Karpathy Used Claude to Sift Through 20 Years of HN Comments and Built a YouTube Without Login

Andrej Karpathy posted a weekend project on X: HackerTube. The motivation was very specific: he was annoyed by YouTube recommendations, so he had Claude go through 43 million Hacker News comments from the past 20 years, filter out 6,454 repeatedly recommended videos, and turn them into a viewing entry point.

Link: https://hackertube.org/

Original post: https://x.com/machinelearnflx/status/2116987936210088264

A few details worth looking at:

• No login required. State stays only in the local browser session.
• Thumbnails do not use the polished covers uploaded by creators; instead, they take a representative frame from the middle of the video. This choice is crude, but useful—at least what you see is an image that actually exists in the video.
• The comment analysis results are packaged into a zip file. You can point Claude at it and have it make another version tailored to your taste.
• There is also HackerTube Longs, dedicated to long videos.

The workflow is not complicated either. He used the Mac desktop version of Claude Code, with Opus 5.5 as default, Medium effort. Most of the time he just talked to Claude and let Claude do the work. There were only three manual parts: creating an API key in the Google Cloud UI to query YouTube metadata, spinning up a VPS, and buying a domain. Once the machine was up, he gave the IP to Claude and had Claude update, harden, deploy, and optimize the code.

There is something worth noting here. Previously, a project like this—from data cleaning to frontend to deployment—would have been enough for several people to grind on for weeks. Now it is more like a configuration task. The barrier has not disappeared; it has just moved from writing code to knowing which services to connect, what data to hand to the model, and where human confirmation is needed. Karpathy's line "many awesome projects... can now be few-shotted for breakfast" is not exaggerated, but don't rush to treat a weekend project as a product. Being able to run and being able to maintain long-term are separated by monitoring, cost, permissions, and exception handling.

HackerTube itself is not complicated either. What it does is more like adding a human-comment pre-filter to the recommendation system. YouTube's recommendations optimize for watch time and clicks, while HN comments leave a different set of signals: which videos are mentioned repeatedly, and which content is still discussed years later. It is not a question of which is higher or lower; but looking at them together, at least you feel a little less led around by the algorithm.

If you have a development background, you can go straight to its data package. If you don't, you can also treat it as a login-free, lower-distraction YouTube entry point. Use it for a few days and see whether the recommendation quality really suits your taste.