Movement Two · The Architecture
Chapter 14
The Language
Sometime around the year 1300, a speaker of English wanted to say that something was pleasant, and reached for the word nice. The word did not mean pleasant. It meant foolish. It had arrived from Latin by way of French, where it described a person who was ignorant — nescius, not-knowing. For two hundred years it meant foolish, and then it began to drift. By the 1400s it could mean fussy, or precise, or finely discriminating. By the 1700s it could mean agreeable. By the time a modern speaker uses it to describe the weather, the word has traveled, across six centuries, from an insult to a pleasantry, with no one ever deciding that it should.
No committee approved the change. No authority recorded the decision, because there was no decision. The word moved because, in a great many small conversations, across a great many years, it was used a little differently each time, and the slightly different uses that caught on were repeated, and the ones that did not catch on were not.
The word changed the way a pheromone trail changes.
The same thing happened to a great many words, and it is still happening now. Silly once meant blessed, then innocent, then weak, then foolish. Awful meant full of awe — worthy of reverence — before it slid to its modern sense of bad. None of these journeys was planned.
A language has the same six things a colony has, and the reader is inside one now. The sentences arriving here are signals. The words in them carry meanings the reader did not assign and could not change alone. The grammar that lets these words combine was built by no one. The containers and the work are there too, in the dialects and registers and in every utterance that succeeds or fails.
The actors are the speakers. Billions of them, each carrying a partial copy of the language in their own head, none of them carrying all of it. Each speaker is, like the single ant in the laboratory dish, almost nothing on their own — a person raised without other speakers does not invent a language; they remain, linguistically, mute. The speaker acquires the language only by being placed back among the speakers who already had it.
The trails are the collocations, the idioms, the worn grooves of habitual combination. A person makes a decision and takes a risk; they do not, in ordinary English, make a risk or take a decision, though both would be perfectly comprehensible. The path between make and decision has been traveled so many times that it is now the path of least resistance, and the path between take and decision has been traveled so rarely that it has faded. No grammarian decreed this. It was deposited, utterance by utterance, the way a trail is deposited, step by step.
A speaker feels the grooves as obviousness. Strong tea, never powerful tea, though the engine that powers a car is powerful and never strong. Ask a native speaker why, and they will have no answer, only a certainty that the other version is wrong. The certainty is the trail. A foreign learner, who has not walked the grooves, produces sentences that are grammatical and comprehensible and unmistakably wrong, and the wrongness is the sound of an untraveled path.
In a market, the single trade was the vivid thing and the trail the faint accumulation behind it. In language it is the reverse. A lifetime spent reading its trails has never included watching a deposit being made, because the marks are laid invisibly, by every speaker, one utterance at a time. That is why convention feels like nature.
What was learned is meaning itself: the accumulated, agreed, hardened correspondence between sounds and the world. The meaning of water, the structure of a question, the way a sentence is built — these are no longer up for renegotiation in any single conversation. And a language carries more than the meanings of words. It encodes how its speakers reckoned time, counted, addressed their elders, named their kin. A speaker inherits all of this without choosing it, the way an ant inherits a foraging landscape it did not survey. It was confirmed so many times, by so many speakers, that it hardened. And once hardened, it was handed to the reader, free, in the cradle.
A word is a path. It survives by being used. When the thing it names disappears, or when a rival word is used a little more often, the path stops being reinforced, and it fades. The English language is a graveyard of words that no living speaker would recognize — swink, to labor; eyne, eyes; fardel, a burden. They faded the way a trail fades when the food at its end runs out. No one mourned them, because no one noticed.
Other words live for a very long time. The most common words in any language — the the, the of, the and, the not — are the most resistant to change. They are traveled so constantly, by every speaker in every utterance, that they harden beyond the reach of ordinary fade. Linguists have traced certain of these words back through thousands of years and across whole families of descended languages, still recognizable, still doing the same work. The words a language uses most are the words it changes least.
Place two groups of speakers in contact, trading and marrying and arguing across a border, and their grammars drift toward each other. Separate them by a mountain range or an ocean, and each community reinforces its own paths, and within a few centuries they can no longer understand one another. Latin did not decide to become French in one valley and Spanish in another and Romanian in a third. It was one language, divided into separate containers, each accumulating its own marks, until the descendants were distinct.
English and German were one language perhaps fifteen hundred years ago; a speaker of one cannot now follow the other. The speakers at each step understood their parents and their children perfectly. The change was never visible from inside any single lifetime, the way the slow widening of a trail is invisible to any one forager and obvious across a season.
A human child can learn any human language. Drop a newborn into any speech community on earth and within a few years they will speak its language natively, with all of its irregularities and its thousand unstated conventions, none of which anyone will explicitly teach them.
A toddler is not handed a grammar before they speak. They are placed in the flow of utterances and begin to produce their own, badly at first, and they are corrected — not by instruction but by response. Within four years, without a single formal lesson, they have recovered a system that no scholar has ever fully written down. They did it the way a young ant learns the trails: not by being told where they are, but by walking until the walking became knowledge.
But designing a language is another matter. Languages have been invented from first principles, with regular grammars and rational vocabularies. A few have attracted hobbyists. Almost none has become the living tongue of a community that raises its children in it and lets it drift and change. The designed languages are too clean. They have no accumulated paths, no hardened meanings deposited by speakers long dead. They are nests with no pheromone in them. An architect can draw a colony. An architect cannot draw the trails.
The reader did not invent a single word they know. They inherited a language that a hundred billion speakers, most of them dead, deposited and reinforced and let fade across uncounted generations. When the reader speaks, they are reading the trails. When they coin a phrase that catches on, they are depositing one.
In the last several years it has become possible to build a model that produces language. Feed it a substantial fraction of everything the speakers have ever written down, and the model learns, from the text alone, which words follow which, which constructions are well-formed, which phrasings land and which do not. It is shown the signals — the accumulated sentences — and from them it recovers the trails. It learns that make goes with decision and take goes with risk, because the paths between those words were the most heavily marked in everything it read.
What such a model has done, mechanically, is extract the path-weight graph of a human language. It has measured how strongly every word is connected to every other — which paths are highways and which are faded. What it learns is a weighting: a vast map of which trails the speakers reinforced and which they let fade.
The model never meets a speaker. It never has a sentence succeed or fail in a real conversation. It is never corrected by a furrowed brow. All of that — the work, the testing, the marking, the fading — happened before the model existed, performed by speakers across the centuries, and was written down. The model arrives after the fact and reads the record. It is good at language for the same reason a detailed map is good at geography. The map did not walk the land.
It is tempting to say the intelligence is in the model. The model is where the impressive behavior appears, the thing one can hold, run, point at. But a model that produces language is a photograph, taken at one moment, of the trail map the speakers had accumulated up to that moment — sharper and larger than any one speaker could hold in their head. A photograph does not forage. It does not deposit new trails or let old ones fade. It holds, frozen, what the living language held at the instant the picture was taken. The intelligence was never a property of the photographing; it was a property of the thing photographed.
This does not diminish the model. It locates it. The model sits in the same place the single ant sits — an agent of great capability, holding a partial copy of something it did not create. The difference is only that this ant has memorized more of the trail map than any ant ever could.
There is a test that makes the point. Train a model on the accumulated text of a language with billions of speakers, and it speaks that language with all the fluency they accumulated. Train it on a language whose trails were walked by thousands rather than billions, and the model is correspondingly thin. The same architecture, pointed at a deep record, is brilliant. Pointed at a shallow one, it is poor.