The pile at zero
We asked a frontier model for the big idea behind an entry to the eVolo Skyscraper Competition, an open ideas competition for vertical architecture, thirty separate times. Twenty-six of the answers were named for the tide: a Tidal Tower, a Tidal Spine, a Tidal Spire. The wording changed; the idea did not.
Thirty answers to the eVolo Skyscraper brief
- 26 named for the tide
- 4 named for something else
Human competitors do not behave this way. On a 2025 Helsinki museum competition, 30 real entries had a mean pairwise similarity of 0.44. Thirty model answers to the same brief had 0.74. Published work finds the same pattern from other angles: post-training narrows output, raters reward the familiar, and different models give answers that resemble each other even more than repeated samples from one model.
How alike thirty ideas for one brief are
Seeds pull ideas apart
We gave every idea its own concrete starting fragment and measured what happened. The test used 24 corpora of life writing, half public-domain memoirs from very different lives and half invented by a model, plus random lists of everyday objects and plain prompting as baselines. Eight competition briefs, four ideas per corpus per brief, 1,536 ideas in all.
| Seed | Similarity between sources | Source identifiable |
|---|---|---|
| None (plain prompting) | 0.793 | 11% |
| Random everyday objects | 0.705 | 8% |
| Invented life writing | 0.697 | 17% |
| Real life writing | 0.701 | 21% |
Lower similarity means more divergent ideas; this is our strictest reading, with the stories stripped out of each idea before it was embedded. Chance for identifying the source is 8%. Any rotating set of concrete seeds escaped the pile, and random objects did as well as real lives. A person’s own writing leaves a fingerprint you can detect at about three times chance, but no automatic measure found that it made ideas better.
Loops drift back
Then we let a model revise ideas over eight rounds. Real Helsinki entries started far from the model’s typical answer. Revised by the model alone, they moved about 70% of the way toward it; revised as a group, further still. Only when a fresh random seed entered every round did they stay apart. The same held for the model’s own answers: with fresh seeds their similarity fell from 0.74 to about 0.5.
What it means
A generator needs new material from outside every round, and it does not need anyone’s private writing to be diverse. Where a person matters is as the judge. That moved our thesis from “seed ideas with a person’s corpus” to “seed widely, then choose by the person”.
Sources
- Feathers & Wax, Corpus seeding divergence test and Corpus seeding test, round 2, Oct 26, 2026.
- Feathers & Wax, weirdness-dial runs on the eVolo brief,
research/cllm-method/weirdness-dial. - Feathers & Wax, human against model drift on the Helsinki brief,
research/cllm-divergence/drift. - Kirk et al., Understanding the Effects of RLHF on LLM Generalisation and Diversity, ICLR 2024.
- Jiang et al., Artificial Hivemind: The Open-Ended Homogeneity of Language Models, 2025.