← Back to feed

In August, physicist and science writer Matt von Hippel issued an unusual challenge to AI labs: compute a nine-loop scattering amplitude in N=4 super Yang-Mills theory, or a seven-loop result in N=8 supergravity. The problems sit at the frontier of theoretical particle physics not because researchers lack methods, but because the computation is so expensive that no academic group could afford to run it to completion.

Von Hippel wanted to see whether AI could do what human researchers could not — use the same academic-level compute budget, apply a smarter approach, and push past a limit everyone assumed was real.

Anthropic published a guest post by von Hippel on Friday confirming the challenge was beaten. Claude computed the nine-loop amplitude.

The problem belongs to a subfield called amplitudeology. Scattering amplitudes describe the probability of particle interactions; computing them at higher loop orders adds precision that can resolve longstanding theoretical questions in quantum field theory. Each additional loop dramatically multiplies the algebraic and numerical workload. Nine loops is a number that the community had effectively treated as unreachable.

Von Hippel is explicit that he was not testing creativity or insight. Those are harder to quantify. He was testing whether an AI with ordinary access — academic, not hyperscaler — could outcompute what the field assumed was computationally gated. The result suggests it could.

Anthropic has not published the underlying technical report or detailed methodology in the guest post, and the amplitude result has not yet been formally peer reviewed. The post is one researcher’s account of issuing a challenge and seeing it met. What it does establish is that Claude, operating on available compute, solved a problem in theoretical physics that the community had not solved — and did so faster than von Hippel expected.

That’s a narrower claim than AGI, but a more verifiable one.