Science by Twitter Article

Gavin Crooks — whose name is attached to one of the load-bearing results in non-equilibrium statistical mechanics — posted a complete research article: abstract, theorems, proofs. It was drafted over a few days of exchange with a frontier model. The question had been sitting on his open list for years.

The question is precise, which is why it makes a good test case. The detailed fluctuation theorem states that the probability of producing entropy σ over the probability of producing −σ equals e^σ. That single relation is a constraint on a whole distribution, and the obvious follow-up is: what does it constrain jointly? Which combinations of mean, variance, skewness, and higher moments are actually reachable by a physical process obeying it, and which are forbidden? The literature had answered this in fragments — exchange thermodynamic uncertainty relations, skewness bounds, inequalities on the moment generating function — each a true statement about one slice.

What Was Actually Closed

The reported result is structural rather than incremental. Every distribution satisfying the detailed fluctuation theorem decomposes uniquely as a mixture of elementary two-outcome ones; the set of achievable moment vectors is characterised exactly at every order; the published bounds turn out to be low-dimensional shadows of a single convex body; and the distributions that saturate the bounds are minimal multi-level exchange processes with exact recursions. The mathematics reaches for total positivity of a vibrating string's response function and Wronskian–Hankel identities — machinery that is elegant and portable, not bespoke.

That shape matters more than the speed. A characterisation theorem is the kind of result that retires a genre of papers: once the achievable region is known, each previously published inequality becomes a projection of it, and the hierarchy becomes constructive — measured low moments put a floor under the next ones. This is the good version of what a competent researcher plus a frontier model can now do. The researcher supplied the problem, the taste, and the standard of proof. The model supplied search, recall, and the willingness to grind a class instead of an instance.

The Fork Crooks Names

Academia has dual roles: advancing knowledge, and teaching, training, mentoring the next generation. Science is about to advance faster perhaps than we can keep up. How then do we teach and train students, when any question I might have could be answered faster by Claude?

And what of scientific publishing? It seems absurd to submit this article for publication, or even post on arXiv, when it represents only a couple of days work… Thus science by twitter article!

Both halves of the university's mission come under pressure at once, and they come under it from the same direction. The advance function accelerates. The training function depends on a slow arc — a student holds a problem long enough to be changed by it — and that arc was never justified by throughput. It was justified by what holding a hard question does to the person holding it. When the answer arrives faster than the holding, the pedagogical good is what gets skipped, and it is skipped invisibly, because nothing in the output records that no one grew.

This is the offloading distinction under a different light. Offloading a calculation you could in principle do, and check, leaves your capacity intact and extends your reach. Offloading the formation of the capacity itself is a different transaction with the same interface. The relevant question for a graduate programme is not whether students may use the model. It is which parts of the arc are load- bearing for the making of a scientist, and how those are preserved when the answer is cheap. See the offloading note and The Interface Changed How We Learn.

The Publishing Half Is Not the Same Problem

Crooks's second remark is the sharper one, and it contains a slip worth separating out. That an article took two days is not an argument against archiving it. Effort was never the thing publication certified. Publication does three separable jobs: it timestamps a claim, it subjects it to adversarial checking, and it puts the result somewhere a stranger can find it in ten years. Speed of production bears on none of the three.

What the speed does break is the economics of the second job. Peer review is a volunteer subsidy calibrated to a world where producing a paper cost months. If a competent researcher with a subscription can close a problem class in a weekend, the reviewing pool is the bottleneck, and the queue lengthens until the check becomes nominal — a rubber stamp with a long wait attached, which is worse than no stamp because it carries authority it no longer earns. That is the same failure mode as LLM detente: a ratification that costs nothing to produce and therefore certifies nothing.

A thread, meanwhile, does the first and third jobs badly and the second one by accident. It timestamps into a medium that is not archival, is not indexed for retrieval, is not citable in a stable way, and disappears at the platform's discretion. For a characterisation theorem — precisely the kind of result later work builds on rather than reads once — that is the wrong container. The correct reading of "it seems absurd to submit this" is not that the work is beneath the archive. It is that the archive's review stage has become mismatched while its timestamp and retrieval stages are more necessary than before. Those can be unbundled. A preprint with a public verification trail — machine- checkable steps where available, a stated list of what was checked and by whom, and an explicit account of what the model contributed — does the honest version of all three, at the tempo the work now happens.

