Crazy 800

Slop isn't volume — it's volume without selection pressure. A day in the field — five hundred generated thumbnails in the morning, four hundred reviewing constraint personas in the afternoon — shows that volume is the new quality principle in design, as soon as you review it as patiently as you generate it: the double coin of generation and verification.

SeriesParadigm ShiftPart 4 of 5
June 18, 2026 18 min
AI agentsPracticeAI & society
This post was originally written in German. You are reading a translation.

The industry has settled on a swear word: Slop — the old word for pig feed, repurposed for mass-generated junk. The term lands, and it rings true. But the same designers who despise Slop keep hand-polishing three drafts a week and calling it diligence. Underneath sits a certainty hardly anyone says out loud because it seems so self-evident: that volume is the opposite of quality.

This certainty is wrong. Volume is the new quality mechanism — under one condition that decides everything. If you merely generate en masse, you really do produce Slop. If you generate en masse and review en masse, you reach a quality manual labor could never touch.

Slop is not a property of volume — Slop is volume without selection pressure.

Watercolor of a Victorian naturalist standing before a large glass cabinet full of lined-up, labeled cards, examining individual specimens with a pen and sample tray; next to it an empty birdcage, in the background bookshelves and a garden.
The breeder in front of the cabinet: hundreds of specimens, one hand that reviews — volume only becomes quality when someone curates it.

The Thumbnail Hour

Five hundred drafts before lunch

One morning from my own practice shows what this looks like. My learning portal had just passed the hundred-published-pages mark, and each page needed a thumbnail: a visual entrance door that condenses the content into a miniature. The assignment for the swarm fit into one sentence — read every page, develop five to six design ideas, build clickable miniatures from them. Behind each tile was a sub-agent that had actually read the content — the text, the diagrams, the tone — and derived its design idea from that instead of rolling the dice for a generic icon.

Then, for an hour, something happened that appears in no textbook of my profession. A demo selection page filled up, a hundred pages times five variants. I clicked on favorites; the selection array moved into the clipboard and back to the agents (the most solemn form a judgment of taste still takes today). For everything that didn't convince me, the instruction was: another five, another ten. After about an hour, five hundred to eight hundred miniatures had spanned out, and from them I condensed the inventory that is online today.

What stays with me from this morning is less the volume than the silence. In the past, productivity had a sound — scratching pens in a workshop, thunderstorms of typing, the babble of voices in a review. This hour had no sound of its own. The most productive design hour of my week sounds like nothing — just the fan.

There's a reason I still don't trust this silence blindly. It wears Bengali characters and will get its say later in this text.

Eight Sketches in Eight Minutes

An appreciation of the tired hand

Before this morning starts to seem self-evident, it's worth a look into the room where I learned my craft. You know the one: workshop atmosphere, a ticking timer, felt-tip pens, a sheet of paper folded three times, eight fields. Crazy 8 — eight sketches in eight minutes. The method was invented so designers wouldn't marry their first idea. It forces the hand to keep drawing when the head has long since been satisfied (the most valuable sketch was rarely the first — usually it was the sixth, just past the point where everyone groaned).

This discipline deserves appreciation, because under its conditions it was the smartest answer available. For thirty years, good design was convergence craftsmanship: one variant at a time, one pair of reviewers at a time, and every additional round cost someone's afternoon. The methods of my guild — paper prototypes, hallway tests, the review before the release — were, on closer inspection, all austerity programs for expensive variants. As a rule of thumb from my own practice: done the classic way, a single thumbnail cost about an hour of manual labor — reading the article, collecting ideas, designing, implementing. For a hundred pages, that's a hundred hours — two and a half work weeks for entrance doors. So we agreed on eight sketches and called it a creativity technique. That sounded better.

Eight was never the number of creativity — eight was the number of exhaustion.

