Taste is the thing you can't download

Why the most defensible skill right now is the one that doesn’t average.

There’s a specific kind of panic I keep watching people go through. They start learning every tool, collecting prompt techniques, subscribing to newsletters about model releases, trying to stay ahead of something that moves far faster than anyone can track.

I understand the instinct completely. I think it’s aimed at the wrong thing, and the reason why sits in how these systems work.

What the machine is built to do

A language model is a probability engine. That’s the design specification rather than a criticism. It takes an enormous body of text and returns the most probable continuation, which means what comes back is, structurally, the middle of the distribution.

The average. By construction, on purpose, as the intended behaviour.

This explains the flat feeling you get sometimes when reading generated text and can’t say what’s wrong with it. Nothing is wrong with it. It’s doing precisely what it was built to do, which is to produce the unremarkable centre of everything that has ever been written on the subject. The absence of a wrong note and the absence of a right one turn out to be the same absence.

For a very large amount of work, that’s not just acceptable, it’s a gift. Most writing doesn’t need to be exceptional. Most first drafts, most summaries, most boilerplate, most of the fifty small obligations that clog a week. Let the machine produce a competent average and take the hours back.

The trouble starts at the edges, which is where the value tends to live. When you’re trying to reach one specific person, or make a call that depends on context nobody wrote down, or stand out in a space where forty other people have identical tools, average is invisible. It reads like nobody was home.

Taste is a small dataset that only you have

Taste is judgment shaped by particular experience, and that’s exactly why it doesn’t average.

It’s assembled out of things that only happened to you. What landed with this team and what died in the room. How this customer talks compared to how that one thinks. Which stakeholder needs context before the decision and which one finds that patronising. The political history everyone knows about that project and nobody has ever written down. The three times you shipped something you were proud of and watched it get ignored, and the once you shipped something rushed that people still use.

There’s an old line that taste is improved by practice and perfected by comparison. I think that’s roughly right, and the comparison half is the part people skip. Practice alone produces habits. Practice plus honest comparison, against what you expected and against what other people achieved, produces calibration.

Your dataset is small, messy, specific to your world, and unavailable to anyone else. It’s also, crucially, the only dataset that contains the outcomes. The model has read everything anyone wrote about what should work. It has read almost nothing about what happened afterwards, because we mostly don’t write that part down.

How it’s really built

Pay attention to what happens after you make something. That’s close to the entire method, and it’s harder than it sounds, because we’re all onto the next thing by the time results arrive.

Did the design get used the way you expected. Did the strategy document change anyone’s behaviour, or did it look professional in a folder for a quarter and then get superseded. Did the feature get adopted, or did it sit there while people carried on with the spreadsheet.

Taste develops in the gap between what you predicted and what occurred. Every time you close that gap slightly, your instinct gets better calibrated, and after enough years of it you start being right about things you can’t fully explain. That’s a strange sensation to sit with, and it’s also the most valuable professional asset most people never notice they’re accumulating.

The part I’m still unsure about

I want to leave this one open, because I don’t think it’s settled.

If a lot of the work that used to build taste is now being done by machines, it’s not obvious where the next generation gets their reps. Taste came from making a hundred mediocre things and watching most of them fail. If the mediocre things are now generated in seconds and shipped without the same investment, the failures might not teach anyone anything, because nobody’s ego was in them.

I don’t have a clean answer to that. It might be that judgment gets built at a higher level, on which things to generate and which to discard, and that’s fine. It might also be that we lose a rung on the ladder and don’t notice for a decade.

What I’m confident about is the narrower claim. When everyone has the same tools, the differentiator is the judgment that the tools can’t reach, and that judgment is made of your specific accumulated evidence about what actually works here.

That part is still yours.