Use case · Idea to test
A feed without repeated prompting
Learn from what people choose to see.
- See an item
- Like, skip or save
- Choose the next item
A feed could learn from explicit feedback instead of asking someone to write a new prompt for every image, clip or story. It could rank existing content or help choose among newly generated items.
Where the small model fits
A personal scorer can rank existing items. With a compatible generator, it could instead select promising internal inputs before paying to render the next item. The original idea included both approaches; a high predicted rating still needs to be checked against the person's response.
What would need to work
Measure satisfaction, variety and unwanted repetition. More time spent scrolling is not automatically a better experience, and a skip is not an unambiguous dislike.
Try the related work
This demo learns from simple drawings. It does not run the application described above.
Related experiments and evidence
Earlier proposal and example code
The earlier page is preserved as a historical proposal; its claims are not new validation.