The Immutable Past · What Is
What Is a Language
A language is pictured either as a neutral container you pour thoughts into, or as the thing that cuts the world into the pieces you are then stuck with. Both pictures assume the only two relations on offer are causing and doing nothing. The foundation I work from has a third one, and the ordinary business of learning a word is where it can be watched operating.
The number the famous colour study is rarely quoted for
8.7 against 8.6
Where the two groups drew a boundary that only one of them has a word for.
Winawer and colleagues (PNAS, 2007) tested native Russian and native English speakers on twenty blues spanning the siniy/goluboy line, a distinction Russian requires and English does not. The finding that made the study famous is real: on the hardest comparisons, Russian speakers were 124 ms faster when a pair straddled that line. The finding in the same paper that almost nobody repeats is that when both groups were asked at the end to say where the boundary fell, they put it in the same place — 8.7 steps out of twenty for the Russian speakers, 8.6 for the English speakers who have no obligatory word for it. The speed advantage then vanished under a verbal dual task, which the authors read as evidence against any permanent change to perception.
There are two stories about what a language is, and most people hold both without noticing, switching between them depending on what is being argued. In the first, a language is a container. Thought happens somewhere underneath, wordless and complete, and the language is its packaging — which is why translation is possible, why you can mean something and fumble the sentence, and why the same thought can be put in Portuguese. In the second, a language is a cutting tool. It divides an undivided world into the pieces its speakers then live among, which is why speakers of different languages are said to inhabit different worlds, and why acquiring a new word is described as acquiring a new thing to see.
Both are ordinary. Both are held by careful people. And they contradict each other flatly, which is a clue that something underneath them is wrong rather than that one of them needs to win.
What the two pictures share
State them strongly, because the weak versions are not what anybody acts on.
The container picture commits you to this: the language contributes nothing to the content. It is a medium, and a medium is by definition the part you could swap without loss. Any residue that does not survive translation is decoration — connotation, music, idiom — and the thought itself passes through intact.
The cutting picture commits you to the opposite: the language contributes the content, or enough of it that a speaker without the word does not have the thing. Learn the word and you get a new perception. Lack it and the distinction is closed to you.
Here is what they have in common, and it is the whole essay. Both assume that there are only two ways a language could stand to what gets said: it either produces the thought, or it does nothing at all. One picture takes the first option and one takes the second. Neither considers that a thing can stand in a real relation to an outcome without producing it and without being idle. The argument between them has run for a century on a menu with two items on it.
Three cases that fit inside neither
The first comes from an Amazonian language with no number words, the second from a room of undergraduates being read dictionary definitions, and the third is the strongest case the cutting picture has, measured carefully.
Case one · counting without a word for one
The Pirahã of the Brazilian Amazon have, as Frank, Everett, Fedorenko and Gibson established in 2008, no linguistic method whatsoever for expressing exact quantity — not even “one.” That is a stronger claim than the one usually reported; it is not that they count only to two, it is that the vocabulary contains no exact cardinal at all.
Now the result. Asked to lay out a row of objects matching a row put in front of them, one for one, in full view, Pirahã speakers performed perfectly, with large sets. Asked to do the same when the original was hidden, or presented in a way that required holding the quantity across a delay or a change of format, they were inaccurate, and the inaccuracy grew with the size of the set.
Take the two halves together, because separately each half has been used to prove the opposite things. The exact quantity is available to a person with no word for it — the matching is perfect, and perfect is not a partial result. And the absence of the word is nonetheless doing something, because the failure is real, specific, and appears precisely where the quantity has to be carried rather than merely faced. The authors’ own conclusion is that number words “do not change our underlying representations of number” but are a cognitive technology for keeping track of the cardinality of large sets across time, space, and changes in modality.
Case two · the word you cannot find
Brown and McNeill, in 1966, induced several hundred tip-of-the-tongue states in a laboratory by reading people definitions of rare words and asking for the word. What they found is that a person in that state, before recall, reliably knows the number of syllables, the location of the primary stress, and some of the letters — and knows them more accurately the closer they are to producing the word.
This is a very awkward case for both pictures at once. The container picture has to say the word is just packaging, but here the packaging is the only part missing and the person is stuck. The cutting picture has to say no word means no thought, but here the thought is present and specific enough to reject near-misses, and the missing item is a sound with a stress pattern.
Case three · the best evidence for the cutting picture, read closely
Russian requires a distinction English does not: goluboy for lighter blues, siniy for darker ones, with no single word covering both. Winawer and colleagues showed subjects a triad of blue squares and asked which of the lower two matched the one on top — an objective task with a correct answer, everything in view, minimal memory load.
