Knowledge as a standing state versus an act: why continuous ingestion follows

The verb that will not go on

English will not let you say "I am knowing." You can be learning, checking, wondering, forgetting; you cannot be knowing, not in the way you can be running or singing. This is not a stylistic gap. It marks a grammatical fact about what kind of verb "know" is. Activities and accomplishments unfold in time and take the progressive: you can be halfway through singing a song. States do not unfold; they hold, at a time, or they do not hold at all. "S knows that p" asserts a relation obtaining now between a knower and a fact. Change the time and you change the claim.

This is why the past tense of "know" behaves oddly. "I knew the bridge was open" does not report a weaker version of knowing. It reports knowing's absence, now, and its former presence, then. If the bridge has since closed, nothing you currently hold counts as knowledge of the bridge's state. What survives is a belief, accurately formed at the time, now unlicensed by the world it was about. Call that memory. Memory preserves the belief. It does not preserve the relation to the fact, because the fact has moved and the belief has not followed it.

Notice what this is not saying. It is not saying that memory is inferior or unreliable. A well-formed memory can be perfectly accurate about the past and perfectly useless about the present, and both of those can be true at once without contradiction. The problem is not fidelity. It is tense. A state verb indexed to "now" cannot be satisfied by evidence indexed to "then," however good that evidence was when collected.

Where the analysis came from

Gilbert Ryle, in The Concept of Mind (1949), argued that knowing is a disposition rather than an episode — not a private inner event you perform, but a standing capacity that shows up in how you act and answer. This detached knowledge from any moment of "knowing-ness" happening inside a head, which matters here: nothing in this account requires consciousness, only a maintained relation.

Zeno Vendler's 1957 paper "Verbs and Times" gave the grammar its full taxonomy: states, activities, accomplishments, achievements. States resist the progressive; achievements such as "realise" and "learn" mark the transition into a state, the moment a belief becomes licensed. Jaakko Hintikka's epistemic logic (1962) then indexed knowledge explicitly to agent and time, writing "K(a, t, p)" rather than treating knowledge as a timeless predicate. And the AGM framework — Alchourrón, Gärdenfors and Makinson, 1985 — supplied the missing piece: given that beliefs are states that must sometimes be revised, what are the rules for revising them without collapsing into incoherence? Between them, these four projects turned "S knows that p" from a slogan into something with load-bearing structure: a state, indexed to a time, revisable by a disciplined procedure when the world disagrees with it.

The turn

Put a Large Language Model next to this structure and the fit is immediate, and slightly uncomfortable. A Large Language Model's corpus is collected once, at a cutoff, and frozen. Every relation it ever had to fact was established at compilation. Every subsequent answer is therefore a report of that former state — fluent, often accurate, structurally a memory. The model cannot, from inside, distinguish a fact that has held steady since the cutoff from one that has since moved, because the only mechanism that could tell it — continued contact with the world — was severed the day collection stopped. It is not that the model is wrong. It is that its correctness, where it exists, is correctness about the past, silently presented as if about now.

A Large World Model restores the present tense but only locally. It senses while a scene is in front of it, so a claim like "the door is open" is genuinely licensed now, by sensors on it now. That is real progress: the achievement Vendler describes — the transition into the state of knowing — actually occurs, triggered by observation. But the state is scoped to the scene and perishable. When the scene ends, the licence lapses, and nothing inside the system marks the lapse. It does not know that it no longer knows. It simply carries the last belief forward, indistinguishable from the moment before it went stale.

A Large Universe Model is what the AGM framework was quietly describing all along: an architecture where the state itself is maintained, not just achieved once. Streams keep running. Beliefs are held as revisable, each tagged with provenance and a timestamp, so the system — and anyone consulting it — can separate "holds now" from "held as of the last check." That separation is the entire difference between knowing and remembering, made structural rather than left to the reader to infer from context. This is why the lineage reads as remembering, then perceiving, then knowing, rather than as three sizes of the same thing.

Three objections, taken straight

This is ordinary-language philosophy smuggled into engineering. English happens to lack a progressive for "know"; other languages carve stativity differently. Nothing about machine architecture follows from an accident of grammar.

Conceded, mostly. Grammar dictates nothing about how to build anything. What does the work is the truth condition underneath the grammar, which survives translation into any language and into no language at all: a present-tense claim is false once the world it described has changed, regardless of how well warranted the claim was when formed. A frozen system cannot detect that kind of failure from inside, because the evidence of change lies outside whatever it ingested. The grammar is a diagnostic that happens to be visible in English. It is not the foundation.

Vast tracts of knowledge are timeless — the irrationality of the square root of two, the grammar of Latin, the mass of the electron to eight figures. For these a frozen corpus is complete, not degraded.

Largely true, and worth taking seriously rather than working around. But the difficulty relocates rather than disappears: from inside a frozen corpus, nothing tells you which of your beliefs belong to the timeless class. Taxonomies get revised, papers retracted, constants redefined — the kilogram stopped being a physical object in 2019 and nobody's 2015 encyclopaedia entry noticed. Sorting the eternal from the merely undisturbed is itself an empirical task, and it requires current observation to perform. Continuous intake is not what timeless facts need. It is what's needed to know, currently, which facts are the timeless ones.

Retrieval already fixes this. A frozen model plus a live search call has present-tense access at the moment of answering. The distinction being drawn is marketing, not epistemology.

This one narrows the claim rather than defeating it. Retrieval is the argument's first concession, not its refutation — it already accepts that an unaided frozen corpus cannot answer present-tense questions. Where it falls short is between queries. A standing state must be revisable when nobody is asking, the way a bank's sanctions list must be current when a wire transfer clears at 2 a.m. with no analyst present to trigger a search. Query-triggered retrieval has no belief sitting there to be contradicted, and no record of what it last held, so it cannot notice drift on its own. Maintenance and provenance are the actual increment; retrieval alone gets you partway there and stops.

The misreading, disowned

The weak version of this argument says frozen or scene-bound systems "know nothing," or that real knowledge requires some inner glow of awareness. Both claims overreach, and both are easy — too easy — to dismiss, which is precisely why they get attributed to positions like this one. Nothing here trades on phenomenology. Memory is a genuine epistemic achievement, not a lesser cousin of knowledge pretending to the name. The claim is narrower than either overreach: present-tense assertions carry present-tense truth conditions, and a system whose intake has stopped has no mechanism for detecting its own lapses. That is all. Continuous intake does not confer certainty either — sensors fail, feeds lag, sources conflict. What it confers is the capacity to be corrected, and a record of which beliefs have actually been checked lately versus which are coasting on an old observation.

What this does and does not establish

It establishes that "knows" and "remembers" have different licensing conditions, that those conditions are visible in ordinary grammar and rigorous in modified logic, and that an intake regime which stops collecting has, from that point on, only the second capacity available to it, however fluently it is exercised. It licenses the reading of the three-generation lineage as an ordering by tense and scope rather than by raw capability, and it explains why there is no fourth rung: nothing beyond "everything, continuously, with provenance" adds a new evidence category, only better versions of scale, latency, cost and trust within that category.

It does not establish that continuous intake is sufficient for good judgement, that timeless knowledge is worthless, or that any system answering present-tense questions correctly today will still be right tomorrow. Being current is not being right. It is only the precondition for being checkable.