Andrey Karpathy introduces HackerTube: An AI-powered YouTube clone for Hacker News

Renowned AI expert, OpenAI co-founder, and former Tesla autopilot head Andrey Karpathy has introduced a new project called HackerTube. This is an AI-driven clone of YouTube that exclusively features video recommendations curated by the Hacker News community over the years.

• HackerTube was built in approximately two hours using advanced AI workflows and the Opus 5.5 model.
• The database includes 6,454 unique videos filtered from 40,000 Hacker News comments spanning from 2007 to 2026.
• The platform operates without user accounts, stores session data locally, and allows full data export.

Background and Inspiration for the Project

According to Karpathy, the idea struck him late on a Friday night when he tried to find something worthwhile to watch on mainstream YouTube. Frustrated by sensationalized thumbnails and repetitive algorithmic recommendations, he decided to take matters into his own hands. Saturday morning, he prompted his Opus 5.5 model to build a custom YouTube clone based entirely on Hacker News recommendations.

The AI downloaded all comments from Hacker News, filtered out spam using the Sonnet model, and extracted mentions of YouTube links. This resulted in 6,454 recommended videos across 150 distinct topics gathered from user recommendations between 2007 and 2026.

Platform Features and Interface

HackerTube mimics the familiar YouTube layout while introducing functional improvements. Instead of relying on creator-uploaded thumbnails, it automatically generates video previews using a frame from the exact middle (50%) of the video.

The "For You" section provides personalized recommendations without requiring users to create accounts, storing all data locally in the browser session. Furthermore, the platform supports complete data exports, enabling developers to easily spin up their own derivative versions of the site.

The Vibe-Coding Workflow

Karpathy emphasized that the entire development process relied on vibe-coding methodologies using the Claude Code desktop application and the Opus 5.5 model configured for medium effort. He noted that he no longer engages in traditional engineering tasks like writing code or executing terminal commands, acting instead as a director communicating with an AI assistant.

The only manual tasks required on his end were creating an API key in the Google Cloud console to fetch YouTube metadata and purchasing a domain name. Once the server IP was provided, Claude independently handled updates, deployment, security hardening, and code optimization.