The question
Our core hypothesis is that each person has a distance from the obvious they like best, and an idea at that distance will feel right to them. The Jester dataset lets us test it: 45,136 people rated the same 100 jokes.
A stable lean
Splitting each person’s ratings in half, how much they favour unusual jokes in one half predicts the other at r 0.33. Most of that is joke length: length alone gives 0.46. Net of length, the lean is real but small, 0.11 to 0.16. People split evenly, about half leaning toward the unusual and half away.
No interior sweet spot
Among 33,360 people with at least 60 ratings, we looked for a peak inside the range: a person who likes jokes a little unusual, and less beyond that. We found no reliable personal curvature, no excess of interior peaks, and people who preferred the middle band in one half did not prefer it again in the other. The small agreement we did see in peak location is fully reproduced by synthetic people who only lean. A planted sweet spot worth about 2% of a person’s rating variance would have shown up clearly.
What it does and does not say
These are 100 ordinary jokes, with unusualness measured by an embedding. The far tail, where our second hypothesis lives, is not in this range, and neither is unusualness as a person judges it. So the sweet spot is untested where it matters most. What the data supports is a direction: each person leans, a little, one way.
For the selector, that means learning a person’s direction and strength from their own picks, then reading the far tail in that direction with their own eyes.
Sources
- Feathers & Wax, personal z in Jester, code in
research/jester-personal-z. - Jester joke ratings, Ken Goldberg, UC Berkeley.