Machines Don't Have Taste

Ask an AI to write you a story and it will, in seconds, and the story will be fine. Ask for a song: fine. A logo, a poem for grandma, a game idea: fine, fine, fine. We have built the first machine in history that produces unlimited fine — and in doing so, entirely by accident, we made fine worthless.

So what's left? What's the thing the machine can't flood?

Knowing which one is good. And it turns out that skill lives only in people — and it can be trained.

To see why the machine's output gravitates toward fine, you have to remember what it is. A model learns by ingesting millions of examples of what people have already made, and its whole method is to produce what's probable given all of it. Probable is a synonym for typical. When you ask for a story, you're drawing from the middle of the distribution of stories — the massive, competent, unsurprising middle. You can prod it toward the edges with a clever request, and the ceiling keeps rising; today's fine would have astonished anyone five years ago. But the pull toward the center never goes away, because the center is what the machine is made of. Average isn't its failure mode. Average is its physics.

Now put that machine in everyone's pocket and ask what happens to the world. The volume of fine goes vertical. Fine essays, fine art, fine apps, fine music — an ocean of competent output rising by the day. And in an ocean of fine, the scarce thing — the valuable thing, the thing that will distinguish one kid from every other kid with the same tools — is the ability to look at twelve fine things and know which one is actually good. And, harder still, to say why.

That ability has an old-fashioned name: taste. And I want to rescue that word, because somewhere along the way it started sounding like a luxury — a thing for wine people and design snobs. It never was. Taste is discernment. It's the trained ability to perceive quality: to hold two things that are superficially similar and feel, with reasons, that one of them is alive and one of them is merely correct. Every field has it. There's engineering taste — this code works, but that code is right. Kitchen taste, coaching taste, comedy taste. The surgeon choosing where to cut and the twelve-year-old choosing which of ten dragon names is secretly the best are exercising the same faculty at different stakes.

Here's what should get your attention: we are handing every child a machine that makes everything except that faculty cheap. Execution: cheap. Drafts: cheap. Options: infinite and cheap. The entire creative process has collapsed in cost except for one step — judging — which has quietly become the whole game. The generation growing up right now will be the first for whom making things is trivially easy and choosing well is the entire job. And I can tell you what we do to prepare kids for that job, institutionally, as a society: almost nothing. There is no class called Knowing What's Good. We grade kids on producing acceptable output — which is precisely the thing the machine just took over.

So how does a person actually get taste? Not by being born with it — that's the myth that makes the word feel snobbish. The path is known; every craft has run it for centuries, and it has three steps.

Step one: volume of exposure, with attention. You develop an eye by looking at a lot of things while actually asking yourself what you think of them. Passive scrolling doesn't do it — we covered that machine in another talk. The attention is the active ingredient: this one works, that one doesn't, and the beginnings of a hunch about why.

Step two — and this is the step almost everyone skips — saying the why out loud. Taste solidifies when the vague feeling gets forced into words. "I like the second logo" is a preference. "I like the second logo because the shape still reads when it's tiny, and the first one turns to mush" — that's taste under construction. The articulation is the workout. It's why critique is a formal ritual in every serious art school and design studio and kitchen on earth: not because young chefs enjoy being told their sauce is thin, but because defending judgments in front of other people is the only known gym for this muscle.

Step three: making things yourself, and living with your choices. The person who has tried to draw the dragon knows what to look for in dragons. Judgment without making drifts into snobbery; making without judgment plateaus at fine. The loop is the thing — make, judge, defend, repeat.

Which brings me to the machine one more time, because here's the twist in this talk: the same machine that floods the world with fine turns out to be the best taste-training equipment ever built. For step one, it's an infinite exposure engine — ten versions of anything, on demand, ready to be ranked. For step three, it collapses the cost of trying, so a kid can make and remake until their hands catch up with their eye. The machine is a menace to taste only when it's used as a vending machine: ask, take the first thing, done. Used as a sparring partner — generate ten, rank them, defend the ranking, tell it what's wrong with number one, make it better — it becomes the gym. Same tool. The difference, again and always, is the posture of the kid.

This is why, at our clubhouse, no build is finished when the machine spits something out. That's when the interesting part starts: which of these is best, and defend it. The defending, out loud, to other kids who disagree — that's the curriculum. Everything else is setup.

I said the machine can't flood the market with good judgment, and that's true. But I want to end one level up from the job market, because taste was never really about employment. A person with taste lives in a different world. Meals are more interesting. Songs have architecture. The difference between the true sentence and the almost-true sentence starts to matter, and a kid who can hear that difference in a sentence can hear it in an argument, and in a promise, and eventually in a life. We are about to find out what happens when everything is easy to make. The kids who'll thrive aren't the ones who can generate the most.

They're the ones who can hear the difference.

Teach them to say why. The machine will handle the rest — and it will never, ever know which one was good.