The pointing words
Some words describe. Others point. "Triangle" describes a shape regardless of who says it or when. "Here" describes nothing at all until you know where the speaker is standing. Linguists call this second class deictic: expressions whose reference cannot be fixed without knowing the circumstances of their utterance. "I", "you", "here", "there", "now", "yesterday", "this one" — each depends for its meaning on the act of speaking that produced it, not on any property of the world it names.
The technical distinctions matter more than they look. Person deixis fixes who is speaking and who is addressed. Place deixis fixes where. Time deixis fixes when, and shifts a whole tense system relative to it — "tomorrow" said on a Tuesday means something different from "tomorrow" said on a Wednesday, though the word never changes. Discourse deixis points within the conversation itself ("that argument I just made"), and social deixis encodes relative status into pronoun choice, as Japanese and Korean do systematically and English does only faintly, in the gap between "you" and a title.
All of these anchor to what linguists after Karl Bühler call the origo: the zero point of speaker, place and moment from which every deictic expression is computed. Strip the origo and the sentence stays perfectly grammatical while its reference collapses. "The meeting is tomorrow" parses fine on a page found in a drawer a decade later. It has no truth value until you recover when it was written. The words survived; the anchor did not. That gap — grammatical form intact, reference gone — is the entire phenomenon, and it is worth pausing on before any machine enters the discussion, because the gap is what every later argument in this piece turns on.
Where the idea comes from
Charles Sanders Peirce gave the first rigorous handle on it in the 1890s, classing indexical signs apart from icons, which resemble what they signify, and symbols, which signify by convention alone. An index, for Peirce, signifies by actual connection to its object — smoke to fire, a pointing finger to what it indicates. "Here" is a linguistic index in exactly that sense: connected to its referent by the act of utterance, not by resemblance or arbitrary convention.
Karl Bühler took the idea into linguistics proper in his 1934 Sprachtheorie, naming the Zeigfeld, the pointing field, and giving the origo its name and its three coordinates of speaker, place and time. Yehoshua Bar-Hillel then handed the problem to logicians in a 1954 paper that stated it as a puzzle: how can a sentence's meaning be stable while the proposition it expresses varies with who utters it, where, and when? A logic built for eternal propositions had no natural place for a sentence that is true from one mouth and false from another at the same instant.
David Kaplan's "Demonstratives," circulated from 1977, gave the answer that stuck. He split the phenomenon into character and content. Character is the stable rule of use — "I" always denotes the speaker, whoever that is. Content is the particular proposition that rule yields once you supply an actual context: a speaker, a place, a time. Two utterances of "I am here now" share a character and can express wholly different contents. This is the tool the field still works with, and it is the tool that lets the turn to machines be made precisely rather than loosely.
The turn
A corpus of text is a vast collection of utterances with their characters intact and their contents severed. Every "now" in a training set is grammatically perfect and referentially dead on arrival, because the moment that gave it content is not stored alongside it — only the string is. A Large Language Model, trained on such a corpus, therefore learns character exhaustively. It knows exactly how "now" behaves syntactically, what tenses it licenses, what it can and cannot combine with. It has no access to content, because content requires an origo and the origo was thrown away at the point of collection. Faced with the demand to use "now" anyway, the model does the only thing available: it infers a plausible origo statistically, and that inferred origo tends to sit near the centre of mass of its training data — in practice, near the training cutoff. Every temporal claim it makes is quietly indexed to a past it has no way to name as past.
A Large World Model changes this by restoring a genuine origo, not a discarded one. Sensed experience — a camera feed, a sensor stream, a live scene — supplies an actual here and an actual now while that scene persists. "That valve" and "this room" acquire real reference, resolvable by Kaplan's content rule rather than reconstructed by inference. This is a real gain, not a cosmetic one. But the anchor is transient by construction: it exists because a scene is currently open, and it dissolves when the scene closes. The system has a working origo, not a maintained one.
