Taste

Lately I keep setting two facts side by side. China's industrialization pushed the price of ordinary industrial goods down to roughly what they cost to make. AI is now doing the same to ordinary intelligence. The mechanism is the same both times: once a capability can be supplied in volume, what it produces becomes an option anyone can have, and its price falls toward its cost.

This essay makes one claim: when options get cheap, the value is in the choosing. Once producing an option costs almost nothing, one more option is not what anyone pays for. What they pay for is the decision, among many options that are all good enough, about which one stays, which ones go, and who answers for it. I want to call that ability taste, and to stretch the word further than it usually goes.

What gets cheap is the option, not the ability

"Devalue" has to be used narrowly, or the whole claim turns into a complaint. Making things has not lost its worth, and neither has thinking. What loses its price is the interchangeable unit: the millionth identical cable, the millionth paragraph that is good enough. Both are still useful, but useful is not the same as valuable. A thing can be useful and still too cheap for anyone to pay a premium, because the next supplier can hand you another at nearly the same cost.

That is what I mean by an option: a unit of supply the next unit can replace. Once options multiply, no single one of them is scarce; what is scarce is the decision made between them. So "industrial goods got cheap" does not mean the factories got worse. They can be better than ever — higher yield, steadier delivery, tighter tolerances — and still the premium does not attach to the product. It attaches to the decision about what the thing should be.

Industrialization moved the bottleneck from making to specifying

In the 1990s Stan Shih drew the smile curve. It is a sketch, not a law: value is high at the two ends and low in the middle, with specification and design at one end, brand and distribution at the other, and manufacturing in the trough. After China joined the WTO in 2001 and entered the global division of labor, that trough became the world price. Within a generation one experience became ordinary: the goods were all still there, they just would not stay expensive. Clothing, chargers, plastic parts, the invisible pieces inside a finished device — whatever anyone could make was priced at cost.

The mechanism is not unique to China. British textiles, and later Japan, each turned some kind of production from a scarcity into a background condition. China simply did it at a scale no one alive today can ignore. The principle is old: when production stops being the bottleneck, the bottleneck moves to what gets produced. Put another way, once many things can be made, what matters is choosing which one to make.

The money does not leave. It stops paying for volume and starts paying for specification, and a specification is a set of choices written down. Give the same molding shop a vague drawing, or one that has already settled the chamfer, the feel in the hand, and the tolerance, and it is in two different businesses. In the first it sells capacity; in the second it carries out someone else's choices. Capacity is priced by the next shop that can do the same work. A choice is priced by whether other people are willing to follow it. Hard and scarce are also different things: however hard a task is, once enough people can do it, the price leaves it.

Intelligence is turning into options

AI is running the same accounting on intelligence, only faster. What a model can now deliver reliably — a usable paragraph, a function that runs, a plan that is good enough, a debugging session that actually reads the error — is turning into an industrial good. The price of a unit of capability keeps falling, past the point where it counts as a variable in any decision. The API prices in the last essay already made me stop once. Looking again, they read even less like a metaphor.

The plainest sign of the change: a plan that used to take an afternoon now arrives in seconds, ten at a time, and all ten are good enough. Plans have gone from scarce goods to options.

So people make the same mistake they made about factories: if making the thing got easier, whoever makes it must be worth more. Not so. The part that got easier is the part whose price follows the ease down. Someone who passes the model's first result straight through is selling one more option, and the competition on that line is everyone with the same model, plus the model itself. There is no premium there.

The premium stays in the layer the model cannot supply, and should not be asked to: among many adequate results, deciding which one should exist.

Why choosing does not get cheap too

The obvious follow-up: if generating got cheap, why wouldn't choosing get cheap along with it? My answer has three parts.

Options add; a choice subtracts. Generating one more version gives you one more option, not a decision. The more options there are, the more attention it takes to read them, compare them, and throw most of them away. The cost of generating is heading to zero; the cost of judging is not following it down. Abundance does not make the choice for you. It makes the choice heavier.

A choice needs a standard, and a standard comes from a purpose. To decide which option should exist, you first have to know whom it is for, what it has to solve, and what you are willing to give up for it. A model can rank options against a standard you hand it. It cannot supply the standard you are prepared to answer for, because the purpose and the consequences sit on your side. Picking without a standard is a lottery.

