On Taste

What survives specification

In 1964 Susan Sontag published fifty-eight numbered notes on camp. The most useful passage has nothing to do with camp. It sits in the preamble, where she makes a claim about taste itself: taste runs under every response a person makes freely rather than by rote. She takes inventory. Taste in people, in emotion, in acts, in morality. Intelligence, she says, is taste applied to ideas.

Taste admits no system and no proof, she writes, and any sensibility you force into one stops being a sensibility. It “has hardened into an idea.”

That is a legibility claim, thirty-four years before James C. Scott gave us the word.

Scott argued in Seeing Like a State that governments can only act on what they can render legible, so they build maps, censuses, standardized surnames, grid cities. Whatever resists transcription gets ignored, or flattened until it fits. The local knowledge that made those systems work, what he called metis, never survives the paperwork.

Sontag found the same failure at the scale of a single mind. Sensibility is the part of judgment that dies in transcription. Write your taste down and you have a rubric.

The legibility engine

For sixty years this stayed an aesthete’s observation. Now it is an economic one, because AI is a legibility engine. Models train on whatever got written down: style guides, process docs, graded examples, logged preferences. The frontier of automation tracks the frontier of legibility. Whatever you can specify, you can delegate. I sell an AI workforce for a living. The entire business is specification.

So the scarce residual is the illegible part of judgment. This is not a capability gap. Taste, on Sontag’s terms, is what remains after specification. Finish the rubric and the sensibility has hardened into an idea. Ideas copy for free.

The objection

Taste distills too. RLHF is thousands of human preference judgments compressed into a reward model. Statistical legibility instead of propositional legibility. The machine never reads your rubric. It absorbs your choices anyway.

Concede all of it. What distills is yesterday’s taste, the frozen average of past discriminations. Taste that everyone shares is convention. Machines execute convention perfectly, and convention pays nothing.

The taste that pays lives at the frontier, where consensus has not formed and the comparisons have not been logged. It is alpha in the trader’s sense. Publish it and it becomes beta. Every act of articulation converts your edge into common knowledge and pushes you back out to the frontier. So taste is a rate, not a stock. Past judgments depreciate on contact with the page. The asset is how fast you form the next one.

Where taste comes from

Kahneman and Klein studied when expert intuition deserves trust and landed on two conditions: real regularities in the domain, and long practice with fast feedback. Taste is System 1 trained by ten thousand System 2 comparisons. A chef gets there tasting two reductions side by side, an editor somewhere around manuscript four hundred, an investor watching a decade of pitches resolve into outcomes.

Three inputs, multiplied: volume of exposure to the best available work, deliberate comparison rather than passive consumption, and pleasure. Pleasure looks like the soft one and does the most work, because nothing else sustains the volume. Nobody grinds through ten thousand comparisons in material they hate. Joy sets the compounding rate.

This explains Sontag’s parenthetical, the one that reads like a throwaway: taste grows unevenly, and fine visual taste rarely travels with fine taste in people or in ideas. Reps are domain-specific, so there is no general taste. Only earned tastes, in narrow lanes, purchased one comparison at a time.

What to do about it

For a person: treat your inputs as training data, because that is what they are. Read the best, not the most. Put two versions of anything side by side and say out loud why one wins. Then notice the trap inside that instruction. The articulation is scaffolding. What you are building is the faster, surer discrimination next time.

For an organization: stop trying to write taste down. The culture doc and the quality-bar memo are useful, and they are ideas. Your competitor can copy them and your model already has. Taste lives in specific people, so the moves are structural. Hire for taste. Station it at the points of highest variance, where one decision swings the outcome, and shorten the feedback loops so it compounds. Taste density beats headcount, and it beats process, because process is legible and legible things are now free.

The free response

Sontag’s opening cut was between the free response and the rote one. Taste governs the first. Everything rote is now spoken for.