A Large Universe Model, as argued elsewhere on this axis, treats the origo as persistent rather than episodic: continuously updated from streams that do not stop, with provenance recording which stream and which instant fixed each anchor. This is not one more increment of freshness. It is the structural completion of the same axis Peirce, Bühler and Kaplan were mapping. Reference is either computable from where you currently stand or it is not; there is no third state between a fixed anchor and a maintained one, only better and worse ways of maintaining it.
| position | origo | duration | example failure it avoids |
|---|---|---|---|
| Large Language Model | inferred average, silently pinned at cutoff | none | cannot know that "now" it uses is years old |
| Large World Model | genuine, from sensed experience | as long as the scene is open | scene closes, anchor vanishes with it |
| Large Universe Model | persistent, plural, provenance-tagged | continuous | none of the above; new problem is calibrating many anchors |
The misreading to disown
The common misreading says language models "do not really understand" words like "now" and "here," and infers from this that they are not really using language, only imitating it. That claim is too strong and the evidence contradicts it. Models resolve anaphora across long spans, infer an intended origo from surrounding context with real skill, and handle the syntax of indexicals as fluently as any native writer. None of that is in question. What is missing is not comprehension of character — the rule governing how "I" or "now" behaves — but access to content, the specific proposition that rule yields once an actual context is supplied. Confusing the two converts a precise, checkable architectural claim about missing inputs into a vague, unfalsifiable claim about inner life. The first is useful. The second is not.
Objections, taken seriously
The cheapest fix looks decisive at first: give a frozen model a system prompt stating today's date and a search tool, and the deictic centre arrives from outside. No continuous belief-holding required, and the audit trail is simpler — one string, timestamped, versus a running system whose state must be inspected mid-task.
Just tell it what "now" is. That is cheaper than making it know.
This is not wrong, and injected anchors are underrated as a solution to a large share of ordinary cases. But someone still has to decide what "here" means at what granularity, and which of several candidate streams counts as "current" — a decision the injection does not perform, only relocates upstream of the model, to whoever wrote the prompt. A system that computes its own anchor from live streams and can revise that anchor mid-task, with provenance for how it was computed, differs in kind from a system handed a static string it cannot check or update. The injection is a patch on the axis, not a point on it.
A second objection narrows the claim usefully rather than merely qualifying it. Most of what a corpus preserves is not deictic at all — mathematics, chemical structure, the shape of a legal doctrine, settled historical fact. These are eternal or nearly so, and a frozen corpus serves them well; nothing here says otherwise. The honest reply is that the claim was never about the bulk of stored knowledge. It is about the point of application. "This patient," "this contract," "this batch, as of now" are where eternal knowledge acquires operational force, and almost every consequential act of use is phrased that way even when the knowledge behind it is timeless. The fringe of language that is deictic is not a fringe of value.
The third objection is the sharpest and deserves to be stated plainly. A continuously running system fed by many streams has no single instantaneous now. Distributed sensing arrives with latency, at different rates, from different vantage points; there is a frontier of arrival times, not one clock. Multiplying streams could fragment the origo rather than restore it, which would mean the terminal claim overstates what continuous ingestion actually buys.
More sensors is more clocks, not one better clock.
The concession is real: there is no single simultaneous now available to a many-streamed system, and treating provenance as if it produced one would be a mistake. But fragmentation is a calibration problem with known engineering answers — hybrid logical clocks, vector clocks, bounded staleness — not evidence that no anchor exists. The right description is a bounded, ordered account of which now applies to which claim, made explicit rather than assumed. That is more honest than a corpus's single silent fiction of a moment, even though it is messier to state.
What this does and does not settle
Deixis establishes that the intake axis is semantic, not merely a matter of freshness. A stale corpus is not simply out of date; a determinate class of ordinary expressions has no content at all without a live origo, and no amount of scale substitutes for one. That is a real result, and it is narrower than it might sound. It does not show that frozen training is useless — most knowledge is not deictic, and character alone carries most linguistic competence. It does not show that continuous ingestion resolves cleanly into a single now — it resolves, at best, into a well-calibrated plurality of nows with provenance attached. And it does not establish that a persistent origo is achieved by any system merely for running continuously; a badly calibrated stream is not better than a well-chosen static anchor. What it establishes is narrower and firmer: that reference is either computable from where you stand or it is not, that there is no fourth kind of anchor waiting past the plural, provenance-tagged one, and that everything past this point on the intake axis is a question of precision, not of category.