Frustrated with YouTube's algorithmic recommendations, Karpathy built HackerTube in two hours using Claude.

OpenAI cofounder Andrej Karpathy said he built HackerTube, an algorithm-free video site, in about two hours using Anthropic's Claude, with the project live and its cleaned dataset open-sourced. The site indexes 6,454 videos across 150 topics, curated from roughly 43 million historical Hacker News comments, with no login, local-only data storage, and static mid-frame thumbnails replacing clickbait covers. Claude Code drove the full build: Opus 5.5 handled architecture, full-stack code, and deployment, while Sonnet performed semantic filtering of about 40,000 candidate recommendations, with Karpathy writing no code and only provisioning API keys, a domain, and cloud infrastructure. The run consumed 504 million input tokens and 7.6 million output tokens, with roughly 90% input cache-hit rates cutting costs; over 90% of spend went to Sonnet-based data classification, while core architecture and deployment used only about 1 million Opus tokens. The project matters as a template for AI-native full-stack development and as a counter to black-box recommendation feeds, showing high-value architecture generation is now cheap while compute skews toward bulk data inference. Watch adoption of the open-sourced dataset and whether similar "vibe-coded" personal software projects emerge from other high-profile developers.