The Taste Economy

Taste is the trained capacity to perceive which output, among many, is worth keeping. As the cost of generation falls toward zero, taste becomes the scarce input that markets price — not because it is rare by birth, but because it is rare by practice. The research on how taste develops has been consistent for forty years; the agent era has made it operational.

This essay is for anyone who has noticed that generating a draft has become trivial, choosing among drafts has become harder, and is wondering whether the second part is a skill or a vibe.

The cost of generating a thousand competent words fell by roughly four orders of magnitude between 2020 and 2026.¹ At publication time, an API call producing a usable first draft costs less than the change in your jacket pocket. The same trajectory has played out across image generation, voice synthesis, code generation, and music composition. Whatever was labor in the early-2010s creative economy is now capital expense — and the capital is becoming nearly free.

This is the operating environment in which the rest of this essay matters. When one input in a production function falls toward zero, value migrates to the next-most-scarce input.² This is not a novel claim; it is what happened to typesetting after Gutenberg, to arithmetic after the spreadsheet, to typing after the word processor. Each time, the new scarce input was editorial judgment one layer up. The pattern is so old it should not be surprising. What is surprising is how rapidly it is happening this time, and how visible the migration is to the people living through it.

The migration this round leads to one specific destination. The next-most-scarce input is taste. And taste, the research will show, is something the productivity literature has consistently undersold.

#Why is taste the scarcity now?

Markets reward what is scarce relative to demand. Demand for high-quality creative output has not decreased with the cost collapse of generation; if anything, it has accelerated, because the floor of acceptable quality has risen as every basic output becomes cheap. What has changed is the bottleneck. When anyone can produce a serviceable image, an essay, a song, a piece of code, the differentiating step is no longer the production — it is the selection. The person who can look at twenty draft variants and reliably pick the one that should ship has the rare skill.

The economic principle is unspectacular. The market consequence is large. People who can curate, judge, edit, and discriminate are commanding pricing that previously belonged to people who could make. The shift is most visible in the paid-newsletter market — Oliver Burkeman, Anne-Laure Le Cunff, and Craig Mod do not write more words per week than they did five years ago, but readers pay them more for the curation around the words than for the words themselves.³ The emerging "AI art director" role at agencies and product teams is the same pattern in business form: the role is not "produce more images"; it is "decide which generated images are usable and how to direct the next batch."⁴

This is the surface phenomenon. The deeper question is whether taste is something one has — in which case the agent era is bad news for most people — or something one develops, in which case the question becomes how.

#Is taste a personality trait or a learnable skill?

The research is unambiguous. Taste is learnable.

The strongest available evidence comes from the deliberate-practice literature, particularly K. Anders Ericsson's work spanning four decades.⁵ Ericsson's studies of musicians, chess players, surgeons, and writers established a consistent finding: expert judgment in a domain is not predicted by raw talent or general intelligence; it is predicted by hours of structured comparison against a standard. What experts do, that novices do not, is develop a more refined internal model of what good looks like — and they do so by repeated exposure plus active evaluation, not by passive consumption.

The seminal cognitive-science study on the structural component of expert taste is Chi, Feltovich, and Glaser's 1981 paper on physics problems.⁶ When asked to sort physics problems into categories, novices grouped problems by surface features (problems with inclined planes, problems with pulleys). Experts grouped the same problems by deep principles (problems about conservation of energy, problems about Newton's second law). The experts were not making aesthetic judgments — they were applying domain knowledge. But the cognitive operation they performed — perceiving the deep structure that surface variations conceal — is exactly what taste is, mechanically.

The contemporary practitioner-tier articulation of this is Rick Rubin's The Creative Act.⁷ Rubin frames taste as "tuning the antenna" — a slow process of developing the perceptual sensitivity to notice signals other people do not notice. His framing is non-technical and his examples are musical, but the cognitive operation he describes is identical to what Ericsson and Chi formalized in academic terms. Rubin's contribution is the operational claim: the antenna can be tuned, but only by doing the work that requires the antenna. There is no shortcut, and there is no substitute for the time.

#What does the brain do when it judges quality?

The neuroscience of aesthetic judgment is younger than the deliberate-practice literature but converges with it cleanly. Anjan Chatterjee and Oshin Vartanian's foundational Neuroaesthetics review describes the brain regions consistently involved when humans evaluate the quality of an artifact: medial orbitofrontal cortex for valuation, default mode network for self-referential evaluation, and executive control network for selection among alternatives.⁸ Roger Beaty and colleagues have shown that the coordination between these networks — particularly the default mode network and executive control network working together — predicts creative-judgment performance more reliably than the activity of any single region alone.⁹

Two findings from this literature matter for the present argument.

The first is that aesthetic judgment is measurable. Skin conductance, heart rate variability, and EEG markers all track aesthetic preference reliably across subjects when the stimuli are properly controlled.¹⁰ Taste is not, in the cognitive-neuroscience sense, a vibe. It is a recognizable cognitive event with physical correlates.

The second is that experts show stronger and faster recruitment of these networks than novices when evaluating domain-specific stimuli. The expert's brain, after years of deliberate-practice exposure to its domain, becomes a more efficient taste-engine. The expert does not deliberate longer than the novice; the expert deliberates less, because the network has been trained. This rhymes with what Ericsson observed at the behavioral level: expert pattern-recognition appears effortless because the effort happened over the preceding years.

#How does taste actually develop?

