The Physicist and the Frustrated Machine

Nobel laureate Giorgio Parisi and Francesco Zamponi used Anthropic's Claude (Sonnet 4.6 and Opus 4.7) to prove a mathematical identity relating critical exponents of the jamming transition, published in the Journal of Statistical Mechanics. The model derived the proof with minimal supervision, though the physicists verified and corrected it over 40 prompts. The result unifies two theories of jamming and demonstrates how large language models can generate novel proofs by composing existing knowledge under human guidance.

In early July 2026, the Journal of Statistical Mechanics published a short paper — posted on arXiv the month before — with an unusual author acknowledgment. Giorgio Parisi — Nobel laureate in Physics, 2021 — and Francesco Zamponi proved an identity, a + b = 1 , relating the critical exponents of the jamming transition. The identity had been observed numerically to arbitrary precision since 2014, in their own work, but never proven. The paper states plainly that the proof “was obtained through interaction with Claude Sonnet 4.6 and Opus 4.7 and verified by us.” The result matters to physicists because it welds together two theories of jamming — the replica approach and Wyart’s marginal-stability arguments — that had arrived at the same numbers by independent roads. But the episode matters to anyone working with large language models for a different reason, and the paper’s own account of it pulls, at first glance, in two directions. On one hand: “The model Opus 4.7 essentially derived the proof by itself, with minimal supervision from us.” On the other, a few sentences later, the same authors: “We checked the proof carefully and pointed out some inconsistencies in an early version, which the model corrected by itself… We eliminated some parts that were, in our opinion, not necessary for the proof and were obscure.” Derived it by itself — yet checked, corrected, pruned by hand. Forty prompts, in a single documented conversation. Autonomy or supervision? The tension is only apparent, and resolving it is the subject of this article — because how those forty prompts were structured turns out to be the interesting part. The right vocabulary for describing it was invented by one of the paper’s authors himself, forty years ago. Start with what Claude did not have: the answer. The proof of a + b = 1 existed nowhere in the training data, because it existed nowhere at all. Yet the model produced it — not flawlessly: the authors report pointing out some inconsistencies in an early version, “which the model corrected by itself,” with Sonnet 4.6 later refining minor steps — but with the core idea intact and reached quickly. Zamponi’s summary is disarming: “The answer was right there, and we simply hadn’t seen it.” How does a system produce something it was never shown? The pieces were all there. The full replica-symmetry-breaking formalism, the scaling equations, the asymptotic techniques — all published literature, all presumably absorbed into the weights. What was missing was not knowledge but composition : which known moves, in which order, assemble into the proof. And some of the constraints that guided the composition were not in any corpus at all: the authors note, in passing, that they had verified “ but never published ” that the identity also holds for a non-positive profile with a single node. Facts they possessed and the weights could not — available to the model only because the humans placed them into the context. Think of the model as an enormous box of puzzle pieces. Generation is the process by which the pieces settle into an assembled picture. When the constraints are sufficient — when the context has pinned down enough of the frame — the assembly is coherent. When they are not, the model interpolates: it fills the gaps with the statistically most plausible piece, producing something locally smooth and globally false. This is one honest definition of hallucination: not a malfunction, but the same operation as generalization, applied where the constraints run out. Under this reading, look at what the paper actually documents about the structure of those forty prompts. The physicists did not open with the question. “In the first part of the conversation, the model had to study the differential equations numerically and produce a C++ code that finds the solution. The aim was to have a high-precision verification of the conjecture. Only at the end did we ask for an analytic proof of the result.” None of that numerical work is new knowledge to the model. It is constraint — edges, not pieces. By the time the real question arrived, the context was saturated with verified numerical facts about exactly the objects the proof would need to manipulate, and the assemblies that contradicted those facts had already been priced out. This was not correction after the fact; it was the frame built first, deliberately, so that essentially one coherent composition remained when the question was finally asked. But the puzzle metaphor, taken alone, predicts too much. If all the pieces are present and the frame is pinned, the picture should assemble cleanly — no hallucinations. That is not quite what happened. The first version contained inconsistencies the authors had to flag; saturating the constraints made the spurious completions rare, not impossible. And the final certification — the careful check, the