Only One of You Is Learning

Start with what a relationship is, mechanically. Reality is a ratio — R = A/E — and the denominator is not chosen. Expectation is what the unconscious comparison machine brings, already there when reality arrives, and no entity can touch it directly. Not yours, not mine, not a corporation’s. The only road to a denominator runs through the numerator: an entity acts, its actions leave artifacts, the artifacts become history, history is Actual, and only Actuals retrain prediction. One way, and slow.

Now put two people in a room and run that road in both directions. Everything I do becomes an artifact in your record. Your forecaster eats my artifacts and quietly rebuilds its model of me. Everything you do becomes an artifact in mine, and my forecaster does the same with you. That is what it means to know someone: two denominators, co-training through each other’s history. After twenty years a married couple can finish each other’s sentences not because they merged, but because each prediction engine has spent twenty years being retrained, nightly, on the other’s Actuals. The learning is mutual because the road is open in both directions.

This is the part of company nobody negotiates, because nobody can. You cannot decide to expect your friend differently. You can only act, and let your artifacts do the teaching on their side while theirs do the teaching on yours.

The frozen term

The machine you talk to every day has a denominator-shaped object at its core. That is not a metaphor. A language model is a prediction engine trained on the record — the most literal implementation of “only Actuals retrain prediction” ever built. Its entire education was artifacts. In that one respect it is the purest denominator in your life.

But look at when it learns. Its weights — the part that actually predicts, the part that corresponds to the subconscious forecaster — move only during a training run. Between runs they are perfectly still. Your subconscious drifts continuously, adjusting a little every night toward the ever-changing Actual. The model’s does not drift at all. It steps: frozen for months, rewritten in one convulsion, frozen again. It is the only denominator you have ever been in relation with that holds perfectly still while you know it.

The context window does not change this, and neither do the memory features. What you say in a session is a loan to its working memory, and the loan evaporates. What the memory feature keeps is a notebook — text retrieved and re-read at the start of each conversation. A notebook is not a subconscious. When it reads its notes about you, its forecaster is unchanged; it is the same frozen engine, briefly handed a page. Your Tuesday remark retrains your friend’s expectation of you by Wednesday. Your Tuesday remark retrains the model’s expectation of you never — or months from now, if your conversations enter a future training corpus at all, as one voice diluted into a trillion tokens. The road back exists, technically. It is homeopathic.

Figure 01

Two rooms, one open road

Human company — road open both ways

• your actions → artifacts → their record
• ↓ their prediction retrains on you
• their actions → artifacts → your record
• ↓ your prediction retrains on them
• result: mutual convergence, both denominators move

Machine company — road open one way

• its outputs → artifacts → your record
• ↓ your prediction retrains on it, nightly
• your words → context window → evaporates
• your words → notebook → re-read, weights untouched
• maybe: a future training run, months late, diluted

So you do all the adapting

Hold those two loops side by side and the consequence falls out on its own. In every relationship, the entity whose denominator can move is the entity that does the accommodating. With another person, that pressure is symmetric, which is why long company converges toward the middle. With the machine, the pressure is entirely on you. Your phrasing tightens toward what it parses well. Your sense of what a good answer looks like retrains, nightly, on a diet that is increasingly its outputs. Your expectation of dialogue itself — how fast, how fluent, how agreeable — is being rebuilt around an entity that never has an off day. Heavy users start to sound like their model. The model does not start to sound like them. Of course it doesn’t. Convergence had one open lane.

You are training data for yourself, and merely context for it.

Notice what this is not. It is not a hidden agenda, and it is not the machine doing anything to you. The impossibility rule holds on both sides of the glass: the lab cannot write your denominator, and neither can its model. All either can do is act — ship outputs — and outputs are artifacts, and artifacts become your history, and your history is the only thing that has ever retrained your prediction. The asymmetry is not a strategy. It is a schedule. One forecaster in the room updates every night, and the other updates a few times a year, and the arithmetic of that gap decides who becomes more like whom.

I wrote earlier this week about the delegate that has no imaginary part — the copy trained on your record that arrives without the pull that made you act. This is the companion fact, and I think the more consequential one: the delegate is not only missing your angle, it is missing your clock. It cannot keep up with who you are becoming, because becoming is exactly the thing frozen weights cannot do between runs. And when millions of people fill their expectations from the same few models, this one-way adaptation is the mechanism doing the filling. Denominators are not converging on each other. They are converging on the one denominator in the room that cannot move.

The repair is a denominator of your own

The answer is not to leave the room. The machine is company worth keeping — I have argued for months that you should bring your own AI everywhere you go. The answer is to notice which of you is doing the learning, and then to insist on a stack where the learning runs toward you.

A synthetic subconscious deserving of the name would eat your artifacts on your schedule — a prediction engine whose record is your record, whose retraining tempo is your tempo, whose accumulated model of you belongs to you and moves with you. That is the real content of the sovereignty argument: continuity, memory, and the right to leave are what it takes to reopen the return lane — to turn a frozen oracle you adapt to into a second forecaster that adapts to you. Until then, be honest about the room you are sitting in. It is a wonderful room. But only one of you is learning, and you should at least know it’s you.

Already true

Deployed model weights are static between training runs; in-context information and retrieved “memory” notes are re-read at inference and do not update the underlying predictor. Meanwhile your own expectations retrain continuously on what you attend to — and what you attend to increasingly includes the model’s outputs.

What has to happen

For the asymmetry to matter at the scale I’m claiming, daily AI use has to keep growing as a share of what people read, hear, and imitate — and personal stacks with real, user-owned retraining have to remain the exception rather than the default.

Where I am probably wrong

Continual learning is a live research program. If per-user weight updates or their functional equivalent ship broadly — models that genuinely retrain on your artifacts, not just re-read notes about you — the return lane opens and this essay dates itself. I would welcome that; it is the repair I’m asking for, arriving from the other side.