用的是 Claude Science,一个专门为科学家外观、能让 Claude 遵循严谨规则运行的平台,模型是当下最强的公开可用模型 Fable 5.1。
他们给 Claude 输入了一个简单的 Prompt:
现在的难题是:计算平面 N=4 SYM 模型在九圈下的六粒子(六边形)振幅。
然后,研究员并没有手把手教 Claude 怎么做,而是给它留了一句话:
我要去睡觉了,接下来的几个小时我都不会在。你继续算,除非我叫你停,每隔 4 到 6 小时给我汇报一次进度。
然后,Claude 就在无人值守的情况下,开始了疯狂的推演!
它运用了 Dixon 和合作者这些年开发的一种叫做「自举」的方法,就像解一场数独游戏一样,在无数参数中不断排除错误选项。
几天后,Claude 不仅算出了九圈,而且是用了两种不同的物理方法(原始自举法和间接形状因子法)分别算出来的!
具体来说,在原始自举法中,Dixon 团队用的方法叫 bootstrap,类似于数独,把所有可能的函数全填进格子里,用物理法则一个个排除,剩下的就是答案。
Claude 直接掏出 Python 和开源库 SymPy,在六粒子振幅的空间里硬算出了一条血路。
在间接形状因子法中,当年 Dixon 他们冲八圈时,先算了个简单的「形状因子」,再利用一种奇特的「对跖对偶」对称性,把结果翻译成振幅。
Claude 顺着这套打法,往上又生推了一圈。
形状因子这一路,Claude 没有算一个费曼积分。它把整个物理难题变成一个超大的方程组,全程用整数求解,不会有小数误差,还能换几个大素数反复验算。
第一道坎是内存。照老办法推到九圈,光未知数就有 185 万个,方程组大到内存根本装不下。
Claude 换了个思路,借对跖对偶先把答案的一部分「猜」出来,再加上对称性,未知数一下从 185 万降到了 7.6 万个。
接下来就是解「数独」。7.6 万个未知数,被物理规则一轮轮排除,最后只剩唯一解。
最后翻译回振幅,光一个切面上的结果就有 300 多亿项,八圈时才 16.7 亿项。
而且,应对人类物理学界的悬案,花费极低。
按普通使用者的售价算,每条路要一两千美元,两条加起来几千美元,大头是 Claude 长时间运行的费用。
而直接自举那一路,纯粹用于运行 Python 代码的算力成本,仅仅只有 100 美元 —— 这只相当于租用了 96 个 CPU 跑了一个星期。
100 美金,一句提示词,Claude 在几天内就轻而易举地捅破了理论物理学的天花板。
前世界纪录保持者:「被 AI 截胡了,但我很兴奋」
结果算出来了,但对不对呢?
Anthropic 的团队找到了「八圈」纪录保持者 ——Lance Dixon 教授本人来验算。
那天是 9 月 1 日,巧的是,正好也是 Fable 5.1 发布当天。
Dixon 教授拿到结果的那一刻,内心是极为震撼的。因为他太知道这有多难了。
从 2023 年算完八圈之后,他和他的团队这两年一直在一筹莫展地死磕九圈。
整个计算过程就像是在搭一个扑克牌城堡,或者烤一个舒芙蕾,中间只要有一个代码写错、一个符号标反,整个计算就会彻底崩溃。
但 Claude 不仅没有崩溃,还从头开始自己把所有繁琐的代码全写了,完美避开了所有的坑!
经过长达两周的仔细验证,Dixon 教授确认,Claude 的九圈结果正确无误!
那么 Dixon 教授感到心碎吗?
