A physics harness ran unattended for a week, and the result held up

Physicist Matt von Hippel publicly challenged AI labs to compute a scattering amplitude to nine loops in a toy theory called N=4 super Yang-Mills, a calculation no human group had finished. Anthropic researchers Liam Fitzpatrick and Siddharth Mishra-Sharma took up the challenge inside Claude Science, a research harness built on Fable 5.1, giving it one prompt describing the problem and then instructions to keep going without them: "Keep working on this until I tell you to stop. Give me updates every 4-6 hours." Claude solved the problem two independent ways. The primary method, a computational bootstrap written in Python with SymPy, ran on 96 CPUs for about a week and cost roughly $100, with the full effort estimated at $1,000 to $2,000. Lance Dixon, the Stanford and SLAC physicist who validated the result, described the setup as one where "if you make any mistake at all in the computational recipe, it all crashes down like a failed soufflé." A separate group led by Song He reached a related result around the same time with more direct human involvement, using GPT-6 for part of the work.

Why it matters: This is a dated, externally checked example of a harness running unsupervised for roughly a week on a task fragile enough that one mistake would sink it, at a total cost of one to two thousand dollars. Anyone deciding how much oversight a long autonomous run needs now has a concrete reference point, even though the domain here is physics rather than software.