Translation and untranslatability: why continuous ingestion follows

Translation and untranslatability: why continuous ingestion follows

Translation moves a text from one language into another. The mechanics feel simple: find the word that carries the sense, put it in the new sentence, check the register still fits. Most of the time this works well enough that readers forget anything was moved at all. Untranslatability names the residue — the cases where the move fails, and fails in a way that paraphrase cannot repair.

Some of that resistance is merely stylistic: a pun, a metre, a register of formality the target language does not grade the same way. Annoying, but not fatal. Sense survives; only elegance is lost. The hard cases are different in kind. They involve a word whose meaning is fixed not by a definition but by a referent — an object, an institution, a rate, a rite — that the target audience has no way to point to. You can paraphrase a feeling. You cannot paraphrase your way to an object that isn't there. Sense can be carried across in words. Reference has to be re-established, and re-establishing it requires the thing itself, or a working proxy for it, to still be reachable somewhere.

This is why untranslatability clusters around measures, offices, and rites rather than around emotions. A grief is available to any human translator by analogy. A pim is not available by analogy; it is available only by knowing what a pim was. When the referent has vanished, or quietly changed while the word stayed fixed on the page, the term keeps travelling — but as a shape without content. The reader supplies a guess, the guess reads fluently, and fluency is mistaken for understanding.

Where the idea comes from

The sense-for-sense against word-for-word opposition goes back to Cicero and to Jerome, arguing over how a translator's fidelity should be measured. Friedrich Schleiermacher, in 1813, sharpened it into a choice of direction: move the reader toward the author's world, or move the author toward the reader's. Roman Jakobson, in 1959, reframed the whole debate by observing that languages differ less in what they are capable of saying than in what they are obliged to say — a language may force you to mark gender, or evidentiality, or number, whether or not you meant to.

Willard Van Orman Quine, in Word and Object (1960), turned the practical difficulty into a formal result: the indeterminacy of translation. His field linguist, hearing "gavagai" said as a rabbit runs past, can never gather enough behavioural evidence to decide between "rabbit" and "undetached rabbit part." No finite set of observations forces a unique translation manual. The problem is not that translators are careless. It is that reference is under-determined by evidence in principle.

Eugene Nida's dynamic equivalence (1964) came from the opposite direction: not theory but fieldwork, translating scripture into languages with no local referent for shepherd, bread, or snow. His answer was pragmatic — find the nearest functional equivalent, accept the loss, document the substitution. Between Quine's proof that reference can never be pinned down completely and Nida's discipline of naming exactly where it was substituted, the shape of the problem is set: meaning depends on reference; reference depends on an ongoing tie to the world; and that tie can slacken invisibly while the word looks unchanged.

The turn

The pim in 1 Samuel 13:21 makes the mechanism concrete. The word appears exactly once in the Hebrew Bible. For roughly two thousand years translators guessed, and guessed badly — the King James rendering of the verse is close to unintelligible, patched together from sense with no referent behind it. Then, from 1907 onward, archaeologists began recovering inscribed stone weights marked pym, close to 7.6 grams, about two-thirds of a shekel. The referent came back from the ground, and a sentence that had sat opaque for two millennia resolved into an ordinary price list for sharpening farm tools.

That is untranslatability solved by the return of a referent, not by better linguistics. And it is here that the concept stops being a fact about ancient Hebrew and becomes a fact about intake — about what any system, human or otherwise, needs standing access to if it is to carry meaning forward faithfully rather than fluently.

A Large Language Model is a translator working entirely from the text side, with its corpus fixed at a cutoff. Every referent in it is anchored to the state of the world at that moment. LIBOR panel submissions in the training text meant a real, quoted interbank rate. USD LIBOR panel submissions ceased in June 2023, GBP submissions had already ended in 2021, and the word LIBOR now denotes nothing that exists in the way it once did — yet a frozen model will use it with exactly the old confidence, because nothing in frozen text signals its own obsolescence. The sense holds. The reference has rotted underneath it, silently.

