Karpathy Builds HackerTube in About 2 Hours Using Claude to Curate Videos
AI assistants · AI coding
Former OpenAI founding member Karpathy stated he used Anthropic's Claude models to process roughly 43 million Hacker News comments, curating 6,454 recommended videos to build a site called HackerTube in about two hours.
Former OpenAI founding member Karpathy posted on social media that, dissatisfied with YouTube's recommendation algorithm, he used Claude models to analyze approximately 43 million Hacker News comments spanning two decades, extracting 6,454 repeatedly recommended videos into a site named HackerTube. According to media reports, Opus 5.5 designed the data extraction pipeline while Sonnet refined the final link list; Karpathy only manually handled API key setup, VPS provisioning, and domain purchase, delegating server hardening, deployment, and optimization to the AI. The site requires no login, keeps history locally, and allows dataset exports. Current details rely on Karpathy's posts and media accounts without independent evaluation.
Key takeaways
- Karpathy reported building HackerTube in about two hours using Claude to curate videos from Hacker News comments.
- The dataset spanned 2007 to 2026 across roughly 43 million comments, yielding 6,454 videos across 150 topics.
- Opus 5.5 established the data pipeline while Sonnet handled filtering; AI also assisted in server setup and deployment.
- The service requires no user login, stores data locally in browsers, and provides dataset exports; independent evaluations are absent.
Source：INSIDE Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.