Competence — the ability to execute a task to a professional standard — is collapsing toward free, and when a capability gets commoditized the premium does not vanish, it relocates. It moves to whatever was scarce all along and merely bundled with the competence. My forecast is that the thing it relocates to is taste: knowing what is worth making, which of a thousand competent options is the right one, and exactly where the line between good and great sits. Taste is the last moat because it is the hardest thing to specify, and the hardest thing to specify is the last thing to be automated. Everything a lab can write down, it will eventually teach a model to do better than you.
That is a bet about the direction of the frontier, not a mystical claim about human souls. So let me define the term precisely enough to defend it, because "taste" is usually invoked as the place where argument stops. I want it to be the place where the argument starts.
Taste is compressed judgment, not magic
Taste is an internalized, high-dimensional sense of quality, built from vast exposure and feedback, that lets you rank options you could not write a rule for. That is the whole definition, and every clause is load-bearing.
Think of it as a learned function. Somewhere in a sommelier's head is a mapping from an enormous input vector — aroma, structure, finish, ten thousand prior glasses — to a single scalar: this is better than that. The sommelier cannot hand you the function. They cannot enumerate the weights. But the function is real, it is stable, and it outperforms chance by a margin that pays their salary. That is taste: a compressed model of quality that runs faster than introspection and produces a verdict its owner often cannot fully justify. "I can't articulate why, but the second one is right" is not a failure of the expert. It is the signature of a model too high-dimensional to serialize into words.
This is why taste and competence come apart, and come apart harder every quarter. Competence is the ability to produce a professionally acceptable artifact. Taste is the ability to tell which acceptable artifact is correct here. A junior designer can now generate a hundred logos that are each, in isolation, competent — balanced, legible, on-trend. Competence at logo production has effectively gone to zero; a model does it in seconds. What has not gone to zero, and is the entire remaining job, is looking at the hundred and knowing that ninety-seven are wrong for a fintech that needs to signal boredom and safety, two are technically strong but will date within eighteen months, and one is right. The generation was the commodity. The selection was the value. The gap between those two is where taste lives, and it is widening precisely because generation got cheap while selection did not.
Why "right" is the part of the spec you cannot write down
The reason a model can generate the hundred but not reliably pick the one is not a temporary capability gap. It is structural, and it connects directly to the thing I have argued is the durable skill of this era: problem specification. A specification says what would count as a correct answer — the goal, the real constraints, the acceptance test. The premise of that argument is that you can write the spec down. Taste is the residual: the part of the specification that resists being written down at all.
"Right" for a given artifact depends on goals and constraints that are only partly specifiable. Some of them you can state — the logo must work in one color, the essay must persuade a skeptical CFO, the checkout flow must clear in three taps. But underneath the statable constraints sits a layer you feel and cannot fully articulate: this brand should read as expensive-but-not-cold; this argument should concede exactly enough to seem fair without going soft; this onboarding should feel effortless in a way that a competitor's, built from the same components, somehow does not. Those are real constraints. They discriminate hard between options. And they are underspecified in the literal sense that the person who holds them cannot serialize them into a prompt. You give the model the ten percent of the spec you can write, it satisfies that ten percent a hundred ways, and taste is the operator that closes the remaining ninety.
Building Kommerce made this concrete for me in a way no design seminar could. It is a cash-on-delivery commerce OS for markets where institutional trust is thin and buyers will not prepay strangers, so the entire product lives or dies on a feeling — does this look like something that will actually show up at my door. A model can generate twenty competent checkout screens. Every one satisfies the writable spec. But whether a screen reads as trustworthy to a buyer in Casablanca who has been burned before is a constraint I can barely put into words and can recognize instantly, because I have watched enough of them convert and enough of them bounce. That recognition is not in the spec. It cannot be, or I would have written it down and automated it. It is taste, and it is the part of the product a competitor with the same design tools cannot copy off the screen.
The training signal is sparse, expensive, and yours
Here is the mechanical reason taste resists automation, stated the way an ML person would state it. To learn a function you need a training signal, and taste's signal has three properties that make it brutal to acquire: it is sparse, it is expensive, and it is contextual.