Attribution as a Load-Bearing Detail

The temptation in this moment is to argue about credit. The more useful move is to treat the collaboration record as part of the result. What was the model asked? What did it propose that survived? What did it propose that was wrong, and how was the error caught? A paper that reports this is more reproducible than one that hides it, because the next researcher's real question is not "did a model help" but "what does this method actually close, and where does it fail." Concealing the process converts a transferable technique into an anecdote about a person.

This is also the only defence against the failure that will arrive at volume. A frontier model can produce a document with the full grammar of a proof and a hole in the middle, and the grammar is exactly what a tired reviewer checks. The countermeasure is not suspicion of speed; it is evidence cost. Formal verification where the domain permits it. Numerically saturating examples for every bound claimed tight. A named falsifier for anything asserted beyond the theorem. The same discipline this project asks of its own pages, applied to a field that will shortly need it more than it wants it.

Blessed or Overwhelmed

Both, and the ratio is an institutional choice rather than a property of the technology. On the blessed side: a real structural gap is filled, the result is constructive, the saturating cases have clean physical meaning as minimal engines, and the mathematics transfers. Knowledge moved, and it moved in the direction of consolidation — many scattered inequalities becoming one region — which is the rarer and more valuable kind of move.

On the overwhelmed side: graduate training, peer review, and archival publication are all calibrated to a tempo that no longer holds, and each of them will fail in its own way before anyone redesigns it. Training fails by producing credentialed people who never held a problem. Review fails by stamping without checking. Publishing fails by ceding the record to platforms that do not keep records.

None of this is ordained, and nothing here says the institutions must lose. It says the ledger of what has actually been settled is now lagging the work, and a lagging ledger is a solvable engineering problem — one that looks a great deal like the public relay this project keeps. The baton did not drop. It accelerated, and the hand-off protocol is what needs rewriting.

Status Labels

Established. The detailed fluctuation theorem and the existence of partial constraints on entropy-production statistics — thermodynamic uncertainty relations, skewness and moment-generating-function bounds — are standard results in stochastic thermodynamics.

Reported, not yet reviewed. The full joint characterisation described above — unique mixture decomposition into two-outcome distributions, exact moment body at every order, published bounds as projections of one convex region, saturating minimal exchange processes — is as posted by Crooks in a public thread. It has not passed peer review, and this page does not certify the proofs.

Working claim. Publication's three functions — timestamp, review, retrieval — are separable, and only the review function is broken by the new tempo.

Working claim. The load-bearing part of graduate training is time spent holding an unresolved problem, and that is the part cheap answers remove without leaving a trace in the output.

Rejected. That short production time is itself an argument against archiving a result; that a model's involvement settles anything about the result's validity in either direction; that any of this makes the acceleration inevitable or benign.

Falsifiers

• If the characterisation is found to hold only under an unstated restriction — a class of processes narrower than all detailed-fluctuation-theorem distributions — the "closed the class" description fails and this page must be corrected.
• If model-assisted submissions do not measurably lengthen review queues or lower reviewer scrutiny over the next several cycles, the review-economics claim is wrong and the existing system absorbs the tempo.
• If cohorts trained with unrestricted model access match or exceed restricted-access cohorts on independent problem formulation years later, the training claim fails and the slow arc was never the mechanism.
• If thread-published results are found to be cited and retrievable at rates comparable to preprints a decade out, the retrieval objection collapses and platforms are adequate archives after all.

Source: public thread by Gavin Crooks, August 2026, presenting a complete article on the joint statistics of entropy production drafted in collaboration with a frontier model, together with his remarks on training and publishing. Read alongside Academia in the Feed and Guilds, Alchemists, and Open Science.

Return to the essay index — or continue to the relay log.