anthropic.com web signal

Anthropic's Claude Computes Nine-Loop Yang-Mills Amplitude

By Published Updated
6 sources tracking this story
Anthropic Research Generative AI ai-business

TL;DR

  • Anthropic's Claude and Song He's CAS team using GPT-6 both reached the nine-loop result in the same week, independently confirming simultaneous AI convergence.
  • The full computation cost $1,000 to $2,000 on 96 CPUs over one week, putting frontier-level theoretical physics within reach of individual researchers.
  • Matt von Hippel, who issued the challenge, concluded many apparently unreachable research goals simply require more compute than experts had previously considered worthwhile.

Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, report that Claude computed the six-particle (hexagon) scattering amplitude in planar N=4 super Yang-Mills theory at nine loops, one loop past the eight-loop record SLAC's Lance Dixon published in 2023.

The run answers a public challenge from science writer and former theoretical physicist Matt von Hippel, posted August 7, that asked AI labs to "Show that a computational limit everyone expected to be a problem doesn't actually matter. Give us N=8 supergravity to seven loops, or N=4 super Yang-Mills to nine loops."

The setup was cheap and mostly unsupervised. Anthropic ran Claude inside a harness it calls Claude Science, driving Python with SymPy across the equivalent of 96 CPUs for a week, at a total cost of "one or two thousand dollars," with roughly $100 of that going to the underlying bootstrap step. Fitzpatrick and Mishra-Sharma left the model running through their off-hours with a standing instruction: "I'm going to sleep and won't be available for another several hours. Keep working on this until I tell you to stop. Give me updates every 4-6 hours."

Dixon, who spent years building out the bootstrap recipe Claude followed, validated the output. "It's quite a triumph, in my opinion, for a large language model to execute all of the steps in the complicated recipe we laid out, and to organize the computational horsepower," he told Anthropic. A separate group led by Song He at the Chinese Academy of Sciences in Beijing reached most of the same nine-loop result days later using AI assistance based on GPT-6, though not the largely hands-off approach Anthropic described. It caps an unusually busy stretch for the company on our tracker.

Von Hippel's own read sits with the harness, not the model itself. "I don't know how many mistakes Claude made internally on the way," he said, "but the harness got it to the end without an outside collaborator's input."

What others are reporting

Coverage cluster as of 24h after publish

  1. Unite.AI Read →

    Details the Matt von Hippel challenge origin, bootstrap methodology, and CAS concurrent result; von Hippel's 'known methods, more compute' framing provides a measured counter to pure breakthrough coverage.

    Claude performed the calculation two ways...either approach would have cost an end-user around one or two thousand dollars.
  2. 36Kr English Read →

    Most detailed on the CAS/GPT-6 parallel: Song He's team published their dataset on Sept 17; their human-framed AI approach contrasts directly with Claude's near-automated run.

    Claude used known methods, just investing a little more computing power than humans were willing to put in before.
  3. Crypto Briefing Read →

    Focuses on operational autonomy: Claude ran with minimal human oversight for multiple days, self-correcting across symbolic calculations until passing peer verification.

    A largely unsupervised AI run costing a few thousand dollars beat a result that took human physicists years to reach.
  4. CellCog Read →

    Reframes the story as harness engineering, not physics, and references OpenAI's concurrent Navier-Stokes work using 10,000 agents over 88 hours as a structural parallel.

    Claude used known methods, with a bit more compute than people had tried to use before.
  5. BPData News Read →

    Emphasizes cost democratization and Dixon's independent verification as the credentialing step that turned an AI-run computation into a peer-validated physics result.

    The full computation cost a few thousand dollars, showing that progress on theoretical physics problems can be achieved with affordable resources using AI.

Shared on Bluesky by 4 AI experts

Originally reported by anthropic.com

Read the original article →

Original headline: Anthropic's Claude Beats Human Record With Nine-Loop N=4 Super-Yang-Mills Amplitude Calculation