A choice worth paying for is not the average. A choice anyone gets by taking the default is just one more option. A choice is worth something because of the gap between it and what is already at hand: it drops what others would keep, keeps what others would drop, and can say why. What carries a price is the part that rises above the default.

Taste is choosing under abundance

The word drifts easily toward decoration: color, type, a prettier sentence. Those are the surface of taste, and the surface happens to be the kind of option a model generates best. If taste were only surface, it would cheapen along with intelligence, and this essay would have nothing to say.

The definition I want is plainer. Taste is the ability to choose under abundance. Abundance is the precondition. When options are scarce, choosing is not taste; it is making do. Choosing gets expensive only when the options are many and most of them are good enough. In practice, this choosing takes at least five forms.

• Omission. Deciding what does not appear. Once generation is nearly free, excess is the default: one more explanation, one more setting, one more feature nobody uses, each with a reason to stay. Taste is the standard that says no, and says it early, before those things turn into maintenance.
• Defaults. Making a choice once so the people who use the thing do not have to make it again. A good default saves a choice on every later use; a bad one is repaid on every later use. Either way, a default is a choice that gets reused, which makes it worth more than any single output.
• Coherence. Getting many small choices to point at one judgement. Each may be fine on its own; together they either undercut or reinforce one another. Coherence cannot be stumbled into by generating more candidates. It needs a single standard that remembers what it has already promised.
• Done. Choosing where to stop. A model stops where the result looks plausible, because its stopping rule is probability, not a standard. The standard has to come from outside: this version can be signed, the last one could not. In volume the difference is almost invisible. In price it is everything.
• A name. Someone has to own the choice. A preference nobody claims is noise, and nobody pays a premium for noise. A name is not a personal brand. It means a mistake has an address and a success can be learned from. A default with no author cannot be relied on, because tomorrow it may be something else.

All five touch aesthetics, and none of them rests on it. The shape of an API, the questions an installer no longer asks, the three paragraphs cut from an essay — all of that is taste. What these share is not that they look good. It is that someone, after the options were already adequate, refused to keep all of them.

Money paid for choices

Linux may be the most extreme case of cheap options. Nearly every component is free, and nearly every component has several alternatives. For decades, assembling them into a machine that feels right has been left to the user. There has never been a shortage of options. What has been short is someone to choose for you.

In August 2026 David Heinemeier Hansson (DHH) set up the Omacom Foundation for Omarchy, his Linux distribution. By 22 September, when Alibaba Cloud joined, the foundation's own page listed pledges and donations of about $21.7 million. That total includes multi-year corporate commitments and model credits; it is not an equity round sitting in a bank account. The number is not what interests me, though. What interests me is what the money bought.

Omarchy is Arch, with Hyprland and Quickshell for the desktop. Its own word for the approach is omakase — Japanese for "I'll leave it to you," the chef ordering for the table. Its reasoning is plain: most people do not know what they want at the start, and they are better off accepting a coherent set of defaults from someone they trust than suffering through the menu. Theme, terminal, editor, and keybindings form one system, not a bag of software.

The criticism is just as plain: this is not a distribution in the traditional sense, only Arch plus one person's config files. I think that describes the product correctly and then stops one step short. The kernel is not new, and he did not write the window manager; every option was already there, and free. What is new is the layer of choices: which defaults deserve someone else's entire workday, and which questions should be gone before the installer finishes. Here the config is not an accessory. The config is the product.

The money went to that layer, and not because Linux suddenly lacked a kernel. What was missing is someone willing to put their name to a way of using the machine, and to maintain it until other people can depend on it. In its 3 September note the foundation described itself as the exclusive sponsor of Hyprland and a premier sponsor of Quickshell and mise, and it also funds resident visual design, kernel work, and the infrastructure a distribution needs. What got funded is how this particular set of choices holds together, not yet another general-purpose runtime.

Paying for a set of choices someone has already made is nothing new. Rails, by the same person, was also a layer of opinion laid over a language. What is new is that the work beneath the choices — getting an environment to where it is usable, debugging it until it works, making the surface consistent — has just become cheap. Once that work is cheap, the choices are nearly the whole product. A pile of donations is not proof, of course; how much of it is fame, I come back to below.

Choosing does not spend efficiency

When people hear taste, they hear slowness: reworking, second-guessing, refusing the first adequate result. If efficiency means output divided by time, all of that does slow you down. But that arithmetic only holds while output is expensive. Once output is itself a cheap option, it is measuring something that is losing its price.

