Karpathy Spent Two Hours Building a YouTube, Using 43 Million Hacker News Comments as Source Material

Andrej Karpathy (former head of AI at Tesla) posted two tweets last weekend. In the first, he said he had long been fed up with YouTube recommendations, so he had Claude go through 43 million Hacker News comments from the past twenty years, filter out the 6,454 videos that were repeatedly recommended, and turn them into a website.

Time: about two hours. Model: Claude Opus 5.5, medium effort setting—he didn’t even turn on High.

The site is called HackerTube, at hackertube.org.

Hacker News is a tech community under Y Combinator. Its comment section has argued just as relentlessly for twenty years as it has recommended things to itself for twenty years. Karpathy’s logic is simple: YouTube recommendations push what the platform wants you to watch, while HN commenters recommend what engineers think is worth watching. The latter has a sample size of 43 million comments—plenty.

Three details about the finished product are worth mentioning.

First, no login required. No account, no recommendation history, no cross-site tracking; watch history exists only in the local browser session.

Second, thumbnails. YouTube covers are the result of uploaders’ repeated A/B testing—one face can’t seem to have fewer than eight expressions. HackerTube simply grabs a frame from the middle of the video as the cover; whatever it looks like is what you get.

Third, all data is public. The analysis results are packaged as a zip file that you can download, feed to your own Claude, and have it remake a version to your taste. Karpathy’s exact words were: point your Claude at it and ask for your own version.

Across the whole project, he did only three things manually: created an API key in the Google Cloud console to query YouTube metadata, spun up a VPS, and bought a domain. After the VPS came up, he handed the IP to Claude and had it handle deployment, hardening, and code optimization—after that, there was nothing left for him to do.

In his second tweet, he wrote: 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. Roughly translated: things that used to take weeks or months can now be built in the time it takes to eat breakfast—what else are we not thinking of?

The first comment under the post: someone posted the first video the algorithm recommended to him and said it was exactly the kind of thing he wanted to watch.

Some people also got nostalgic for the old arxiv sanity. Others felt sorry for the agent that had read 43 million HN comments, saying it was probably traumatized beyond repair by 'well, actually.'

Someone posted an animation depicting a viewer in front of a YouTube recommendation page, with YOU WON'T BELIEVE written on the screen; five minutes later the person is already looking utterly shocked. By contrast, HackerTube’s page is as quiet as some tech blog from 2010.

One thing is worth pondering. YouTube recommendations optimize for watch time and click-through rate, and neither metric has any necessary connection to whether the content is useful. HN commenters are the exact opposite: when a video keeps getting posted, it’s usually because someone actually used it to solve a problem, or because it’s worth watching again and again. The former is a platform’s fight for attention; the latter is peers endorsing value.

Karpathy didn’t invent a new algorithm. He just switched data sources and then outsourced the filtering work to Claude. The truly scarce stuff may have been there all along, just buried in 43 million comments with no one bothering to fish it out.

By the way, this process can be replicated. Export the conversations from communities, forums, or mailing lists you care about, throw them at a model, and ask which things get mentioned repeatedly—what you get may be cleaner than any recommendation feed. Karpathy has already put the zip out there.