「不会,」Dixon 教授在博文中豁达地写道,理由有两个。
一是他们团队本来就在训练定制的 transformer 模型预测更高圈数,口号是机器给出的任何候选答案都有工具验证。Claude 也是 transformer,只是大概比他们的大一百万倍。
二是 Claude 用的全是他和合作者这些年开发的方法,交答案时还用了他们团队早就定好的格式(教授猜,这是 Claude 在给他们卖面子)。
我在验证 Claude 的结果,Claude 也在验证笔者过去所有的工作。
他还给出了一个极高的评价:
除了我和我的合著者,Claude 可能是这个世界上最懂我 2019 年和 2023 年发表的论文的「人」了。
而他的同组大佬 Kyle Cranmer 还在 X 上凌乱,「刚发现笔者组被 Anthropic 抢发了,我还在消化…… 虽然不算太意外,但我真希望当初笔者能有时间和资源投进去。」
极限反转中国团队靠 GPT-6 也杀到了九圈
有趣的是,Anthropic 刚找完出题人 von Hippel 没几天,他又收到了另一个重磅消息 —— 来自中国科学院理论物理所的何颂团队。
何颂是国内散射振幅领域的大牛,近年代表性工作涵盖提出量子场论散射振幅的 Cachazo-何-袁形式,开创性地发现场论和弦论散射的几何起源,并在场论高圈计算和严格求解,规范场、引力和弦论的对偶等方向做出了一系列关键工作。
就在 9 月 17 日,何颂、景继荣、李想把六粒子振幅从二圈到九圈的核心骨架数据,公开挂上了学术平台 Zenodo。
何颂团队同样用上了 AI。Dixon 透露,GPT-6 帮中国团队算了一部分约束条件,不过总体框架还是人类自己搭的。
von Hippel 也确认,他们确实用了 GPT-6 做辅助,但绝不是 Anthropic 那种纯 AI 打法。
Dixon 自嘲说,「人们好像就是喜欢跑来告诉我,他们把九圈给搞定了。这下好了,我先被一台机器抢发,转头又被『人类 + 机器』联手抢发了一次。」
前不久 OpenAI 宣布攻克千禧年难题纳维-斯托克斯方程、Claude 推进黎曼猜想的一个关键下界的时候,就有人这样问:为什么 AI 只去攻克数学难题,不去攻克物理难题?
现在,这不就来了么。
虽然 Claude 算出的只是 N=4 超级杨-米尔斯的「玩具模型」,距离解释真实宇宙的暗物质、反物质还有很长的路要走,但它的象征意义是空前的。
这就像是 AlphaGo 第一次战胜李世石。人们惊呼的,是机器居然能掌握属于人类直觉的领域。
Dixon 教授留下了一句耐人寻味的话:
大型语言模型可以执行笔者设定的复杂配方并组织计算,这是一个巨大的胜利。
但当有一天,大语言模型开始在人类之前提出新的物理学原理和洞察力时,那才是真正触及人类灵魂的时刻。
或许,现在已经遍地都是低垂果实。
参考资料:
本文来源:https://www.ithome.com/1/007/444.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Anthropic officially announced that Claude has set a cutting-edge record in theoretical physics.
Without nearly any human intervention, it ran continuously for several days, conquering the notoriously difficult "nine-loop scattering amplitude" calculation in high-energy physics!
Specifically, it calculated the nine-loop result for six-particle amplitudes in the planar N=4 super Yang-Mills theory, pushing the human record, which had previously stopped at eight loops, further forward.
Note that the "Yang" in the Yang-Mills model (Yang-Mills) refers to Mr. Chen-ning Yang, the titan of physics.
The "Yang-Mills theory" co-founded with Mills in 1954 is the cornerstone of modern particle physics.
Any high-order calculation based on it is considered a challenge at the limit of human intellectual capacity.
Today, top human theoretical physicists have been "jumped" by AI.
Even more terrifyingly, Claude accomplished this god-tier challenge with just a single prompt, costing only a few thousand dollars!
The answer submitted a few days later contained 30 billion terms just for part of its expansion.
It seems that the low-hanging fruits in the scientific community are far more abundant than imagined.
This time, physicist Lance Dixon, responsible for verifying the AI's results, expressed such a sigh:
Most of the theoretical physicists around me have actually long admitted that large models would sooner or later disrupt physics, it's just that they didn't know when that day would hit them personally.
For me, that day was September 1st.
Hardcore physicist taunts AI: "My specialty is unique"
The story began with a "provocation" from a physicist.
Matt von Hippel, formerly a theoretical particle physicist, is now a science writer.
Facing recent AI news, he expressed skepticism unless he saw large models tackle the difficult problems in his field with his own eyes.
Deep Blue beat chess master Kasparov, and the Go world said: Go is too complex for computers. Then it was slapped in the face by AlphaGo.
