Karpathy Built a YouTube in 3 Minutes, and Some Commenters Called It "AI Slop"

Karpathy shipped another project over the weekend. He turned programmer Tristan Hume's YouTube tier list into TristanTube: 47K videos, with TV mode, built in one three-minute shot with Claude Code.

Tristan Hume's list is worth a look. In 2020 he wrote a version of his favorite YouTube channels and updated it in 2026, split into five tiers: S/A/B/C/D, with 5 in S, 14 in A, 12 in B, 33 in C, and 92 in D. The S tier includes Applied Science, which does electron microscopes and plasma sputtering; Clickspring, which makes clocks and restores the Antikythera mechanism; Stuff Made Here, which makes basketball backboards and automatic apple peelers; Captain Disillusion, who debunks visual-effects hoaxes; and Kiwami Japan, who makes kitchen knives from seawater and shells.

In his tweet, Karpathy said the list looked too good, so he had to turn it into TristanTube. The link is https://karpathy.ai/tristan-tube/ . A user screenshotted the interface, which looks like this:

TV mode is made for vegging out. Press 1 through 9 to jump to 10%–90% of the video, 0 to replay, N for next, P for previous, Space to pause, left/right arrows for ±10 seconds, M to mute, and Esc to exit. Karpathy said the prompt was roughly "Here's the link to Tristan's top channels; make a YouTube-style player with an About page, dark mode, and TV mode."

This isn't his first weekend project. The earlier HackerTube was even more hardcore: annoyed by YouTube recommendations, he had Claude filter 43M comments from 20 years of Hacker News and surface 6,454 repeatedly recommended videos. The site is https://hackertube.org/ ; no login is required, and state is stored in the local browser session. Thumbnails are representative frames from the middle of the video, not uploader-optimized covers from A/B tests. The underlying comment-analysis data is packaged as a zip so you can have your own Claude build a custom version from it. Karpathy said the project took about 2 hours with Opus 5.5.

He wrote about what he did by hand: created an API key in the Google Cloud UI to query YouTube metadata, spun up a VPS, and bought a domain. Once the VPS was up, he gave the IP to Claude and had it update, secure, harden, deploy, and optimize the code. Everything else was handed to Claude Code. His exact words were: "Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of."

The comments split into two camps. One user posted the first video recommended by the HackerTube algorithm and said it was exactly what they wanted to watch:

One user said, 'YouTube spent billions building a recommendation algorithm, and you beat it twice in one weekend.' That's an exaggeration: HackerTube and TristanTube have a much narrower recommendation scope, drawing from HN comments and Tristan's list rather than site-wide behavior prediction. But the effect is indeed direct: one user said that by the fifth or sixth video, a favorite creator they hadn't seen in a long time showed up.

Another user said flatly that this was just another kind of AI slop. Karpathy didn't respond to that. He replied to the user asking about the prompt, saying he didn't remember exactly, but it was roughly the sentence above.

Other people in the comments also posted projects they'd made. Robert Lukoszko used AI to make a 3blue1brown-style channel, where the AI researches superconductivity, makes animated videos, publishes them, and sends daily notifications. Raymond made a women's history video site for his sister using Crash Course content:

These projects share one thing: the model didn't generate content out of thin air. It organized things humans had already recommended, written, and filmed. HackerTube mines the archaeological layers of the HN comment section, while TristanTube moves Tristan's taste. The model does the scraping, filtering, deployment, and adding a usable interface.

One user asked whether indexing 47K videos hit the YouTube Data API quota, or whether he just pulled channel RSS. This question is more specific than "AI slop," but Karpathy didn't reply to that either.

As for whether a website built in one three-minute shot counts as AI slop, the comments never finished arguing. Karpathy had already moved on to replying to the next person.