keyboard: ← pass · → like · S skip · U undo
The ring is the loop from the 2019 app: every 20th rating, your current ideal is generated. Keep going after each reveal — it sharpens.
When on, the next cards are drawn against your current model — a deliberate mix of predicted likes, uncertain cases, and pure exploration. Turn it off for uniform random sampling and feel how much duller (and slower to learn) the deck gets. The 2019 patent calls this out directly: bored users stop training.
The model, live
1 · Reverse classification — generate your ideal
Instead of asking "how would you score this sample?", we solve for the latent point your classifier would score highest, and decode that. With a linear model this is exact math, not search.
scores — on your model
The slider is the truncation trick: a 100% optimum sits at the edge of the latent box and can look caricatured; pulling toward the middle trades a fraction of a percent of score for plausibility. And because "ideal" isn't unique, relaxing the target by a couple of logits yields endless distinct near-ideals — the patent's 99%-vs-100% point:
2 · Forward pass — score anything
The same model runs forward to score samples it has never seen — that's matching and ranking. In the dating design, two people's models score each other's encoded photos, so matching happens without anyone's photos being browsed.
3 · Minimal-change transform
Take an average sample and apply the smallest latent change that reaches a target score — the makeup-recommendation use case: what little would have to change, and in which direction? Closed-form for a linear model.
Target score: 90%
What changed
- —
What's real here, and what's a stand-in
Real: the classifier (logistic regression — the patent's "single dense layer with sigmoid"), the per-rating retraining with honestly measured times, the active-sampling strategy, reverse classification, and the minimal-change transform. This is the actual loop, end to end.
Stand-in: the decoder. These faces and compositions come from a hand-built parametric renderer with 12–16 interpretable dimensions, so you can see the weights make sense. The 2019 app ran the identical loop against StyleGAN's 512-dimensional latent space, where dimensions are learned rather than labeled — see the original demo videos, the provisional patent, and the open-sourced engine.
Also real: the point. A taste model this small trains in about a millisecond on a phone, which is what makes per-user personalization of generative models economically boring — in the good way.