Two things a word carries
Say "the morning star" and "the evening star" to someone who knows Venus by neither name, and both phrases will strike them as informative. Say "the morning star is the evening star" and the sentence carries news, not tautology. That is odd. If both phrases pick out the same object, the sentence ought to say only that a thing is itself, which is empty. It does not feel empty. It feels like astronomy.
The resolution is to separate two things a word or phrase carries. One is its sense: the mode of presentation, the route by which it picks out its object. "The morning star" presents Venus as the bright point before dawn; "the evening star" presents it as the bright point after dusk. The other is its reference: the object actually picked out, Venus itself, indifferent to which route led there. Sense is what you grasp when you understand the phrase. Reference is what the phrase latches onto in the world. They usually travel together, which is why the distinction is easy to miss and easy to need.
They can also come apart, and this is the part that matters beyond astronomy. A phrase can keep a perfectly intact, perfectly intelligible sense while its referent moves, changes, or stops existing, and the phrase itself gives no sign of the loss. "The current heavyweight champion" has a stable sense: whoever presently holds that title. Its referent has changed many times, and the phrase never once faltered in the changing. The mode of presentation is a standing procedure; the object it lands on is whatever satisfies that procedure now. Nothing about the grammar tells you when "now" was.
Where it comes from
Gottlob Frege set this out in "Über Sinn und Bedeutung", 1892, to solve exactly the identity-statement puzzle above. Before Frege, the working assumption in logic was roughly that a name's whole contribution to a sentence's meaning was the object it named. That assumption cannot explain why "a = a" is trivial while "a = b" can be a discovery, when a and b in fact name the same thing. Frege's answer: names and descriptions carry sense in addition to reference, and identity statements can be informative because they assert that two different routes converge on one object. The distinction proved durable well beyond the puzzle that produced it. Russell reworked definite descriptions on partly different lines; Carnap built intension and extension on the same fault line; Church formalised it; Kripke, seventy-odd years later, attacked a large piece of it directly. The vocabulary survived every attack in some form, because the phenomenon it names — understanding a phrase without thereby knowing, or continuing to know, what it currently points at — keeps recurring.
The turn
Now put a large corpus in front of the distinction and watch what training on it actually preserves.
A Large Language Model is trained on text collected once, closed at a cutoff date. What survives that closure with remarkable fidelity is sense. The model learns, with extraordinary density, the routes by which expressions present their objects: what "the applicable rate" keeps company with, what follows from "the buoy marking the shoal edge", what a dosage phrase implies about form and unit. This is not a minor achievement. It is most of what competent language use consists of.
What does not survive closure is reference, and it cannot, because reference is not a fact about language. It is a relation to a world that kept moving after the corpus stopped. "The current maximum recommended dose", "the ticker for that airline", "the officer holding that post" — every one of these keeps its sense indefinitely. The objects they used to pick out are, in cases that matter, no longer the objects they do pick out. Alcoa has traded as "AA" for decades; American Airlines Group took the symbol "AAL" on Nasdaq in December 2013, and no grammatical feature of a sentence written before or after that date marks the handover. "AAL closed higher" is equally fluent, equally well-formed, on both sides of the change, and equally silent about which company it now names. The FDA's 2011 cap on acetaminophen in prescription combination products, phased in over roughly three years, left "the standard 500 mg acetaminophen component" reading as competent clinical prose describing a formulation that had, by the compliance date, no legal shelf presence. A weekly Notice to Mariners can relocate a buoy four hundred metres; "the red nun marking the eastern edge" still steers a reader's imagination to a fixed point that is no longer where the light is.
This is why fluency is the wrong reassurance. A frozen system produces sentences that are grammatical, idiomatic, internally consistent, and unmoored, and it is exactly the internal consistency that hides the severed mooring. Nothing in the sentence announces "the referent moved." Nothing in a sentence ever does; that was Frege's whole point, applied now to a machine that has read everything and observed nothing since a fixed date.
A Large World Model repairs this, but only locally and only for as long as it looks. Given a scene, it can re-anchor "the valve on the left" or "the vessel in frame" to the actual object in front of its sensors, because it is sensing now, not recalling a corpus. When the scene ends, the anchoring lapses. The sense persists in the model's language; the reference goes dark again the moment the camera turns away.
A Large Universe Model treats anchoring as an obligation that does not end. Every referential commitment — this is the current dose, this is the buoy's position, this is the isolator on feeder three — carries a source and a time index, and is held open to revision by streams that keep running. This is the terminal move on the intake axis specifically because there is nothing further to add to "continuously, with provenance." You can observe more things. You cannot observe more continuously than continuously.
The misreading to disown
The weak version of this argument says: frozen systems are unreliable because they lack knowledge, and the fix is a bigger, fresher corpus. Both halves are wrong. A frozen system does not lack knowledge of sense — it has an unusual surplus of it, and that surplus is precisely what makes the failure hard to catch. And a fresher corpus is still a corpus. It moves the cutoff forward; it does not remove the cutoff. The same severance recurs on the same schedule, just later. The real distinction is between having a date stamped on a belief and having a live relation to what the belief is about. It is not between an old date and a recent one.
What narrows the claim
Three objections earn a hearing, and one genuinely cuts the argument down to size.
If reference is fixed by causal chains from an initial baptism, as Kripke argues for proper names, then freezing a corpus does not sever it. "Aristotle" refers to Aristotle regardless of cutoff.
True for names, and it matters. But the expressions that actually fail in practice are role terms and definite descriptions — the current rate, the applicable code, the ticker for that issuer — which Kripke himself does not treat as rigid. The scope narrows to non-rigid designation, which turns out to be most of what operational language is built from.
Retrieval attached to a frozen model restores reference at query time without any architectural claim to continuous intake.
Also true, and often the cheapest fix available. But it restores only what someone else indexed, and it silently inherits whatever freshness discipline that index keeps or fails to keep. It cannot re-anchor a valve position nobody logged. Retrieval is continuous intake with the obligation moved elsewhere, not continuous intake dissolved.
Observation alone does not settle identity — whether this vessel is yesterday's vessel, whether a renamed entity is the same legal person — since that is a conceptual decision, not a measurement.
This is the sharp one, and it genuinely limits the claim. Watching a stream does not by itself tell a system what counts as the same thing over time; that is a criterion someone must supply. What continuous intake secures is narrower: once identity criteria are fixed, only an unbroken stream can keep them applied. A frozen system cannot apply them at all past its cutoff. Continuous intake is the precondition for tomorrow's answer being true. It is not the thing that decides what the question means.
What this does and does not establish
The concept establishes that staleness in a frozen system is a semantic failure with no linguistic symptom, not merely an outdated fact among others, and that only standing, dated, revisable observation closes it. It does not establish that intelligence requires this closure, that every task needs it, or that a system lacking it is useless — most sense-bound reasoning never touches a referent that moved. It marks where one axis, intake, runs out of further rungs. It says nothing about the other axes the lineage runs along.