The Double Price Drop

Two price tags fall at the same time

This exhaustion rested on an economic foundation, and that foundation has crumbled. Everyone now tells half the story: generating has become almost free; ten drafts cost barely more than one, fifty barely more than ten. If you only see this half, you land squarely in Slop. The second half gets told less often, and it's the more consequential one: in the same stroke, reviewing has become almost free. The same agents that design can review — clicking through, comparing, checking against rules, justifying their judgments (a reviewer who doesn't get tired and isn't offended when you discard their draft). Two price tags fell at once: the one on the variant and the one on the eye that reviews it.

And with that, the business foundation of convergence craftsmanship tips over. It was never so much a method of conviction as a method of scarcity. The scarcity the entire craft was calibrated to vanished overnight — on both sides of the desk. What remains is the question that governs the rest of this day: if generating and reviewing both drop toward zero — who or what holds the quality?

The Aviary

Darwin at the desk, part one: variation

What takes the place of scarcity has a historical predecessor, and it coos. Charles Darwin kept pigeons1, crossbred them, compared the litters — the dovecote was his laboratory, selective breeding his method of insight.2 That very method has moved into my professional life, with one difference I want to state plainly for once: nothing that breathes gets bred here. What gets bred are layouts, forms, entrance doors to learning pages.

The breeding goal first. None of this begins with a swarm; it begins with a document. Requirements, style rules, target audiences, taboos — I define what a draft must fulfill before any agent starts; the breeding goal lives in intent and tests instead of pixels. Only then comes the staging: ten drafts, twenty, thirty, fifty — Crazy 8, expanded by two zeros. The swarm knows three tones here. Mute: “Make ten variants.” Discursive: “Make ten and discuss pros and cons for each variant.” Primed: “Give me a recommendation beforehand.” The third tone is the most comfortable and, in critical decisions, the most dangerous, because a recommendation upfront narrows the exploration before it has begun. The swarm doesn't have to guess what I want: it's told before it moves its first feather.

Mutation and recombination. Three survive the first relay — let's say: variant 2, 7, and 13. What follows is the actual process, and it consists of two parallel assignments: sub-variants of the survivors, meaning deep drilling in one direction, and recombinations between them, like the grid of 7 with the color logic of 13. Then select again, span out again. The loop breathes: the solution space expands, contracts, expands anew along the survivors, contracts again. So that the density converges toward quality instead of toward the baroque, two judge passes with opposing objective functions run over the favorites at the end: the first checks if everything is watertight; the second deletes everything that is over-engineered — both poles are in the same manifesto.3 One more word on modality: the eye scans twenty layouts next to each other in seconds, it reads twenty code variants one after the other — which is why breeding unfolds its full effect in the visual. With every breath, the solution space becomes denser around the good and thinner around the arbitrary.

The feed costs. Before the next relay starts, my eye falls on an inconspicuous number: the morning's token budget. An aviary this size eats, and it eats money. So the variants draw on a library of reusable components instead of building every subroutine from scratch — the variance lives in the arrangement; the foundation stays paid for and tested. This budget management is a survival condition: without it, the method dies at rollout, killed by the token bill. It has a name: Generation Economy. If you can't feed the aviary, you aren't breeding lines — you're collecting beginnings.

Variation is only half of breeding — without an environment that sifts out, the aviary is merely full.

Four Hundred Eyes

Darwin at the desk, part two: the environment sifts

In the afternoon, the swarm reverses direction. In the morning, it designed; now it reviews. On the screen, instead of miniatures, there are now logs, and the same economy that allows fifty drafts at once allows four hundred reviewers at once — the four-eyes principle becomes a four-hundred-eyes principle.

