Karpathy Built a Video Site in Two Hours, Filtering Out YouTube's Recommendation Algorithm; the Homepage Is Full of Long Videos That Few People Watch

Over the weekend, Karpathy built a website called HackerTube.

The motivation was very specific: he was fed up with YouTube's recommendations. The approach was just as specific: he fed 43 million comments accumulated over twenty years on Hacker News to Claude, filtered for videos that were repeatedly mentioned, and ended up with 6,454. The whole process took about two hours, using Opus 5.5.

The site is hosted at hackertube.org. The interface is quite clean, and a few design choices are telling:

No login required; all state is stored locally in the browser.

Thumbnails are taken from a frame in the middle of the video, not the face the uploader carefully A/B-tested.

The underlying analysis data is packaged into a zip file and made public; you can throw it at Claude and have it remake a version to your own taste.

There's also HackerTube Longs, focused specifically on long videos.

Someone asked what the first item recommended by the algorithm was, and a reply included a screenshot: a video explaining how a USB keyboard works. Made by Ben Eater, 34 minutes, 3.8 million views. The person asking said this was exactly what they wanted to watch. The list also includes Freya Holmér's 1-hour-13-minute talk on spline continuity and Gil Tene's famous talk on latency measurement, 42 minutes.

Opening the categories page makes the taste behind this filtering even clearer:

Satire & parody: 362; Functional programming: 225; Programming language design: 193; Media, journalism & propaganda: 334; Tech pioneers & interviews: 369; History of computing: 231. It reads like a sample survey of the tech world's self-perception.

How he describes the making process is more worth reading than the result. Using the desktop version of Claude Code, Opus 5.5 at medium effort, he talked only to Claude the entire time. There were only three parts where he actually did something by hand: go to the Google Cloud console to apply for an API key for looking up YouTube metadata, spin up a VPS, and buy a domain. After the server was up, he gave the IP to Claude and had it update, harden, deploy, optimize—and it did all the rest.

In his words, many projects that used to take weeks or even months can now be finished over breakfast.

But the comments section wasn't entirely convinced. Someone said bluntly: without short videos, why would I go out of my way to a separate website to watch long videos? That's not balanced.

This rebuttal hits the mark. HackerTube's homepage is indeed packed with content that often runs half an hour or more, while most people's patience for scrolling videos doesn't last three minutes. Filtering solves inaccurate recommendations; it doesn't solve no one wanting to watch. Those are two very different things.

There's an even subtler problem. Being repeatedly recommended across 43 million comments doesn't mean it's good; it only means many people mentioned it over these twenty years. This approach naturally makes the site's content skew old, classic, and already-famous. For people looking for new things, this may just be another form of recommendation failure.

The most interesting comment says: I really feel sorry for that agent that has to read through 43 million HN comments; no one should have to endure that much “well, actually.”