Parisi and Zamponi prove a+b=1 after 40 Claude prompts for jamming math

Giorgio Parisi and Francesco Zamponi did not just get a better answer to a decades-old jamming puzzle. They used Anthropic's Claude to mathematically prove the relation they had long suspected from their earlier work: a+b=1. And they needed 40 Claude prompts to get to a verified publishable analytical solution, described July 1 in the Journal of Statistical Mechanics: Theory and Experiment.

Here is the actual stakes moment: a and b are not random parameters. In jamming physics, they “dictate exactly how the distribution of contact forces and small gaps scales” as a physical system hits its critical jamming point. Parisi and Zamponi had seen in 2014 numerical solutions that a and b always add up to 1, but they “had never been able to mathematically prove the relation a+b=1,” Zamponi told Live Science. After a decade with no progress, Claude produced a LaTeX-based proof strategy that Zamponi could verify, even though the team had to revise for initial errors.

To understand why executives should care, you have to understand why this problem was sticky in the first place. Jamming is the sudden transition from a fluid-like state to a rigid-but-disordered one. A simple mental model is a pool table covered with billiard balls: as you keep adding balls, you eventually reach a point where there is no room left for another ball, and every ball is locked in place by its neighbors. That final configuration is disordered and completely frozen, a jammed state. In the research, the authors (Parisi, the 2021 Nobel Prize winner in physics, and Zamponi at Sapienza University of Rome) had mathematically described jamming and offered numerical solutions in a 2014 paper. During that work, they noticed the mysterious “a and b would mysteriously always add up to 1.”

The frustrating part is that multiple groups were seeing the same pattern but not the same bridge between approaches. Separate work by Matthieu Wyart at the Swiss Federal Technology Institute (EPFL) used a completely different approach but yielded the same relation, a+b=1. That overlap suggested something big was missing. For Zamponi and collaborators, it pointed to the need for “entirely new physical concepts” that could link their work with Wyart's and simultaneously explain why a+b=1. In other words, it was not just a gap in calculation. It was a gap in explanation.

So Parisi looked for an external “path forward” after concluding that their domain knowledge had hit a wall. He turned to Claude, Anthropic’s generative AI, and used it first in a pragmatic way: to reproduce the 2014 numerical result. Once Claude could do that, Parisi prompted it to prove why a+b=1. Zamponi described what happened next with the kind of specificity that makes this story feel less like a technology demo and more like research execution: when Claude output arrived while Parisi was traveling, Zamponi reviewed it “on an airplane.” As he read the LaTeX file Claude generated, it was “immediately clear that the core idea was correct,” though the initial output contained errors that required revision.

The key detail for leaders is that they are not claiming Claude invented “new physics.” Instead, the paper’s payoff was sharper: the solution was “hidden directly within the equations themselves.” The researchers “didn't need any external physical assumptions or deep connections between functions.” They got a verified analytical result by prompting Claude 40 times to reach publishable math. That matters because it frames the contribution as a speed-up of proof engineering and formal reasoning, not a free-for-all where AI hallucinations get a Nobel-themed paint job.

This is also where the second-order implications show up for boards, investors, and operators. In the story, Zamponi explicitly hedges the mechanism: it is possible that Claude “simply trawled the vast mathematical literature and used pattern matching,” or it may have applied something like creativity. For him, the practical difference is “moot,” because what counted was that the AI provided a workable route. He put it bluntly: “we could not see the path forward, and Claude did.”

Then he adds a nuance that should land with anyone building AI workflows: human guidance remains indispensable, at least for conceptual direction. He said interacting with AI forces him to reconsider definitions of reasoning, intuition, and creativity. Yet he intends to collaborate with the technology to accelerate “mundane tasks” and generate fresh perspectives on hard problems. And he is already stress-testing it on another math front: a problem involving “random sequential addition of hard hyperspheres.” There, Claude accelerates “writing and optimizing code,” but Zamponi says he has had to provide “the vast majority of the conceptual ideas.” That is a critical operational lesson. Even when the AI can produce the formal machinery, teams still need domain instincts to decide what is worth proving in the first place.

For peers in theoretical research and for decision-makers in adjacent fields, the July 1 result is a proof that a generative model can be more than a draft partner. Under the right conditions, with iteration and verification, it can help produce an analytical solution that survives publication. If AI can reliably move you from “we have a pattern and can’t prove it” to “here is a validated proof,” then the competitive edge is not just better models. It is tighter research pipelines, faster formal verification loops, and teams that know when to ask the machine to do the hard symbolic work.