Andrej Karpathy launched HackerTube, a platform featuring 6,454 videos from Hacker News comments spanning 2007 to 2026. The site organizes content into 150 topics using AI models. Karpathy said, "The entire project took him about two hours to complete." Users can browse these videos easily without needing an account today.

Renowned artificial intelligence (AI) researcher Andrej Karpathy has launched a new platform called HackerTube. The innovative site is a YouTube-style discovery platform where users can explore video recommendations by topic. The launch comes as part of an effort to demonstrate how AI can handle large community archives. The project is based on years of video recommendations shared in Hacker News comments and features 6,454 videos spanning from 2007 to 2026.

The vast video collection on HackerTube has been organized into 150 different topics. The project employs a total of 39,834 comments to curate this extensive catalog.

Karpathy explained that the process of creating this catalog was done in two stages with the help of Claude Opus 5.5 and Claude Sonnet models.

The first stage of the process involved downloading and filtering comments using Claude Opus 5.5. In the second stage, Claude Sonnet was used to screen for spam and identify video recommendations in a Claude Code workflow.

Karpathy said that the entire project took him about two hours to complete, with three manual tasks involved: creating a Google Cloud API key for YouTube metadata retrieval, starting a virtual private server (VPS), and buying the domain name.

HackerTube offers a seamless user experience, with each video page featuring a frame from the video's midpoint as its thumbnail.

The site also provides keyboard shortcuts and data export options for added convenience.

Users can browse without creating an account, and their recommendation state is stored in the local browser session for easy access.

Karpathy revealed that the project involved 504 million input tokens, with 90% of them cached, and 7.6 million output tokens.

He also noted that Sonnet's comment-labeling calls made up over 90% of the project's usage.

The building and deploying process of the site consumed about a million Opus 5.5 tokens, although Karpathy didn't specify a dollar amount for this usage as he was on a subscription plan when he asked Claude to total it up.