Andrej Karpathy Launches HackerTube With 6,454-Video Catalog

AI researcher Andrej Karpathy launched HackerTube on Saturday, Oct. 31, organizing years of video recommendations in Hacker News comments into a 6,454-video catalog that shows how AI can process a large community archive.

The videos span 2007 through 2026 and are sorted into 150 topics. The project uses 39,834 comments, or about 40,000 when rounded. HackerTube is a YouTube-style discovery site where visitors can browse the recommendations by topic.

Karpathy described a two-stage process for building the catalog. Claude Opus 5.5 downloaded and filtered the comments, then Claude Sonnet, in a Claude Code workflow, screened for spam and identified video recommendations.

Karpathy estimated the project took about two hours. He identified three manual tasks: creating a Google Cloud API key to retrieve YouTube metadata, starting a virtual private server and buying the domain. “This takes ~2 hours now. I can breathe again,” Karpathy wrote.

Each video page uses a frame from the video’s midpoint as its thumbnail. The site also offers keyboard shortcuts and a data export. Visitors can browse without an account, and recommendation state is stored in the local browser session.

Karpathy reported 504 million input tokens, 90% of them cached, and 7.6 million output tokens for the project. Sonnet’s comment-labeling calls accounted for more than 90% of the project’s usage. Building and deploying the site used about 1 million Opus 5.5 tokens. Karpathy said he was on a subscription when he asked Claude to total the usage; he did not give a dollar amount.

The catalog turns recommendations from one online community into a topic-based video index. Karpathy’s two-hour estimate describes this project and does not establish a typical development time or cost for similar sites.