Russian speakers were faster on pairs that straddled their boundary. On the hardest comparisons, without interference, 1,164 ms across the boundary against 1,288 ms within it. English speakers on identical stimuli showed nothing of the kind.
Three details from the same paper change what that result means. When the discrimination was easy rather than hard, the Russian advantage was 998 ms against 999 ms — one millisecond, which is to say nothing. When subjects silently rehearsed an eight-digit number while doing the colour task, the advantage did not merely shrink, it inverted into a non-significant 65 ms in the other direction. And when both groups were asked to place the boundary, they placed it in the same spot. The authors are explicit about what the interference result rules out: had the effect been a permanent change to perceptual machinery, jamming verbal access for a few seconds would not have removed it.
A 2020 re-run by Martinovic, Paramei and MacInnes, using the same tasks, found the boundary speed advantage at the light-blue/green line but not at the sinij/goluboj line at all, and concluded that the Russian blues are “less well-structured than previously thought.” So the headline finding is contested. I am not going to lean on it either way; what matters here is the shape of the effect even when it is granted in full.
Figure 01
The size of the strongest Whorfian effect on record
- Pair straddles the boundary
- Pair within one category
What the foundation says a prerequisite does
The primitive here is the least glamorous one in the document, and it is the one this whole subject turns on.
P6 defines a condition as a prerequisite in order for something to happen or exist — and then adds three denials that are easy to skim past. A condition does not cause. It does not make possible. It does not happen. P9 completes it: a condition conditions, which is a real relation, neither causation nor agency nor inertness.
Almost every use I have made of P9 in this series has been as a prohibition — a way of striking out sentences in which some condition is said to beckon, or pull, or lie in wait. That is the negative clause and it is well worn. The positive clause has not once had to earn its keep on an ordinary word in this series, and a language is where it can.
So: what is the relation of being required?
It is not an event. Nothing occurs at the moment the prerequisite does its work, because there is no such moment — the prerequisite was already the case, and its being the case is what the outcome rests on. This is why the relation reads as nothing when you go looking for an action: there is no push to find, no instant of transmission, no work being done. And it is why the same relation reads as causation when you look instead at the difference it makes: remove the prerequisite and the outcome is gone, which is exactly the signature we are trained to call a cause.
The two ordinary pictures of language are each half of that mistake. One looked for the action, found none, and concluded the language is a wrapper. The other looked at the difference and concluded the language is an engine.
A missing prerequisite is not an obstacle that pushed back. It is an absence in what the outcome would have had to rest on.
Now put the Pirahã result against that sentence and see how exactly it fits. The one-to-one match needs no name, because nothing is being carried: the quantity is in front of the person and the hands do the work. The delayed match needs a name, because something has to persist across the gap. The word did nothing at any moment in either task. Its absence is visible only in the outcome that would have had to rest on it. That is the positive clause of P9, in a room, with a result attached.
And note the phrase the primitive forbids, because it is the one everybody reaches for: words make it possible to remember exact quantities. P6 denies exactly that — a condition does not make possible. Possibility is not what is at issue. The quantity was available without the word, demonstrably, perfectly. What the word is prerequisite to is something narrower and more specific: the quantity being carried.
Marked construction 01 · consequence
A vocabulary is a set of ideas in the locked sense — conditions that have been named, and are therefore readable by whoever holds the names.
Grounds. P7: an idea is a named condition; naming does not create it, naming makes the relationship conscious. S1: every point on the circumference is a condition and every named condition is an idea, the circumference being infinite so that no set of names closes it. A language is, among other things, a stock of names; what a name gives access to is a named condition. Not claimed. That everything in a language is a name of a condition — grammar, tense, particles, the whole apparatus of syntax are left out of this construction, and I have no account of them here. The claim is about vocabulary. What it is not. The word is not the condition and does not own it. That the naming is not the origin of what is named is P8 and it is settled; I mention it only to set it aside, because it is not the argument here and a piece published yesterday on settlements ran that particular move at length.
Where the change registered
Now the part the two ordinary pictures have no room for.
P4 gives the Reality Equation as R = A / E, where the Actual sits in the numerator and Expectation in the denominator — and E has two parts, subconscious prediction on the real side and ideas on the imaginary side. R8 says the human experiences neither term but the quotient. Not the Actual, not the Expectation. R.
