Zvi's AI newsletter covers HackerTube, noting the project as an example of AI-assisted content curation.
Karpathy's HackerTube built with Claude
Latest: Karpathy Built HackerTube from 43M HN Comments in Two Hours With Claude
The project is a concrete example of an AI model handling a full software pipeline, from data analysis through deployment, with minimal human input. Karpathy said, 'Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast.'
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
Andrej Karpathy built HackerTube, a curated video site drawn from two decades of Hacker News community recommendations, using Claude Opus 5.5. The full pipeline, scanning 43 million comments, surfacing 6,454 repeatedly recommended videos, writing and deploying the site, ran in approximately two hours on Claude's Medium effort setting. Karpathy described his own role as minimal: he created a Google Cloud API key, spun up a VPS, and bought the domain, then handed the server IP to Claude with instructions to 'update, secure, harden, deploy and optimize the code,' which it did. The site deliberately sidesteps YouTube's algorithmic recommendations and uses a representative mid-video frame as the thumbnail instead of A/B-tested uploads. Karpathy made the underlying comment-analysis data available as a zip file so others can build their own versions using Claude.
How it developed
Andrej Karpathy built and deployed HackerTube, a curated video site drawn from 6,454 YouTube recommendations extracted from 43 million Hacker News comments spanning 20 years, in roughly two hours using Claude Opus 5.5.
Karpathy described his own role as creating a Google Cloud API key, spinning up a VPS, and buying the domain; Claude handled data analysis, code writing, security hardening, and deployment. The site uses a mid-video frame as its thumbnail rather than A/B-tested uploads and skips algorithmic recommendations; Karpathy released the underlying comment-analysis data as a zip file for others to build their own versions.
Sources
- DeepSeek V4-Pro open-weight release
- Runway's Solaris frame-by-frame UI model
- Google DeepMind agentic video understanding for Gemini
BackgroundFrontier models, open models, and how they are compared
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.