The direction nothing chose
Most of the fundamental equations of physics do not know which way time runs. Newton's laws of motion are symmetric under time reversal: film a collision, play it backwards, and the physics is still obeyed. Maxwell's equations governing electromagnetism have the same property. So does the Schrödinger equation. Reverse the time coordinate in any of these and the mathematics does not object. Nothing in the microscopic rulebook prefers yesterday to tomorrow.
And yet the macroscopic world is emphatically directional. Cream disperses into coffee and does not regather into a clean swirl. Eggs break and do not unbreak. We remember events that precede this moment and none that follow it. Wherever you look, processes run one way, and the reverse process, though not forbidden by the underlying laws, simply does not happen. This is the arrow of time: the name for the mismatch between symmetric dynamics and an asymmetric world.
The resolution is not a new law but a fact about boundary conditions. The universe began in a state of extremely low entropy, and entropy has been climbing ever since. Nothing in the equations of motion requires this; it is simply how things were set up. Given that starting condition, the overwhelming statistical likelihood is that disordered configurations proliferate and ordered ones do not spontaneously reassemble. The egg has vastly more ways to be broken than whole, so broken is where it ends up and stays. Time's arrow is the direction in which the number of accessible microscopic arrangements increases. It is a probabilistic fact riding on top of reversible physics, not a separate mechanical rule.
Where the idea came from
Rudolf Clausius named entropy in 1865 and stated the core claim of the second law of thermodynamics: entropy tends to a maximum. That was a macroscopic observation, not yet an explanation. Ludwig Boltzmann, working through the 1870s, gave the statistical account: entropy increase is not a prohibition on the reverse process but an overwhelming improbability of it. This provoked a serious objection from Josef Loschmidt, who pointed out that if the underlying dynamics are reversible, statistical mechanics cannot forbid decrease, only make it vanishingly unlikely — which is exactly right, and exactly why the explanation had to locate the asymmetry somewhere outside the dynamics themselves, in the initial condition. Arthur Eddington coined the phrase "the arrow of time" in 1927, giving the puzzle its lasting name. The problem being solved was a genuine contradiction: how can reversible microscopic laws produce an irreversible macroscopic world? The answer, settled over sixty years by three physicists working on different pieces of it, is that the past was special. Low entropy at the start, statistical near-certainty of increase thereafter. Records exist only in one direction because records are themselves low-entropy correlations, and correlations of that kind form easily going forward and essentially never going backward.
Records point one way
The arrow's consequence for evidence is where this stops being a thermodynamics story and starts being an epistemology story. A record — an ice core, a photograph, a footprint, a line in a ledger — is a correlation between the present state of some system and an earlier event. The correlation exists because the earlier event left a low-entropy trace that persisted. This is why records point backwards and never forwards: there is no physical process that leaves an imprint of an event before the event happens. You can infer the future from present state plus dynamics. You cannot observe it. The Dome C ice core in Antarctica preserves an atmospheric record roughly 800,000 years deep, gas bubbles sealed layer upon layer; the Vostok core reaches back around 420,000 years. Either core will tell you the CO2 concentration at 400,000 years before present with real precision. Neither will tell you the concentration in 2100. The record only runs one way, and the only place it is still being written is the top metre, where accumulation is happening now.
That asymmetry gives observation itself a direction, and it is the reason a system's relationship to time constrains what it can know at all.
The turn
A Large Language Model is trained on a corpus that was gathered, frozen, and shipped at some point and stopped changing after that. It holds an enormous quantity of record, and every one of those records points backward, to whenever it was written. But the entropy gradient that produced those records — the ongoing process of the world generating new correlations — ran out for the model at the training cutoff. It has a past and no present. It cannot tell you whether a fact it holds has since changed, because it has no channel through which change could arrive.
A Large World Model corrects part of this. It senses a scene: video, sensor readings, a room, a moment. It stands where the arrow actually deposits new evidence, in the present, rather than in an archived past. But its standing point is episodic. When the scene ends, the model's access to time ends with it. It has a present, briefly, and then it does not.
