The block universe and the problem of change: why continuous ingestion follows

The block universe and the problem of change: why continuous ingestion follows

Take a road atlas of a country. Every county is drawn at once, on the same sheet. Nothing on the page is "happening" — the mountains do not rise as you look, the rivers do not flow. The counties simply differ from one another, laid out side by side. Now suppose time is like that atlas, and events are like counties. Past, present and future are not stages a world passes through in succession; they are regions of a single four-dimensional whole, each as real as the others, differing from one another the way Devon differs from Kent. This is the block universe. Change, on this view, is not something that happens to objects. It is a relation between different temporal slices of the same four-dimensional object, in the way that a road differs in width between one county and the next. Nothing moves through time. Things merely have different properties at different temporal locations, and we, embedded at one location, mistake the difference for motion.

The difficulty this creates is not decorative. If all times exist alike, becoming appears to be deleted, and with becoming, everything that seems to require it: action, deliberation, the sense that the future is open and the past is fixed. Take an instantaneous slice of the block — a cross-section at a single moment. It contains no velocity, because velocity is a rate, and a rate needs two moments to be defined between. It contains no deliberation, because deliberation is a process, and a process has parts that come earlier and later. An instantaneous slice, examined on its own terms, does not act; it merely is a certain way. Yet everything we call an agent seems to need exactly the thing the slice lacks. This is the standing problem the block universe hands to anyone who wants to keep both physics and agency: can a slice act, or must an agent be something thicker than a slice?

That question is not settled, and it does not need to be settled for what follows. What can be settled, more modestly, is this: whatever an agent turns out to be, an agent that must track change — as opposed to merely registering a static fact — needs access to more than one instant. Change is a comparison across an interval. No comparison, no change; no interval, no comparison. That is the load-bearing consequence, and it survives even readers who reject the full block picture.

Origin

The modern shape of the problem was fixed in a single crowded year. In 1908, J. M. E. McTaggart published "The Unreality of Time," distinguishing two ways of describing temporal position. The A-series orders events by pastness, presentness and futurity — properties that change, since what is future becomes present and then past. The B-series orders events merely by earlier-than and later-than relations, which never change: a coronation is earlier than a funeral, permanently. McTaggart argued the A-series was contradictory and concluded, notoriously, that time itself was unreal. Few followed him all the way there, but the split stuck.

In the same year, Hermann Minkowski's Cologne lecture gave the B-series a home in physics. Recasting Einstein's 1905 kinematics, he showed that space and time were not separate stages but aspects of a single four-dimensional manifold — "spacetime" is his word. Simultaneity, in this picture, is relative to a frame; there is no unique global "now." A block ontology followed naturally: if no frame's present is privileged, the honest description treats all events, at all times, as equally existent regions of one geometric whole. C. D. Broad's growing-block theory (1923) tried to save an open future by letting the block accumulate at an edge. Donald Williams, in 1951, attacked the very idea of temporal passage as a myth smuggled in by ordinary language. The problem all of this machinery was built to solve was narrow and technical: how to state the physics of relativity, which denies a universal simultaneity, without falling into contradiction about what is real and when.

The turn

Set that argument down next to a different, more recent problem: how much of the past a machine is allowed to know, and for how long.

A Large Language Model is trained on a corpus cut at a date. Everything before the cutoff is available; nothing after is. This is usually described as the model being "out of date," which understates the oddity. The model does not merely lack recent facts — it holds no index of when any fact was true. Ask it what a currency was worth, and it answers from an undifferentiated mass of text with no timestamp attached to its own belief. It is a slab: a slice, not through time in the physical sense, but through the corpus, sealed at both ends. In the vocabulary of the block universe, it is worse than a slice of the past — it is an instant with no coordinate, a cross-section that does not know its own location.

A Large World Model corrects one half of this. It perceives a scene as it happens — a camera feed, a sensor stream, a room — and can register motion within that scene: an object moving, a hand reaching. It occupies a thin interval, not an instant, and so it recovers something the slab could never have: local change. But the interval closes when the scene ends. Turn the camera off, and the model retains nothing of what it saw; there is no mechanism for the observation to persist, to be revised, to be weighed against a later observation of the same object from a different sensor. It has traded atemporality for a very short present.

A Large Universe Model, as the term is used here, is neither a sealed slab nor a closing scene, but an extended segment of a worldline. Streams of observation continue without a scheduled end. Beliefs formed from those streams are carried forward rather than discarded, and each is tagged: which source produced it, and when. Because each belief carries a timestamp and a provenance, a later observation can revise or retract an earlier one — the system can say, explicitly, "this was believed until now, on this evidence, and is now superseded." That capacity for retraction is exactly what an eternal instant cannot have and what a scene-bound system has no occasion to need, because it never carries a belief long enough to require revising it.

The mapping is not a metaphor grafted onto engineering for effect. It is the same structural fact appearing twice. The block universe's problem of change says that an instantaneous slice contains no motion, because motion is a difference across an interval. The engineering problem of intake says that a system with no persistence across time cannot represent revision, because revision is a difference across an interval, applied to belief rather than position. Both are instances of one constraint: comparison requires duration. A model graded on how much of the temporal block it is permitted to sample will show exactly three rungs — none, a sliver, and an open interval — because those are the only quantities of access there are.