editing for accuracy — remained external, human. Why? Because there are two distinct ways a puzzle can go wrong, and only one of them is fixed by adding information. The first is missing pieces : the training data simply doesn’t cover the region, and the model interpolates. More information helps here. This is the failure mode RAG and long context are designed to address. The second is subtler: conflicting pieces . Training corpora contradict themselves — source A asserts X, source B asserts not-X. Gradient descent cannot satisfy both. The model converges on a superposition, holding both patterns with relative weights. And this is where the vocabulary turns back on its inventor. In spin-glass physics — the field for which Parisi received the Nobel — the central concept is frustration : constraints that are mutually unsatisfiable. Spin A wants to align with B, B with C, but C wants to oppose A. No configuration satisfies everyone. Parisi’s foundational contribution was showing what frustration does to the energy landscape: it shatters it into an enormous number of nearly degenerate metastable minima, none of which resolves the conflict. Each is merely a different local compromise. Contradictory training data is, precisely, a set of mutually unsatisfiable constraints. The result is a glassy landscape of plausibility: many minima of comparable depth, each corresponding to a different “resolution” of the conflict. Generate without sufficient constraint, and the system falls into one of them more or less at random. Fall into the wrong one — or worse, into a chimera stitched from incompatible resolutions — and you get a hallucination of the second kind. This distinction is operational, not decorative. Interpolation hallucinations respond to added information. Frustration hallucinations do not respond to more information of the same kind — piling on further contradictory sources makes the landscape rougher, not smoother. What they respond to is targeted, disambiguating input: information selected to break the tie, not to enlarge the pile. Much of today’s standard tooling quietly assumes every hallucination is of the first kind — retrieve more documents, extend the context, stuff the prompt — which, applied to a frustration problem, is the wrong cure for the wrong disease: it treats a tie-breaking deficit with volume, and the volume brings its own contradictions in with it. Which is, as we are about to see, precisely the mistake the two physicists did not make. Honesty requires a disclosure here: the marriage between spin glasses and neural networks is not my metaphor. It is the founding literature of the field. Hopfield built his 1982 associative-memory model directly on the physics of disordered systems; Amit, Gutfreund and Sompolinsky analyzed it in 1985 using precisely Parisi’s replica method; and the 2024 Nobel Prize in Physics, awarded to Hopfield and Hinton, certified that genealogy at the highest level. When a physicist reads “LLMs have glassy landscapes,” the correct reaction is: of course they do — that is where they came from. But the classical connection lives almost entirely on one side of the house: training . The loss surface over the weights is the landscape physicists have studied — its ruggedness, its degenerate minima, the surprising benevolence of overparameterization. What the Parisi episode illuminates is a different, less-mapped room: inference . At generation time there is a second landscape — over output configurations , conditioned on the context — and it is a distinct mathematical object. It inherits its ruggedness from the first the frustration frozen into the weights shapes which completions compete , but it is not the first. The weights are fixed; only this second landscape can be deformed at runtime, and the context is the only handle available to deform it. Keeping the two landscapes separate matters, because it locates the claim precisely. I am not saying anything new about training dynamics. I am saying that the conditional landscape of inference behaves, phenomenologically, like a frustrated system under an external field — and that an expert steering a model is doing physics on the second landscape while the first stays frozen. The old marriage, a new room. Here is the part I initially got wrong, and the correction is the heart of the argument. My first instinct was to say the physicists pruned the bad branches. They did not. The weights are untouched; the frustration remains carved into the model permanently. What their context did was act as an external field . In frustrated systems, an applied field breaks the degeneracy: it raises the energy of some minima relative to others, making one state preferred where before there was ambiguity. Read the documented conversation structure in this light and it snaps into focus. The long numerical preamble — the ODEs studied numerically, the C++ code, the high-precision verification of the conjecture — was not a warm-up. It was the field being installed : every verified numerical fact deposited in the context is a bias term that conditionally suppresses, for that conversation, the completions inconsistent with it. The flagged inconsistencies in the early draft were the same mechanism applied reactively — the paper documents one flagged round that the model corrected by itself, followed by refinement of minor steps and a final human edit for accuracy and readability. Close