Mind the Gap — Between taste and talent

For most of the history of creative work, talent and taste traveled together, almost by necessity. To paint convincingly, you needed years at the easel. To write a working piece of software, you needed to master a language's syntax and its failure modes. To design a typeface, you needed a hand steady enough, and trained enough, to draw a stroke that felt inevitable. The skill required to execute an idea acted as a filter. If you had bad taste, you rarely had the patience or technical grounding to produce much of anything — the two were bottlenecked through the same narrow door.

Generative AI removes that bottleneck. A model can now produce a competent draft, a plausible layout, a passable riff, on command. The barrier that used to separate people with ideas from people who could realize them has effectively disappeared. This is, on balance, a remarkable democratization — more people than ever can externalize what’s in their head. But it has also exposed something that was previously hidden by the difficulty of execution itself: most of what made “good” work good was never really the execution. That collapse is not abstract in my own practice: on Lumasis, a design system built for Glydways’ autonomous vehicles, apps, and kiosks, a model can now draft a dozen plausible token structures in an afternoon — the actual project went instead into deciding which handful of decisions should hold across all three surfaces at once, which was never the part a faster draft could shorten.

Call this taste: the capacity to discriminate, reliably and under pressure, between the merely adequate and the genuinely right. Taste is not an aesthetic preference in the abstract sense of liking certain colors or fonts — it’s a compressed, internalized model of an audience, a medium, and a moment. Nor is it display, the board or feed held up to prove an eye. It is a decision you commit to, ship, and live with. On Medusa, the system I built in 2019 and the one this portfolio still runs on, that model argued for five core colors where the source material — jellyfish and deep-sea light — could have justified dozens. Everything since is a blend of neighbors: four named gradients, three spectrums, and no color picked by hand again in seven years. Restraint, not a technical ceiling.

What makes taste hard to manufacture is that it is formed by consequence. You develop it by shipping something that didn’t land and understanding why, by watching a decision you were confident about fail in front of real people, by sitting with the gap between what you intended and what you actually made. A Portrait took years for that reason and no other — not because the code was difficult, but because you cannot compose a portrait out of your own material until you have accumulated enough of it to recognize the shape. A model has no equivalent process. It has seen an enormous amount of finished work, but none of the discarded versions, none of the arguments, none of the moments where someone chose the harder option because the easy one was hollow. It learns the surface of good judgment without ever having borne the cost of bad judgment. It can imitate discernment statistically, and where a domain keeps score it can arrive at judgments no one held before — Go has a win condition, and the machines found moves the theory had ruled out. Design doesn’t. Nothing tells you the fifth color was one too many. That verdict comes years later, indirectly, from people who won’t explain themselves.

This has a practical consequence for how creative careers and organizations should now be structured. The premium is moving away from people who can produce and toward people who can select, edit, and direct — toward the person on a team who looks at fifty AI-generated options and immediately knows which three are worth iterating on, and precisely why the other forty-seven are wrong. It’s why, on Glydways’ Metro Visual System, wayfinding and accessibility standards were fixed before a single screen existed — the selection had to happen before there was anything yet to select from. This role used to be implicit, buried inside the act of production itself, because the person doing the producing was, by necessity, also the one making the judgment calls. Now the two can be separated, and increasingly are. A person with strong taste and weak technical skill, augmented by AI, can now out-produce a person with strong technical skill and weak taste. That reversal is new, and it’s not fully priced in yet — not in hiring, not in education, not in how creative teams assign status.

Taste shows up as subtraction at least as often as selection, and subtraction is the part a model will not volunteer. Nearly every word in this site’s navigation carries one blackletter letter as a quiet signature — Journal, which is its own thing, deliberately carries none — and the ornament never reaches an aria-label, a page title, or anything else a search engine or a screen reader depends on. Type Cast tells that story in full. The same rule governs motion: one organic shape on a page of straight lines and hard edges, the cursor, taking the gradient of whatever ground it crosses. Everything else moves only to reveal.

Which leaves the question this argument used to end on rather than answer. If taste cannot be installed by a prompt, can it be handed over at all? Portfolio System is where I’ve been testing that. Twenty years of work are held by five constants — structure, type, color, motion, language — and every one of them began as a judgment call. The system’s only job is to write those judgments down as values someone else can check, so a person, a script, or a model can carry them and they still arrive intact. Medusa supplies the material in three layers: Foundations sets the color and type, Components turns them into parts, and Motion gives those parts movement. Figma, Webflow, and code are joined through Claude and MCP servers, so a decision made in the morning is live by the afternoon, across six templates and two grounds. What that arrangement buys is not taste. It is transmission. The judgment still has to be made once, by someone who has been wrong often enough to make it well; the system only ensures it never has to be made twice.

None of this means talent stops mattering. Deep technical skill still matters enormously at the edges — in solving genuinely novel problems, in knowing when a tool’s output is subtly, dangerously wrong, in doing the last ten percent of refinement that separates competent from exceptional. But talent is no longer the gate. Taste is. And taste is not something a prompt can install. It’s built the slow way: by making things, showing them to people who’ll tell you the truth, being wrong in public, and paying close enough attention to absorb why. AI can compress the time between having an idea and seeing it made real. It cannot compress the years it takes to know, with any reliability, which ideas deserve that treatment in the first place. What it can do — and this is the part I underrated when I first wrote this — is carry a judgment already made, faithfully, at a speed that used to be out of reach. The gap doesn’t close. It moves upstream, to the one decision that still has to be yours.