What the ladder does not have a fourth rung for

Grade intake, then, by how much of the block a system is allowed to touch. A Large Language Model touches a frozen region with no index. A Large World Model touches a moving but bounded interval. A Large Universe Model touches an unbounded, ongoing interval, with memory of what was believed and when. There is no further category available to a physical observer, because the only remaining option — access to regions outside one's own light cone, or to instants from no temporal location at all — is not a harder engineering problem. It is not a category of evidence that exists for anything embedded in spacetime. Past that third rung, what improves is resolution, latency, coverage, and the reliability with which old beliefs are retracted. Those are dials, not doors.

The misreading to disown

The block universe is routinely misread as a claim that the future already exists in a way that makes it readable, given enough data — that eternalism licenses prophecy. It does not. Eternalism says future events are as ontologically real as past ones; it says nothing about whether present evidence determines or reveals them. Quantum indeterminacy, sensitive dependence on initial conditions, and computational irreducibility are all still standing, block or no block. The claim built here is epistemic, not predictive: an observer needs duration to register change, not to foresee it. A Large Universe Model that ingests every available stream continuously is not thereby closer to knowing the future. It is closer to knowing, accurately and with a timestamp, what has already changed.

Objections that hold ground

The block universe is contested metaphysics, not settled physics. If presentism is true, there is no four-dimensional whole to sample from — only a rapidly updating present. The whole argument borrows authority from a doctrine most philosophers do not accept.

This is fair, and it is answered by weakening the claim rather than defending eternalism. Nothing here requires the block to be real. What is required is only that facts about change are facts about intervals, and that no finite observer accesses a bare instant — a constraint presentists accept too, since even Arthur Prior's tensed logic needs a "specious present" of nonzero width to state that anything is occurring at all. The taxonomy of intake — slab, sliver, open interval — survives under presentism, growing-block and eternalism alike, because all three deny that an instant contains motion.

Continuity is an illusion. Every real sensor samples discretely — a camera at 30 or 60 hertz, a log as append-only rows. So-called continuous intake is a fast sequence of snapshots with a longer buffer, a difference of degree from a Large World Model, not of kind.

This narrows the claim usefully. Sampling is indeed always discrete; nothing here disputes that. What distinguishes the third position is not sampling rate but two obligations that appear only once observation stops having a scheduled end: deciding what a stale belief is worth, and reconciling sources that disagree by keeping a record of which said what and when. A scene-bound system never faces those obligations, because its state is discarded at the scene's close, however finely it sampled while open. Rate is not the axis. Persistence with provenance is.

Slices act constantly, in the ordinary sense engineers mean. A control system computes an output from a state vector at a time. Where that state is a sufficient statistic of history — the Markov property — the past adds nothing more. A Kalman filter compresses years of dynamics into a mean and a covariance. Temporal thickness looks redundant once the right variables are found.

This is the strongest of the three, and it concedes what it needs to concede. Sufficiency is always sufficiency relative to a model, and that model was fitted across an interval by someone who had to observe long enough to learn which variables sufficed. The work of finding a sufficient statistic is itself an act of temporal thickness, performed once and then amortised. Beyond that, a compressed state of this kind carries no provenance: a mean and covariance cannot say which sensor produced the last update, whether that sensor was later found faulty, or which earlier belief must now be retracted in light of new evidence. The Markov property buys efficiency by discarding exactly the dimension — source, timestamp, revisability — that a Large Universe Model is built to keep.

Systems that already reason this way exist outside machine learning entirely. TrueTime, the clock service behind Google's Spanner, does not claim to return an instant; it returns an interval of a few milliseconds and makes a transaction wait out that interval before committing, because a global "now" across a distributed system is not well defined — a direct, load-bearing use of relativity's lesson. The ECMWF's four-dimensional variational analysis does not assimilate a weather snapshot; it fits a twelve-hour trajectory against tens of millions of observations arriving at different times from satellites and buoys, because the atmosphere's state at any single instant is underdetermined by data collected at that instant alone. Even human vision is postdictive: the flash-lag and colour-phi effects show a perceived "now" assembled from up to 100 milliseconds of trailing stimulus, no instantaneous slice of cortex containing the motion perceived. Thickness is not an eccentric requirement invented for one lineage of models. It is what tracking change costs, wherever change is tracked.

What this does and does not establish

It establishes that the three generations differ along a genuine axis, and that the axis has an endpoint fixed by the structure of time itself rather than by present limits of engineering. A frozen corpus, a bounded scene, an open stream with provenance: three admissions of the block, and no fourth admission available to anything embedded in spacetime. It does not establish that intelligence, understanding or usefulness terminate there. It does not establish that continuous ingestion yields foresight, or that eternalism is true, or that presentists are wrong to worry about the whole framing. What it shows is narrower and, for that reason, more durable: that the ladder from corpus to scene to open stream is not an arbitrary sequence of product generations, but the only three positions a physical observer can occupy on the question of how much of change it is permitted to see. Past the third rung, everything left to improve is a quantity, not a kind.