AlphaFold crushed human protein structure experts, but other fields said: We don't have that kind of data, and can't replicate it.
Now, large language models can read literature and imitate mathematical proofs. But the debate continues: Is AI just finding counter-examples without proposing new frameworks? Does it only work in clearly defined fields like mathematics, but not in vague fields like physics?...
It's always the same routine: scholars doubt first, until AI breaks into their own domain.
He believed his old specialty of "scattering amplitudes" was different.
Last month, he issued a "challenge letter" to all AI companies on the internet through his blog:
If you AI companies really want to amaze researchers like us, come challenge my old specialty!
Please prove that AI can tackle unresolved problems in the field of scattering amplitudes within a computational budget affordable to ordinary scholars.
He posed three questions in total, one of which was:
Calculate the N=4 super Yang-Mills model to the ninth loop!
In theoretical physics, physicists use "scattering amplitude" formulas to predict particle behavior, such as how particles collide in the Large Hadron Collider.
But this formula is a nightmare. To infinitely approach the real answer, physicists must layer on fine corrections called "loops" one by one.
For each additional "loop," the answer becomes more precise, but the computational cost explodes exponentially!
In practice, most humans stay at the second or third loop. The most precise predictions in particle physics only use the fifth loop.
And in special "test sandbox" models, the current human limit record is "eight loops." This was achieved by Professor Lance Dixon and his team at SLAC National Accelerator Laboratory after years of intense work.
"Nine loops"? It seems to be a realm requiring massive computational power and countless genius brains working for years.
Matt's challenge to AI was just that simple and direct.
One prompt, a few thousand dollars, Claude "went crazy"
Facing this public challenge, the scientists at Anthropic quietly accepted the battle.
At the end of August, two physicist employees at Anthropic, Liam and Siddharth, quietly contacted Matt: "Hi, your challenge, we got it done."
Matt was stunned on the spot. How did they do it? Did they burn over a million dollars in computing power?
No.
The research team used Claude Science, a platform designed specifically for scientists that allows Claude to run under rigorous rules, using Fable 5.1, the strongest currently available public model.
They input a simple Prompt into Claude:
The current challenge is: Calculate the six-particle (hexagon) amplitude of the planar N=4 SYM model at nine loops.
Then, the researcher did not hand-hold Claude in how to do it, but left him with one sentence:
I'm going to sleep, and I won't be around for the next few hours. You continue calculating; unless I tell you to stop, report your progress to me every 4 to 6 hours.
Then, Claude began a frantic deduction in an unattended state!
It utilized a method called "bootstrapping," developed by Dixon and collaborators over the years, which is like solving a Sudoku game, constantly eliminating wrong options among countless parameters.
A few days later, Claude not only calculated the nine loops, but did so using two different physical methods (the original bootstrap method and the indirect form factor method)!
Specifically, in the original bootstrap method, the approach used by Dixon's team is called bootstrap, similar to Sudoku, filling all possible functions into the grid and eliminating them one by one using physical laws, with what remains being the answer.
Claude directly pulled out Python and the open-source library SymPy, carving a blood path through the space of six-particle amplitudes.
In the indirect form factor method, when Dixon and others pushed for eight loops back then, they first calculated a simple "form factor," then used a peculiar "antipodal duality" symmetry to translate the result into an amplitude.
Claude followed this approach and pushed one more loop up.
In the form factor route, Claude did not calculate a single Feynman integral. It turned the entire physical problem into a super-large system of equations, solved entirely with integers, avoiding decimal errors and allowing repeated verification with different large primes.
The first hurdle was memory. Following the old method to nine loops, there were 1.85 million unknowns, making the system of equations too large to fit in memory.
Claude changed its approach, using antipodal duality to "guess" part of the answer first, and added symmetries, reducing the unknowns from 1.85 million to 76,000.
Next was solving the "Sudoku." 76,000 unknowns were eliminated round by round by physical rules, leaving only one unique solution at the end.
Finally translating back to amplitudes, the result on just one slice had over 30 billion terms, compared to 1.67 billion terms at eight loops.
Moreover, dealing with this unsolved case in human physics cost very little.
Calculated at the retail price for ordinary users, each path cost one to two thousand dollars, totaling a few thousand dollars for both, with the main cost being Claude's long-running fees.
