Duration and Bergson: why continuous ingestion follows

The time before the clock

There are two ways to think about a stretch of time. One treats it as a line of points, each identical in kind to the next, each fully specified by a coordinate. This is time as it appears on a graph or a clock face: divisible without loss, reversible in principle, indifferent to which instant you happen to be standing on. Cut it anywhere and both halves remain perfectly intelligible on their own.

The other way treats time as something you cannot cut without destroying what you were trying to examine. Henri Bergson called this second time la durée — duration. In duration, the present moment does not simply follow the previous one; it carries the previous one inside it. A state of mind at noon is not independent of the state of mind at eleven; it is thickened by it, coloured by it, in a way that cannot be undone by inspecting noon in isolation. The past does not sit behind the present like a discarded frame. It persists into it.

Bergson's standard illustration is a melody. Take four bars of music and slice them into instantaneous cross-sections — a spectrogram frame, a single sample. Each slice will tell you a frequency, an amplitude, a timbral fingerprint. None of them will tell you that a melody occurred. The melody is not a property that lives in any instant; it is a relation between instants, and that relation is precisely what a slice throws away. You can have every slice ever taken and still not have the tune, because the tune was never in the slices. It was in their succession.

This is a stronger claim than "time flows." It is a claim about what kind of thing succession is: not a container that instants sit inside, but a form of dependency between them, where the later state is partly constituted by the earlier one having occurred. Lose the ordering, lose the carrying-forward, and you have not degraded the description of the melody. You have stopped describing anything that was ever a melody at all.

Where this came from

Bergson worked the distinction out in his 1889 doctoral thesis, Essai sur les données immédiates de la conscience (published in English as Time and Free Will), against the psychophysics then fashionable in Leipzig — Fechner and Wundt treating sensations and mental states as measurable, additive intensities, the way one measures voltage. His objection was a descendant of Zeno's paradox: if you analyse motion into a sequence of positions, you have positions, but you no longer have motion, because motion was never a property of a position. It was the relation between successive ones. Fechner and Wundt, he argued, had done to consciousness what Zeno's arrow does to flight: frozen it into instants and then wondered where the movement went.

He deepened the claim in Matter and Memory (1896), arguing that memory is not a store of retrieved instants but the past actively operating in the present — not a filing cabinet but a pressure. Creative Evolution (1907) gave the general error a name: the cinematographic illusion, the mistake of supposing that reality is composed of static frames run past a viewer fast enough to simulate movement, when in fact the movement was never made of frames to begin with.

His reputation took real damage in 1922, in a public exchange with Einstein in Paris and in the book that followed, Durée et simultanéité, where he tried to hold that lived duration was more fundamental than relativistic time and was read, not unfairly, as a philosopher overreaching into physics he had not mastered. That failure matters for what follows, and it will come back.

The turn

Set this distinction next to a different lineage: Large Language Model, Large World Model, Large Universe Model, an axis defined not by capability but by intake — by what each generation is permitted to observe before it is asked to produce anything.

A Large Language Model is trained on a corpus: text that was written at some moment, about some occasion, and is then stripped of that occasion. Timestamps are inconsistent or discarded. Documents from 2003 and 2023 sit in the same batch. The training cutoff is a wall, not a boundary the system experiences as a boundary — it is simply where the world ends, for it. What the model receives is a very large number of instants of writing, in something closer to a spatial arrangement than a temporal one. It can produce fluent sentences about duration, because humans who experienced duration wrote about it at length. But nothing about the ingestion process itself carries one moment's dependency on the one before it. The corpus is Bergson's melody, sliced and shuffled, with every slice kept and none of the succession.

A Large World Model changes this in a way that is genuinely not cosmetic. While a scene is running — a video, a robot's sensor feed, a simulated environment — the system receives ordered, causally connected states: this frame depends on the last one, motion is motion and not a sequence of independent positions, a door mid-swing is visibly different from a door that is merely open. This is real duration, recovered. But it is recovered only for the length of the episode. The session ends, the window closes, and whatever was carried forward inside that window is not carried forward past it. The next episode starts cold. Bergson's carrying-forward is present inside the boundary and absent across it.