the session, the field switches off, the landscape relaxes back to degenerate. Compensation, not surgery. Rent, not purchase. And here the metaphor can be cashed out line by line, because the paper lets us trace exactly what the numerical preamble made available. The entire proof hinges on one integral, K, being strictly positive — which in turn requires selecting the physical branch of the equations, the one where an auxiliary function f stays between 0 and 1. The paper is explicit that this is “the only tricky point” , and equally explicit about what the numerics had already shown: “Numerically, we find 0 < f t < 1, so there is no problem in concluding that K 0.” In other words, the C++ phase did not merely verify the target conjecture; it deposited into the context the precise property — the positivity and boundedness of f — that the analytic argument a Fisher–KPP identification plus the parabolic maximum principle would later have to establish rigorously. The field was not generic “verified facts.” It was the load-bearing fact, placed in advance at the exact point where the proof would need to lean on it. That is what targeted compensation looks like when you can read the receipt. Two consequences follow, and both are testable against everyday experience with these systems. Compensation decays. If the field is an additive term rather than a structural change, then as the corrective signal dilutes — the conversation grows, the constraint stated two thousand tokens ago loses attentional weight — the spurious minima become competitive again. This is the familiar failure where a model “forgets” a correction and repeats an error already fixed; in this frame it is not a mysterious bug but relaxation toward the uncompensated landscape as the field weakens. No such relaxation is reported in the Parisi–Zamponi paper — and the frame suggests why it would be unlikely here: a field made of verified numerical facts about the specific objects at hand is about the densest, most targeted field one can build, and the question was asked while it was at full strength, not after two hundred digressions. Whether the transcript bears this out in detail is, happily, a checkable question. Compensation requires a map. The right field is not “more information”; it is information targeted at the specific wrong minima . You must know which resolutions are spurious in order to penalize them — or, as here, you must know which verifiable facts will price them out in advance. This is why domain experts succeeded where a longer prompt would not. A generic user can raise the field everywhere — more context, more instructions — but without a map risks tilting the landscape in the wrong direction, or injecting frustration of their own. Domain expertise, in this picture, is literally the map: the decade the authors spent on the problem — including, by their own account, years of unsuccessful attempts before it faded into a drawer — was the surveying that made the field placement possible. You cannot route a system through territory you have not charted. One more thing distinguishes this episode, and it deserves more attention than it has received: the authors deposited the full text of the conversations in a Zenodo repository DOI: 10.5281/zenodo.20478427 , cited as a reference in the paper itself. The forty prompts are not a press-release claim; they are a published behavioral record. Anyone — including anyone skeptical of the framing I’ve offered here — can read exactly what was injected, where the model’s early draft went wrong, and how the correction propagated. Documented AI-assisted discoveries will be argued about for years; this one, unusually, can be argued about from the transcript . That standard — claims about model behavior licensed by logs, not by recollection — is one the rest of us writing about these systems should be held to as well. One honest caveat about the result itself, since precision cuts both ways: the paper is explicit that the existence and uniqueness of the fullRSB profile, the convergence of the scaling expansion, and the rigorous existence of the matching region are taken as given — well established numerically, but a fully rigorous construction would require, in the authors’ own words, extending the work of Talagrand and Panchenko on the Sherrington–Kirkpatrick model to the hard-sphere setting. This is a physics-grade proof conditional on the CKPUZ framework, not a theorem in the strict mathematical sense, and the authors say so plainly. It does not diminish the episode; it locates it. Before the closing image, the strongest objection deserves the floor: this is a metaphor, and metaphors from physics have a long history of flattering AI systems they do not describe. An LLM’s inference is not literally a spin glass minimizing a Hamiltonian. To be precise about what carries over and what does not: sampling temperature is, remarkably, one of the exact correspondences — it enters the softmax precisely as an inverse temperature enters a Boltzmann distribution, and setting it to zero is a quench, the degenerate limit any physical annealer also possesses. What does not carry over is anything quantum: the probability is classical, there is no interference, no non-commuting observables — so “collapse” language, if used at all, must mean concentration of a classical distribution, nothing more. And unlike the training-side connection, which is