The test
We took four frameworks from very different fields: the architect Alberto Campo Baeza, the cook Anthony Bourdain, the comedian George Carlin and the artist Yayoi Kusama. Each was rewritten into the same 380 to 450-word shape, with the name removed. A model then used each one to answer six briefs: a cartoon caption, a one-liner, a five-course tasting menu, a chair, a song concept and a brand line for a city bike-share. Each person had at most one brief from their own field.
A judge model saw four anonymous summaries and one output, and picked whose it was. Chance is 25%.
Results
| Outputs | Picked right |
|---|---|
| All 96 | 88% |
| Restated in plain words at one length | 86% |
| Briefs away from the person’s field | 86% |
| Using none of the person’s home words, and not named | 67% |
Every framework also moved its outputs well away from what the model wrote with no framework, most where the plain answers were most alike: tasting menus sat 9 to 16 standard deviations out.
Mostly their world
The commonest route was to carry the person’s world into the new field. Bourdain’s chair is a line cook’s stool. Kusama’s song is called “Ten Thousand Dots”. Campo Baeza’s menu opens with a course called “Oculus”. That is easy to recognise and probably reads as pastiche to a person.
The more interesting cases carry a way of thinking without the props: Kusama’s banana “weighed, and weighed, and weighed”, or Bourdain’s bike-share line, “Ride it like you live here.”
Where it fails
Humour transferred least. Short jokes converged on the same punchline, and two different frameworks produced the same caption word for word. On a cartoon, the pull of the best-known joke beats the framework.
The open question
The model named all four people from their frameworks alone, so this test cannot separate the framework from the model’s memory of a famous person. The follow-up, Method, not motif, answers that with invented people.
Sources
- Feathers & Wax, framework transfer, round 1, code and data in
research/framework-transfer. 712 calls, pinned models.