Sparse: the label is a single comparison — "this one is better" — and it carries almost no information about why. One bit, no gradient, no explanation. Expensive: producing that one bit requires a genuine expert to actually attend to the options, which does not scale like scraping the internet; the web is full of competent artifacts and almost empty of trustworthy verdicts about which was right for which context. And contextual: the same artifact is excellent for one goal and wrong for the adjacent one, so a label collected in one setting does not transfer cleanly to the next. A model can ingest every logo ever made. It cannot easily ingest the ten thousand private moments where a great creative director looked at two near-identical options and knew, for reasons tied to a specific client on a specific day, which one shipped. That data was never written down. Much of it was never even conscious.
This is the same structure that makes it hard to evaluate an agent you cannot specify: the bottleneck is not the doing, it is the grading, and taste is the grading function. When the reward signal is cheap and dense, automation arrives fast — that is the story of every benchmark that fell. When the reward signal is one expensive bit per expert-hour, automation crawls. Taste sits at the far, slow end of that spectrum, which is exactly why I am willing to forecast it as the last moat rather than a temporary one.
The editorial seat becomes the valuable seat
Follow the labor consequence and it lands on a clean prediction. As execution tasks get automated, a job does not disappear; it dissolves into its component tasks, the automatable ones fall to near-zero cost, and the human premium concentrates in whatever is left. In field after field, what is left is the editorial layer: choosing, curating, directing, killing. The person who decides which of the model's outputs ships, and against what standard, and why.
You can already see the pattern. A film director's competence — operating a camera, cutting a sequence — is not why they are paid; the taste to know which take, which cut, which score is the job, and it was always the job. What AI does is strip that pattern out of its expensive wrapper and stamp it on every knowledge field at once. The editor was always worth more than the writer per hour; now every writer has infinite competent drafts and the scarce act is editorial. The art director was always worth more than the illustrator; now everyone has infinite illustration. The premium migrates to the seat that says this one, not those, because that seat holds the one function that did not get cheap.
| Layer | What it does | Direction of value |
|---|---|---|
| Generation | Produce a competent artifact | Collapsing toward free |
| Selection / editorial | Judge which artifact is right here | Rising, concentrating |
| Specification | State what "right" would mean | Rising where writable |
| Taste | The part of the spec you cannot write | The residual moat |
What to actually do about it
A frame that ends in "so cultivate taste" is useless, because taste feels innate and therefore un-actionable. It is neither. It is a learned function, and learned functions have known training procedures. Three of them.
First, buy exposure by volume, deliberately and outside your own output. The function is fit from examples, and most people starve it. They make a lot and look at little. Invert that. Consume great and terrible work in your domain at ten times the rate you produce it, and — this is the part people skip — study the ones that are almost right, because the decision boundary is learned at the margin, not from obvious masterpieces versus obvious garbage. The signal lives in the near-misses.
Second, get reps at judging, not just making, and get the label attached. Before you look at the answer, commit to your verdict — this is the better option, and here is the standard it wins on — then find out from someone with more taste than you whether you were right. That is the expensive bit, deliberately farmed. Making without judging trains your hands and leaves the quality function untouched. A weekly ritual of ranking options and then getting graded by a superior eye will move your taste faster than another year of production.
Third, position yourself and your product one layer up, at selection. If your value proposition is "we generate competent output," you are selling the thing whose price is going to zero. If it is "we tell you which output is right, and we are reliably correct," you are selling the residual. For a founder that means building the curation and judgment layer on top of cheap generation, not competing inside generation. For a career it means engineering your role so your hours flow toward the editorial calls and away from the production your competitor's model now matches for free. And when you hire, hire for judgment over output — a portfolio of finished artifacts tells you someone is competent, which is now abundant; a track record of choices that aged well tells you someone has taste, which is not.
Be honest about the two risks, because a forecast without its failure modes is marketing. Taste can ossify: a quality function fit on last decade's data confidently rank-orders a world that has moved, which is how senior experts become reliably wrong with total conviction. The defense is the same as the acquisition — keep the exposure fresh, keep taking the graded reps, treat your own certainty as the first thing to test. And the model itself may erode taste at the bottom: if everyone offloads the near-miss judgments to a tool that is decent-not-great, a generation may never fit the function at all, and the median eye drifts toward the model's mean. That is a genuine open question, and I will not pretend to know how it resolves. But it cuts the same way as the forecast: it makes the people who did cultivate taste rarer and more valuable, not less.
The machines will make everything competent. The question that will sort the next decade of winners is smaller and older than it sounds: out of everything that is now possible and passable, do you know which one is right — and can you tell before the market does.