Anthropic's Claude computes nine-loop scattering amplitude, breaking three-year record
Anthropic announced that its AI model Claude, running nearly unsupervised on the Claude Science platform for several days, computed a nine-loop six-particle scattering amplitude in planar N=4 super-Yang-Mills theory, surpassing the previous eight-loop record set by Lance Dixon and collaborators in 2023. The result was independently verified by Dixon, who praised Claude's execution of the complex calculation. The computation cost roughly $1,000-2,000 for a regular user, with only about $100 for the numerical part, and was initiated by a public challenge from physicist Matt von Hippel.
Coverage timeline
机器之心机器之心
编辑｜Panda 理论物理中一项保持了三年的计算纪录，被 AI 刷新了。 几个小时前，Anthropic 宣布：Claude 在科研平台 Claude Science 中基本 无人监督 地连续运行数天，算出了平面 N=4 超对称杨-米尔斯理论（planar N=4 super-Yang-Mills）中的 九圈 六粒子散射振幅。 这是理论物理「 散射振幅 」领域公认的前沿难题，此前在该模型中的最高纪录停留在八圈，由 SLAC 国家加速器实验室的 Lance Dixon 与合作者于 2023 年创下。 这次突破由八圈纪录的创造者亲自把关 。Dixon 独立验证了 Claude 的结果，并评价称，一个大语言模型能执行完这套复杂配方中的每一步、并把算力组织起来，在他看来「 是一场相当了不起的胜利 」。他甚至直言，除了他的合作者之外，Claude 比任何人都更懂他们团队 2019 年和 2023 年的那两篇论文。 更令人意外的是过程和成本。研究者交给 Claude 的只有一句任务描述，此后的「指导」几乎只剩下「 继续 」； 整个计算折合到普通用户身上约一两千美元 ，其中真正用于数值计算的部分只有约 100 美元，相当于 96 个 CPU 跑一周。而 Dixon 原本认为，直接计算九圈振幅太难，他的团队为此已经筹备了好几年。 这项突破的起点，是一封公开的「战书」。 8 月 7 日，前理论物理学家、科普博主 Matt von Hippel 在自己的博客 4gravitons 上发了一篇带点挑衅意味的文章，标题叫「只有 AI 进入我的领域才算数」。他在文中点名向 AI 公司下战书： 用一名学者能负担得起的算力，去解决散射振幅领域的一个悬而未决的大问题 。 https://4gravitons.com/2026/08/07/it-only-counts-when-ai-gets-to-my-field/ 他给出了两个选项：要么把 N=8 超引力算到七圈，要么把 N=4 超对称杨-米尔斯理论的六粒子振幅算到九圈。 一个月后，AI 完成调整。Anthropic 刚刚刊出的，正是 von Hippel 本人撰写的客座文章，标题就是：「 是的，Claude 能算九圈 」。 https://www.anthropic.com/research/yes-claude-can-do-nine-lo
Related stories
Anthropic: Claude Traces NIST ECDSA Curve Seeds to PRNG, Easing Backdoor Fears
Anthropic reported that its Claude research model reverse-engineered the generation of five NIST elliptic curve seeds in under half an hour, without prior hints. The seeds were traced to the lrand48 pseudorandom number generator, seeded by a Unix timestamp, challenging decades-old suspicions of an NSA backdoor. The finding was independently verified by the bounty sponsor and lowers concerns about hidden weaknesses in the curves.
4 reports · 4 sources1058 points · 388 comments
Yau Shing-Tung's paper on exotic spheres acknowledges GPT 6 Astra and Claude Pro
Mathematician Yau Shing-Tung, who once said AI could not affect top mathematicians, has co-authored a paper that thanks GPT 6 Astra and Claude Pro for helping explore proof ideas and calculations. The paper addresses a conjecture about seven-dimensional exotic spheres, listed by Yau in 1982, and provides an answer to whether these spheres can have positive sectional curvature everywhere.
Anthropic: Claude formalizes Fermat's Last Theorem proof in Lean over 11 days
Anthropic announced that its AI model Claude worked largely autonomously over 11 days to produce the first complete computer-checked proof of Fermat's Last Theorem, formalized in the Lean programming language. The effort involved about 13 million lines of Lean code and over 30,000 intermediate theorems, with the final proof using roughly 29,500 of them. The project, which did not introduce a new proof but translated the existing human-readable proof into machine-checkable form, is more than five times the size of Lean's core math library Mathlib.
4 reports · 4 sources766 points · 509 comments
Claude optimizes 30+ biomolecular models, ~4x speedup; ScienceIDE open-sourced
Anthropic reported that Claude optimized over 30 open-source biomolecular models in under four weeks, achieving an average speedup of about 4x with minor numerical precision changes and nearly 2x with identical outputs, using custom GPU kernels such as FlashPairformer. The code was open-sourced. Separately, AItonomy Foundation released ScienceIDE, an open-source infrastructure that converts scientific codebases into agent-learnable environments, sponsored by PhAI Labs and Alibaba Qwen.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe