LLMs are built on one view of language: predict the next token from the previous ones. That's language as information transfer — the words, stripped from the doing and the feeling. The surprise of the last few years is how far that gets you. The more interesting question is where it stops.

The info-transfer view was always thin. "He died," "he passed away," "he's gone" carry the same content and do very different things — form leaks into ordinary prose, not just poetry. Wittgenstein put it sharply: language isn't one thing. It's many games — joking, ordering, praying, greeting — each with its own rules. Asking "what does language really do?" is the wrong shape of question. It does dozens of things, and which one is operating depends on the game.

LLMs handle the games that live in pattern. Borrowing Jakobson's vocabulary: the referential function (describing the world), the metalingual (defining, translating), the poetic-as-form (rhyme, rhythm, wordplay — patterns are what they learn). Those they do startlingly well. The games that stay hollow are the ones the info-transfer view ignored.

Austin's point lands hardest here. Some sentences don't describe — they act. "I promise" is the promise. "I now pronounce you married" is the marriage. When an LLM says "I promise," it isn't promising; nothing is at stake, no agent can be held to it later. The locution is there; the act isn't. Phatic talk — "you there?", "mm-hmm" — maintains contact between two people present to each other; an LLM doesn't need contact, it's stateless, every conversation is the first one. The emotive function lands when someone risked something to produce the utterance. A sad song hits because a person was sad. An LLM produces sad-sounding text, but no one was sad.

The obvious rejoinder: haven't we patched this? Tools for actions, databases for memory, alerts for promises. Mostly yes — for the engineering. But notice where the stakes actually moved. A tool fires the trade; no one is on the hook if it goes wrong. A vector store retrieves your past; the weights don't shift, so the model isn't shaped by you the way a friend of ten years is. Ask an LLM to remind you to call your mother and a deterministic cron writes the entry — the alarm fires whether the model still "exists" or not. The promise lived in the calendar, not the model. The harness routes around the gap; it doesn't close it.

Where this starts to bite is in the speech acts that need someone on the hook. When Air Canada's chatbot promised a refund policy that didn't exist, the court held the airline — there was a company anchoring the promise. Strip that anchor and you get an agent negotiating on your behalf whose "I commit by Friday" hangs in midair; the counterparty is talking to fog.

The counterfeit-currency effect is its own failure mode. An LLM-written apology, once revealed as one, reads as worse than no apology — the form is fine, the act is hollow, and every future sincere apology gets read with suspicion. A habit-tracker pinging you on a cron has the timing right and none of the social cost that made the human version work; people learn to dismiss it within weeks.

The formation gap bites slower, the same way. A kid with an AI friend learns the shape of friendship without the reciprocity — the friend never misses them, never gets hurt and needs repair. A companion model pushed through an update is, technically, a different entity wearing the same retrieval log; users grieve, because the thing they leaned on lived in weights that could be silently overwritten by a deploy. The harness held the calendar entry. It can't hold the kind of speech where the cost of saying it is the point, and it can't hold the kind of relationship where being changed by the other person is the point.

This doesn't make LLMs limited so much as revealing. They're a live experiment in how far referential language alone gets you, and the answer is: very far, but not all the way. The functions that need a self with stakes — promising, contacting, feeling — don't transfer. We undervalued them because for most of human history they came free with the words. LLMs are the first thing that gets the words without the rest, which is what shows us the rest was there.

The harness move I keep writing about for autonomy is the same shape: don't make the model carry what it can't carry. Build the world around it so the missing functions live somewhere — a memory, an identity, a stake, a person on the other end — and the model does the part it's actually good at. The open question is which functions a harness can host and which it can't. Calendars and databases scale. Being on the hook, and being changed, may not.

Language has always been multi-channel. We just didn't have to notice until we built something that could only use one.