theorem-grade physics, the inference-side picture I’ve drawn has no replica calculation behind it — it is a structural analogy, not a derivation. What I am claiming is correspondingly narrow: the structure of the failure modes matches. Constraint satisfaction over conflicting evidence generically produces rugged landscapes with degenerate compromises — true of spin glasses, of SAT problems near the satisfiability threshold, and plausibly of a conditional distribution shaped by a contradictory corpus. The value of the frame is not that it is literally true but that it earns its keep three times: it separates two failure modes that are usually conflated, it predicts the decay of corrections over long contexts, and it locates exactly what the human contributes — the field design and the final verification. Where it breaks — and it does break, because latent space lacks the univocal geometry of real puzzle pieces — the break itself is informative: it tells you the final certification can never be fully internalized. There is a passage in the paper’s final section that has gone almost entirely unquoted in the coverage, and it is the key that turns the whole story symmetric. The authors ask themselves the obvious question and answer it with startling candor: “Why have we not seen the proof? Difficult to say. We have not even tried to use this approach. We thought there was a deep, hidden, direct relationship between the functions p1 t and J t that we were unable to see. We were looking for something deeper, and we neglected the conceptually simple case hard to see due to the many algebraic cancellations .” Read that in the vocabulary this article has been building. The experts had a frustrated landscape of their own. The decade-long conviction that the answer must be deep was a self-applied field — one that raised the energy of the simple basin and kept two of the best minds in the discipline circling the wrong region of configuration space for ten years. The model, arriving without that prior, descended into the conceptually simple minimum they had priced out of their own search. So the compensation ran in both directions: the physicists compensated the model’s frustration with verified numerical constraints, and the model compensated the physicists’ frustration with the absence of the prejudice that it had to be profound. Neither system could have reached the answer alone — and the trap was not the same trap. The physicists were caught by a misplaced field, a bias pointing at depth; the model, left to itself, would have been caught by the absence of one, adrift among degenerate completions with nothing to break the ties. Too much bias on one side, too little on the other: the collaboration worked precisely because the pathologies were opposite, each the cure for the other. And this is where annealing, banished earlier for imprecision, legitimately returns. The ten years of failed human attempts were the high-temperature phase: broad, expensive exploration of the deep basins, none of them the right one — but exploration that surveyed the landscape and produced, as a byproduct, the map. The forty-prompt conversation was the quench: field installed, temperature dropped, the system settling in hours into the minimum the survey had finally made findable. The decade was not wasted; it was the annealing schedule, run on the humans, that made the final quench possible. The man who invented the formalism for rugged landscapes spent a decade lost in one, prepared a field, and watched a machine fall straight to the bottom. It is tempting to say Claude helped Parisi. It is more accurate, and stranger, to say that two differently frustrated systems compensated each other’s fields — and that reading which is which, in the transcript now sitting on Zenodo, is left as an exercise. The proof exists. Both landscapes are still rugged. The next question is who de-biases whom. The paper: Parisi & Zamponi, “A proof of an identity for the critical exponents of jamming,” J. Stat. Mech. 2026 073301, DOI: 10.1088/1742–5468/ae7bd7 arXiv:2606.03300, posted June 2, 2026 ; the framework it completes: Charbonneau, Kurchan, Parisi, Urbani, Zamponi, J. Stat. Mech. P10009 2014 . The full conversation transcripts are deposited by the authors at Zenodo, DOI: 10.5281/zenodo.20478427 reference 14 of the paper . All quotes attributed to the paper are verbatim from Sections 1.2, 1.3 and 7 of the arXiv version. Zamponi’s “The answer was right there, and we simply hadn’t seen it” is from the Sapienza University press release, as carried by Phys.org. On the spin-glass/neural-network lineage: Hopfield, PNAS 79, 2554 1982 ; Amit, Gutfreund & Sompolinsky, Phys. Rev. A 32, 1007 1985 ; Nobel Prize in Physics 2024. A note on method: this essay grew out of an extended dialogue with Claude Anthropic . The central ideas — reading context as an observer, the two classes of hallucination, the shift from “pruning” to an external field — are the author’s, developed in conversation; Claude supplied the physical formalism the spin-glass vocabulary, the frustration mapping , structured the argument, and drafted and revised the prose across several passes. All factual claims were checked against the primary sources cited above. The Physicist and the Frustrated Machine https://pub.towardsai.net/the-physicist-and-the-frustrated-machine-2f2fc952328c was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.