Tired of YouTube recommendations? Andrej Karpathy builds HackerTube with 6,454 videos to explore
The website uses years of Hacker News discussions to curate videos across 150 topics, while also demonstrating how AI coding tools can speed up website development.
- Nov 2, 2026,
- Updated Nov 2, 2026 5:00 PM IST
The project brings together nearly two decades of community-recommended videos, offering an alternative to YouTube’s recommendation feed and showing how quickly AI can help build a working website.
Finding worthwhile videos on YouTube can be difficult when its recommendations do not match what you actually want to watch. Andrej Karpathy, an AI expert, decided to approach the problem differently by turning to a source that already had years of recommendations from a technology-focused community. He asked Anthropic’s Claude Opus 5.5 to scan 43 million comments posted on Hacker News between 2007 and 2026. The exercise produced a collection of 6,454 recommended videos spanning 150 topics, which now powers a website called HackerTube. The project offers another way to discover videos, drawing on recommendations people have shared in online discussions rather than relying only on YouTube’s usual feed
MUST READ | ‘Stop AI development if….”: Microsoft AI chief warns against machines gaining consciousness
How HackerTube turns Hacker News discussions into video recommendations
Karpathy shared the project on X, describing it as a weekend hack that he built after becoming dissatisfied with YouTube’s recommendations. The resulting website, available at hackertube.org, presents videos through a simple interface with thumbnails and requires no login. Their browsing state is stored locally in the browser.
One detail Karpathy highlighted is how the thumbnails are selected. Instead of using creator-uploaded thumbnails that may be heavily optimised for clicks, HackerTube uses a representative frame from the middle of each video. The project also includes a section called HackerTube Longs.
The underlying data from the analysis is available as a ZIP file. Users can provide it to Claude and ask the AI to create their own version of the website.
DON'T MISS | Embracing AI advantage: 7 in 10 Indian businesses plan increased investment, finds study
Claude handled most of the development work
Karpathy said the project took around two hours to build with Opus 5.5. He used Claude Code on his Mac, with Opus 5.5 set to Medium effort by default, and relied on the AI for most of the development work.
His manual involvement included obtaining a Google Cloud API key to access YouTube video metadata, setting up a virtual private server (VPS) and purchasing a domain. Once the server was ready, he gave Claude its IP address and asked it to update, secure, harden, deploy, and optimise the code.
The project received praise for its quick development and community-driven recommendations, although some users felt the video selection was dated. Karpathy said projects that would normally take weeks or months could now be built much faster with the help of AI.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
The project brings together nearly two decades of community-recommended videos, offering an alternative to YouTube’s recommendation feed and showing how quickly AI can help build a working website.
Finding worthwhile videos on YouTube can be difficult when its recommendations do not match what you actually want to watch. Andrej Karpathy, an AI expert, decided to approach the problem differently by turning to a source that already had years of recommendations from a technology-focused community. He asked Anthropic’s Claude Opus 5.5 to scan 43 million comments posted on Hacker News between 2007 and 2026. The exercise produced a collection of 6,454 recommended videos spanning 150 topics, which now powers a website called HackerTube. The project offers another way to discover videos, drawing on recommendations people have shared in online discussions rather than relying only on YouTube’s usual feed
MUST READ | ‘Stop AI development if….”: Microsoft AI chief warns against machines gaining consciousness
How HackerTube turns Hacker News discussions into video recommendations
Karpathy shared the project on X, describing it as a weekend hack that he built after becoming dissatisfied with YouTube’s recommendations. The resulting website, available at hackertube.org, presents videos through a simple interface with thumbnails and requires no login. Their browsing state is stored locally in the browser.
One detail Karpathy highlighted is how the thumbnails are selected. Instead of using creator-uploaded thumbnails that may be heavily optimised for clicks, HackerTube uses a representative frame from the middle of each video. The project also includes a section called HackerTube Longs.
The underlying data from the analysis is available as a ZIP file. Users can provide it to Claude and ask the AI to create their own version of the website.
DON'T MISS | Embracing AI advantage: 7 in 10 Indian businesses plan increased investment, finds study
Claude handled most of the development work
Karpathy said the project took around two hours to build with Opus 5.5. He used Claude Code on his Mac, with Opus 5.5 set to Medium effort by default, and relied on the AI for most of the development work.
His manual involvement included obtaining a Google Cloud API key to access YouTube video metadata, setting up a virtual private server (VPS) and purchasing a domain. Once the server was ready, he gave Claude its IP address and asked it to update, secure, harden, deploy, and optimise the code.
The project received praise for its quick development and community-driven recommendations, although some users felt the video selection was dated. Karpathy said projects that would normally take weeks or months could now be built much faster with the help of AI.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine