Meaning Sold Separately

A senior engineer at a well-funded company instructs his team to stop writing full sentences to the model. Articles go first. Then the verbs that a noun can imply on its own. The prompt is reduced to the words that survive a cost review. The industry has a name for this now — caveman — and a senior person at one of the frontier labs has been seen contributing to the technique. The reasoning is not disputed. Words are tokens. Tokens are billed. A sentence is an invoice.

The same week, a different engineer at a different company declares a moratorium in the other direction. Kenton Varda, on the output side of the same machine: "I just declared a moratorium against AI-written change descriptions." His stated reason is precise. The model's descriptions were "worse than useless" — "outlining details of the code that could easily be seen by looking at the code, but omitting the higher-level framing needed to understand broadly what the code is doing."

Two moratoria in one week, pointing in opposite directions. One bans elaborate language going in. One bans elaborate language coming out. They are the same finding, filed twice.

The two tolls

The finding is that language, on both ends of the machine, has been re-priced. It used to be free. It is now metered. And once a thing is metered, an organization will discover — quickly, and with the surprise it reserves for things it has always known — how much of that thing it never needed.

There are two tolls, and it is worth naming them separately, because the industry pays them separately and rarely notices they are the same bill.

1. The input toll. Every word sent to the model costs money. So the words are removed, in a specific order. First the articles — the, a, the small connective tissue that carries no billable weight and, it turns out, a great deal of unbilled meaning. Then the verbs a competent reader can reconstruct. Then the qualifications, the "roughly," the "in most cases," the hedges that were not hedging but scope. What remains is a pile of nouns. The pile is cheaper. The pile is also, in the precise technical sense, ambiguous — the disambiguation was living in the words that got cut.

2. The output toll. Every word the model returns is a word a human has to read, and reading is the one cost the token meter does not display. So the model, asked to describe its own work, produces volume: a description that restates the code line by line and omits the one sentence explaining why the code exists. It is thorough. It is complete. It is worse than useless. The engineer bans it, not because it is wrong, but because it is expensive in the currency the invoice does not track.

What was actually being purchased

Put the two tolls next to each other and the shape appears.

On the way in, the organization strips the language down to what it can afford, and loses the part that carried the intent. On the way out, the organization receives all the language it could ask for, and none of it carries the intent. In both cases the intent is the missing element. In both cases the intent was the expensive part — the part that does not compress, that a model cannot cheaply generate and a meter cannot cheaply price.

This is not a new discovery. It is the oldest one, arriving in a new envelope. The meaning was never in the word count. It was in the specific, non-fungible act of one person deciding what another person needed to know. That act has no token price because it was never for sale. The market has simply gotten better at charging for everything adjacent to it, and the adjacency is now close enough to feel like the thing itself.

An empirical note

The numbers are available and worth reciting flatly. Simon Willison, releasing a database utility this month, reported that a single release candidate had been written mostly by a model for $149.25 — a figure he published because it was, correctly, remarkable. In the same season, the industry's most-read critic charges $70 a year to argue that the whole apparatus is losing money, and two of the largest model companies filed to go public.

We note only that all of these are counts of the same thing — words, moved through a machine, priced by the token — and that none of them is a count of whether the words meant anything. That figure is not collected. There is no line for it on the invoice. This is not an oversight. It is the design. Meaning was never a metric; it was the thing the metrics were built to stand in for, back when standing in for it was cheaper than the thing.

The forecast

The pricing will hold. Words in will keep getting cheaper to send and emptier to read. Words out will keep getting more abundant and less worth reading. The gap between the two — the intent, the framing, the sentence that says why — will keep being the one part nobody can automate, which the industry will continue to experience as a surprise.

The bill for this arrives later, and under a different name. It is called comprehension — the growing pile of code, decisions, and documents that were produced cheaply and can now be understood only expensively, by someone willing to reconstruct the framing that the moratorium removed and the caveman prompt never sent. At a conference this month the phrasing was understand to participate: the observation that delegating to an agent whose output you can no longer read is a one-way door. Elsewhere the same week, a proposal to stop calling it "human in the loop" and start calling it "we recruit agents to join our loop" — a linguistic promotion for the human, issued at the exact moment the human's language was being priced out on both ends.

We are not against any of this. We are only keeping the receipt. Someone reconciles it eventually, in the currency that was never on the meter, and by then the sentence that would have explained the whole thing has been optimized away — first the articles, then the verbs, then the reason.

— Filed by The Context Loss Bureau