Mills 九圈散射振幅计算，物理学家 Matt von Hippel 撰文回顾挑战过程
Key Highlights
Physicist Matt von Hippel recounted a result that caught the research world's attention: Anthropic's Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude Science, built on Fable 5.1, on a budget of roughly one to two thousand dollars, to compute the six-particle nine-loop amplitude of planar N=4 super-Yang-Mills theory. In plain terms, AI pulled off a piece ofextremely difficult theoretical physics that used to demand huge money and years of human effort.
What Happened
A "nine-loop amplitude" is an extraordinarily hard object in quantum field theory, traditionally requiring large collaborations, supercomputing, and years of accumulated technique. Advancing it with an AI science agent inside a few-thousand-dollar budget hints that the cost structure of frontier research may be rewritten. Von Hippel's recap focuses less on perfection and more on whether the method can be reproduced and checked.
Technical Details
N=4 super-Yang-Mills is among the most computable quantum field theories and a common testbed for new scattering-amplitude techniques. Nine loops means pushing perturbation theory to an extreme tier where algebra and integration grow explosively. AI's value here is partly automating the strategic intuition and symbolic manipulation that human experts normally chain together by hand.
Comparison with Alternatives
Versus pure hand derivation or pure symbolic software such as Mathematica with specialized packages, an AI science agent's edge is that itprobes, plans, and borrows from existing literature patterns. But it currently behaves more like an augmented human collaborator than a full replacement, and reliability still depends on domain-expert review.
Industry Impact and Use Cases
For research, this signals "AI research assistants" moving from writing literature reviews to actually tackling hard calculations. For underfunded groups, cheaply reproducing high-end computation is a real benefit. Yet the field must build new verification norms: how is an AI-derived amplitude independently confirmed?
What to Watch
Watch whether other labs replicate the result with different systems, and whether a verification standard emerges for AI-generated physics. Reproducibility, not novelty, will decide if this becomes routine.
Bottom Line
The headline is not "AI solved physics" but "AI made a forbidding calculation affordable." That reframes who gets to do frontier science, not just how fast it goes.
One More Angle
Von Hippel's own recount matters because he is a skeptic insider, not a vendor. When a careful physicist validates the method's plausibility, the claim gains credibility that a company blog post alone would not.
Looking Forward
Expect more "nine-loop-class" results attempted by agents, and a parallel race to build the tooling that lets humans verify them quickly enough to trust them in print.
The Road Ahead
Expect more nine-loop-class results attempted by agents, plus a parallel race to build tooling that lets humans verify them fast enough to trust in print.
Practical Takeaway
For research groups, treat AI science agents as force multipliers on the algebra, not as proof engines; keep the expert review loop firmly intact.
The Stakes
The headline is not "AI solved physics" but "AI made a forbidding calculation affordable." That reframes who gets to do frontier science, lowering the cost barrier that used to reserve such work for richly funded groups.
A Closer Look
Von Hippel's own recount matters because he is a skeptic insider, not a vendor. When a careful physicist validates the method's plausibility, the claim gains a credibility that a company blog post alone would never carry.
Final Note
Reproducibility, not novelty, will decide whether this becomes routine. If other labs repeat the nine-loop result on different systems, a new normal arrives; if not, it stays a tantalizing one-off.
Why It Matters
For underfunded research groups, cheaply reproducing high-end calculation is a genuine equalizer, letting smaller labs attempt problems once reserved for well-resourced centers with supercomputing access.
One More Angle
The method's reliance on expert review is a feature, not a bug. AI did the tedious algebra; humans supplied the judgment. That division, not full automation, is what made the result trustworthy enough to recount.
Practical Note
Research groups should pilot AI science agents on the algebraic grind, not on the final claim. Keep a human in the verification loop, and treat the agent's output as a candidate to be confirmed, never as a conclusion.
Takeaway
The takeaway for physics-adjacent AI is that domain specialists can now pressure-test models against textbook results; when a model reproduces a known amplitude, that is a stronger signal than a benchmark score, because it tests reasoning rather than pattern matching.