Matt von Hippel argues in Anthropic that Claude’s calculation changed where he sees the limit

Matt von Hippel argues in Anthropic that Claude’s calculation changed where he sees the limit

Originally published September 25, 2026. Matt von Hippel, writing for Anthropic, describes how Anthropic physicists used Claude to calculate a nine-loop amplitude in the toy model N=4 super Yang-Mills. The result was independently checked by Lance Dixon, and von Hippel says Claude used known methods with more computing power than people had tried before—not a super-intelligent new approach. His argument is that the apparent computational barrier was lower than he had assumed: the calculation was within reach on a reasonable budget, and Claude carried it through without outside scientific input. Dixon adds that the calculation used methods developed by his collaborators, and says Claude had to build the code for a fragile, complicated recipe. Von Hippel’s takeaway is that experts may overlook achievable progress, and that scientists should consider whether AI can tackle other frontier calculations while planning how to check results. But he says this case does not answer his broader question about AI’s future or superintelligence; he had hoped to see new methods and a glimpse of how far AI could push computational limits. Dixon, meanwhile, says the more searching moment will come if large language models begin producing new physical principles and insights before humans.

Publication: ea3e1e25-82d9-70d9-45dd-eef2a61576c9

attribution — pass: The reading correctly attributes the post to Matt von Hippel, writing for Anthropic at Anthropic's invitation. It attributes the addendum views to Lance Dixon. It keeps the claims as the authors' assertions and notes the compensation disclosure.

Anthropic invited Matt von Hippel to write this post and compensated him for his time.

summary — uncertain: The central terms 'nine-loop amplitude', 'loop' and 'N=4 super Yang-Mills' are never explained, and the definitions cover only superintelligence and LLM. A general reader cannot tell what was computed or why nine loops is hard. Repair: add a plain gloss, e.g. scattering amplitudes predict particle reactions and loops measure their complexity. The summary also omits that Song He's group nearly reached the result, which grounds von Hippel's point that humans could do it too.

Instead, it did something it turned out humans were also able to do.

date — pass: The reading's September 25, 2026 date matches the UTC day of originalPublished. The source's timeline (end of August contact, a September 1 notification, two weeks of validation) is consistent with it. There is no hindsight framing, and the kind is publisher-page, so no archive disclosure is needed.

For me, it happened on September 1, when Liam Fitzpatrick and Siddharth Mishra-Sharma at Anthropic told me

support — uncertain: Most claims match the source. However, claim 4 and the summary turn Dixon's confident 'when' into a conditional 'if', which weakens his forecast that LLMs will produce new physical principles. Smallest repair: write 'will come when'. Also, the page was edited after publication (dateModified 2026-10-21), so it is not plainly confirmed that the cited passages, including the disclosure, are original.

But the more soul-searching moments will come when large language models start to come up with new physical principles and insights before humans.

media — pass: No media was selected, and no verified media candidates were supplied, so leaving media empty is appropriate.