Back home

HackerTube

sitemap 0.78
Visit source

Intro

HackerTube (hackertube.org) is a personal YouTube alternative built by Andrej Karpathy and launched on October 31, 2026. Instead of an engagement-driven recommendation algorithm, its catalogue is made of 6,454 videos that people recommended in Hacker News comments between 2007 and 2026 — talks, lectures, documentaries and "timeless" tech videos.

Karpathy wrote on his Bear Blog that the YouTube homepage had become an "enshittified wall of slop" that "definitely does not work for me," so he spent a Saturday morning directing Claude Opus 5.5 to mine the full Hacker News comment archive, filter recommendations with a model pipeline and wrap the result in a familiar video-site interface. The whole project took roughly two hours. His HN post "HackerTube: YouTube built from videos recommended by HN" reached 592 points and 214 comments, and outlets such as RuntimeWire covered it as a showcase of agentic coding.

Explain

How it was built

  1. Data: the agent pulled Hacker News comments from 2007 to 2026 and extracted YouTube links people recommended.
  2. Filtering: a model pipeline (Karpathy mentions Sonnet-class models) scored and filtered candidates, leaving a curated set of 6,454 videos.
  3. Product: Claude Opus 5.5 built a YouTube-like front end — Home, Longs, Liked, History, Watch later, Settings and Explore — and deployed it.

Why it got attention

  • It is a concrete example of a senior AI researcher replacing a mass-market recommendation feed with a personal feed built from a trusted public corpus.
  • It demonstrates one agentic model handling data processing, product building and deployment in a single short session.
  • It revived the debate about algorithmic feeds versus community curation; HN users immediately began sharing individual HackerTube watch pages.

HackerTube is a personal side project, not a company product, so features and availability may change.

Usage

How to use it

  • Open https://hackertube.org and browse like YouTube: Home, Longs (long-form), Explore, and personal lists such as Liked, History and Watch later.
  • Use it as a low-noise discovery feed for talks, lectures and documentaries the HN community has vouched for.
  • Builders can treat it as a template: pick a trusted corpus (a forum, newsletter archive or Discord), let an agent extract and rank recommendations, and ship a personal front end.

Links