← Writing

STO · AI

When machines do the execution, taste is the job

Rick Rubin cannot play an instrument, does not operate the mixing board, and has produced some of the most important records of the last forty years. For most of that time this looked like a curiosity. In a year when a model can generate a competent draft of almost anything, it looks like a preview. As execution gets cheap, the scarce and defensible thing is the one Rubin has always sold: judgment about what is worth making and what to cut.

The producer who can't produce

Rubin is disarmingly direct about it. Asked what he actually does, he has said, in effect, that he has no technical ability and knows the least about music of anyone in the room — what he offers is that he knows what he likes, and he trusts that reaction completely. Artists from Johnny Cash to the Beastie Boys to System of a Down did not hire him for his hands. They hired him for his ear and his presence: someone who could sit in a room, feel where a song went dead, and protect the part that mattered.

His 2023 book, The Creative Act, is usually shelved as a book about art. Read it again and it is something stranger — a manual, in the lineage of Taoist and Stoic texts, about how to make things without clinging to how they turn out. It almost never talks about technique. It talks about attention, non-interference, and getting the self out of the way so that judgment can operate cleanly. That is not a book about music production. It is a book about the faculty that survives when production is automated.

What taste actually is

It is easy to treat taste as a mystical word, which lets people who have it feel special and people who don't feel excluded. The mechanical description is less flattering and more useful. Taste is a compressed prior over quality — a fast, mostly non-verbal model of what is good in a domain, built by long exposure and thousands of small acts of discrimination. It is the pattern-recognition that remains after you have seen enough that you no longer need to reason your way to the judgment.

Two properties follow, and both matter commercially. First, taste is expensive to acquire: it is compiled slowly from immersion, which is why it cannot be downloaded and is unevenly distributed. Second, it does not transfer by copying. You can hand someone your conclusion — "cut the second verse" — but not the model that produced it. That combination, costly to build and hard to copy, is the shape of a durable advantage.

The best art divides the audience, where if you put out a record, and half the people who hear it absolutely love it, and half the people who hear it absolutely hate it, you've done well, because it's pushing that boundary. If everyone thinks, oh, that's pretty good, why bother making it?Rick Rubin, The Tim Ferriss Show, ep. 76, 15 May 2015

Notice what that judgment requires. It is not a technical skill; it is a settled point of view about what you are willing to make, held firmly enough to reject the safe, averaged option that a committee — or a model trained to predict the most likely next token — will always drift toward.

Why this is suddenly the founder's problem

For most of business history, execution was the bottleneck. Building the thing was slow and hard, so the people who could build were scarce and the edge lived in their hands. Generative models are collapsing that. When a competent draft of the code, the copy, the design, or the analysis is minutes away and nearly free, being able to produce is no longer where the scarcity is. The scarcity moves upstream, to the two decisions a model is worst at: what to make, and what to throw away.

This is Naval Ravikant's point about judgment as the leverage that outlasts effort, arriving on a faster clock than anyone expected. In a solo studio like mine, the AI is the band — it can play anything I can specify. That makes me, whether I like it or not, the producer. The value I add is no longer the volume of output; it is the quality of the selection: which of the ten generated directions is alive, which is competent and dead, and the discipline to delete the nine.

The practical form is a rule I keep: use the model for execution, never for judgment. Let it generate widely; do the choosing yourself, from a point of view you can articulate. The moment you let the model decide what is good — averaging its way to the least objectionable option — you have outsourced the one thing that was yours to keep.

A worked session: ten drafts, one judgment

Here is the abstraction made concrete, from an ordinary afternoon in a studio of one. The task is a name for a new product. The old workflow was to sit and strain for the one right word; the current one is to brief a model with the constraints — what the thing does, who it is for, what it must not sound like — and ask for forty candidates. Ninety seconds later there are forty names, and the part that used to be the whole job, generation, is over and worthless. Everything that matters now happens in the rejection.

Most of the forty are competent and dead. They are the averaged answer, the name a committee would approve and no one would remember, and a model trained to predict the likely next token will always drift toward exactly that. Three or four have something — a friction, an odd angle, a word that is slightly wrong in an interesting way. The work is to know which, and it is not a technical skill. It is a point of view held firmly enough to say no to thirty-six competent options and defend the one that divides opinion, for reasons you can articulate out loud.

The failure mode is precise and tempting: to ask the model which name is best. It will answer fluently, and its answer will be the averaged, least objectionable one, because that is what "best" means to a system with no taste and a loss function that punishes outliers. The moment you accept that ranking, you have handed the one irreplaceable act — the judgment — to the tool, and kept only the typing. The value you add is not the forty candidates. It is the selection of the one, from a point of view the model does not have and cannot manufacture.

Scale that single decision across a day — the copy, the layout, the feature to cut, the direction to kill — and the shape of the job in an AI studio comes into focus. The machine is the band; it can play anything you specify. You are the producer, and the entire value of the role is the quality of the no.

Where "taste is the moat" goes wrong

This is a comfortable thesis, and comfortable theses deserve the hardest look. It flatters exactly the people most eager to believe it — those who would rather not learn to build. Three objections are serious.

One: taste without craft is usually hollow. Rubin is the seductive exception that hides the rule. His ear rests on decades of total immersion in the craft, even if he never learned the console; his taste is compiled from proximity to the making. Taste that has never been near execution tends to be opinion with good lighting. The honest version of the thesis is not "skip the craft." It is "the craft is now a means of building judgment, and judgment is the output."

Two: taste is a weak moat in the strict sense. Porter's moats — scale, switching costs, network effects — are structural and defensible on a balance sheet. Taste is personal, hard to measure, and it walks out of the building when the person does. It can produce an edge, but calling it a "moat" borrows a solidity it does not have. A company whose only advantage is one founder's taste is one bad year, or one departure, from having no advantage at all.

Three: we canonise the taste that won. Survivorship bias is doing quiet work in every story like Rubin's. We build the theory of taste from the hits and never see the equally confident taste that produced flops, because those people did not get books and profiles. Some of what we admire as prescient judgment was, in part, luck that got narrated as skill afterward. Hold the thesis, but hold it knowing the evidence is selected.

So yes — as the machines take the execution, cultivate the judgment. Just don't mistake the flattering half of that sentence for permission to stop making things. Taste is compiled from the craft you are tempted to skip.

Sources

  1. primaryRick Rubin, The Creative Act: A Way of Being (2023).
  2. primaryRick Rubin in conversation with Tim Ferriss, The Tim Ferriss Show, ep. 76 (15 May 2015) — the passage on art that divides the audience, quoted from the published transcript.
  3. primaryRick Rubin in conversation with Anderson Cooper, 60 Minutes, CBS News (15 January 2023) — the remark about having no technical ability and trusting his own taste; paraphrased in the text.
  4. primaryLaozi, Tao Te Ching — non-action (wu wei) and non-attachment.
  5. primaryMarcus Aurelius, Meditations — attention and the discipline of the self.
  6. secondaryNaval Ravikant, "How to Get Rich" (podcast/essays, 2019) — judgment as leverage.
  7. secondaryMichael E. Porter, Competitive Strategy (1980) — the strict sense of a durable moat.