HackerTube: YouTube built from videos recommended by HN
A community-built site called HackerTube curates YouTube videos that have been recommended in Hacker News comments, offering a stripped-down, ad-free interface and filters for length, recency, and quality signals as an antidote to YouTube’s engagement-driven “slop.” Commenters praise how closely the feed matches their interests compared to YouTube’s algorithm, suggest features like playlists, RSS, and alternative thumbnails, and raise questions about long-term sustainability and YouTube’s API terms. Many are also struck by how quickly the project was assembled using modern AI coding agents, seeing it both as a showcase of new development workflows and as a glimpse of more user-controlled recommendation systems.
Overall reception and purpose
- Many commenters praise the site as a “breath of fresh air” vs. YouTube’s “enshittified” homepage.
- People like that it’s effectively a HN‑tuned “video front page” of long‑form, technical, and educational content rather than influencer or ragebait material.
- Some say the feed looks remarkably close to what they wish their YouTube home page showed.
Recommendations vs. YouTube
- Repeated sentiment: YouTube’s algorithm optimizes for engagement and ad views, not user value, leading to clickbait, repetition, ragebait, and trashy kids’ content.
- Several note that well‑trained YouTube accounts can work decently, but require active likes, dislikes, and “don’t recommend” usage.
- Others describe radical tactics (turning off history, RSS, external clients, ad blockers) to escape the default feed.
- A few think HN recommendations still surface some polarizing or low‑quality channels, underscoring that HN is “just another algorithm”.
Features, filters, and UX
- Strong appreciation for: duration filters, upload vs. “recently recommended” sorting, type filters (talk/lecture/documentary), like‑ratio sorting, “Hidden gems” (HN‑loved but low‑view), “Evergreen only”, and a “Longs” section.
- Requests include: better prominence of filters, keyboard navigation, random video, RSS feeds, dislike counts and SponsorBlock integration, plus Apple TV / tvOS apps.
- Some users want playlists or a way to push the curated set back into YouTube (e.g., auto‑generated playlists, “Play all on YouTube”).
Technical implementation and LLM usage
- Discussion notes that the pipeline was built almost entirely via chat with an LLM agent: scraping HN, classifying comments, building backend, frontend, deployment, hardening.
- Reported usage: ~504M input tokens (mostly cached) and 7.6M output, with the bulk spent on classifying comments using a cheaper model; estimated on‑demand API cost around $250–300, but effectively “free” under a flat subscription.
- Some are impressed by how fast and cheap this is; others argue similar could be done with smaller or cheaper models.
Data, openness, and forking
- The underlying dataset (video list, metadata, labels) is downloadable as a ZIP, explicitly intended for others to fork into custom front ends or pipelines.
- Commenters discuss building variants: different topics (food, art, documentaries), alternative interfaces, email digests, or desktop/TV clients.
Legal, ToS, and longevity concerns
- Some warn that previous YouTube wrappers have faced pushback for acting like alternative search indices or caching metadata beyond allowed durations.
- Others note this site embeds official YouTube players and doesn’t host video itself, but ToS compliance and long‑term viability are seen as uncertain.
Broader reflections and critiques
- Thread contains broader critiques of engagement‑driven design, addiction, and recommendation systems that narrow horizons instead of expanding them.
- Some worry that any successful meta‑feed (including this one) could eventually attract spam, astroturfing, and “slop”.
- A minority criticize the hype level around a curated wrapper and question how much novelty the project really adds beyond HN + YouTube.