When AI makes competent output cheap and abundant, the competitive advantage shifts from production capability to judgment — the ability to distinguish what is generic from what is true and worth pursuing. "Taste" in this context means distinction under uncertainty: what you notice, what you reject, and how precisely you can diagnose what feels wrong. However, taste alone is insufficient; it must be combined with authorship, stakes, and the willingness to build something that could not have emerged from the statistical average.
Why AI flattens the middle
LLMs are pattern-compression engines that recombine absorbed patterns at speed, naturally trending toward the safe center of the distribution. The result is a "crowded 7 out of 10 world" where:
- Landing pages share the same structure with different logos
- Product copy could describe almost any app
- Visual design looks modern but not memorable
- Essays have clean headings and little lived judgment
Average used to be hard enough to create separation. Now it is abundant.
The new bottleneck: judgment
AI compresses the cost of first drafts, moving value downstream. The scarce skill is now refusal — the ability to say:
- "This looks fine, but it is too generic"
- "This sounds impressive, but it hides the real trade-off"
- "This interface is polished, but it doesn't fit how the user actually thinks"
Taste becomes useful when it moves from vibe to diagnosis.
AI as a mirror for taste
Generating 10+ versions of an artifact with AI reveals how clear your own judgment is. If your critique stays vague ("this feels off"), your taste is underdeveloped. If it becomes precise ("this collapses a regulatory constraint into marketing language"), your judgment is stronger than the model output.
Why taste alone is not enough
If humans reduce themselves to selecting from AI outputs, they become discriminators in a machine-driven loop — useful but fragile. Important work historically emerged from co-creation under constraint: arguing with reality, collaborators, budgets, materials, and consequences.
What humans still own that models cannot:
- Holding the stake — real consequences (trust, regulatory exposure, team capacity) that don't fit in a prompt
- Working with the truly new — genuinely novel ideas look wrong at first because they don't resemble the training set
- Choosing direction — what problem is worth solving, what trade-off is acceptable, what you refuse to optimize for
A practical loop for training taste
- Pick one high-leverage artifact from your week
- Generate 10–20 versions with AI
- For each, write one sentence starting with "fails because..."
- Rewrite the strongest version with hard constraints (no buzzwords, one idea per sentence, must acknowledge a real trade-off)
- Ship the final version and observe what happens
The goal: build a sharper rejection vocabulary.
Subtractive taste is the moat AI cannot copy
When code generation is cheap, the scarce skill is knowing what to leave out. Ethan Ding's analysis of why coding agents have not produced runaway product advantages lands on the same conclusion from the software-engineering side: at the frontier, the bottleneck is "tastemakers," not tokens. The evidence is comparative — Linear (178 people, $100M ARR) scores higher on consumer quality than Jira despite Jira having ~56× more cumulative engineering effort; the difference is not better-drawn boxes but a specific creative vision executed with restraint over years. Quality and codebase mass are not the same thing.
- The durable advantage lives in specific people who hold a theory of what to delete, compress, or refuse — Nan Yu at Linear, Kelly Johnson at Skunk Works (whose SR-71 was fast because Johnson had a theory of what to leave out, not because his team produced more blueprints)
- tinychat (comma.ai) made the principle operational: an alarm that triggered when the codebase exceeded a size threshold, and a culture that celebrated deleted code
- This taste-to-delete "isn't on any frontier model's roadmap" — and it becomes more valuable as the floor beneath it rises, because the statistical average is exactly what AI now produces for free
- Product-quality improvements at the frontier are bounded by how fast you can come up with ideas good enough to push it, not by how fast you can write code — and those ideas come from slow, uncomfortable thinking, not from sprinting through a backlog