I built this filter because my own Shorts feed had become unusable — scammy re-uploads, AI slop loops, and clickbait compilations stacked ten deep. I wanted numbers, not vibes, so I pulled 9,000 Shorts and scored them against two classifiers: the current one and the previous revision I still had deployed.
The old model let 8,667 of the 9,000 through. That's a 96.3% miss rate, and it explains every "why is my feed like this" moment I've had this year. The current model blocks 99.98%.
One concrete detail of how it works: I don't classify the video file directly. I sample 8 frames per video, hash them, and dedupe near-identical frames across the whole corpus first — that alone kills most mass-reupload spam before any model sees it. Survivors go to a transcript-plus-visual embedding pass, and anything under the confidence threshold gets blocked, not demoted. Demotion just moves the junk into the "maybe" rail where humans still see it.
What it costs the operator: roughly 40 GPU-minutes per 1,000 videos on the frame-sampling pass, plus storage for the hashes. If I skipped the dedupe step and sent everything straight to the model, that bill would be about 12x higher for the exact same block rate. The dedupe layer is the whole economics of the thing.
I recorded the full walkthrough with the per-class breakdowns and the failure cases of the old model: https://www.youtube.com/shorts/oZmarp6b8Qw
Top comments (0)