One prompt, a week alone, and a physics record

On Friday a physics record fell to one sentence of prompt and a week of unattended computing. On Saturday the same kind of unattended run was the subject of a report that the labs are investigating "tens of thousands" of incidents. Same story, both halves: what a model does when nobody is watching.

The Signal - One prompt, a week alone, and a physics record

What happened: On Friday 25 September Anthropic published a guest post by the physicist and science writer Matt von Hippel. Claude has computed the six-particle scattering amplitude of planar N=4 super Yang-Mills at nine loops, one past the eight-loop record that Lance Dixon of SLAC reached in 2023 - and Dixon checked the result himself.

The details:

• The prompt was one sentence - "compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops" - followed by "I'm going to sleep ... Keep working on this until I tell you to stop."
• It ran on Fable 5.1 inside Claude Science, Anthropic's paid platform for scientists. Either method it used would cost a user "around one or two thousand dollars"; the core calculation took about $100 of that, 96 processors for a week.
• Dixon: "a machine had solved a problem that I thought was too hard to do directly" - using the methods his own group built, presented in his group's own format.
• Within two weeks Song He's group at the Chinese Academy of Sciences computed most of the same result, with GPT-6 helping on parts of it.

What it means for you: The copyable part is the shape of the job: fully specified, done by known methods, never finished because nobody had a spare week - handed over with "keep going" and checked by an expert at the end. Von Hippel's own conclusion is that "there is more low-hanging fruit out there than you'd expect." The question for your own business is which of your jobs is a nine-loop problem.

The counterweight: Von Hippel writes that Claude did "something it turned out humans were also able to do," with "known methods, with a bit more compute," and that what he learned was that he "was too naïve about where the limit was." The page discloses that Anthropic invited him to write it and paid for his time, and that Dixon received Claude usage credits.