Where we stand

What holds

• Models collapse toward one typical answer, and a distinct seed per idea prevents it. Measured on the raw idea text, plain ideas had a mean pairwise similarity of 0.77 to 0.81; ideas that each started from their own concrete seed, 0.42 to 0.58.
• Closed loops drift back to the typical. Asked to revise real entries to a Helsinki museum competition, a model pulled them about 70% of the way to its own typical answer within eight rounds. Only a fresh random seed each round kept them apart.
• Creative taste is mostly personal. In 1,759 people rating art and songs, about 23% of a rating is shared, 44% is stable and personal, and the rest is noise.
• People agree on the extremes, not the middle. Across 14 datasets, shared taste nearly vanishes among middling items. Filtering out weak ideas can be generic. Choosing among good ones has to be personal.
• A generic judge only partly predicts one person. A model judge matched 9 of the founder’s 14 picks, above the one-in-three chance rate; a written profile of the founder matched 6, no better than chance. Fourteen picks is too few to rank the two.
• Liking the unusual is a lean, not a sweet spot. Among 45,136 people who each rated at least 40 of the same 100 jokes, each person leans slightly toward or away from the unusual, and the lean repeats. No one shows a favourite distance beyond chance.

What does not hold yet

That people prefer what we make. In the one blind human rating so far, the founder scored 24 ideas without knowing where they came from. Ideas from our recipe averaged 2.62 out of 5 on “would pursue”, against 3.25 for the best plain ideas, with eight ideas per arm. That difference is not significant, and it runs the opposite way from the model judge, which preferred our ideas 72% of the time.

Three more claims are untested. We have not run our generator against Verbalized Sampling, the strongest cheap way to restore diversity, which asks a model for several answers with their probabilities. We have not built a selector trained on one person’s picks. And everything so far was judged as text, while an architect judges a building by its drawings.

What we read from it

The generator is not the company. Tricks that restore diversity are being absorbed into frontier models, and we expect that to continue. What no one we have found is building is the per-person target: ideas unfamiliar to one person that the same person still picks. The selector, and the pick data behind it, is the work.

What we are running next

• A visual blind round: four briefs, plain ideas against ours at four distances, drawn as plan and section, judged by forced choice rather than scores.
• A Verbalized Sampling arm in every round from now on.
• A selector trained on one person’s blind picks, scored on picks it has not seen. The target is 80% of the rate at which that person agrees with themselves, after at most 40 calibration picks.

Next: What juries reward