The mechanism is straightforward and, by the standards of cognitive science, well-understood. Three components must be present simultaneously:

The first is time-with-the-work. The novice cannot become an expert by reading reviews of the work; the novice must spend hours with the artifacts themselves, doing the slow exposure that builds the perceptual baseline. Csikszentmihalyi's interviews with creative practitioners across domains found this with eerie consistency — every expert reported a stretch of years in which they did nothing but absorb the canonical artifacts of their domain.¹¹ No shortcut.

The second is active evaluation. Passive consumption does not train taste. The brain has to do something with the artifact — judge it, compare it, write about it, mimic it, criticize it. The active step is what Ericsson called "deliberate practice" and what the cognitive-science literature calls "elaborative processing." Reading a thousand poems passively builds little taste; reading a hundred poems and writing a sentence about why each one works or does not work builds a great deal.

The third is a stable internal standard that evolves with use. The standard cannot come from outside; if it comes from outside, the person is following someone else's taste, not developing their own. But the standard cannot start fully formed; it has to be built by the work. The bootstrap problem — how do you have a standard before you have done the work that builds the standard? — is solved by aspiration. The novice imitates the work of practitioners they admire, and through the imitation discovers what about the admired work they actually value. Andy Matuschak has written extensively about this for tools-for-thought practitioners; the dynamic generalizes.¹²

These three components — time, active evaluation, an evolving standard — are what taste-development is. They have not changed in the agent era. What has changed is that one of them is now harder to come by.

#What did the agent era make harder?

Time-with-the-work is the casualty.

The agent has not removed the artifact. There are more artifacts than ever — generated images, drafted essays, suggested code completions. What the agent has removed is the slow time during which a person sits with an artifact without immediately producing a next one. The default behavior of every productive surface now is: hit the generate button, get a candidate, evaluate it briefly, accept or regenerate. The cycle is fast. The active evaluation that taste requires has been compressed to seconds. The fact that the evaluation is happening at all is the only reason taste is even maintained, let alone developed.

The cost of generation is not only that more artifacts exist. The cost is that the temptation to skip the evaluation step has become structural. Why dwell on this draft when another costs nothing? Why argue with this output when regeneration is a click? The economics of attention-with-the-artifact have changed even where the user is willing to spend the attention, because the alternative — generate again — is so cheap.

This is the operational problem. Taste is trainable. The conditions under which it trains have been quietly degraded by the cost collapse that simultaneously made taste more valuable.

The product implication is clear, and it is the implication that informs every consequential Particle decision.

#What does the taste economy look like in practice?

The market has begun pricing this.

The paid-newsletter market is the most visible evidence. Burkeman's Imperfectionist, Le Cunff's Ness Labs, Craig Mod's Roden — these are not businesses that sell quantity. They sell curation. Their readers pay for the trust that someone whose taste they have come to know has done the selection work the reader does not have time to do. This is taste-as-economic-position made operational at the writer scale.¹³

The agency-side evidence is the "AI art director" role, which moved from novelty to standard between 2024 and 2026. The role is not "produce more images." The role is "look at the fifty images the team has generated, pick the three that will ship, and brief the next batch." The role pays well because the supply of people who can do it well is small relative to the demand.

The same pattern is appearing in engineering teams. The 2025-2026 wave of "AI-augmented developers" did not result in headcount expansion at any major company; it resulted in selection-skill becoming the differentiating axis. The developer who can write specifications such that the agent produces correct code, evaluate the agent's output for subtle errors, and decide which generated approach to keep — that developer is now more valuable than the developer who can write the code by hand.¹⁴ The skill has migrated up one layer.

This is not a prediction. It is a description of an already-happening 2026 reality. The taste economy is here. The question for any individual knowledge worker is whether they have been developing the capacity it rewards.

#Where does this leave Particle?

Particle has refused to ship suggestions inside the CAPTURE input, refused to auto-fill the planner, refused to write the user's intentions for them. Every one of these refusals was a feature decision that competing products treated as obvious functionality. The reason we made the opposite choice is the argument this essay has now laid out: an empty input is the surface on which taste develops. Filling it is the productivity-feature that subtly degrades the very capacity the product is supposed to support.

The Coach operates on the same logic. The Coach observes patterns and surfaces them; it does not tell the user what to do. This is not a hedge. It is the direct consequence of believing taste is trained, taste is the scarce skill, and taste develops only when the human does the judging. Any feature that decides for the user gets in the way of the practice the user is here to do.

The product position is not that AI assistance is bad. It is that certain specific places in the work pipeline must remain the human's. CAPTURE is one. REFLECT is another. The session itself — the slow time inside which the user actually sits with the artifact — is the third. These are the surfaces where taste is being trained. Touching them with automation, even helpful automation, would be the productivity-feature that breaks the product's reason for existing.

The next essay in this series, The Judgment of When to Stop, treats the second of the five capacities the agent era keeps on the human side: knowing when the current output is enough. Taste tells you what good looks like. Stopping is the moment of acting on the signal. They are companion skills; you cannot do one well without the other.

The third essay, The Prompt as Document, turns to how the human renders intent into the form an agent can execute — the writing layer of the agent era. The fourth, The Architecture of Trust, addresses the calibration problem: when can you stop verifying agent output? The finale, The Last Instruction, synthesizes all five capacities into the single argument that strategic re-orientation — knowing when to abandon a plan — is the meta-skill the agent era leaves with humans.

The five capacities are the human layer. Taste is the first one. The research on it has been quietly consistent for forty years, and the agent era has finally made the question of taste's development feel urgent in a way it never quite did before.

Practice the taste. Use the architecture that protects the conditions of its development. Watch the next layer of work move up to the place where you have been preparing for it.

#References