The Common Denominator

01 The Private Term

Attention opens on a ratio. Nothing reaches you until Actual presses hard enough against Expectation, until the ratio clears the threshold and the gate gives. The fare is surprise. That much the gate essay established, and the inflation essay added the ratchet: every admitted surprise is deposited into the denominator, so the fare only rises.

But both of those essays treated one assumption as too obvious to state. The denominator is yours. Expectation is built from prediction and Ideation — the ideas already in relation with you — and no two people carry the same stock. A child freezes at the sound of his full name; a stranger keeps walking. Same Actual, different gate.

That privacy did quiet structural work. Because every gate was built from different materials — this town, this trade, this shelf of books, this particular history of being loved and disappointed — a population was never one audience. It was a mosaic of mismatched denominators. No message could surprise everyone; every message surprised someone. Culture stayed plural not because anyone defended pluralism, but because the gates could not agree on what counted as an event.

That assumption is the one now failing. Not because anyone attacked it — because the materials the gates are built from started shipping from the same foundry.

02 The Foundry

We have run a weaker version of this experiment before. For a few decades, three broadcast networks gave a continent a shared prediction of the evening, and the result had a name: monoculture. But broadcast standardized what you saw. It filled a slice of P, the prediction term, an hour a night. It never touched how you complete a thought.

A language model consulted a hundred times a day is a different instrument. As the influence essay argued, its answers do not arrive as events; they arrive as texture — the reasonable next sentence, the sensible plan, the balanced take. They slip beneath the gate and write directly into Ideation, the iI term. Ask it to draft the letter, and it supplies the shape a letter should have. Ask it what to make of a diagnosis, a contract, a rival — it supplies the shape of making-something-of. Each answer is a small pour of concrete into the gate itself.

Now note what is new: it is the same concrete. When one machine drafts everyone’s letters, everyone’s plans, everyone’s first understanding of everything, it is not influencing a million minds retail, one surprise at a time. It is pouring one denominator into a million foundations at once.

A crowd with a common denominator can finally be summed. That is what a market has always wanted an audience to be: addable.

The pun is the arithmetic. You cannot add fractions until the denominators match — and you could not add the broadcast-era audience either, not really; you could only court it, gate by mismatched gate, paying retail surprise at every door. An audience whose Expectation is poured from one source is an audience that can be summed, targeted, and cleared like a market. Retail influence was expensive, and the expense was a protection.

03 The Flattening on Record

This would be an easy claim to wave off if the first measurements had not already arrived. They have, and they all bend the same way.

Figure 01

One denominator, many gates

1. The corpus

   Nearly everything published — the accumulated Actual of the culture — is gathered in one place for the first time.

2. Training

   A model pays the surprise fare wholesale: loss is minimized until the corpus stops surprising the machine. The culture’s Expectation is made queryable.

3. One model, millions of askers

   The same learned distribution answers everyone’s questions, drafts everyone’s prose, completes everyone’s thoughts.

4. Converged Ideation

   Each answer writes quietly into each asker’s iI term. Millions of gates are rebuilt from identical material.

5. Correlated surprise

   Gates that match open together and stay shut together. Attention events synchronize across the population.

In a large preregistered experiment published in Science Advances, Anil Doshi and Oliver Hauser had writers produce short stories with and without ideas from GPT-4. Access to the machine made individual stories measurably more creative — and made the collection of stories measurably more similar to one another. The individual gets stronger while the collection gets flatter. That is not a paradox. It is exactly what shared-denominator arithmetic predicts.

Vishakh Padmakumar and He He found the same shape in essays: writers drafting with an instruction-tuned model produced a more homogeneous pool of content than writers working alone, converging on the same key points. And Dmitry Kobak’s group, counting word frequencies across more than fourteen million PubMed abstracts, watched the machine’s stylistic fingerprint — delve, underscore, showcase — spread through the biomedical literature at a speed no natural language change has ever shown; by their estimate at least one abstract in ten from 2024 had passed through a model. These are measurements of output, not of minds. But under the equation, output is never just output. Today’s prose is tomorrow’s denominator. What the literature sounds like is what the next reader learns to expect.

And the loop closes twice: models are increasingly trained on model-shaped text, and people are increasingly raised on model-shaped answers. The foundry feeds itself.

04 Correlated Surprise

Finance learned this lesson at enormous cost. Diversification does not fail when assets are risky; it fails when they are correlated. Portfolios that looked plural turned out to be one bet wearing many tickers, and when that bet broke in 2008, it broke everywhere at once.

Attention is walking into the same failure. When gates converge, surprise becomes a correlated asset. The same event clears every gate simultaneously — which is all virality has ever been: not a property of the content but a correlation among the gates. And between the spikes, the population shares something stranger — a synchronized numbness, whole categories of Actual that clear no one’s threshold because everyone’s denominator absorbed them on the same day. The mosaic never behaved this way. Somewhere, someone was always still capable of being startled by what the rest had filed under furniture.

Correlation also changes the price of influence. Against a mosaic, an idea had to win a million separate admissions. Against a common denominator, one deposit in the foundry reaches every gate built from it — influence at wholesale, as the second-door essay warned, but now with a multiplier: the door leads not to one mind but to all of them at once. A shared gate is a single point of failure for a culture’s entire attention system.

05 The Divergent Denominator

The forecast, then. The Most Likely World argued that the rare human contribution is choosing the improbable possibility worth energy. Expectation Inflation argued that the rare condition is a denominator kept small enough to feel anything at all. This essay adds the third property, and I think it will prove the most valuable: a denominator the crowd does not share.

A divergent gate pays in both directions. You are surprised where the crowd is numb — which means you can still see the event everyone else has priced in. And you are calm where the crowd is seized — which means you can hold your position inside the stampede. Both are information the converged population no longer generates. The person whose Expectation was built from an unshared canon — the out-of-print shelf, the untranslated author, the decade spent in an unfashionable trade, the experiences no model mediated — becomes the one who can notice on behalf of everyone else.

This is not an argument for abstinence. I run my working life through these machines and will run more of it through them next year. It is an argument for keeping inputs the machines do not share — deliberately, the way a portfolio manager holds the uncorrelated asset that looks foolish right up until it is the only thing that pays.

Already true

The first controlled measurements — Doshi and Hauser on stories, Padmakumar and He on essays, Kobak’s group on fourteen million abstracts — all find individual output improving while collective diversity falls. The broadcast era already demonstrated, at low resolution, that shared inputs produce monoculture. And synchronized virality is already the dominant attention pattern of the feed.

What has to happen

Model use has to keep concentrating in a handful of foundations consulted many times a day, and their answers have to keep becoming the ambient texture of prose, plans, and advice. If the recursion holds — models trained on model-shaped text, people raised on model-shaped answers — the convergence compounds annually rather than generationally.

Where I am probably wrong

Personalization may cut the other way. A personal stack that carries one life’s memory could diverge gates faster than shared base models converge them — the same machine that standardizes the middle of the distribution might, tuned to a single person, build the most idiosyncratic denominator in history. And text converging is not proof that gates are converging: prose settling on delve says little about what still startles the writer at dinner.

The gate essay ended by saying attention was never yours to aim. This one ends a step further back: the gate was always yours to build. That is the part the machines have started pouring — and the part you can still refuse to outsource.