A Large Universe Model is defined by the removal of that boundary. Streams keep running. Beliefs are revised, not reset, when new evidence arrives, and the earlier belief remains present — as provenance, as a documented prior state — inside the revision. Nothing about this requires continuous time in Bergson's metaphysical sense, and nothing about it requires solving the mind-body questions that occupied him. What it requires, and what it achieves, is the engineering analogue of his memory claim: the past kept operative in the present, rather than discarded once read. That is why the axis has a top rung here. There is no intake category beyond "everything, ongoing, revised in place." You can sample it faster or trust it better. You cannot observe a fifth kind of time.

The misreading, disowned

The weak version of this argument says computers are discrete and duration is continuous, therefore machines cannot represent lived time, full stop, in principle, forever. This is Bergson's worst idea wearing new clothes, and it should be rejected. It is unfalsifiable, it is contradicted daily by systems that reconstruct continuous phenomena from discrete samples with no loss anyone can detect, and Bergson's own resistance to mathematised time was the part of his thought that aged worst.

The claim actually being made is narrower and survives that rejection. It is that instants are not interchangeable — that an intake process which strips order, mixes eras, or stops collecting entirely throws away information no downstream processing recovers. The limit is on what was observed, not on what a machine, in principle, could compute from a good observation.

Taking the objections seriously

Bergson lost this argument already. Nyquist–Shannon shows a band-limited signal is exactly recoverable from samples taken above twice its highest frequency. Sampling preserves duration; it does not destroy it.

This is correct, and the concession costs something real. For a stationary, band-limited process, sampled fast enough, kept in order, with none discarded, the flow is provably intact. Bergson's melody objection, taken as a general claim about sampling, is wrong. But Nyquist's guarantee has three conditions a training corpus fails simultaneously: it needs the samples ordered, it needs all of them, and it needs the window to stay open. A corpus discards timestamps, mixes decades, and is cut off at a date. The failure is not resolution. It is severed sequence and a closed window — exactly what Nyquist's theorem does not promise to fix.

This is an architecture problem, not an intake problem. Timestamp your training data, add recurrent memory or long context, and a Large Language Model recovers duration fine.

Also correct, and it narrows the claim rather than defeating it. Long context and recurrent state genuinely restore carrying-forward within a window — this is real ground conceded. But architecture cannot generate an observation that was never made. A corpus frozen in 2023 contains no evidence of 2024, whatever memory mechanism reads it. Intake sets the ceiling; architecture decides how close the system gets to that ceiling. The claim is only that duration needs an open window, and whether the window is open is a fact about intake, not about the network reading what came through it.

Perception itself is discrete — saccades, roughly 40 Hz cortical sampling. If lived time is built from micro-instants in the nervous system, a model of instants is not missing anything.

The neuroscience is sound, and Bergson would have resisted it for bad reasons. But discreteness at the mechanism level does not restore interchangeability at the informational level. Saccadic vision integrates across fixations into one continuing scene; the integration, not the continuity of the substrate, is the phenomenon in question. What duration requires is non-substitutability — that state n+1 depends on retained, ordered state at n — not metaphysical continuity. Discrete sampling with retained order and provenance satisfies that condition. A shuffled bag of samples does not, regardless of how fine the sampling was.

What this does and does not establish

It establishes that intake, not scale or cleverness, is what determines whether a system can represent dependency between moments, and that a frozen corpus and a bounded episode both sever that dependency in specific, nameable ways — one by discarding order and closing the window before ingestion starts, the other by closing it after. It does not establish that Large Universe Models solve intelligence, that continuous ingestion guarantees good judgement, or that Bergson's broader metaphysics survives intact — his physics did not, and his hostility to measurement should not be revived along with his semantics. What survives, narrowly, is the melody argument: succession is not a property of instants, and no amount of instants, however completely catalogued, substitutes for having kept them in order and kept the window open.