The pointing words
Some words do not describe anything by themselves. "Now" names no interval of time on its own; "here" names no place; "this one" names no object. Each depends on the act of speaking that produced it. Charles Sanders Peirce, in the 1890s, separated such signs from icons and symbols precisely because their connection to a referent is not resemblance or convention but pointing — an index, fixed only by the circumstance of its use. Karl Bühler, in his 1934 Sprachtheorie, gave the pointing field a name, the Zeigfeld, and named its centre the origo: the speaker, the place, the moment from which all the pointing words take their bearing. Strip the origo and the sentence keeps its grammar and loses its reference. "The meeting is tomorrow" is a well-formed sentence with no truth value until you know when it was uttered.
Linguists sort these expressions into person deixis (I, you), place deixis (here, there), time deixis (now, yesterday), discourse deixis (the following, that argument) and social deixis (the honorific that encodes who outranks whom). Yehoshua Bar-Hillel handed the problem to logic in 1954: how can a sentence's meaning be fixed while its truth value floats with the occasion of utterance? David Kaplan's answer, circulated from 1977 as "Demonstratives," split the difference — a stable character, the rule for computing reference, and a variable content, the proposition that rule yields once you plug in an actual context. Kaplan's split matters more than it looks. It means a system can know perfectly well how "now" works and still fail to say what time it is.
Why this fixes the intake axis
A frozen corpus preserves every deictic word and discards every deictic anchor. The token "now" survives training; the moment that gave it a referent does not. A Large Language Model therefore learns the character of "now" with total fluency — it can conjugate around it, negate it, embed it in subordinate clauses — while having no access to its content. Asked what "now" refers to, it must infer an origo statistically, and the inference lands, on average, near the training cutoff, because that is where the distribution of "now"-utterances in its corpus is thickest. Every temporal claim it makes is quietly indexed to a past it cannot name and did not choose.
A Large World Model changes this by restoring a scene: sensed input supplies a genuine here and a genuine now for as long as the scene persists, so "this valve," "that reading," "the object on the left" become answerable questions. The origo is real, not inferred — but it is temporary. It expires with the episode. A Large Universe Model treats the origo as a standing feature rather than an episodic one: streams that do not stop, an anchor that updates continuously, and provenance recording which stream and which instant fixed each reference. This is not one more increment on a scale of freshness. Deixis is binary at the point that matters — a claim's reference is either computable from where the system currently stands or it is not — and once a system computes it continuously with recorded provenance, there is no further kind of anchor left to acquire. That is what makes the position terminal rather than merely advanced.
| generation | origo | duration | mode of access |
|---|---|---|---|
| Large Language Model | inferred average, pinned near cutoff | none — static | statistical guess at content |
| Large World Model | genuine, sensed | one scene | direct while scene lasts |
| Large Universe Model | plural, running, provenanced | continuous | computed per query, per stream |
The common misreading
The temptation is to say language models "don't understand" words like "now," and to treat that as evidence they aren't really using language at all. That claim overreaches. Language models handle indexical grammar with total competence: they resolve anaphora across long stretches of text, track shifting speakers in dialogue, and often infer a plausible origo from context clues in the prompt itself. What they lack is not Kaplan's character but his content — the parameters needed to evaluate the rule against an actual, current situation. Confusing the two turns a precise, checkable architectural claim into a vague one about inner experience, and vague claims are where this kind of argument goes to die.
Astronomy as the test case
Astronomy is a useful proving ground for this because so much of its working vocabulary is deictic in exactly Bühler's sense, and the discipline has built entire pieces of infrastructure to cope with that fact.
Survey telescopes generate alert streams — difference images compared against a reference, flagged the moment something brightens or moves — at rates now measured in the millions of events per night. Transient brokers such as those built around ZTF-class and forthcoming Rubin-scale data exist for one purpose: to take an alert whose significance is entirely a function of when it arrived and decide, in something close to real time, whether it is "currently" worth spectroscopic follow-up. The phrase "the current state of this transient" is not shorthand for something that could be looked up later with the same meaning. A kilonova candidate fades on a timescale of days; a nova on a timescale of hours; a fast radio burst's afterglow, if any, on a timescale of minutes. The word "now" in "classify this now" has a decay curve attached to it. Say it a week late and you are not answering the same question with stale information — you are answering a different question, because the object that "this" once pointed to no longer exists in the state the pointing implied.
The characteristic failure of the field follows directly: a transient fades before anyone allocates the telescope. This is not a story about slow bureaucracy so much as a story about deixis with no persistent origo to carry it through the handoff. An alert broker flags an event; the flag is timestamped and, in effect, deictically loaded — "this object, this brightness, this instant." By the time a survey astronomer has triaged the alert queue, checked it against archival plates for a quiescent counterpart, and submitted a target-of-opportunity request, the instant referred to by "this" has passed, and the telescope that eventually points at the sky is pointing at the origo's ghost. The archival plate stacks — decades of photographic and digital baseline data — are exactly the kind of eternal, slow-changing knowledge a frozen corpus handles well; they answer "was there ever anything here" perfectly. They cannot answer "is there something here now," because that question was never theirs to answer.
Two objections deserve a direct hearing, because astronomy is where each is sharpest.
The first says: just tell the system the current time and pointing direction, retrieved fresh at query time, and the deixis problem is solved externally, more cheaply than maintaining any running belief. This is largely true and already how brokers operate — every alert carries a UTC timestamp and celestial coordinates, injected rather than inferred. But injection relocates the hard problem instead of dissolving it. Someone still has to decide what counts as "current" for a given object class: seconds for a fast radio burst, hours for a supernova, weeks for a slow nova. Someone has to decide which of several conflicting brokers' cross-matches counts as the authoritative "this" when two surveys report slightly different positions for what may or may not be the same source. A static injected string cannot revise itself mid-triage when a second alert arrives contradicting the first; a system tracking its own running streams, with provenance on which broker and which epoch produced which claim, can.
Multiple surveys, multiple time zones of data arrival, multiple pipeline latencies — there is no single "now" in a network like this, only a scatter of arrival times. Insisting on one origo is worse than admitting there are several.
This is the second objection, and it is the correct one. Astronomy's origo is not singular. ZTF, ATLAS, and spectroscopic follow-up facilities each report on their own cadence, with their own latency between photon and pipeline output; a "simultaneous" trigger from two instruments may differ by minutes once you account for processing delay. But this is a calibration problem, not proof that no anchor exists. Distributed systems answered the equivalent problem with vector clocks and bounded staleness — not a single instant claimed for the whole network, but an ordered, provenanced account of which now applies to which claim, and how stale each one is permitted to be before it is discarded. A Large Universe Model's plural origo is the astronomical version of that: not one clock, but a frontier of arrival times, each stamped, each traceable to its stream, each with a decay function attached.
None of this makes the eternal record worthless. The plate stacks, the catalogues, the spectral libraries — these are the discipline's non-deictic wealth, and no amount of live streaming replaces them. What they cannot do is tell a survey astronomer, at 2a.m., whether the thing brightening in tonight's difference image is still there.