99.98% blocked: I ran 9,000 Shorts through my filter so the old model could fail in peace
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…
A journalist built a filter to combat scammy re-uploads, AI slop loops, and clickbait compilations that were clogging up their own Shorts feed. They ran 9,000 Shorts through two classifiers: the current model and an older revision still in use. The previous model allowed 8,667 of the 9,000 videos through, resulting in a 96.3% miss rate and explaining their recent frustration with the feed's quality.
The new model, however, blocks 99.98% of the videos. The mechanism behind this high blocking rate involves sampling 8 frames per video, hashing them, and deduplicating near-identical frames across the entire collection before they reach the model. This preemptive step eliminates most mass-reupload spam. The filtered videos then undergo a transcript-plus-visual embedding pass, and any entries under the confidence threshold are blocked rather than merely demoted.
The latter approach ensures that junk content remains in a "maybe" queue where humans can still identify and address it. The operator's costs are roughly 40 GPU-minutes per 1,000 videos during the frame-sampling phase, along with storage for the hashes. Skipping the dedupe step would nearly triple the operating costs for the same level of blocking.
The journalist has detailed the entire process, including per-class breakdowns and the old model's failure cases, in a YouTube Shorts video available at https://www.youtube.com/shorts/oZmarp6b8Qw.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
