The Producer That Couldn't

A while back, I came upon a few clips from "60 Minutes" where American record producer Rick Rubin was being interviewed by Anderson Cooper. In the interview, Rubin self-attests to having no technical abilities whatsoever. He states himself that he knows nothing about music!

I know what I like and what I don't like. I am decisive about what I like and what I don't like.

Cooper, quite reasonably, pressed him on it:

So what are you being paid for?

The confidence that I have in my taste and my ability to express what I feel have proven helpful for artists.
Rick Rubin

An aside, if you are unfamiliar with Rick's work

Slayer, Tom Petty, System of a Down, Johnny Cash, Slipknot, Linkin Park, the Red Hot Chili Peppers. The range is the point. I grew up listening to a good number of those records, all shaped by the same pair of ears, unbeknownst to me at the time of course. The full discography is a rabbit hole worth falling into.

Paid For Taste

So here is a man who, by his own cheerful admission, cannot play the instruments, cannot work the mixing desk and holds not a scrap of music theory, and who has nonetheless spent four decades handsomely paid to sit in a room and pronounce on what is good. Four decades of tasting other people's cooking, in effect, at the most rarefied table the industry can lay.

Watching him work is telling in itself: eyes closed, swaying, letting the melody and the rhythm move through him, and somewhere in all of that a verdict quietly forms, long before there are any words to hang on it.

Something measurable is happening in that stillness. The pleasure of music, it turns out, lives less in the sound than in the prediction. Your reward system quietly wagers on where a piece is headed and pays out when it lands somewhere better than you guessed, and the dopamine even arrives in the beat before the peak rather than on it. Rubin's ear is just decades of those wagers, settled so many times that the verdict shows up before the words do.

Interactions Between the Nucleus Accumbens and Auditory Cortices Predict Music Reward Value

Anatomically Distinct Dopamine Release During Anticipation and Experience of Peak Emotion to Music

Parts of the interview struck a chord with me, and I found myself drawing parallels to what I see in the tech industry today. But before we get there, it is worth asking a more basic question. What actually is taste?

Sweet Wine And Other Contrasts

Begin with the literal sort. Ask most people what wine they favour and a fair number may admit to a preference for something sweet. Put that to a sommelier and you will naturally get a different response. Dry reds are meant to be the "grown-up" answer (or so they say), all tannin and structure and hard-won appreciation, and sweetness is where one is presumed to have loitered before one knew better.

Except no one nursing the sweet glass is having a lesser evening for it. Their wine is every bit as vivid to them as the dry red is to the connoisseur doing the wincing.

The wince has a rather human explanation, mind you. Nobody is born liking tannin; bitterness only turns pleasant after enough exposure, so a taste for dry reds is mostly proof that the training happened. And preferences that take years to acquire have a way of becoming badges for the people who acquired them. The French sociologist Pierre Bourdieu wrote a whole book on exactly this, Distinction, taste operating as a marker of class. Quoting him here is, I am aware, precisely the sort of badge-polishing he was describing, but the line is too good to leave out: taste classifies, and it classifies the classifier.

Distinction: A Social Critique of the Judgement of Taste

Trained By Contrast

I believe taste, in the literal sense, is built out of contrast. Savoury means almost nothing until there is sweetness to lean against; salt to sharpen the caramel, acid to cut through the fat. And we are the luckiest generation ever to sit down to dinner, with chillies from one continent, citrus from another and some fermented oddity from a third all crowded onto the same supermarket shelf. Our palates are schooled by more of that collision than anyone before us, which is why "good taste" refuses to hold still long enough to be pinned to a wall. Each new contrast quietly re-teaches the palate that will judge the next one.

Some of it is just hardware. Around a quarter of people are "super tasters" whose bitter-taste receptor gene TAS2R38 makes coffee and tonic water far harsher than the rest of us feel them, and you will not talk them out of it, so the wine drinker reaching for something sweet might not be lazy or untrained at all. They could just be running different hardware. The rest of it gets written by exposure: pair a flavour with enough repetition and the machinery that decides whether you like it quietly rewrites itself, and it will rewrite again the moment the bargain changes. Nobody issued you a palate. You have been cultivating one your whole life, every time you sit down to eat.

