There is a moment anyone who makes things will recognize. Two versions in front of you, both competent, neither wrong, and you have to pick. No metric settles it, and if someone asks you to justify the choice afterwards, the honest answer is usually something feeble: this one has more life in it. Some people make that call well, over and over, across years. Most of us don't, or only manage it inside one narrow specialty.
The old word for making it well is discernment. I think it is about to become one of the most valuable things a working person can have, and I want to try to say why without pretending I can define it precisely. Part of the point is that nobody really can.
Judgment you can't show your work for
The clearest illustration I know comes from publishing. Raymond Carver's early stories, the spare ones that made his reputation, were edited by Gordon Lish, and when the original manuscripts surfaced years later it became clear how heavy Lish's hand had been. Some stories lost close to half their length. Endings changed. People still argue about whether Lish perfected Carver or overwrote him, and I don't think that argument has a clean answer. What nobody disputes is that Lish could look at a page and see a version of it that didn't exist yet. He wasn't applying a rule. There is no rule that says cut this ending, and any rule that tried would wreck a hundred other stories.
That's the shape of the thing. Discernment is judgment that runs ahead of your ability to justify it. You know the sentence should go; asked why, the best you can offer is that it doesn't belong. I realize how mystical that sounds, and I want to be careful with it, because 'I can't explain it' is also what people say when they're guessing. The difference is the track record. Discernment is intuition that has been wrong in public often enough to have learned something from it.
What it gets confused with
It isn't having opinions. Everyone has opinions, instantly and in bulk. It isn't criticism either, and this distinction took me longer than it should have: finding the flaw is the easy half, it feels like intelligence while you're doing it, and it never requires you to commit to anything. The hard half is choosing one imperfect thing over another and living with the consequences. As for the snobbery version, dismissing things to advertise your own refinement, that's mostly about status. The genuinely discerning people I've met tend to be less certain than average, not more, because they remember their misses.
And they do miss. Decca turned down the Beatles after their audition on the first of January 1962 and signed Brian Poole and the Tremeloes instead. The line everyone quotes, that guitar groups were on the way out, comes to us through Brian Epstein and may well have been sharpened in the retelling. But the rejection itself is fact, and it was made by professionals whose entire job was this kind of call. I find that oddly reassuring. A skill you can fail at publicly is a real skill, not a mystique.
The bottleneck moves
For most of working history, the scarce thing was the output. Producing good work was slow and expensive, so the people who could produce were the valuable ones. That constraint is dissolving in front of us. A model will hand you forty drafts of anything before lunch, most of them competent, none of them obviously wrong.
I see this every week building an assessment platform. Generating plausible test questions used to be most of the work; now it is nearly none of it. The job has quietly collapsed into the other step, deciding which questions actually measure something, and that step used to be a quick pass at the end. Now it is the whole day. I don't think my situation is unusual. When production is free, choosing is what's left, and choosing well among many decent options, with no formula to lean on, is the thing I'm describing.
Here is where I should concede something. The standard claim, and one I used to make, is that machines can't do this part because taste lives outside what can be measured. I'm no longer sure that's true. Models are getting noticeably better at taste-shaped judgments, and I've watched one pick the stronger of two paragraphs, correctly, when I would have gotten it wrong. What I'm more confident of is narrower: if you can't make these calls yourself, you have no way of knowing whether the machine is making them well. You can delegate the choosing. You can't delegate knowing whether the choosing was any good.
When the numbers stop meaning anything
The tempting move, once options are free, is to let the metrics decide. Highest score wins, most engagement wins, best test numbers win. It feels rigorous. My favorite counterexample is the Netflix Prize, partly because nobody in it behaved badly. Netflix offered a million dollars for a ten percent improvement to its recommendation algorithm; a team won it in 2009; and Netflix never fully deployed the winning system, because the accuracy gain wasn't worth the engineering cost and the business had shifted toward streaming, where the problem looked different. The metric got optimized flawlessly. It had simply stopped being the thing that mattered, and no metric can notice that about itself.
I'm not against measurement. I use numbers constantly, and people who ignore them entirely are usually protecting an ego. But metrics can only count what someone already decided was worth counting, which means they are always facing backward. (Goodhart's law usually gets quoted at this point; I'll spare you.) Somebody in the room has to be able to tell when the number and the reality have come apart, and that somebody is having a very good decade.
The slow way in
As for developing it, I only know slow answers. Make calls whose outcomes you will eventually see, then actually go back and check; most of us make small predictions all day and audit none of them. Spend unreasonable amounts of time with good work in one narrow domain, since taste seems to come from comparison rather than theory. If you can find someone whose judgment reliably beats yours, argue with them about specifics until you can predict what they'll say.
None of this is efficient, and there is no certificate at the end. For a long time I filed that under drawbacks. Lately I've started to think it's the reason the skill will hold its value: the slow, uncompressible way you acquire it is the one part of the process nobody has figured out how to speed up. Though given the decade we're having, I'd write that in pencil.