Research
19 papers from the studio’s experiments, including the ones that did not go our way. Every number traces back to a run we can reproduce, and the claims ledger keeps the caveats.
Timeline
Framework begins. The first version writes thinkers’ ways of reasoning down as files a model has to reason from.
The studio begins. A team of model roles takes on its first competition brief, the eVolo Skyscraper Competition.
The experiments behind these papers. Why model ideas converge, how personal taste is, what juries reward, and how far one person’s framework carries. Framework grows past a thousand people.
Most results here come from model judges and embeddings, which are screens, not evidence that people prefer something. Each paper says which kind of evidence it rests on. The founder is one rater among many to come; names of private participants are never published.
Essays
Longer arguments about where design goes when making is cheap, and our own status reports.
- What breakthroughs have in commonWe coded the documented processes of 74 creators in 11 fields. No single move is universal, but four jobs recur in every field: feed, reframe, import, and vary and select.
- What art survives automationIf making becomes nearly free, what is left for art? History suggests five traits that machines do not erase, and one risk that they create: everything starting to look the same.
- The open loopIdeas revised in a closed loop drift back to the typical. People escape that by taking in real outside material. Living abroad is the clearest natural case, and it suggests a residency for our workflow.
- Where we standWhat four days of tests confirmed, what they did not, and the one test that decides the product.
Frameworks
Writing one person’s way of thinking down, and testing what it does and does not carry.
- Method, not motifFrameworks of invented people carry as well as famous ones, so the framework text works without the model’s memory. Told to carry how someone thinks and nothing from their field, frameworks made the designs a critic preferred.
- Two masters in one roomTwo architects’ frameworks critiquing and building on each other make stronger-looking blends, not ideas in territory neither reaches alone.
- The shape of a frameworkFive layouts of the same evidence barely change how well a model generates as the person. What a layout can change is how hard it pushes, and every shape pushes past where the person really went.
- Sounding like him is not predicting himWe wrote an architect’s framework from his work up to 2008 and asked it to forecast 56 projects he designed after. It made the model sound like him. It did not make the model predict him, and it showed where he actually sits.
- Far from himself, still himselfSeeding ideas with an architect’s own contradictions takes them far from his typical work while they still read as his. A random seed does just as well.
- Can a critic think like an architect?A critic holding Campo Baeza’s framework did no better than a plain critic at telling his competition winners from his losing entries.
- A way of thinking travelsWritten down as a framework, one person’s way of thinking carries into fields they never worked in. A judge picked the right framework out of four 88% of the time. Most of that was their world, not their method.
- A loop does not make a masterworkWe ran an architect’s framework through nine rounds of critique to see if it would reach his mature work. It improved, then plateaued, and every judge could still tell it from the real thing.
Generation
Why model ideas converge, and what keeps them apart.
- Closed loops convergeRevision pulls ideas toward the model’s typical answer: half the way in one round on real competition entries. Only fresh outside material in every round keeps ideas apart.
- How far from the obviousA single scale for weirdness, and why the right distance belongs to the judge, not the brief.
- One idea in thirty outfitsModels collapse toward one typical answer. A distinct seed per idea is what pulls them apart.
Taste and judging
How personal taste is, what juries reward, and how far a model judge can be trusted.
- What juries rewardAcross three kinds of contest, juries and crowds do not simply reward the unusual. They punish unusual presentation, favour typical jokes and, in short fiction, slightly favour atypical stories.
- A model judge is one voterOn New Yorker captions, a strong model agrees with the crowd about as often as one random voter does. Fine for screening, not for choosing among good ideas.
- A lean, not a sweet spotIn 45,136 people rating the same jokes, each leans slightly toward or away from the unusual, and the lean repeats. We found no sign of a personal sweet spot inside that range.
- Taste is mostly personalIn 1,759 people, the stable personal part of taste is about twice the shared part.