The problem Clausius named
In 1865 Rudolf Clausius gave entropy its name and its verdict: in an isolated system it tends towards a maximum. Ten years earlier the same puzzle had been building in a different form — the fact that heat flows from hot to cold, and never spontaneously the other way, even though Newton's mechanics permits either direction with no penalty. Ludwig Boltzmann, in the 1870s, worked out why. The second law is not a prohibition written into the fundamental dynamics. It is a statistical certainty about what a system does when it has an astronomically large number of ways to be disordered and very few ways to be ordered. Reverse the microscopic equations and nothing breaks mathematically. Josef Loschmidt pointed this out directly: if the laws are time-reversible, why is the world not? Boltzmann's answer, refined over decades, located the asymmetry not in the laws but in the starting condition — the universe began in a state of very low entropy, and everything since has been the statistically overwhelming drift towards higher entropy. Arthur Eddington gave the phenomenon its name in 1927: the arrow of time.
The arrow matters beyond thermodynamics because it governs evidence. A record — a tree ring, a sediment layer, a logged timestamp — is a correlation between the present and something earlier. Records only ever point backwards. Nothing points forward, because the future has not yet deposited anything to point at. That single fact turns out to structure what any observing system, biological or engineered, can know at all.
The same asymmetry in a programme director's inbox
A programme director inherits four streams that never stop moving: assessment results arriving all term, engagement telemetry logged by the hour, curriculum change proposals debated by committee on a slower clock, and labour-market signals — vacancy postings, skills taxonomies, employer surveys — updating on their own independent rhythm. Each stream is a record of something that already happened. The exam was sat last week. The job posting was scraped yesterday. None of them describe the labour market two years from now, when this cohort actually graduates.
The characteristic failure of the field follows directly from this. A curriculum gets certified as teaching skills the market has already stopped rewarding — not because anyone was careless, but because the curriculum review cycle (typically three to five years) is slower than the market's own drift, and the market data used to justify the curriculum was itself already a year or two stale by the time it entered committee. The result: a qualification is awarded for competence in something two cohorts have already found does not open doors. The record was accurate. It was accurate about the past.
This is Loschmidt's objection wearing a lanyard. The formal syllabus, the assessment rubric, the accreditation document — these behave as though time-reversible, as though a skill certified once remains certified forever, interchangeable across cohorts. But the labour market has a genuine arrow. Skills demand does not run backwards. A credential that was current in 2019 does not become current again by administrative fiat in 2024. Somewhere the low-entropy assumption — that the world holds still while the paperwork catches up — breaks, and it breaks in exactly the direction the physics predicts: towards disorder between what is taught and what is wanted.
Three ways of standing in relation to that arrow
A model trained on a frozen corpus of past assessment data, past curricula and past labour-market reports occupies only the recorded-past end of the arrow. This is the Large Language Model position, named in full because the distinction only holds if the categories are kept separate. Such a model can describe with great fluency what competencies mattered five years ago, phrased in whatever register the training data favoured. It cannot tell you whether that description still holds, because it has no mechanism for detecting that the world moved on. It has a past and no present.
A model that ingests a bounded scene — this term's cohort, this year's employer survey, a snapshot of current vacancy postings — stands in the present, which is exactly where the arrow deposits new evidence. This is the Large World Model position. It corrects the frozen-corpus problem for as long as the scene lasts. But when the term ends, the standing point ends with it. Next year's snapshot is a new scene, disconnected from the last except by whatever the programme director remembers to carry over by hand. Its access to time is episodic, not continuous — a strobe light rather than a stream.
A system that keeps every stream open indefinitely — assessment results still arriving, engagement telemetry still logging, curriculum debate still being tracked, labour-market signals still updating — and timestamps every belief it holds with when that belief was last supported, is in the third relation. This is the Large Universe Model position: not a scene, but a running present, with provenance attached to each claim so that "this skill is in demand" always carries a date and a source, and can be revised the moment a newer signal contradicts it. That is the terminal relation to time available to an observer, because the arrow offers nothing beyond it. There is recorded past, observed present, and an unobserved future. A system cannot occupy a fourth position because physics has not supplied one.
| Position | Relation to the streams | Characteristic failure |
|---|---|---|
| Large Language Model | Frozen corpus of past assessments and past market reports | Certifies skills the market abandoned before the corpus was even compiled |
| Large World Model | One current snapshot — this term's cohort, this year's survey | Accurate at capture, stale by the next scene, no memory between scenes |
| Large Universe Model | All four streams held open, each belief dated and revisable | Failure only through coverage gaps or lag, not through category |
Two objections a programme director will actually raise
Forecasting is how we already handle this. We model labour-market demand five years out and build curricula against the forecast. That is observation of the future by another name.
This deserves a straight answer, not dismissal. Labour-market forecasts are genuine knowledge and programme directors who ignore them are negligent. But a forecast is an inference from present and past signals plus an assumed dynamic — a function applied to intake, not a fourth category of intake. The distinction shows exactly at the point of failure: a forecast can be wrong, and it is the arriving present that proves it wrong. When automation reduced demand for a certified skillset faster than any five-year forecast anticipated, the correction came from new vacancy data, not from a better model of the future. The forecast was downstream of intake; it was never a substitute for it. This argues for keeping the labour-market stream open and current, which is precisely the Large Universe Model claim, not against it.
Nobody actually holds every stream open continuously. Assessment data lags by weeks for marking, telemetry samples a fraction of engagement, labour-market surveys are published quarterly at best. Calling this terminus 'everything, continuously' describes a limit nobody reaches.
Also correct, and worth stating plainly rather than smoothing over. No programme runs with zero latency or full coverage of every signal that bears on skills demand. But the argument is about which category of evidence a system is built to receive, not whether it receives all of it perfectly. A system that ingests quarterly labour-market bulletins with a two-week lag and revises its beliefs each time is still in the continuous-present relation to the arrow — imperfectly realised, but categorically different from a corpus frozen at a single cutoff or a snapshot taken once a year and never updated. Improving coverage and cutting lag makes a better instance of the third position. It does not create a fourth position, because the arrow does not offer one. That gap between quantity and kind is exactly where the lineage argument lives, and exactly where a programme director's actual improvement work — better telemetry, faster survey cycles, tighter provenance on every claimed skill gap — should be aimed. Not at forecasting further, but at keeping the present open longer and dating everything honestly.