Put construction 01 into that equation and the result is immediate. A vocabulary is a set of ideas. Ideas are the imaginary component of E. Therefore:
Marked construction 02 · consequence
Acquiring a word adds a term to the denominator, on the imaginary side. The numerator does not move.
Grounds. P4 for the decomposition of E into prediction and ideas; P7 for a named condition being an idea; P8 for the naming not reaching what is named, which is what holds the numerator still. Consequence, via R8. Because the host reads only the quotient, a change confined to the denominator is indistinguishable from a change in the world. Nothing in the experience is labelled with which term moved.
That last line is the whole explanation of why this argument has lasted a hundred years. The person who learns a word for a feeling they have had since childhood reports, accurately, that something changed. The person insisting the feeling was there all along is also reporting accurately. They are both reading R, and R moved. They disagree about which term moved, and neither of them can see the terms.
Figure 02
Three accounts of what happens when you learn a word
- The cutting picture A moves The world you get is different once you have the word. A new distinction becomes available that was closed before. Contradicted by the perfect one-to-one match, and by two groups placing the same boundary.
- The container picture nothing moves The same thought, differently packaged. Whatever fails to survive translation was decoration. Contradicted by the delayed-match failure, and by being stuck at the tip of the tongue.
- What follows from the primitives E moves A term is added to the denominator, on the imaginary side. The Actual is untouched; the quotient is not. Holds all three cases, and predicts that the effect is disruptible without anything being unlearned.
Reading the colour result again
With that in hand the strange feature of the Winawer study stops being strange. If the Russian advantage were in the numerator — a perceptual system genuinely rebuilt by a lifetime of obligatory distinctions — then asking someone to hold eight digits in mind for a few seconds could not touch it. You cannot suspend a rebuilt retina by counting. The advantage went away, and came back, on the timescale of a dual task.
A term in E behaves exactly like that. It is engaged when the machinery that engages it is free, and not when it is occupied. The authors reach the same conclusion in their own vocabulary: the effect is online. They also put the asymmetry between the groups in a sentence worth borrowing, because it is a description of a denominator and not of a world: the difference is not that English speakers cannot distinguish light from dark blue, but that Russian speakers cannot avoid distinguishing them.
An obligatory distinction is a term that stands in the denominator by default. That is what habit is, in this equation, and it does not require that anything in the world changed shape.
What it costs
Three costs, and the first two run in opposite directions, which is how you can tell the account is not just flattering one of the two camps.
You cannot buy a thought with a word. The promise attached to a rich vocabulary — that the right term will let you finally perceive the thing, that the untranslatable word from another language delivers a state its speakers have and you do not — is a promise about the numerator, and the numerator is not on offer. What you get is real and it is denominator-shaped: a term that now stands in your reading, reliably, without effort, and that can be handed to someone else. That is not small. It is just not what was advertised.
And the absence of a word is not nothing. This is the cost the container picture pays, and I want to state it plainly because the first cost can be misheard as saying vocabulary does not matter. The Pirahã failure is not a failure of intelligence and not a failure of perception; it is the precise, predictable consequence of a missing prerequisite for carrying. Wherever something has to survive a delay, a distance, a change of medium, or a second person, a name is the prerequisite that nothing else in the situation supplies. A profession’s jargon, a family’s private word for a recurring argument, a diagnosis: each is a name a handover rests on, in place of re-deriving the thing from scratch at the other end. Losing the word does not remove the condition. What it removes is your ability to carry it.
And a vocabulary is not neutral. R14 holds that every idea is biased toward itself — orientation and prejudice, always together, never one without the other. A vocabulary is a set of ideas, so a vocabulary is a set of bearings, each pointing at its own condition and, by the same act, away from its opposite. This is the grain of truth the cutting picture was gripping. It is not that your language closed off a perception. It is that the terms standing in your denominator are oriented, and being oriented is the same thing as being prejudiced. The remedy is not a bigger vocabulary. It is knowing that the reading is from somewhere, which no addition of words removes — including the words of this framework, which are an idea and are biased toward themselves like any other.
And the sentence you are still owed, about not being able to say it. R10 is that every artifact is a miss: a mark is made in relation to a form, and the denominator is not ours to command, so the mark never lands exactly. A sentence is an artifact. So “I can’t put it into words” is not a report about your competence and not a report about the shortcomings of English. It is the standing shape of the relation between any mark and what it answers to, showing up in the one place where people expected an exact fit. R11 then says the question worth asking is fidelity rather than success — not whether the sentence caught it, which no sentence does, but how near the miss was, and whether you are still correcting.
