Andrej Karpathy Uses AI to Build HackerTube, a Video Site Made From Hacker News Recommendations

The author of karpathy.bearblog.dev, who signs his blog as karpathy, published a post on 31 October about a new video site called HackerTube (hackertube.org). It resembles YouTube, but contains only videos that Hacker News users have recommended. Hacker News is a discussion platform focused on technology, where users share links and comments. The project matters as an example of one person using AI to build an alternative to a major platform in a few hours.

Why the idea appeared

The author writes that on a Friday night around 11pm he opened YouTube and felt he was looking at overly exaggerated thumbnails and repeated topics. In his view, the YouTube algorithm may work for the median user, but not for him. So on Saturday morning he asked the AI model Opus 5.5 to write his own version of YouTube, based on videos recommended on Hacker News.

How the site was built

According to the author, Opus downloaded all Hacker News comments and filtered them. The Sonnet model was used in a workflow to remove spam and low-quality content. About 40,000 comments remained, each recommending a YouTube video. In total there were 6,454 videos, recommended by various users between 2007 and 2026 and grouped into 150 topics. The content was then put into a YouTube-style interface, a domain was purchased, the site was deployed to a virtual private server, and it was secured and optimized.

Features of the site

The author lists these features of HackerTube:

• quality content;
• thumbnails that use the 50% frame of a video instead of the uploaded thumbnail;
• customized recommendations in a ForYou section without user accounts, with state kept in the browser's local session;
• keyboard shortcuts;
• an export of the site's data, which can be given to one's own agent to create a personal version.

The author says the whole build took about two hours and thanks Opus 5.5.