1kB site is new status symbol. LLM's can't make it.

Taste got invisible

We are in the age of slop. Any site can now be generated for free, in seconds, by a machine trained on all sites. The page no longer proves anything about its maker.

Craftsmanship has been obfuscated, not just faked, but hidden. Skilled people work through the model, and their choices vanish under its text. In a lemons market, invisible quality is priced as zero; good sellers exit. Slop hides the honest article, and hides it more completely the more faithfully you use the tools.

We sold skills to LLM's. And lost taste.

It used to be genuine. Developers bragged about scale and beauty, and the brag couldn't be faked; the artifact proved it. Skill was the visible half of a pair.

Taste rode along with skill. Skill and taste were trained together under apprenticeship, so high skill implied high taste, and nobody had to look for taste separately.

Then the visible half went free. Skill is now something you ask a machine for; the artifact no longer demonstrates it. And taste, the half nobody ever trained alone, is the only signal left. It was always assumed, always covered by skill. So we face a world where easily-seen skill is free and taste was never grown, and the decoupling is complete: skill is cheap, taste is scarce, and nobody trained to spot it.

So the question is narrow: where does taste still surface?

It surfaces where the generator cannot follow. At the constraint.

Constraints surface taste.

Bloat is the model's home turf. Models sample what is typical; verbosity is their drift. They cannot default-produce the region where nothing is typical, where every byte carries meaning.

Compression makes the page brief; atypicality makes it unlikely under the model's distribution. Both come from the same act: removing redundancy.

A hard budget forces decisions. A 3MB site can absorb two hundred generations with no trace of a choice. A kilobyte cannot. Every byte is a decision that survived.

The site is the result of choices, not generation.

This is the handicap principle: a signal is credible when it costs the faker more than the honest. The currency is attention, the one thing still scarce.

Every tool that makes production cheap moves legible skill to the tool's edge. The camera moved painting toward abstraction. The LLM moves skill to the constraint. Status lies at the boundaries of possibility.

You can't learn to reduce until you've overflowed. Bloat is apprenticeship, the beginner's necessary stage, the fat you need before you can learn to cut. The sequence is the whole story: overflow, then reduce, then squeeze, then stop when nothing can die.

LLM's iterate, Humans steer.

Minimizing is a hard problem, by definition, because it is branching. Every feature is a fork: pay 30 bytes for it, or pay 70 bytes for a version that unlocks a second feature at only 10. Take the cheap fork, and the second feature now costs 300. The branches compound with every byte, and the arithmetic is global; the price of a phrase depends on the whole arrangement. But the space is not infinite. For a given outcome, the paths are finite, and one of them is best: the one with maximum entropy, every byte load-bearing, nothing removable without the meaning dying with it.

That branching is what makes the handicap credible. Effort alone proves nothing; anyone can grind, and grinding can be claimed. But this effort cannot be faked, because it is spent where the machine cannot follow: valuing the forks. The cost is attention spent on valuation, the one currency that cannot be generated and cannot be outsourced, and it costs the tasteful less per unit of quality than it costs anyone else. A person with taste can value a fork early and reject it cheaply. A person without it pays for every branch.

The whole trick only works if taste makes the work cheaper, not harder. Someone with taste sees a fork and knows the answer from experience. Someone without it has to try every branch. If cutting were just as hard for everyone, a dense page would prove nothing about anyone. Taste is the thing you build once and use forever.

The site is the residue of that attention: hours per byte, spent where nothing else could spend them. That is the handicap, and it is the whole game.

Two pages at the same byte count are not the same object. Small-by-empty: nothing there, trivial. Small-by-dense: every byte load-bearing. Test it: delete a byte. If the page survives, you had slack. If the meaning dies, it was reduced.

The deletion test is the falsification instrument of the whole aesthetic.

AI generated human product.

A better model can approach the craft. It cannot arrive.

A model learns from what's already there. The interesting pages are the ones nobody has judged yet, and where nothing has been judged, there's nothing to copy. The model can only do the usual, or take a human's judgment as input. And that judgment is exactly what the site is supposed to show.

The hard part is not generation: it's valuation.

The cost of a feature is global; its worth is not computable from the bytes alone. It depends on the reader and the author. A model could learn to judge, but only from human judgments, and those judgments are the scarce thing. Valuation is an input, not an output.

The machine can't tell its good guesses from its bad ones. Making is cheap; judging is the expensive part, and it's the human part. So faking the signal takes the very thing the signal proves. If you can fake it, you already have it.

This is not a hard optimization. It is an incomplete specification. The master doesn't optimize against an objective; the master constitutes it.

Teaching machine taste, won't help.

Taste-training widens the gap. Models trained on many readers move toward the average, and status lives at the extremes. The better the model gets at taste, the more useless its taste is where craft is maximal.

A good prompt can steer a model anywhere, but only if the person writing it already knows where to go. The prompt has to be invented first, and inventing it costs the same taste the site is supposed to show.

The division of labor: the model iterates; the human selects. The output is cheap to discard; the selection is binding. Bearing the consequence is the one thing that can't be delegated: the site is the residue of selections.

LLMs obfuscate skill and taste. Constraints surface it.

Thus Sub-1kB site is the new status symbol.