Andrej Karpathy built a video discovery site over a weekend by using Claude to sift through 43 million Hacker News comments. HackerTube presents 6,454 videos that users repeatedly recommended across roughly 20 years of discussion, replacing personalized engagement rankings with community history.
The site requires no account, stores viewing state in the browser and uses a frame from the middle of each video instead of an optimized thumbnail. Karpathy says this produces a better viewing experience than YouTube’s recommendations, although that is his assessment rather than a comparative study.
Claude Opus 5.5 handled most of the data analysis and Claude Code handled work through server setup, according to Karpathy. He said his manual contributions were obtaining a Google Cloud API key, renting a server and buying the domain. He then made TristanTube, a separate collection of 47,000 videos from a list assembled by engineer Tristan Hume, in about three minutes with Claude Code.
The projects are small demonstrations, not evidence that coding agents can autonomously maintain a production platform. They do show how inexpensive it has become to turn a large public archive into a working interface. The main human contribution was choosing a useful ranking signal: repeated links in a technically focused community. That editorial choice, more than the volume of generated code, determines whether the result is meaningfully different from another recommendation feed.