HackerTube: YouTube built from videos recommended by HN

The thread treated HackerTube less as a novelty video site than as an indictment of engagement-driven recommendation systems. Many readers found an HN-curated, unpersonalized feed more calming and more useful than YouTube because it strips away aggressive thumbnails, repetitive topic loops, influencer bait, and the sense that every click is being optimized to prolong a session. The striking claim was that a lightweight social filter can match or beat a platform with years of behavioral data, suggesting YouTube may optimize watch time rather than genuine viewer satisfaction. Requests for recency sorting and topic playlists reinforced that people want transparent controls over discovery, not another opaque feed. Several commenters argued that recommendation AI has narrowed rather than expanded their horizons: it repeatedly predicts familiar preferences when they go online specifically to encounter surprise, disagreement, and new fields. One proposed a local AI curator that learns private preferences without exporting them to an advertising system. Others pushed back indirectly, noting that careful long-term feedback training can yield an excellent YouTube feed and that curation by HN merely substitutes one community’s biases for algorithmic ones. The project also became evidence of agentic coding’s reach: its creator reportedly built the pipeline, search, recommendations, and deployment largely through Claude Code, though one frustrated user said current LLM agents often quit, reverse themselves, or fail to sustain complex work.