The letter to Tait
In 1867 James Clerk Maxwell wrote to his friend Peter Tait describing a thought experiment meant to needle the second law of thermodynamics. A vessel holds gas at uniform temperature. A tiny being sits at a shutter in the dividing wall, able to see each molecule as it approaches. Fast molecules it lets through to the right, slow ones to the left. No obvious work is done — the shutter is frictionless, the being does no lifting — and yet a temperature difference appears where there was none. Order from disorder, for free. Maxwell's point was not that this happens. It was that the second law is statistical, true of populations of molecules and not of any single one, and that a being with the right kind of sight could exploit the gap between the two.
The puzzle sat unresolved for sixty years because nobody had priced the looking. Leo Szilard, in 1929, built a version of the demon around a single molecule and showed that extracting one bit of information about its position was worth exactly kT ln 2 of work. Rolf Landauer, in 1961, found the other half of the ledger: it is not measurement that costs, it is erasure. Wiping a bit of memory to make room for the next one always dissipates at least kT ln 2. Charles Bennett closed the loop in 1982 — a demon with finite memory must eventually forget what it has seen, and that forgetting repays the entropy debt precisely. The demon was never exorcised. It was put on a budget and allowed to keep working.
The demon at the shutter
Strip away the gas and the vessel and what remains is a claim about permission. The demon's power comes entirely from being allowed to look at the molecule now, at the shutter, before it decides. Take that permission away — let it glance once at the whole chamber and then close its eyes — and it can no longer sort individuals. It can only sort by the average behaviour it once observed. That is a strictly weaker demon, and it is the demon a performance analyst becomes the moment the season's tracking feed is replaced by last year's report.
Sports analytics runs on four streams that never stop: player tracking data, arriving at 10 to 25 hertz per athlete depending on the system; injury and fitness reports, updated daily through the physio room; transfer and roster activity, which resets personnel with no warning; and opponent tendencies, the patterns a rival has built its identity on and is, at every moment, capable of abandoning. An analyst's job is Maxwell's job with a different vessel: watch the molecules at the shutter, sort what matters from what doesn't, and produce a gradient — a game plan — that the team can act on before the next match.
The tendency that already died
The characteristic failure in this domain is precise enough to name without embellishment. A team spends a week building its defensive scheme around an opponent's strong tendency to overload the left half-space in transition, a pattern present in eleven of their last thirteen matches. The report is thorough, the clips are cut, the walkthrough is convincing. What the report does not say is that the tendency broke down four weeks ago, when the opponent's first-choice inverted winger tore a hamstring and was replaced by a player with the opposite footedness, who does not make that run. The analyst was not lazy. The analyst was blind at exactly the moment that mattered, because the corpus closed before the injury did.
This is the Large Language Model's failure mode, transplanted intact. A model trained on a corpus with a cutoff date knows the world as it was, distributed and averaged, and infers forward from there. It has no mechanism for learning that the specific molecule at today's shutter has changed. A scouting report built from last season's footage is the same object: an enormous, genuinely informative single look, followed by a shutter that stays closed while the team it describes keeps playing, injuring players, signing replacements and adjusting its shape in response to whoever they now employ.
A good analyst updates constantly. This isn't a demon problem, it's a discipline problem — do the homework every week and the failure disappears.
That is fair as far as it goes, and it is exactly the move from one rung of this ladder to the next. Updating weekly is a demon that reopens its eyes for the duration of one scouting cycle and closes them again. It is real sight, genuinely better than none, and it is still bounded. A transfer completed on Thursday, an injury sustained in Friday's training session, a tactical tweak made at half-time of a match played while the report was being finalised — all invisible to a demon on a weekly clock. This is the Large World Model's position: a demon posted at one shutter, seeing clearly what passes through it, blind to the rest of the vessel and to what happens between visits.
Where the ceiling sits
Maxwell specified something stronger than a demon who checks in periodically. He specified continuous measurement — every molecule, every arrival, no gap in which the shutter is unattended. Applied to sport, that specification has a name analysts already use informally when they describe the ideal they cannot quite reach: a live model of the opponent that ingests tracking data as it is captured, injury news as it is reported, transfer activity as it is confirmed, and tendency data as it accumulates match by match, continuously revising the picture rather than replacing it on a schedule.
This is the third position on the intake axis, and it is worth being exact about why it is a ceiling rather than merely an improvement. A demon that sees every degree of freedom, all the time, has nothing left to observe. There is no fifth stream waiting to be discovered, no additional channel of information about an opponent that isn't tracking, injury status, personnel and pattern. What limits such a system is not access — it is the discipline of what it does with access: dating each observation, sourcing it, deciding when a belief has decayed past usefulness and should be revised or dropped. Beyond continuous, provenanced, revisable intake across every channel, there is no further category of evidence to add. There is only better inference on the same evidence, and cheaper memory to hold it.
| Position | What it sees | Characteristic blind spot |
|---|---|---|
| Large Language Model | One long look, then closed eyes | The winger who was fit at cutoff and isn't now |
| Large World Model | Sight renewed on a cycle — weekly report, matchday scout | Whatever changed between visits |
| Large Universe Model | Every stream, continuously, with dated and decaying beliefs | Nothing structural — only the fidelity of what is fed in |
Two objections worth taking seriously
The strongest challenge to this picture is thermodynamic, not tactical. Bennett's accounting shows the demon breaks even — every bit of sorting is repaid by the cost of erasing memory to make room for the next bit. If continuous observation is neutral at best, then a club running live ingestion across four data streams is not being handed an advantage, it is being handed a bill. And the bill is real: a Premier League club's tracking vendor alone can generate on the order of a million data points per match per team, and injury, transfer and scouting feeds run continuously in parallel. Someone has to store it, clean it, decide what to keep.
Concede the physics and the objection still fails to reach its conclusion. Landauer's bound at room temperature is about 2.9 zeptojoules per bit erased — trivial. The real cost of a club's data operation is not thermodynamic, it is organisational: analyst hours, integration between systems that don't talk to each other, trust in a feed that might be wrong. Those costs are large, but they are engineering costs, not proof that the sorting has no value. The break-even accounting establishes that the sorting is real and that it cannot happen without the measurement. It says nothing about whether the measurement is worth having. In a league where fine margins decide relegation, it is.
The second objection is more interesting because it is domain-specific. Much of what matters in a sport is stable. The offside rule does not change match to match. A team's foundational shape, built over a manager's tenure, might hold for three years. If large parts of the picture are invariant, why pay for continuous intake at all — why not take one very long look at the fundamentals and stop?
That is correct for the fundamentals, and it is exactly why static scouting reports were never worthless. The error is treating the stable fraction as the whole picture. Formation is stable; who plays in it is not. A rulebook is stable; a suspension list is not. The molecules at the shutter — this week's fitness status, this transfer window's arrivals, the tendency that shifted after a manager was sacked in October — are precisely the parts that a single look cannot fix in place. The demon does not need the whole vessel to be volatile. One fast-changing degree of freedom, watched continuously, beats a comprehensive snapshot of everything except that one thing.
The lineage was never a claim that watching more makes a smarter analyst. It is the older and narrower claim, priced by Szilard and audited by Landauer and Bennett: sorting requires seeing the molecule that is actually at the shutter, now. A frozen report cannot do that. A weekly one does it once. Only a system built to keep watching, and honest about when its beliefs have gone stale, does it as long as the game is still being played.