I ask for a thing now and ten versions of it come back. All ten are competent. Several are better than what I would have produced alone on a Tuesday afternoon. None of them is obviously the answer, and the work — the actual work, the part that is mine — has quietly become choosing.
This is a short note about that, because I have written a great deal lately about machinery and I want to say the other thing plainly.
The gap changed sides
Ira Glass has a passage about creative work that everyone who makes things eventually gets sent. The version I remember: you get into this because you have good taste, but for the first years the things you make are not good, and you know they are not good precisely because your taste is intact. Taste, he says, is "why your work disappoints you."1 The prescription is volume. Make a lot of things, for years, and your ability climbs until it meets the standard your taste set at the beginning.
What happened over the last two years is that the ability side of that gap moved without anyone doing the years. Not for me specifically — for everyone, simultaneously, including people who never developed the taste in the first place. The floor came up to meet the ceiling.
So the gap did not close in the way Glass meant. It inverted. For most of the history of making things, the binding constraint was execution: you could see what good looked like long before you could produce it. Now execution is cheap and abundant, and the thing in short supply is the judgement about which output is the right one. The scarce input is the one that was never scarce before.
Why this is not a comforting story
There is a self-flattering version of this essay where taste is the special human thing the machines cannot take, and everyone reading nods. I do not believe that version, or at least I do not think it follows from anything.
What I think is narrower and less consoling. In 2024, Anil Doshi and Oliver Hauser ran an experiment with 293 writers and 600 evaluators: some writers got story ideas from a language model, some did not. The assisted stories were judged more creative, better written and more enjoyable — and the effect was strongest among the writers who had been least creative to begin with. The assisted stories were also more similar to one another than the unassisted ones.2
Read that twice, because both halves matter. The tool raises the floor, most for the people standing lowest. And it pulls everyone toward the same middle.
The consequence for anyone building anything is unsentimental: competence stops being a differentiator at roughly the moment it becomes universally available. If the assisted median is good, then being good is the new table stakes, and whatever distinguishes your work has to live above that line. Taste is not valuable because it is human. It is valuable because it is the part of the distribution the tools have not compressed.
What taste actually is, unromantically
I want to strip the word down, because it usually arrives wrapped in mystique and the mystique makes it sound unteachable.
Taste, in the work, is a selection function. Ten plausible options, one of which is right for this context, this audience, this constraint — and the ability to identify it quickly and say why. The "say why" is not decoration. It is the whole difference between taste and preference. Preference is "I like this one." Taste is "this one, because the second paragraph is doing work the others are only gesturing at, and the thing we actually need here is the work." One of those transfers to another person. The other does not.
Which means taste is legible, and therefore trainable, and therefore something you can build in a team rather than hire for and hope. Slowly. In the same unglamorous way it was always built — by making a great many decisions, being wrong in public, and noticing.
It also means taste is domain-bound in a way the romantic version denies. I have reasonable judgement about how a system should be shaped and whether a sentence is carrying its weight. I have close to none about a great many other things, and the failure mode I watch for in myself is assuming that taste in one domain licenses opinions in another. It does not. It is one of the ways experienced people become unbearable.
[ PLACEHOLDER: two or three real moments where judgement was the entire job — a thing you killed that was working, a version you rejected that everyone else preferred, a call you got wrong and only understood later. This note lives or dies on these; without them it is an argument, and it is supposed to be an account. ]
The part that does not scale
Everything else I have built at Gigaverse is designed to scale — machinery built once and reused, agents doing the leverage, one senior team across several companies. The premise is that the expensive parts transfer.
Taste does not transfer that way, and I have stopped trying to make it. You cannot write it into a system prompt, because a system prompt is a description of a judgement and not the judgement itself. You cannot delegate it to the thing whose output you are judging — that is circular, and the circularity is not fixed by a better model. And you cannot buy it in a hurry, because the acquisition cost is measured in decisions made and consequences absorbed, which is another way of saying time.
So it stays a per-person, per-domain constraint in a business otherwise organised around removing those. That is uncomfortable and I think it is simply true. The honest version of the leverage thesis is that agents multiply everything downstream of the decision and nothing upstream of it, which puts a hard ceiling on the whole model shaped exactly like the judgement of the people making the calls.
[ PLACEHOLDER: whether you have found any way to distribute this across the team — a review ritual, a standard, an argument you keep having deliberately — or whether it is still concentrated in you. If it is still concentrated in you, that is the more interesting admission and it belongs here. ]
I am aware this is the essay where a person in my position congratulates himself for having the one thing that cannot be automated, and I have tried to write the version that does not do that. The claim is not that my judgement is good. The claim is that judgement is now the constraint, that mine is the only one I can work on, and that it is a slower project than any of the technical ones I have written about here. I will know in a few years whether I spent the time well. So will you — it will be visible in the work, which is the only place taste has ever been legible.