SkinDeep
Research notes and experiments by Steve Seguin.
SkinDeep started with a simple idea: show people generated images, learn which ones they like, then use that feedback to suggest new ones.
The work now includes small models that return a decision or a click location, experiments that test whether a model can stop processing early, and ways to keep a long job going when there is more text than a model can read at once.
Learn what you likeUse likes and dislikes to guide suggestions.Return a decisionGive a label instead of writing a reply.Find where to clickPoint to a button in a screenshot.Stop earlyTest when fewer processing steps are enough.Work past the context limitKeep long jobs going without losing track.
Two tools came out of this work: ctxkeeper keeps long AI conversations going, and our own inference runtime runs one model family fast and checks that its output matches a reference run.