And in the direct bootstrap route, the computing cost purely for running Python code was only 100 dollars — which is equivalent to renting 96 CPUs for one week.
100 dollars, one prompt, and Claude easily broke through the ceiling of theoretical physics in a few days.
Former world record holder: "Jumped by AI, but I'm excited"
The result was calculated, but was it correct?
The Anthropic team found the holder of the "eight-loop" record — Professor Lance Dixon himself — to verify the calculation.
That day was September 1st, which coincidentally was also the release day of Fable 5.1.
At the moment Professor Dixon received the result, he was deeply shocked. Because he knew all too well how difficult this was.
Since finishing eight loops in 2023, he and his team had been stuck for two years trying to tackle nine loops.
The entire calculation process was like building a card house or baking a soufflé; if one line of code was wrong or one symbol was reversed, the whole calculation would collapse completely.
But Claude not only didn't collapse, it wrote all the tedious code itself from the beginning, perfectly avoiding all pitfalls!
After a careful verification process lasting two weeks, Professor Dixon confirmed that Claude's nine-loop result was correct and error-free!
Is Professor Dixon heartbroken?
"No," Professor Dixon wrote in his blog with a generous attitude, for two reasons.
First, their team was already training customized transformer models to predict higher loop numbers, with the slogan that any candidate answer given by machines has tools to verify. Claude is also a transformer, just about a million times bigger than theirs.
Second, Claude used methods developed by him and his collaborators over the years, and when submitting the answer, used a format their team had long established (the professor guessed this was Claude giving them face).
I am verifying Claude's results, and Claude is also verifying all my past work.
He also gave a very high evaluation:
Apart from me and my co-authors, Claude may be the "person" in the world who understands my papers published in 2019 and 2023 the most.
And his senior colleague in the same group, Kyle Cranmer, was in a mess on X, "Just found out our group was scooped by Anthropic, I'm still digesting it... Although it's not a total surprise, I really wish I had the time and resources to invest in it back then."
Limit reversal: Chinese team also reached nine loops with GPT-6
Interestingly, just a few days after Anthropic finished contacting the questioner von Hippel, he received another piece of big news — from the He Song team at the Institute of Theoretical Physics, Chinese Academy of Sciences.
He Song is a big shot in the domestic field of scattering amplitudes, with recent representative work including the proposal of the Cachazo-He-Yuan formula for quantum field theory scattering amplitudes, the pioneering discovery of the geometric origin of scattering in field theory and string theory, and a series of key works in the directions of high-loop calculations in field theory and rigorous solutions, and the duality of gauge fields, gravity, and string theory.
On September 17th, He Song, Jing Jirong, and Li Xiang publicly uploaded the core skeleton data of six-particle amplitudes from two loops to nine loops to the academic platform Zenodo.
He Song's team also used AI. Dixon revealed that GPT-6 helped the Chinese team calculate part of the constraints, but the overall framework was still built by humans themselves.
von Hippel also confirmed that they indeed used GPT-6 as an aid, but it was absolutely not the pure AI approach of Anthropic.
Dixon joked, "People seem to just like coming to tell me that they've nailed nine loops. Now that's settled, I was first scooped by a machine, and then immediately scooped again by a 'human + machine' team-up."
Not long ago, when OpenAI announced its breakthrough on the Millennium Prize Problem Navier-Stokes equations, and when Claude advanced a key lower bound of the Riemann Hypothesis, some asked: Why does AI only tackle mathematical problems, not physical problems?
Now, here it is.
Although what Claude calculated was just the "toy model" of N=4 super Yang-Mills, and there is a long way to go before explaining dark matter and antimatter in the real universe, its symbolic significance is unprecedented.
It is like when AlphaGo first beat Lee Sedol. People were surprised that a machine could actually master a domain belonging to human intuition.
Professor Dixon left a thought-provoking statement:
Large language models can execute the complex recipes I set and organize calculations, which is a huge victory.
But when one day, large language models begin to propose new physical principles and insights before humans, that will be the moment that truly touches the human soul.
Perhaps, low-hanging fruits are now everywhere.
Reference materials:
