And the palate you school at dinner does not have the courtesy to stay at the table. The same faculty, forever appraising each new thing against everything it has already met, turns its attention to whatever you spend your days among.

Taste In Software & Systems

For a good many of us, that turns out to be software. Now and then you cross paths with someone who simply has it: taste in design, or its near relation, a feel for standards and process and the unshowy excellence of a system built with care.

Apple is the obvious example, and obvious is exactly why it works. You can level all sorts of criticism at them, and plenty is fair, but pick up one of their devices and you feel it anyway. The weight sits right in the hand. The box opens with that slow, deliberate exhale of air, like it was engineered to make you wait half a second longer. All of that is the residue of thousands of small moments where somebody who trusted their own palate looked at a nearly-finished thing and said no, not yet. I still cannot bring myself to throw the packaging away, which is a slightly embarrassing thing to admit about a cardboard box.

Software perhaps sometimes admits the same flair. A codebase whose naming is coherent, an API that reads as though a single mind presided over it, an interface with no discordant seams (although granted more mathematical in nature in this sense than food, who says maths can't be tasteful ☺️).

Taste in a system is almost never perceived head-on, any more than you consciously admire the equilibrium of a well-composed dish. You simply register the absence of friction, the merciful sense that nothing is clamouring for your attention.

The Producer's Chair

Chatting with a colleague of mine recently, he phrased something rather eloquently: "LLMs can only look backwards, never forwards."

Well, me being myself, I thought "huh", that is a rather profound statement... Neurological fireworks commence!

The models are trained on everything already made; they interpolate the past. There is a half truth to this, and the interesting part is the half that's missing. Because Rubin looks backwards too, thousands of records deep, and the whole first half of this essay is backward training of one sort or another. A palate is only the food already eaten. So the gaze cannot be the thing that separates us from the machine, since we are all of us looking the same way. The half it misses is what the looking produces.

Rubin's thousands of records forged one particular pair of ears that had made their mind up about things, while the model swallowed every palate at once and committed to none. You could almost say it has too much taste, everyone's at the same time, and a taste that excludes nothing has nothing to say, since on this essay's own terms taste is exclusion. He told us as much: decisive about what he likes and what he does not. And the other half is that Rubin's backward look was felt, the verdict forming in the body before the words arrived, which is where a commitment comes from at all.

Suno is where this stops being abstract for me. In the right hands it can reverberate ideas and expressions into something audible far sooner than the crafted way ever could, and fair play to anyone using it like that. My own interest dies about ten seconds in, because what comes out never quite carries a groove or any reason to be, which is about what you would expect from every palate at once.

The model has looked at more than he could hear in a thousand lifetimes, and not one second of it ever landed as liking, no wager placed and none paid out. What they furnish instead is speed, and speed unaccompanied is not the same as taste. Given time, it is your judgement that decides whether the output was worth anything. I dwell on this at greater length in this article.

And here is where it landed for me. With "teams" of agents now doing the making, the engineer's hands are coming off the instruments. We are shifting, whether we asked for it or not, into the "producer's" chair. Call it the "Rubin role". You sit in the room while the band plays, you decide what is good, decisively, and you express it clearly enough that the band can act on it.

Now just imagine, as a fun mental exercise, if real-life Rick were to employ --dangerously-skip-permissions in his processes. With Claude Code, we can pass that flag and let the agent do whatever it likes without ever stopping to ask us first. Perhaps, in some strange organic way, Rubin already does?

Software engineers need good taste. Arguably we always did, but AI has just promoted it from a nice-to-have to the differentiating skill, because when any fella and his laptop can generate the notes (or 20k lines of code in a day), all that is left to be paid for, really, is knowing which ones to keep.

A producer does more than pick the good takes though. He tells the band what to play in the first place. These days, I spend most of my time writing specs and documents that get turned, near enough directly, into code, and I have written plenty of bad ones. When a spec is vague about the hard case, the code is vague about it too, and you find out much later than you would like. So the taste has not gone anywhere. It just has to show up earlier now, in the thing you write before any code exists.