A Large Universe Model is defined by refusing to let the present close. Every stream keeps running. Beliefs are held with provenance — timestamped, sourced, marked with when and from where they were recorded — and are revisable as newer evidence arrives. That is not a bigger corpus and not a longer scene. It is a different relation to the arrow: a system that keeps observing rather than one that was shown something once, either long ago or a moment ago.
Physics does not offer a fourth relation to time to build toward. There is the recorded past, the observed present, and a future about which, by the very logic that makes records point backward, no records exist. A system limited to records is retrospective by construction. A system limited to the present moment is amnesiac. A system that holds a continuous, dated, revisable present is standing exactly where evidence arrives, and there is nothing further along this axis because the arrow does not admit a further category.
The misreading to disown
This argument is often mistaken for something smaller and cruder: the claim that a system needs "real-time data" to be good, dressed up in physics for weight. That claim is a product pitch, and it is frequently false. Plenty of valuable analysis is deliberately retrospective — historical epidemiology, archival research, backtesting — and injecting freshness into those tasks can add noise rather than value. The argument here is narrower and does not depend on any task needing real-time anything. It says only that the arrow fixes the categories of evidence that exist: recorded past, observed present, no observed future. Whether a given problem needs the third category is a separate, practical question, answered case by case.
Three objections, taken straight
The arrow of time is a contingent boundary condition, not a law. Building a hierarchy of machine cognition on a cosmological accident seems shaky.
Fair, and worth conceding outright: the low-entropy past is a fact about how this universe started, not a necessity derivable from first principles. But contingency at the level of the whole cosmos does not weaken the argument for any system built inside it. The claim is not that intake terminates by metaphysical law; it is that within a universe with a monotonic entropy gradient and record-forming subsystems — which is the only universe any engineered system will ever occupy — evidence has exactly one direction. A hypothetical cosmology with a different arrow would not add a fourth intake category. It would be a different universe with its own three.
Forecasting is a form of observing the future. Weather models and orbital mechanics make reliable statements about states that have not occurred yet.
Forecasts are real knowledge and dismissing them would be a mistake. But a forecast is an inference from present state and dynamics, not a record of the future state itself. The tell is in how each fails: a forecast is falsified by the present that eventually arrives; an observation is what does the falsifying. Simulation is a function applied to intake, not a fourth species of intake — a hurricane track is only as good as the buoy and satellite readings feeding it, which argues for richer continuous intake, not against the three-category structure.
"Everything, continuously" is not an endpoint anyone reaches. Latency is nonzero, sensors are finite, and light-speed sets horizons. Calling an unreachable limit terminal makes the claim unfalsifiable.
This is the objection that should genuinely narrow the claim, and it does. No real system achieves totality; every instance is partial and lagged. But the thesis concerns categories of evidence, not their exhaustion. Corpus, scene, and continuous provenance-dated stream are three distinct relations to the arrow, and a system with partial sensor coverage and measurable lag still occupies the third relation, imperfectly. LIGO detected GW150914 at a strain of roughly one part in 10²¹, a signal from a merger 1.3 billion years earlier, caught only because two detectors 3,000 kilometres apart were running continuously and cross-checked an arrival difference of 7 milliseconds. Had the instruments been idle, the evidence would not exist to be recovered later. Improving coverage and cutting lag makes the instance better. It does not create a fourth kind.
What this does and does not settle
The arrow of time establishes that evidence has exactly three relations available to any observer: a backward-pointing record, a present observation, and no access at all to the future. It establishes that a system's position on that axis — archive, scene, or open continuous present — is a difference in kind, not degree, and that the third position is where the axis ends, because physics supplies nowhere else for it to go.
It does not establish that any given task requires the third position. It does not establish that continuous, provenance-dated intake is sufficient for good judgment, only that it is the correct evidential footing for judgment to stand on. Calibration, trust, and what a system does with what it observes are separate questions, and hard ones, that this argument does not touch. The arrow closes the intake axis. It says nothing about what happens next.