When Everyone Can Build, Taste Is the Bottleneck

If AI made building cheap, what got expensive?

Production got cheap, judgment did not. It is worth being precise about what AI changed and what it left untouched. It changed the cost of generating things to almost nothing. It did not change the cost of knowing whether a thing is any good.

A model will generate the wrong idea beautifully. It will write a clean, well-structured feature that should never have existed. It has no stake in whether the thing helps anyone, no memory of the user you are actually building for, no instinct for the difference between clever and right. It answers the question you asked, not the question you should have asked.

That gap is the whole game. The generation is free. Deciding what to generate, and what to do with what comes back, is where the work quietly relocated.

What taste actually means here

When I say taste I do not just mean whether something looks nice, though that counts. I mean the accumulated judgment you apply across an entire build, most of it invisible in the final product.

• Which idea, out of the ten you could ship this month, actually deserves to exist.
• Which feature to cut so the thing stays sharp instead of sprawling.
• Which of the model's twenty suggestions is right and which nineteen are merely fine.
• Whether the wording on a button is honest or just filler.
• When a thing is genuinely good, and when it only looks finished.

None of those are questions a model can answer for you, because none of them have a single correct output. They are calls. Taste is what you use to make them, over and over, faster than you can explain why.

Abundance makes selection the job

When options were scarce, producing more was the win. The person who could ship two more features or write two more pages was ahead, because supply was the constraint. That world is over.

Now supply is infinite and nearly free. You can have as many versions of anything as you are willing to sit through. The constraint flipped. The scarce, decisive skill is no longer making more, it is choosing well from the flood: looking at ten plausible versions and knowing, quickly, which one is actually good and why.

This is why I think taste gets more valuable in an AI world, not less. The easier it is to produce options, the more the whole outcome rests on the person doing the selecting.

There is some evidence for the flattening. Astudy of AI-assisted short storiesfound that individual pieces were rated higher while the set as a whole grew less diverse. And METR found thatmany benchmark-passing pull requests would not be mergedby the maintainers whose repositories they targeted.

Can you learn taste, or do you just have it?

Taste is built, not downloaded. You cannot prompt your way to it. There is no setting for it and no model that installs it. Taste comes from shipping real things to real people and then living with the results, which is slow and a little uncomfortable and does not compress.

Every small tool I have put out taught me something the last one did not. Building Comoji, Burly, and Premail meant making a hundred tiny judgment calls each, about what to include, what to refuse, how a thing should feel when you use it for the tenth time. StockCar was the same, a long series of decisions about what belonged and what did not. You do not get those reps from reading about building. You get them from releasing something and watching what happens.

That is the part AI cannot hand you. It can accelerate the making. It cannot give you the scars that tell you which making was worth it.

Where I actually spend it: the last ten percent

A model gets you to ninety percent fast. That part is real and I am not going to pretend otherwise. The rough draft, the scaffolding, the plausible first pass, all of it arrives quickly now.

But the last ten percent is entirely taste, and it is the ten percent people feel. It is signing and notarizing a Mac app so it opens without a scary warning. It is a picker that responds the instant you hit the key instead of a beat later. It is filtering someone's mail quietly enough that they forget the tool is even running. None of that shows up in a feature list, and all of it is the difference between something that works and something people trust.

That last stretch is exactly where the model stops helping and you start. It will not tell you the animation is a hair too slow or that the empty state feels cold. You have to notice, and noticing is taste.

The part that does not commoditize

Here is the uncomfortable, freeing truth of it. When everyone has the same generation tools, the raw output converges. Ask a hundred builders to make the same app with the same model and you get a hundred versions that rhyme. The tool is no longer the edge, because everyone has it.

What does not converge is the judgment layer on top: what to build, what to leave out, what to polish until it disappears. Caring about those things turns out to be a competitive advantage precisely because it cannot be automated. The model does not care, and it never will. You can.

In a moment when it is trivial to flood the world with generic, plausible software, the scarce thing is work that was clearly made by someone who gave a damn about the details. That is not nostalgia. It is the only part of the process the machine cannot copy.

The test I keep applying

Cheap generation changed how much I can make. It did not change what is worth making, and if anything it raised the premium on getting that call right. So before I ship anything now, I run it through the same four questions:

1. Should this exist at all, or am I building it only because I can?
2. What here can I cut without losing the point?
3. Is this actually good, or just plausible?
4. Would I be proud to put my name on it?

When the answers hold up, the thing tends to be worth shipping. When they do not, the model was happy to help me build something I should not have. The making is cheap now. Knowing what is worth making is the whole job, and that has always been taste.

← All postsNext: Building a Nonprofit →