surprisal

Your moots' posts become a trigram Markov model and a likes regression, trained in this tab. Then you type a post: it scores well if the regression thinks it would earn likes and the Markov model didn't see it coming.

1 · gather a corpus

Reads who you follow and who follows you, pulls each moot's whole repo, then looks up like counts. Big circles take a while — you can stop any time and train on what's in.

what the models learned

average likes lift (vs. the same author's typical post) by how surprising the post was — left is boilerplate, right is wild

predictablesurprising

2 · the contest

Entering teaches the Markov model your post, so the same trick is less surprising next time. Points = predicted likes lift × surprise credit (credit starts above the median surprise, full at the 90th percentile; mostly-unseen-word gibberish is docked). 100 points is an average post that is just surprising enough.

surprisal

1 · gather a corpus

what the models learned

2 · the contest

leaderboard (this browser)