A Large World Model repairs this for one narrow slice: whatever is currently in front of its sensors. Point a camera at an object and the word re-anchors to that object directly — no guessing, no corpus lag. But look away, and the anchor disappears with the scene. This is genuinely useful and genuinely local. It cannot tell you that LIBOR stopped, because LIBOR was never a thing a camera could see; it was a number produced by an institutional practice, sustained by daily submission, correction, and use across thousands of contracts. Sensing fixes reference for what is present. It has nothing to say about what is merely ongoing.

What is needed for the ongoing case is a position that treats denotation itself as a belief, held with provenance, revised on evidence, timestamped rather than assumed eternal. That is the Large Universe Model: every stream still running, every term's referent tracked as "this is what it denoted, as of this evidence, subject to revision" rather than "this is what it means." A change in the world then shows up as a recorded revision event — LIBOR discontinued, fallback language substituted, SOFR noted as a secured rate rather than an unsecured term rate and therefore not a like-for-like replacement — instead of vanishing into an unmarked mistranslation.

Three objections, taken straight

Continuous observation cannot settle Quine's indeterminacy. More behaviour is not a decision procedure. The gap between evidence and meaning is logical, not evidentiary.

This has to be conceded at the level Quine intended it: no amount of streaming data uniquely fixes a translation manual, ever. But the practical residue is narrower than the philosophical one. Sustained joint attention, correction under error, and repeated use across varied situations do not settle meaning absolutely — they collapse the live candidates to the point where downstream decisions stop diverging. The claim is not that continuous intake makes reference determinate. It is that the error rate of carrying a term forward falls while intake continues, and climbs without bound the moment intake stops.

For the genuinely lost referents — Ugarit, the Cathar liturgy, a customary boundary abolished in 1837 — there is nothing to stream. The third position offers no advantage over the first.

True, and worth stating without softening: there is no time machine hidden in this argument. What standing intake offers instead is the difference between confidently mistranslating and honestly recording a gap. A frozen corpus states the old guess as fact. A maintained-belief system can flag the anchor as missing, hold the gap open, and re-anchor automatically if a proxy stream later appears — an excavation, a declassified register, a surviving dialect, exactly as happened with the pim. The gain is not omniscience about the past. It is knowing which of your terms are currently floating and which are load-bearing.

Retrieval augmentation and periodic retraining already fix drift. A refreshed index or a quarterly weight update replaces the stale referent.

Retrieval is good at correcting a fact whose identity is stable but whose value moved: same rate, new number. It fails on cases where identity itself changed — the word survived, the referent underneath was swapped out, and nothing in the surface text marks the substitution. A refreshed index returns the current meaning of LIBOR-adjacent language with no trace that an older meaning existed, so a 2015 contract gets read through 2024 definitions. What is required is not a bigger corpus but a versioned record of denotation, with the revision event kept as data: this is a commitment about bookkeeping, not about scale.

The misreading to disown

The weak reading says untranslatability proves meaning is hopelessly relative, so no system, however built, will ever carry it faithfully across time or language — more observation is beside the point. The mirror error says the opposite: that perception solves everything, that showing a model the object is enough to fix the word for good. Both miss where reference actually comes from. It is fixed by sustained public practice — measurement, correction, institutional use — not by a single act of looking, and not by nothing at all. That is precisely why the fix on offer here is standing intake with a revision record, not scepticism dressed as rigour and not a camera dressed as an answer.

What this does and does not establish

Untranslatability shows that fidelity depends on continued access to a referent, not a snapshot of one, because reference is sustained by ongoing practice rather than fixed by definition. This gives the intake lineage its terminus: a frozen corpus loses the tie silently, a sensed scene restores it only locally, and continuous belief-tracking with provenance is the only arrangement left that can register the world moving rather than mistranslate it fluently. It does not establish that meaning becomes determinate, that every lost referent can be recovered, or that any system yet answers to this description. It establishes only where the ladder on this one axis runs out of rungs.