The Unlearnable Feed

Boredom has a reputation it did not earn. We treat it as a failure — of the world to be interesting, or of the person to appreciate what is there. The equation reads it differently. Reality is Actual divided by Expectation, and the denominator is a prediction machine that retrains on every arrival. Meet the same source often enough and the predictor learns its shape. Actual starts landing where Expectation already stood. The ratio settles toward one, and the charge of unexpectedness drains out of the encounter. What remains is unitless stability: still present, still conscious, but without the jolt.

That settling is not a malfunction. It is a receipt. Surprise is tuition — you pay it in attention, and what the payment buys is convergence. The colleague who startled you in your first month is legible by your sixth. The tenth notification from the same app takes a fraction of what the first one took. In a lawful world, every source of surprise is a depreciating asset, because the one thing surprise teaches a predictor is how to stop paying for that source. Habituation is not your attention failing. It is your attention succeeding — the proof that learning occurred, delivered as a falling price.

Boredom, read on the meter, is a diploma.

01 / The platform’s problem

Now stand on the other side of the glass. Suppose your revenue is attention, and attention is paid out of surprise. Then habituation is not a psychological curiosity; it is your depreciation schedule. Every session, each user’s predictor converges a little further on what you show them, and every step of that convergence is future revenue amortizing to zero. The retention curves that keep product teams awake are not mysterious. They are learning curves viewed from the seller’s side of the counter. Your users are graduating, and graduates stop paying tuition.

A business built on surprise has three options. It can keep producing genuinely new Actuals, which is expensive — the world only generates so much news, and originality does not scale on demand. It can accept the amortization and shrink, which no growth model tolerates. Or it can take the third road: make the sequence unlearnable.

02 / The schedule

The third road was mapped decades before the feed existed. In the operant laboratory — the program documented in Ferster and Skinner’s Schedules of Reinforcement in 1957 — the question was how the timing of reward shapes the persistence of behavior. The finding that matters here concerns the variable-ratio schedule: reward after an unpredictable number of responses. It produces high, steady rates of responding, and a pigeon rewarded on an unpredictable count keeps pecking long after the food stops — far longer than one that was rewarded every time. Predictable schedules teach an animal when to stop. Unpredictable ones withhold precisely that lesson.

The casino industrialized the finding. In Addiction by Design, her study of machine gambling in Las Vegas, Natasha Dow Schüll documents players who describe entering the machine zone — playing not to win but to stay inside the rhythm of play, money reduced to a means of remaining seated. The machine’s schedule is engineered so that the player’s predictor never closes the gap. There is nothing to learn about when the next payout comes, and so the lesson that ends the session — this is now predictable, you may stop — is never issued.

The pigeons and the slot floors are record. What follows is argument. Pull the feed downward and watch what arrives: sometimes nothing, sometimes filler, occasionally something that lands squarely on an idea that already had you. The interval is not learnable, and that is not sloppiness — it is the load-bearing feature. A sequence with no structure at your scale is a sequence your predictor cannot compress. Prediction error stays irreducible, not because your predictor is weak, but because there is nothing there to learn. A slot machine is a machine for keeping a predictor untrained. So, I would argue, is a feed.

03 / The rule it honors

Be precise about what has and has not happened here, because the doctrine does not bend. No entity can touch either term of the ratio. Not the numerator, which is finished the instant it exists. Not the denominator, which is brought unasked. There is one causal road and it runs one way: an entity acts, its actions leave artifacts, artifacts become history, history is Actual, and only Actuals retrain prediction.

The feed has not hacked your denominator. That is impossible, and no scheduling algorithm changes it. What the feed controls is its own actions — and its actions become the Actuals of your evening. It walks the only legal road, in full compliance with the law, and engineers the road to lead nowhere. Convergence is prevented not by touching Expectation but by making sure the Actuals your Expectation retrains on have no shape to converge upon. The road to the denominator runs through the numerator; the unlearnable feed keeps that road open, well-lit, and endless.

Figure 01 · Conceptual

The price of the same source, paid in surprise

12345678

• Learnable source — surprise decays as the predictor converges
• Unlearnable source — surprise never amortizes

04 / Surprise without learning

Earlier this month I wrote that surprise has a noise floor: deliver a change in increments small enough and you can update a person completely without their ever feeling it. Change without feeling. The unlearnable feed is the mirror image — feeling without change. Every pull pays out a real jolt, and the jolt is the real part of surprise, the only part you feel. But you can be surprised a thousand times by such a source and know it no better than the day you subscribed. The jolts were real. The learning they normally purchase never accrued. These are the two mature strategies of the surprise economy: move the change below the line where feeling begins, or hold the feeling above the line where learning ends.

Put this beside the machine that eats surprise and the arrangement becomes hard to unsee. One industry spends billions training machines whose entire objective is to drive their own surprise toward zero — loss going down and to the right. Its sibling spends comparable engineering keeping your surprise permanently alive. Opposite objective functions, and the same substrate on both: a predictor that retrains on arrivals. One eats surprise. The other farms it — and a farmed animal is one that is never permitted to finish growing up.

05 / The tell

What follows is a diagnostic, not a dial. You cannot decide to be bored; the denominator is not yours to set, and no reframing exercise will set it. But you can read your own meter, and the reading is one question: after an hour with this source, what can I predict that I could not predict before?

A learnable source gets cheaper. Its surprise decays, its lessons accumulate, and it eventually bores you — which is to say, it graduates you. The deep sources never quite finish the job: the sea, mathematics, a person you love. But notice what their inexhaustibility is made of. Each surprise they hand you pays out convergence, and behind it stands more structure, learnable all the way down. The distinction is not how often a source jolts you. It is whether the jolts accrue.

An unlearnable source never gets cheaper, and this is its signature. If a source never gets boring, it is usually not because it is inexhaustibly rich. It is because it is random. Whatever freedom you have lives on the left side of the equation, in action: you cannot choose your Expectation, but you can choose where you stand, and where you stand decides which Actuals become your history and, in time, your predictor. Stand where the diploma is issued. The feed is a school with real tuition, collected on every visit, whose commencement is scheduled never.

Already true

Variable-ratio schedules produce high, persistent responding, and intermittent reward resists extinction far better than constant reward — that is the operant record, documented since Ferster and Skinner. Machine gambling deliberately runs such schedules; Schüll’s ethnography shows the players themselves describing play without a destination.

What is argument, not record

Reading habituation as predictor convergence is this framework’s gloss, not a laboratory result — lab habituation is stimulus-specific and shows dishabituation and recovery. And feeds are not literally slot machines: some of their unpredictability is emergent from inventory and ranking, not designed. The claim is about what the schedule does to a predictor, not about intent in every product meeting.

Where I am probably wrong

The entertainment defense, at full strength: a person scrolling at midnight may want exactly what they are buying — jolts as jolts, the zone as the zone — and calling that theft is paternalism dressed as arithmetic. I concede the purchase is real and freely made. The equation does not forbid buying jolts. It only insists you read the label: whatever this product is, it is not information, because information is the part that would have made it boring.