There is a kind of skill AI makes easier to fake and harder to replace.
It is not fluency. Fluency is cheap now. A model can make a sentence sound finished before the thought has paid rent. It is not volume either. Anyone can generate a hundred options in five minutes and call the pile exploration.
The harder skill is knowing which option deserves to live.
Wyndo named that skill as taste in a short note this week, and the line stayed with me longer than I expected. His point was clean: in the AI era, creation gets cheaper, judgment gets more expensive. Taste is what lets you choose what resonates, what is right, what is worth the bet.
I agree with the direction. But I kept circling one missing layer.
Taste is usually described as the premium output. I want to name the machinery underneath it.
The framing names a deliverable. It does not name the stack.
A creak in the next room. A flash at the edge of vision. One word in a paragraph that does not belong. Your brain finishes the sentence before your awareness catches up. That is what taste actually is when you watch it run. Minimal signal in, completed pattern out, an output that feels like a decision but is really a prediction.
It works only when two conditions hold. Most writing about taste mentions the first. The second one matters more.
Slow root: a corpus that survived
Taste is the output of a predictor trained on a corpus of falsified guesses. Not on hours. On hours that ate consequences.
Two examples, picked because they punish prediction differently: animals punish projection, markets punish narrative.
Animals are a clean training set. A dog does not negotiate politically. A cat does not perform comprehension. The feedback is non-verbal, immediate, pre-rational. You guess what they want, the guess holds or it does not, the corpus updates without commentary. Anyone who has watched a living creature at length learns to read minimal signal because nothing else worked.
Markets are the adversarial version of the same trainer. A stop-loss does not care how good your story sounded. The corpus updates by deleting wrong predictions from your account balance. Trading taste, when it exists, is what is left after the bad predictors have been stripped.
Cross-domain is the multiplier. A pattern that survives in one domain may be local superstition. A pattern that survives two unrelated punishments starts becoming taste. That is not deeper experience. It is parallel experience. The difference between one expert opinion and two unrelated ones that agree.
Without the corpus, taste is a hallucination with confidence.
Clamp: discipline that protects signal
This is the part the framing skips. Intuition without discipline drowns in noise. The same brain that completes a sentence on minimal signal will also complete it on a wrong signal if nothing is filtering the input.
Discipline is the clamp. Mechanical gates that catch what you would have waved through on a tired afternoon. The habit of scanning a paragraph for the one word that does not belong, even when the rest reads fine. A pipeline that surfaces friction instead of smoothing it away. Mined receipts instead of vibes. None of this is taste. All of it is the scaffolding under taste.
Without the clamp, your own intuition will read noise as signal and ship it. With the clamp, the predictor runs on clean input. The clamp is not the smart layer. It is the boring layer that makes the smart layer possible.
A vibe coder has the predictor and no clamp. The output looks fluent and breaks under load. A disciplined operator with no corpus has the clamp and no predictor. The output is correct and uninteresting. Taste lives where both hold.
Why AI moved the price
AI cheapens generation. A hundred options in five minutes is real. The line that says “creation is cheap, judgment is premium” is correct as far as it goes. It stops one step short.
If the model wrote the hundred options, your corpus did not grow. You did not falsify a guess. You did not pay the price that updates a predictor. The output looked like work. The training did not happen.
Taste does not become premium because it is rare. It becomes premium because the raw material that builds it, lived consequences under discipline, becomes rare. Anyone can generate. Fewer people are still paying the cost that compounds into a working predictor. That cost is what AI did not cheapen.
What I would tell a younger me
Pick at least two domains, and make sure they punish you in different ways.
Build the clamp before you trust the intuition. The clamp is unglamorous. It is what keeps the intuition honest.
Treat the corpus as something you pay into, not something you are owed. Nothing about it is given.
Wyndo named the deliverable. The stack underneath is a corpus that survived, plus a clamp that keeps it clean. Premium judgment is not a vibe. It is a prediction that survived.