Swap output for value in the numerator and the arithmetic flips. In the same hour, output nobody chose is priced near the cost of generating it; output that went through a choice is priced by whether other people can rely on that choice. Choosing does not subtract anything from efficiency. It replaces the numerator, trading a cheap good for something that still has a price. Without that swap, going faster just means arriving sooner at a result the market will price at cost. That is not efficiency going up. It is waste, arriving faster.

This shows up in three places.

Choose upstream, and downstream never has to choose again. Omakase is itself a kind of efficiency: the guest does not have to study forty dishes, and part of the bill pays for not choosing tonight. Omarchy's installer makes the same offer. Linux used to be slow because every step asked you a question; those questions have now been answered, and the answers fit together. One person's choices, made once, become everyone else's speed. Calling taste private fussiness sees the one exercise and misses how many times it gets reused.

The waste moved. When production is scarce, waste is what you failed to make. When options are abundant and generation is nearly free, waste is shipping something that should not exist, or generating ten versions with no standard for which one stays. The ten versions are not the expensive part; reading them and deciding is. Skip that step and throughput rises, but the price the market puts on throughput is sliding toward the cost of generation, and that cost is sliding toward zero. Throwing nine away looks slow. Shipping the average of the ten is what actually spends your time on something whose price is vanishing.

Throughput without a specification saves the buyer money. A factory that pushes undifferentiated goods out at astonishing speed is efficient and thinly profitable. The margin is thin not because it is slow, but because anything it makes can be swapped for the next unit. China's industrialization did not punish careful factories. It punished output with no specification: the faster you ran, the more money you saved whoever bought your capacity. AI carries that sentence over to cognitive work. Multiply throughput by ten without the standard keeping up, and all you have done is multiply a commodity by ten. Taste is that standard. It does not compete with efficiency for a place. It decides whether the same efficiency lands on a commodity or on something that still has a price.

Where this could be wrong

There are a few places where this argument could fail. I want them written down, or the claim is only a nice-sounding sentence.

Models choose too. The most direct objection: models do not only generate, they also pick, and they pick better every year. Won't choosing get cheap as well? I agree that it will, and it already is: whatever picking a model can do for you will become an option, just as generation did. But that is the mechanism above at work, not a counterexample to it. Choosing supplied in volume becomes the new default, and the value moves to the layer above it: setting the model's standard, deciding which of its suggestions to reject, putting your name on the result. The waterline keeps rising, but there is always a waterline.

Fame. Omarchy's money may mostly be DHH's audience rather than any law about taste. If someone without an audience ships defaults of the same quality and nobody pays, my example was just celebrity. My answer for now: fame explains speed and size, not the category. An audience sets how large this particular sum is, but the category "choices can be sold" existed long before it. It already holds Apple, Rails, and restaurants with no menu. Omarchy is simply the case in that category that became easiest to see once intelligence got cheap.

Bad taste. A wrong default is slower than flexibility, because everyone has to route around it. That is true. Taste is a bet that can lose, not a guarantee. I am not saying every choice wins. I am saying that once the generic unit is free, whoever refuses to bet is holding nothing that can carry a price. A wrong choice at least has a shape: people can leave it, and they can fix it. An average has no shape. It has only a price, and the price keeps falling.

Status. Sometimes people pay for taste as a signal: I can afford this judgement. That is the weakest form, and the first to die in a price war. The form that survives is not a signal but work taken off someone else's plate. You adopt the choices, you get your time back, and the result is more coherent than anything you would have pieced together under deadline. That is productivity, not luxury.

Finally, parts of making and thinking have not cheapened at all. Process, materials, and yield still pay. So do new problems, responsibility, and the context that exists only because you were in the room. If "the choosing is what is valuable" is heard as "craft and thought no longer matter," I have said it badly. The expensive part of craft and thought was always the choosing: knowing how far to go, and where to stop. What got cheap is only the part the next model can redo.

The step I am keeping

I am not going to use these models less. They have made options nearly free, and free options should be used to the full: generate more, compare more, try more. But using them does not mean handing the output straight over. Before it goes out, it has to pass through one step that is still expensive: the choice.

The standard I hold myself to is narrow. Before something carries my name, I have to be able to say what it left out, and why its defaults are what they are. If I cannot, it is still just an option, however finished it looks. That step is slower than generating, and the slowness is exactly where the price still lives. The cheaper the options, the more the choice is worth.