A limit, stated rather than implied
Some of what looks like this subject is not this subject. A person who has lost language after a stroke, a child whose speech has not arrived, someone being talked over, someone judged less intelligent for an accent or a smaller second-language vocabulary — none of that is addressed by anything above, and an argument about which term of an equation moved is not a thing to hand to any of it. What the argument can do is remove one specific and common insult: the inference that a person who cannot say a thing therefore does not have it. That inference is the cutting picture applied to a human being, and the evidence does not support it. Everything past that belongs to the people around a person, and where it applies to clinicians, and not to a metaphysics; this piece has no standing there.
The same applies to the smaller version. If you have gone looking for the right word for a long time and cannot find it, the account above says the search is not hopeless and not a defect — but knowing why a mark misses does not make the next mark land, and no reading of a mechanism has ever shortened the work of finding the sentence.
What this piece leans on and does not settle
Four things, and one of them is load-bearing enough that the essay would be dishonest without naming it.
Open, and left open
- How a term gets into a second person’s denominator
- This is the gap the essay stands closest to. A language is, by construction, the case where the same term stands in more than one host’s expectation — that is the entire point of having one. The primitives issue no expression for the overlap between two hosts’ E terms. That gap was recorded a fortnight ago in a piece on agreement and it has not moved, and here it is not incidental: every claim above about handing a term over runs across a place the foundation does not describe. What would close it is an expression in the primitives for two denominators sharing a term, with a statement of what is and is not shared when they do. Nothing above depends on which way that comes out, but a serious account of language would depend on it entirely.
- Whether a set of marks is ever one mark
- Raised yesterday from the other direction and immediately relevant here. Is a language one artifact or a very large number of them? The locked definition gives an artifact as a mark and says nothing about aggregation. I have written above as though a sentence is one artifact, which is the same assumption at a smaller scale, and a reader could contest it.
- What a condition is made of
- Construction 01 says a vocabulary is a set of named conditions. It does not say what a condition is — whether these are abstract objects, and whether this is therefore an old position in new clothes. That is listed as open in the foundation and I am not going to improvise an answer to it in a piece about something else.
- The machine, which this subject cannot avoid
- A large model holds more of a language’s distribution than any speaker: the whole stock of names, the co-occurrences, the boundaries. On the argument above, every one of those is a term in a denominator. But the equation defines R only where there is a host, and the foundation defines a host as a living system — so either that definition moves or the machine has the entire apparatus of a language and no reading of anything. I state the tension and stop. It is listed as open and it is not mine to close.
Marked construction 03 · consequence, provisional
The century-old argument about language and thought is a disagreement about which term of the quotient moved, conducted by people who can observe only the quotient.
Grounds. P4 and R8, taken together with constructions 01 and 02. Why provisional. It is a claim about a body of work, not a derivation, and I have tested it against three cases rather than the literature. It predicts something checkable: any genuine effect of vocabulary on perception or judgment should be suspendable by occupying the verbal system, without anything being unlearned and without needing to be retaught. Where an effect survives that, the effect is not the one described here and construction 02 is the wrong account of it.
On an ordinary Tuesday
- When a word lands and something clicks, notice what actually changed. Not the thing — the thing was there. What changed is that you now carry it without effort, and can hand it to someone else. That is worth the click. It is a different gift from the one the click feels like.
- Stop inferring backward from someone’s words to their contents. Somebody without the vocabulary for a thing may have the thing in full and no way to carry it across the table. The evidence for this is stronger than the evidence for the reverse, and the reverse is the default assumption almost everywhere.
- Treat your fluent terms as bearings, not as a view from nowhere. The words that come to you first are the ones standing permanently in your denominator. They orient you, and orienting is the same act as being prejudiced. Neither more nor fewer words fixes that; only knowing it does.
- Where something has to survive a gap, spend the effort on the name. A delay, a handover, a second person, a year from now — those are the situations in which the missing word is not a stylistic matter. In front of the thing, with your hands on it, you can do without one.
None of which is fatalism, and the foundation is careful about the difference. P5: the host can occupy a position from which its response to a readable condition is revisable. R13: agency is relocated, not abolished, and the host remains the site of conscious responsibility. Which terms you let stand permanently in your denominator, which ones you go and learn, and whether you read a person by their fluency, are decisions available to a person on a Tuesday. The equation says where the change registers. It does not say the change is out of your hands.
A language does not build your world and does not merely wrap it. It is a stock of names for conditions that were there either way — and carrying one of them across a delay, a distance, or a table is the thing that rests on having the name. Judge a vocabulary by what you can hand over with it, not by what you expect to see through it.