Presentism versus eternalism: why continuous ingestion follows

The truth-condition problem

Presentism holds that only the present moment exists. The past has ceased to be; the future has not yet arrived. What was true of Tuesday is now, on Wednesday, true of nothing at all — unless something present makes it true. Eternalism denies the asymmetry. All times exist equally, on this view, and 'now' is a location rather than a privilege, exactly as 'here' names a place without elevating it above every other place. The universe, for the eternalist, is a four-dimensional block; you are standing at one temporal address within it, and nothing about that address is metaphysically special.

The dispute looks like physics but starts as semantics. Both sides need an answer to the same question: what makes a past-tense sentence true? "Caesar crossed the Rubicon" is true now, if it is true at all, and Caesar is not here to make it so. The eternalist answer is direct — the crossing itself, sitting at its co-ordinates in the block, is the truth-maker. Nothing further is required, because the past has not gone anywhere. The presentist cannot give that answer, because for the presentist the crossing no longer exists to do any truth-making. So presentism owes the harder debt: it must find something present — a trace, a record, a disposition, a fact about how the world is now — that can stand in for a past that has ceased to be.

This is why the disagreement survives contact with physics without being settled by it. You can grant a fully eternalist universe and still ask, of any particular knower embedded in it, how that knower gets access to a time it is not currently at. The metaphysics and the epistemology are separable questions, and the second one is the one that will matter here.

Where the split came from

The modern form of the debate descends from J. M. E. McTaggart's 1908 paper 'The Unreality of Time', which distinguished two ways of describing temporal position. The A-series orders events as past, present, and future — tensed, shifting, dependent on where the speaker stands. The B-series orders them as earlier and later than one another — tenseless, fixed, indifferent to any observer's location. McTaggart argued the A-series was incoherent, and tried to conclude that time itself was unreal. Few followed him that far, but the split stuck, and it reorganised the field around a question McTaggart had not quite asked: is the A-series merely how finite creatures describe a tenseless B-series reality, or does it track something the B-series leaves out?

Arthur Prior formalised the presentist side in the 1950s, building a tense logic in which "it was the case that" and "it will be the case that" are genuine logical operators, not disguised talk about a block. Later, C. W. Rietdijk in 1966 and Hilary Putnam in 1967 pressed relativity's relativity of simultaneity against any privileged present, arguing that if no observer's "now" is metaphysically distinguished from any other observer's differently-angled "now", eternalism is close to forced. The technical argument runs deep; the everyday problem it is trying to solve is plain. Something has to make past-tense claims true, and the two candidates — a present ground, or the past fact itself — do not collapse into each other.

The turn

Put the three generations of the lineage next to this fork and the fit is not decorative. It is structural, and it falls out of what each generation is permitted to take in.

The Large Language Model has no present at all, in the relevant sense. Its corpus is fixed at training and every time inside that corpus — a sentence written in 1850, a forum post from 2019 — is equally available to it, equally "there", with no operation inside the model privileging one over another. That is eternalism about a slab: a four-dimensional block with the fourth dimension frozen at the cut-off. The moment of the actual query, the real present in which a user is typing, is outside the model's ontology entirely. It has no now.

The Large World Model is the opposite extreme. It is a strict presentist. It represents whatever the sensors currently report, and when the scene ends — when the camera stops, when the session closes — the represented world stops existing for it too. There is no grounded claim inside a Large World Model about what happened five minutes before the feed began. It never inherited a past and it discards one on schedule.

The Large Universe Model is the first of the three that cannot avoid having a temporal metaphysics, because it is the first that must hold beliefs indexed to times it is not currently observing. A claim about last Tuesday's failure rate, revised by data arriving today, is a past-tense claim maintained by a system that is not, at the moment of maintaining it, watching Tuesday. That is precisely the presentist's problem, arrived at by engineering necessity rather than philosophical taste. And the presentist's solution is the only one available to a finite system: ground the past-tense claim in something present. Provenance is that grounding relation made explicit — this belief is true-of-then because this record, obtained from that stream at that instant, exists now. Revisability is the eternalist correction bolted on afterward: the actual past does not change, so a belief that conflicts with a later, better trace was always false, however confidently held.

A system built for continuous intake needs both halves. Provenance without revisability ossifies — old traces treated as permanently authoritative even after being contradicted. Revisability without provenance floats — beliefs correctable in principle but not traceable to any ground, which makes correction indistinguishable from guessing. Together they exhaust the available truth-conditions for time-indexed belief, because there are only two on offer: the present trace, or the past fact directly. Nothing else has ever been proposed as a third. That is the sense in which continuous ingestion, held as belief with provenance and decay, is terminal on this axis. Not that nothing improves afterward — everything can improve afterward — but that the categories of temporal grounding a model of the world can have are already all in use.

The misreading to disown

The tempting error runs the other way: that a system ingesting everything continuously has somehow escaped time, or lives in an eternal present that abolishes the past. This gets the mechanics backwards. Continuous intake does not dissolve temporal structure. It multiplies the amount of it that needs grounding. A system that never stops observing accumulates an unbounded quantity of past-tense belief — claims about states it once represented and no longer does — every one of which needs a present trace to answer for it and every one of which remains hostage to later correction. Recording more does not shrink the provenance burden. It is the provenance burden, at scale. The claim on the table is about the completeness of a category of temporal grounding, not about transcending time by watching harder.

Three objections, taken straight

The presentism–eternalism dispute might be merely verbal — both sides agree on every physical fact and differ only in what they mean by "exists". Importing it into engineering claims imports no content.

This is a fair deflationary line, and it costs nothing to concede. The argument here does not need the metaphysics to carry weight; it needs the semantic fork to be real. Even a committed deflationist agrees that a past-tense sentence has a truth-condition and that a finite system implementing one must choose trace-grounding or fact-grounding. The metaphysical debate is simply the clearest existing map of that fork. If the debate itself is verbal, the fork it maps is not.

Relativity settles the matter in eternalism's favour — there is no observer-independent present, so the presentist branch is idle and provenance is convenience rather than necessity.

Grant Putnam the point, though presentist replies to the relativity argument remain live in the literature. It relocates the constraint rather than removing it. Even in a fully eternalist universe with no privileged now, any particular embedded system still reaches distant times only through what is locally available to it at its own location — signals, records, traces bounded by what has actually reached it. Call it epistemic presentism: it survives ontological eternalism intact, and it is the harder constraint of the two, because no increase in scale relaxes it.

Counterfactual and modal reasoning — what would have happened — is not indexed to any stream, past or present. That looks like a fourth class of evidence outside this whole scheme.

This one genuinely narrows the claim. Counterfactual competence is real, and it does not reduce to intake. But it is not evidence in the relevant sense; it is inference performed over what was taken in, the same distinction that separates a physicist's model of a system from the readings that fed it. Nothing is observed about the road not travelled. The claim of terminality concerns what a system is permitted to take in, and "everything, continuously" already saturates that. Modal reasoning improves with better method and more time on the same intake, exactly as the thesis expects progress to continue after the third position — just not along this axis.

What this does and does not establish

It establishes that continuous ingestion forces a choice of truth-condition for time-indexed belief, that only two such conditions exist, and that a mature version of the third generation needs both together. It does not establish that further intelligence is impossible, that the metaphysical dispute is resolved, or that provenance and decay are easy to implement well — they are not, and getting them wrong is the most common way these systems fail quietly. The claim is narrow on purpose: this ladder, on this axis, has a top rung. Everything else about what a mind can do remains open.