Which is really only to say that a document is a piece of writing, and writing has been a matter of taste since long before any of us turned up to fret over it. A spec can be lucid or muddled, generous to the poor soul who inherits it or quietly hostile, in precisely the way a novel can. We seldom think of a design doc as prose, yet prose is all it is, and the very instincts that make a novel sing or fall flat are the ones presiding over whether the thing you handed the machine was any good.

But Is Your Taste Any Good?

There is, of course, an obvious objection loitering here. How do you know you possess good taste? What even is good taste? The easy answer is that it is subjective the whole way down, with no bedrock to strike. Only that is not quite right, and the rest of this piece keeps quietly contradicting it, because in software taste turns out to be judgement on a very long feedback loop. The bill does come, just years late, and it tends to land on whoever inherited the code rather than whoever wrote it.

Culture and creed aside, a human is a human. A jarring system or product leaves a sourness in the mouth, and you know the flavour at once: the settings screen that behaves unlike every other screen, the API that reports its errors in three contradictory dialects, the app that contrives to make you feel stupid for holding it wrong. And yet there will always be those who take to it regardless, and sincerely so. Some people simply prefer the sweet wine, and, as with the drinkers from earlier, they are having their experience.

Hard To Teach

So taste cannot be graded, and it cannot really be taught either. Rubin did not take a course. He listened to thousands of records until the contrast between them trained his ear, and then, the part most people skip, he became decisive about what he felt. It is the same for us: read a lot of code, use a lot of products, feel enough jarring seams and enough seamless ones that the difference becomes something you sense long before you can explain it. Nobody hands that to you in a curriculum, which is why so much of it fails to transfer in one. You have to go and get it, one contrast at a time.

The Awkward Thing About Software

Software is an awkward place to talk of taste, mind you. The thing is mathematical underneath. Code compiles or it throws, the tests turn green or red, and beneath it all lies a hard floor of correctness that no amount of elegance will excuse you from. Set against something so binary, taste can look like ornament brushed on after the fact.

Except anyone who has lived in a codebase long enough knows the poetry is real. I have deleted a working function and rewritten it, changing nothing a test could see, purely because the first version made me tired to look at. Two solutions pass every check and one of them still makes you exhale while the other makes you brace for the day you have to change it. The good one names things well enough that you stop carrying them around in your head, and it leaves a shape the next edit drops into. A green test suite has nothing to say about any of that, which is how it gets waved off as fussiness.

The Taste Of Another

The trouble is that "it works" is only true on the day you measured it, and code has to keep being true for years afterwards, usually in the hands of someone who was not in the room when you wrote it. I learned this the hard way at Amazon, and not from my own code. Most of the pain came from systems I inherited, things built by people long gone, for which I carried the pager without ever having the capacity to change them. You get woken at three in the morning by a decision someone made years ago, in a service you barely understand, and there is no fixing it tonight. You are just holding it upright until sunrise. That is the part nobody warns you about.

That is where the value of this subject actually hides. It is the outage that never happened because the failure mode was sitting right there where nobody could miss it, the page that never came. Nobody thanks you for it, and that is rather the point, the same way nobody compliments a dish for not being too salty, or the way I still cannot bring myself to bin the box the phone came in.

And the absence compounds in a way a single bad decision never hints at. Each shortcut works, in the narrow sense the tests care about. Stack a few thousand of them and you get a system nobody dares touch, where a one-line change takes a week because it could be anywhere and the code refuses to tell you. That is what taste quietly prevents.

Of course, in reality, many of those prior decisions were never about taste at all. They were made under pressure, against time constraints that were often arbitrary. You could argue that undoes the whole premise. Or you could argue it is the better case for the machines: let them absorb the pressure and weed out the rushed decisions earlier, so we can "Rubin" our systems up to a standard we actually chose.

So no, taste is not decoration. It is most of the reason software you shipped a year ago is still standing, and still bearable to work on, for whoever inherited it. It is a hard thing to teach, yet one must learn. Experience is the learning, and education, more often than we would like to admit, just gets in the way. And if you don't fancy good wine or music, then at least have good taste in your own work.

But hey, that's just my opinion after all.

via GIPHY

P.S. pun was intended.