Karpathy built a better YouTube with Claude in two hours

Tired of YouTube's noisy recommendations, Andrej Karpathy decided to build his own clean alternative over a single weekend afternoon.

Karpathy revealed HackerTube, a custom video recommender built in about two hours using Claude Opus 5.5.

Mining 20 years of hacker taste

To bypass YouTube's engagement-driven algorithm, he used Claude to filter down 43 million Hacker News comments spanning 20 years. The result is a curated collection of 6,454 videos that technical users repeatedly recommended over two decades.

The tool is stripped of all modern dark patterns. It requires zero login, keeps your watching state saved locally in your browser, and actively ignores clickbait thumbnails by simply grabbing frames from the middle of each video.

The building process itself is a testament to current AI capabilities. Karpathy ran Claude Code on his Mac to generate the app, only stepping in manually to set up a Google Cloud API key, configure a VPS, and purchase the domain.

Why it matters

We think HackerTube is a brilliant blueprint for how internet users can fight back against platform algorithms. Instead of accepting what an ad-driven recommendation engine pushes, it replaces the black box with twenty years of proven human taste from a high-signal community.

More broadly, it highlights the dawn of truly disposable, hyper-personalized software. A single engineer equipped with a frontier model can now spot an annoying digital friction point and ship a fully functional, custom-tailored web utility in just two hours.