Dev builds Chrome extension that filters YouTube feed with a small local model via WebGPU, no server
manjunath_shiva · reddit · 2026-10-06
Developer manjunathshiva released an open-source Chrome extension that filters your YouTube feed using a small decision model (opendecider-nano, ONNX) running entirely in the browser — no server, no API key, offline after one download.
Features
- Hides 11 video categories (music, gaming, comedy, vlogs, news, etc.) or follows natural-language rules like "hide crypto" / "only cooking"
- One-switch Shorts hiding
- Bonus: right-click any text to check it for prompt injection with the same model
How it runs
- ONNX Runtime Web in an offscreen document; fp16 on WebGPU (755 MiB), q8 on WASM without GPU (569 MiB)
- 40 titles classified in 1.1s on WebGPU, 15s on CPU (M4 Max); 3 GiB RAM while loaded, unloads after 10 idle minutes
- Weights pinned by revision and SHA-256; only network calls are to Hugging Face
Accuracy: on 400 videos, rules like "hide music" average 0.934 balanced accuracy; 11-way classification is harder at 0.780. Evaluated on English titles only. Apache-2.0, Web Store version under review, loadable manually from GitHub Releases.
More from coding & agent
- Her game hit 2M players and 100k+ concurrents in 4 days, yet people still call her a vibecoder — FrankFelixAI · 2026-10-06
- Creator shares 120-prompt workflow building an AI video game with Seedance 2.5, Godot and Magnific — techhalla · 2026-10-06
- Power users nuke all memory, workflows and setups whenever a new model ships — nateliason · 2026-10-06
- Cognition's SWE-2 outperforms Fable-5.1 and Opus 5.5 on Animation Bench, testers report — himanshustwts · 2026-10-06
- OpenAI makes Codex auto-review free, no longer counting against usage limits — mark_k · 2026-10-06
- DIY ESP32-S3 pocket voice assistant with e-paper display and Hermes agent routing — K8s-monk · 2026-10-06