Environments instead of reviewers. The first reflex would be to start the same reviewer four hundred times. That finds the same thing four hundred times. The variance comes from constraint sets (rulebooks that each lock in one condition of use): you may only use the keyboard; you operate everything by voice; you're on a smartphone, in Dark Mode, with these personality traits and this patience. Each set is less a reviewer than an environment a draft has to survive in — an artificial habitat that sharpens exactly one survival condition. The load-bearing axes here are functional: language, cognitive load, tech familiarity, age — they're dictated to the model as explicit specifications and override what it would assume on its own. So a hundred, two hundred, four hundred agents go over the software like a shotgun blast, deliberately spread wide because collective coverage is what counts; a consolidation agent then condenses the findings into patterns. And so no caricatures end up doing the clicking, role and reasoning are separate layers: beneath every persona role runs an internal monologue that argues every action against the rulebook and role plausibility before the persona executes it. Without this layer, the agent is acting in a play; with it, it simulates. The workshop poster has become an environment with mandatory logging.

The log as artifact. The actual test artifact of these swarms is the reasoning chain of the monologue, the reasoning trace — the click path is only its shadow. Friction patterns can be bundled from reasoning chains: where a persona hesitates, why it aborts, which phrasing it reads twice. This is the place for the primary evidence — an excerpt from a real run, anonymized, with date:

Persona run · March 11, 2026 · Constraint set: exclusively keyboard, no mouse

Task: open the term “Adversarial Examples” in the glossary.

“Tab — immediately ‘Skip to content’, good. But do I want the content or to search directly? Tabbing through the A–Z list costs 26 stops, I won't do that. So the search field, one tab further. Except the field carries no label — just the instruction ‘Enter text to find pages. Use arrow keys to navigate.’ Only by reading it do I understand that I switch to the hits with the arrow keys after entering. I type ‘Adversarial’, arrow down, Enter. Goal achieved — but I had to piece the path together myself.”

Friction: The search field has no label of its own; how to get from the input to the hit list is only in the placeholder instruction — whoever skims over it is stuck.

What you're reading there is the difference between a metric and a witness: the run documents its own hesitation, justifies its decision, and names the point where the interface lost it. And that reviewing is structurally cheaper than generating isn't just a claim from my own desk: AI safety research builds entire oversight processes on this asymmetry. One of them, the debate procedure, I described in my book: two AI systems compete against each other, each trying to convince a human judge — the core assumption being that it's easier to recognize a lie than to find the truth yourself.4

The click says what happened — the reasoning chain says why it will happen again.

The stack of disputes. No human reads four hundred review reports, and nobody has to. The consolidation sorts the findings into two classes: things the brigade agrees on — those move into the pipeline in bulk — and things it argues about. My reading time belongs to the argument. Where four hundred environments come to the same finding, my nod is enough; where they diverge, there's information. Except this sounds like less work than it is: whoever decides which of two well-reasoned contradictions gets to change the product needs more judgment than before, and needs it more often. The bottleneck hasn't disappeared — it has migrated to the judgment over dissent.

That leaves the objection that comes up first in every discussion about simulated users: personas from a language model are supposedly shaped by its biases and therefore unrepresentative. The representation debate asks a real societal question — it just answers one that UX testing never asked.5 What's being tested is whether a form survives under the condition “small screen, low tech familiarity, Easy Language”; the condition is dictated — set as a constraint for the model, not read out of it. The hardest demonstration of this logic, though, waits somewhere else: in languages I can't decipher three words of. The swarm reviews even where I can't so much as read what it's reviewing.

Bengali, Latin

The story of a gap

My portal — basiswissen-ki.de, a freely accessible AI learning offering — appears in 28 languages — as of this article's publication, because it keeps growing whenever the backlog is idle and a few tokens are left over. Every page exists twice: standard language and Easy Language, even where the target language has no Easy Language tradition of its own. The total volume today is around 35 million words — about 275 times as much as is in the book (350 pages, excluding the roughly 650 sources) that I sat over during those same years, by hand, night after night. A stable five percent or so of visitors have Easy Language switched on, says my own portal's telemetry. The whole thing was set up as a roadmap: each language prioritized by reach, complexity, writing direction, and the training-data quality of the translation models; at the turn of the year 2025/26, when the translation wave ran, the portal stood at a good 250 book pages per language. All of this is only possible because the same breeding logic is at work here as with the thumbnails: agents generate, brigades sift — AIs review AIs, against guidelines I set up in two variants per language, with idiom rules, a glossary, consistent naming of UI elements.

