How Karpathy Built HackerTube YouTube Clone in Two Hours with AI

HackerTube is a YouTube-style video site at hackertube.org that shows only videos recommended in Hacker News comments. Andrej Karpathy, a co-founder of OpenAI and former head of AI at Tesla, published it on October 31, 2026, and said Claude built most of it in about two hours.

Instead of an algorithm deciding what to watch, thousands of thoughtful commenters did.

Karpathy wrote that YouTube’s recommendations had stopped working for him, describing the homepage as a wall of low-quality videos. He wanted a feed built from recommendations by people he trusts more than an engagement algorithm.

• Claude downloaded a public dataset of Hacker News comments
• It found about 40,000 comments that recommend a YouTube video
• A Sonnet-powered workflow in Claude Code reviewed them and removed spam
• YouTube metadata came from the YouTube API
• The result: 6,454 videos from 2007 to 2026, sorted into 150 topics

• Claude Opus 5.5 at medium effort, in the Claude Code desktop app on Mac, ran the project
• It processed the data, built a YouTube-like interface, and wrote the site
• It deployed the site to a server, then secured, hardened, and optimised it

• Created a Google Cloud API key for YouTube data
• Started a virtual private server
• Bought the domain name
• Gave Claude the server address and directions

He said he did not write the code or run terminal commands himself; he worked in between Claude’s tasks over a Friday night and Saturday morning.

• Browse 150 topics
• A “For You” feed without accounts, with your history kept in your own browser
• No login
• Keyboard shortcuts
• A data export so others can have their own agent build a version

It does mean

An expert with clear direction got a working, deployed site in about two hours of his time

It does not mean

Anyone can build any app in two hours, or that costs and time will be the same for you

The two-hour figure is Karpathy’s own estimate.

• Start from a real frustration you personally feel
• Use public data creatively
• Let a strong model run the project; use a cheaper model for bulk filtering
• Keep the human jobs: accounts, money, direction, and judgment
• Ship small and personal first

• Pick a trusted source of recommendations, such as a forum or newsletter archive you are allowed to use
• Ask Claude Code to collect and filter the data
• Use a cheaper model to classify items in bulk
• Ask for a simple interface and test it locally
• Deploy to a static host or server, and review security before you share

Build a site that collects every book recommended in this public dataset, removes spam, groups books by topic, and shows them in a clean, searchable interface. Plan first, then build step by step.

• Follow the terms of any API and dataset you use
• Link to original videos rather than copying them
• Keep API keys secret and out of your code