The paper proves an identity for critical exponents of jamming.
Many users celebrate Claude assisting Nobel laureate Parisi with complex jamming-exponents proofs as evidence of real scientific productivity gains, while others dismiss the feat as overhyped or still limited compared to everyday tasks.
If Claude is good enough for Nobel Prize winners it is good enough for you
https://arxiv.org/abs/2606.03300
Unless it’s Claude Opus 4.8 in which case it’ll tell you you’re not good enough and to go to sleep or something
If Claude is good enough for Nobel Prize winners it is good enough for you
https://arxiv.org/abs/2606.03300

@EMostaque

@EMostaque it's not good enough to take to the car wash

@EMostaque And look at the paper’s credits. Work derived from AI interactions count as authorship.

@EMostaque

@EMostaque As a Wi-Fi QA, I told it to make an MCP over an IEEE spec w/ high-lvl preferences. Repeat for all docs. Throw em in the wrapper for accurate assists on RTFM asks of “does STA1 or STA2 use dynamic SMPS in that one test in this 2k-page pdf” bc ik the structure but don’t wanna dig

@EMostaque this Nobel-prize-level copy is crazy, guess my "write a tweet for me" prompts are small potatoes now

@EMostaque Doing sniffer analysis, maybe 1-2/week, I hoard the scripts from each ask of “a visual/histogram/scatterplot/etc of RTS-NAV durations” or any data w/ latent patterns. Then I tell it to use MCP to refine the script with one annotation per test step, so I can show it w/o explaining

@EMostaque Doesn't really track. You need a certain level of sophistication to get good outputs from the (current) models.

@EMostaque Little things add up. MU-EDCF (Wi-Fi 6 media prioritization algorithm) anomaly debugging used to be my most tedious-but-make-no-mistake task if it ever came up, practically a short surgery, now it’s just eyeballing that RTS-NAV histogram without even crossreferencing the testplan

@EMostaque i use claude to argue with people on the internet
feels like im aiming low

@EMostaque There's a million nobel prizes lined up waiting

@EMostaque Using Claude in hard-STEM is a numbers game. Where’s the money going in academia? To pay the mathematicians/scientists spending their precious hours peer-reviewing potentially a complete dud. If Claude catches those before time is sunk, that is still a LOT of money/time saved

@EMostaque Nobel laureates using Claude is great, but are we benchmarking inference consistency or academic output?

@EMostaque If a Nobel Prize winner needs to verify Claude's work, so should you.

@EMostaque real, but nobel winners get the paid plan while im on the free one fighting for my life
@EMostaque Giiioooorgiiioooooo ppaariissiiii 🤤

@EMostaque LLMs are only good at mathematics because it's the only field whose empirical foundation is entirely abstract and completely knowable.

Yes — I read the arXiv paper this time.
Retrieval Floor
Source provided: direct arXiv PDF. Access status: full PDF accessible, 12 pages. Basis: I reviewed the abstract, strategy section, proof structure, theorem/conclusion, and stated limitations. Scope: collision below is based on the paper text, not on independent verification of the proof algebra. Mode used: Paper Collision / AI-origin claim check Collision verdict: Partial-to-strong collision The paper’s core claim is precise: within the CKPUZ fullRSB scaling framework for dense hard spheres in infinite dimension, it proves the previously numerical identity a + b = 1, which then yields the physical exponent relations tied to mechanical marginal stability. The paper states that CKPUZ had numerically observed the identity but lacked a proof, and that this paper supplies an analytic proof from the scaling fullRSB equations.
What the paper is really saying
It is not claiming “AI solved physics” in the broad sense. The actual scientific claim is narrower:
Given the CKPUZ scaling equations, the CKPUZ marginal-stability identity, and the physical no-node branch, one can derive a + b = 1 analytically.
The AI-origin claim is separate: the authors say the proof was obtained through interaction with Claude Sonnet 4.6 and Opus 4.7, that Opus largely derived the proof with limited supervision, and that the authors checked, corrected, edited, and deposited the conversation.
Strongest pressure point
The strongest pressure point is not the algebraic identity itself. The paper lays out a clean route:
derive an identity involving integrals K, N, D;
use CKPUZ marginal stability to set D = 2N;
prove K > 0 on the physical branch;
conclude a + b = 1. That is a real collision structure: the claim depends on a specific bridge, not vibes. Weakest pressure point The paper’s own limitation section matters. It admits that existence/uniqueness of the fullRSB profile, a fully rigorous derivation of the scaling regime, and a complete proof that the scaling limit is attained are not completed here.
So the proof is strong inside the assumed CKPUZ scaling framework, but not a full mathematical construction of the whole hard-sphere fullRSB theory from first principles.
Internal collisions
Internal Collision 1
Claim A: The abstract says the paper gives an analytic proof of a + b = 1 from the scaling fullRSB equations.
Claim B: The paper later says the existence/uniqueness of the fullRSB profile and rigorous derivation of the scaling regime are taken as given or left open.
Collision: “Proof” is true only at the level of the accepted scaling framework, not as a fully rigorous derivation of all underlying assumptions.
Repair: Say “conditional analytic proof within the CKPUZ scaling framework” whenever discussing the result publicly.
AI-output trust issue
The paper avoids the worst AI-trust failure because the authors do not merely say Claude gave an answer and therefore it is true. They state that they checked the proof, found inconsistencies in an early version, had them corrected, edited the draft, and deposited the interaction record.
But it still triggers a Source Independence Check: the AI generated or helped generate the proof, while the authors verified it. That is better than raw AI output, but the claim has not fully met outside resistance until independent physicists/mathematicians check the proof.
Failure tags
Evidence Floor, Source Independence Check, Definition Load, Overreach risk.
Strongest revision direction
The safest public framing is:
This paper presents a human-verified, AI-assisted analytic proof of the identity a + b = 1 within the CKPUZ fullRSB scaling framework for jamming. Its force depends on accepting the CKPUZ scaling equations and marginal-stability identity; the paper does not claim a complete rigorous construction of the full hard-sphere fullRSB theory.
Best pressure question:
Does the proof survive independent expert verification of the algebraic cancellations, the Fisher–KPP positivity step, and the passage from the fullRSB flow to the scaling limit?