Karpathy used Claude to build HackerTube from 6,454 videos recommended on Hacker News

On Friday evening, Andrej Karpathy opened YouTube and saw a feed of clickbait thumbnails and near-identical recommendations. The next morning he asked Claude to build him his own YouTube, one that would include only videos that Hacker News readers had recommended to each other. That is how HackerTube came about: a catalog of 6454 videos across 150 topics, built from recommendations HN users have left since 2007.

How the agents processed the Hacker News comment archive

The work was split between two Anthropic models. Opus downloaded the entire HN comment archive. Sonnet, running in a workflow (Workflow), then went through that data, discarded spam and kept about 40 thousand comments in which someone recommends a specific video. These became the basis of the database: 6454 videos that people on HN have recommended to each other over almost two decades.

The source data is open. The Hacker News archive is on Hugging Face in the open-index/hacker-news dataset, so anyone can repeat the filtering from scratch or take their own subset, for example only comments from recent years.

Karpathy did only two things by hand, both of which the agent could not do:

1. Created a Google Cloud API key to fetch video metadata from YouTube. 2. Set up a VPS and bought a domain.

The models did everything else, from parsing the archive to building the site itself.

What's inside HackerTube

The main difference from regular YouTube is visible right away in the thumbnails. The image is taken from the middle of the video rather than from a specially prepared cover. As a result, open-mouthed faces and red arrows disappear without any extra moderation: a cover's author cannot influence a frame from the middle of the video.

For You recommendations work without an account. Preference data is stored directly in the browser and is not sent anywhere. Hotkeys make navigation easier. A topic grid of 150 sections lets you move from one interest to the next instead of scrolling an algorithmic feed.

There is also a nod to agents themselves: the data of the entire site can be exported and handed to your own agent. It will build a similar service for your needs, for example from the recommendations of another community you trust.

What this means and where the limits are

The story shows how the cost of small personal products has changed. An idea that used to take weeks of parsing, data cleaning and layout work has become a task of a couple of hours of agent work plus manual infrastructure setup. In this arrangement, the human is responsible for taste and access: deciding which source to build the collection from, and providing the keys and servers.

There are limitations. The catalog reflects the tastes of the Hacker News audience, mostly developers and technically savvy people, so its topical range is narrow. Sonnet filtered the recommendations in the comments, and the author has not published any independent check of that filtering's quality. In addition, the service does not replace YouTube: videos are still hosted and played from there, and HackerTube only offers a different way to find them. The dependence on a Google Cloud key also remains.

The practical takeaway is simple. If a platform's feed annoys you, you can take an open dataset from a community you trust, have an agent do the filtering and build yourself a storefront. All that's left is to find a weekend for it, as the author of the retold post remarks with irony.