The Taste Gap

A modern coding agent can produce ten working versions of almost anything in the time it takes to read this sentence. Ten landing pages, ten database schemas, ten ways to phrase the same paragraph. Each one runs. Each one looks finished. The machine has no opinion about which is worth keeping, and it never will on its own.

That silence is the whole story of the next few years. For most of computing history the expensive step was making the thing. Writing the code, drafting the copy, rendering the design. We organized teams, careers, and salaries around that cost. Generative systems have driven it close to zero. The step that did not get cheaper is deciding which output is good, and that gap is where human value is now concentrated.

When production is free, the scarce skill is no longer making. It is choosing. Taste is the name for choosing well under abundance, and it does not come bundled with the tools that create the abundance.

Why making got cheap and choosing did not

A generative model is an averaging engine. It returns the most statistically plausible continuation of your prompt, which is exactly why its output looks competent and rarely looks great. Andrej Karpathy, describing his own agent-assisted work in 2026, put it bluntly: the code these systems write "can be bloated, copy-pasted, awkwardly abstracted, brittle. It works, but it is gross." It works and it is gross is the defining texture of cheap generation. The floor went up. The ceiling did not.

Choosing did not get cheaper because it depends on something the model has no access to: what this specific output is for, who will live with it, and what will still feel right in six months rather than this afternoon. Those are not retrieval problems. No larger context window solves them. A venture investor writing about the same shift called it the move from a creation economy to a taste economy, where filters beat generators and curators beat creators. When everyone can generate, the person who can tell the good one from the merely plausible one holds the rare position.

Taste is calibrated judgment

The word taste invites a lazy reading, as if it meant personal preference that no one can argue with. In professional work it means almost the opposite. Taste is judgment that has been calibrated against thousands of real outcomes until it becomes fast and mostly unconscious. The editor who feels a sentence is flabby before she can explain why. The engineer who distrusts an abstraction on sight and is usually right. That speed is earned, and it rests on a model of quality the person can defend when pushed.

Karpathy framed the current arrangement precisely. The agents, he said, are like interns: "You still have to be in charge of aesthetics, judgment, taste, and oversight." He went further on why this is structural rather than temporary. AI code does not improve in taste because "there may be no aesthetics reward" in how these models are trained. Quality that no one scored cannot be optimized toward. Until that changes, and it has not, discernment stays a human seat, and the people who hold it set the standard everyone else's output is measured against.

Why a machine cannot hand you its taste

The tempting shortcut is to ask the AI to judge its own work, or to absorb good judgment by reading enough polished output. Both fail for the same reason, and learning science named it decades ago.

Robert Bjork and Elizabeth Bjork showed that the feeling of fluency during study is a poor guide to actual learning. Smooth, easy input produces a warm sense of understanding that is largely an illusion. The conditions that build durable judgment feel harder and slower at the time. This is the trap with frictionless AI output. Reading a clean answer feels like learning, so you accept it, and you build no independent model of why it is good or whether it is. You can collect a thousand finished examples and still have no taste, because taste is the residue of having struggled to evaluate, not of having been shown the answer.

This is why the team that leans hardest on generation can quietly lose the ability to judge it. Every output looks fine because nothing built the muscle that would notice when it is not. The gap closes only when a person does the evaluative work the machine skipped.

How to train taste on purpose

Taste is trainable, but only through the friction that fluent tools remove. A few practices put the friction back deliberately:

• Force the comparison. Have the AI generate three to five genuinely different versions, then write one sentence on why one wins. The choosing is the exercise, not the generating, and the sentence is what turns a hunch into a defensible standard.
• Critique before you accept. Spend the first pass listing what is weak in an output rather than what is usable. Naming the flaw is what calibrates the eye over time.
• Keep a standard you can point to. A reference example of work you consider excellent gives judgment an anchor instead of a vibe, and makes it teachable to others.
• Protect the reps that feel slow. The moments where you slow down to evaluate by hand are where taste is built. Automating them away feels efficient and quietly removes the only thing that was training you.

None of this rejects the tools. It points them at generation, where they are extraordinary, and keeps the evaluative seat human, where the leverage now lives.

What this means for the people who lead

The competitive picture has quietly inverted. When a competitor can clone your feature set in a weekend of prompting, the durable advantage is no longer what your team can build. It is what your team can tell apart. Two organizations with identical tools will diverge on the quality of the thousand small choices nobody can outsource, and that divergence compounds.

The uncomfortable part is that taste is slow to grow and easy to lose. It does not arrive with a license or a model upgrade. It is the one capability that gets stronger the more your tools improve, because the better the generator, the more everything depends on the person deciding what is worth keeping.

What leaders should do

1. Measure and reward judgment, not output volume. Track how often work is improved or rejected in review, not how much was generated, so the people with the best eye are visible and valued.
2. Make engineers and writers defend their choices in one written sentence. Requiring a stated reason for the version that ships converts private instinct into a standard the team can learn from.
3. Protect the slow evaluative steps from automation. Identify the review and critique moments that build judgment, and keep them human even when a tool could skip them.
4. Hire and promote for discernment. When generation is commoditized, weight a candidate's ability to critique and choose at least as heavily as their ability to produce.