And right there sits the story I've owed you since this morning. A single paragraph was missing from the guidelines: “Use this character set for this language.” Nobody had written it, so nobody reviewed it. Bengali and Russian flipped into the Latin alphabet in places — transliteration instead of script, right in the middle of the body text. The swarm didn't catch it, and neither did I, and it would have been easy to catch by eye; exactly that is what makes the matter so instructively embarrassing. I only discovered it when everything seemed finished.6 The consequence was a massive refactoring in April 2026: all language guidelines revised, newer models sent over all 28 editions once more, this time with a character-set clause. This second pass had a price that no design discussion ever mentions, and it varies wildly: a Germanic language — Dutch, Swedish — runs cheapest, the model commands it in its sleep; Romance and Slavic sit above that; an Indic one like Bengali costs three and a half times as much, because there the model wrestles for every syllable. In numbers: around 42 million tokens for the cheapest, a good 150 million for the most expensive; summed across all twenty-eight, a pass comes to almost two billion tokens — at list prices, about $2,700.7 And the most expensive languages are, of all things, the ones I don't read: the missing line about the character set was, counted in tokens, the most expensive paragraph I had never written.

I will leave the lesson unvarnished: the swarm reviews the environments you have built for it — an environment that is missing does not sift. Scaling also scales mistakes; it just does it more quietly. A swarm of four hundred eyes is as farsighted as the environments you build for it — and as blind as the ones you forget.

The Embossing

What actually constituted work this day

By early evening, the day lies spread out like a bank statement: five hundred generated miniatures in the morning, four hundred reviewing environments in the afternoon, in between the memory of two alphabets nobody ordered. It takes the whole sum to see what connects the two movements.

The word hardly needs coloring in; the day itself delivered the object lesson. What stays open is the question that occupies me as I shut down the computer: I was responsible for quality today as rarely before — but I drew nothing, pushed nothing, designed nothing. What was my work on this day, when not a single draft came from my hand?

From Draftsperson to Breeder

The method gets its name

The answer appears in no job profile, but it has a name, and it's time to set it down: Generative Design Exploration — not a new procedure, but the name for what this day was from start to finish. My work consisted of four things: the intent that preceded every relay; the constraints that turned review agents into environments; the gap in those environments, for which I am liable; and the judgment on the disputes the swarm submitted to me for decision. Darwin would have had a sober word for it: selective breeding — the selection happens in the dovecote, the breeding goal is made at the breeder's desk. The UX professional breeds: they design requirement spaces, survival conditions, testing climates, and leave the drawing to the swarm.

If that sounds like loss, like a job that has coagulated into administration: the opposite is true. I wrote in my book about an artist who went through 900 prompt attempts for a single image: whoever wrestles for nuances that long isn't playing the lottery — they are creating. The threshold of originality lay in the relentlessness of the vision.8

The relentlessness has remained. Only its subject has moved — from the single stroke to the conditions that judge thousands of strokes.

The draftsperson designed solutions — the breeder designs the conditions under which solutions survive.

Border Control

The same swarm, a different front

One question remains, and it leads out of the studio. Interfaces are increasingly rendered in the moment of their use, tailored to the individual. In such a world, a manipulated variant is no longer an A/B artifact that could be found in aggregates: it only exists for its victim, for a single moment, and deletes itself with the page change. Adversarial Hyperpersonalization leaves an impact, but no evidence.

“The Foreign Service”, the kickoff of this series, left a question open exactly here: how do you recognize systems that work against their users when every manipulation only has a single witness? The operational answer has been on my screen since this afternoon. A brigade of hundreds of constraint personas each sees its own freshly rendered interface; the consolidation overlays the renderings and makes drift visible — systematic deviations that pile up where a provider treats tired users differently than wide-awake ones. No single human can provide this proof. A swarm can — as the only actor with enough eyes in enough places.

And so this methodology hinges on a question larger than any design team. That swarms do the reviewing will, in a few years, be as self-evident as the spam filter is today. Who directs them, who owns the breeding goal and the judgment — that's the real distribution question, and a market will be decided by it too: whoever gets to define selection pressure holds the monopoly on quality that, just yesterday, lived in agencies and design departments. The same brigade that sifts layouts today is, tomorrow, the only entity that even gets to see a manipulation in the first place.

End of the Workday

Who was missing today

Evening is here, the fan spins down, and for a few minutes the room sounds the way the end of the workday has always sounded. On the second monitor are still the miniatures from the morning, five hundred attempts, a curated inventory. It was a good day: fast, thorough, in a quality that I alone wouldn't have reached even in a month. The profession has switched sides in the process — bred instead of drawn, and the judgment was my only manual labor.

Only when I stand up do I notice who was missing from this scene all day. In every “studio”9 where I learned, a second person sat somewhere: the intern, the working student, the new hire — someone who hung up the sketches, logged the reviews, and along the way, simply by being there, got the map of the profession into their head. Today there was nobody. The swarm needs no errand boy — and precisely because of that, nobody stood next to me who could have learned by watching.

Nobody learned anything today — and that is the only line in the log that worries me.

What becomes of a profession that no longer trains its beginners anywhere, and which software company survives a world without this foundation — “The Book of Hours” will explore that.

Until then, what this day has shown holds true: volume became quality on the day we started reviewing it the way we generate it — en masse, patiently, and with a breeding goal written by a human.

  1. In 1859, a reviewer for his publisher advised Darwin to write a book just about pigeons instead of one about the origin of species — “everyone is interested in pigeons”, and it would end up “on every table in the kingdom”. Darwin didn't do it.
  2. The mechanics are significantly older than their current appearance: John Holland described variation, selection, and recombination in 1975 in Adaptation in Natural and Artificial Systems as a formal search procedure — genetic algorithms. So the algorithm is not new. What is new is the phenotype: finished software.
  3. This footnote originally had four sentences, an excursion on mutation testing and a second example. Then I treated it like my layouts. This is what survived.
  4. Basiswissen Künstliche Intelligenz, Chapter 10: Alignment and Safety.
  5. The objection honors the discipline with a standard it never applied to itself. In classic usability practice, “representative” often enough meant: the eight test subjects who had time on a Thursday afternoon and came down to the lobby for a voucher. We took that seriously, and it worked. Only the simulation gets held to representativeness in earnest — a compliment disguised as an objection.
  6. It seemed.
  7. These dollar amounts are list prices — what such a run would have cost via the API. I never paid them: I run the pipeline via a flat-rate subscription (Claude Max, 20×) into which the same tokens disappear. The effort is real, my bill was not.
  8. Basiswissen Künstliche Intelligenz, Chapter 7: Creative Machines.
  9. “Studio” is putting it kindly — I learned in software sweatshops. The second chair was there all the same.

Plain Language

This preview is in Plain Language. The whole portal is also available in Plain Language. You can change the language at the top. Or use the button below. Our Plain Language texts are AI-assisted and editorially reviewed. They follow the core principles of easy-to-read language, but they are not certified Easy Read: a review panel of people with learning difficulties does not test them.

Preview: Make a Lot and Check a Lot

What it's about

Many people complain about AI slop. Slop means: bad content made in huge amounts. I say: Large amounts can bring good quality. For that, you must check the large amounts just as carefully.

A day from my work

In the morning, the AI makes 500 small preview pictures. I pick the best ones. In the afternoon, 400 AI test users check my portal. Each test user has their own rules. For example, one only uses the keyboard.

A mistake with the writing system

Once, a rule for the writing system was missing. Then parts of the Bengali text appeared in Latin letters. Bengali is a language from South Asia. Nobody noticed. The checkers only check what you tell them to check.

What changes

The designer no longer draws everything by hand. The designer sets the rules. Then the AI creates and checks. One worry remains. Beginners no longer learn anything this way.