Karpathy built HackerTube with Claude over a weekend: 6454 videos from 43 million comments

It is a familiar frustration: YouTube recommendations are optimized for watch time, not quality. You want "what smart people actually recommend," but that feed does not exist. Andrej Karpathy tackled the problem over a weekend. What makes this interesting is less the site itself than how little manual work it required.

What he actually built

HackerTube is a curated video catalog. Its data source is unusual: 20 years of Hacker News comments, around 43 million in total. People there regularly post things like "you have to watch this lecture." Claude, running Opus 5.5, filtered that archive down to 6454 videos that the community recommends again and again.

Think of it as voting with links. If the same talk keeps surfacing in discussions across different years, it is probably good. That is far more reliable than an algorithm serving up the next video to keep you from closing the tab.

The product choices are revealing, too. There is no login. Instead of optimized thumbnails, the site shows a single frame from the middle of each video, eliminating clickbait. State—what you have watched and saved for later—is stored locally in the browser, so the backend needs neither accounts nor a user database.

Immediately afterward, Karpathy built TristanTube: 47 000 videos drawn from a list of YouTube channels compiled by Tristan Hume. He described the project as "one-shotted with Claude Code in 3 minutes," meaning it was built from a single prompt in three minutes.

Where the real news lies

Karpathy says Claude Code did almost all the work, from data analysis to server configuration. He manually completed just three steps that the agent could not: obtaining a Google Cloud API key, renting a server and buying a domain. The agent handled everything else, including processing the enormous comment archive.

By his estimate, projects that once took weeks can now be completed in a few hours. That is one person's assessment, and he is highly experienced and has privileged access to tools, so it should not be treated as a universal benchmark. But the direction is clear: the cost of a data-driven side project has fallen to a weekend's worth of work.

Limitations

Several points deserve a clear-eyed assessment. First, the account comes from Karpathy himself on X; there has been no independent assessment of the selection's quality. The claim of a "significantly better experience than YouTube recommendations" is subjective. Second, the sample is biased: Hacker News caters to programmers, so the catalog will focus on technology, mathematics and science rather than everything under the sun. Third, the videos come from YouTube, making the platform dependent on third-party links, API quotas and terms of use. Fourth, Karpathy is not a typical user: he knows how to brief an agent and can spot mistakes quickly.

What to do with this

For developers: look at the data you have accumulated, whether comments, tickets, support chats or logs. The pattern of having an agent process a large archive and build a curated layer on top of it is now inexpensive to implement. Treat the first prototype as a weekend experiment rather than a quarterly project.

For product managers and founders: curation based on community signals is becoming a viable niche again. If audiences are frustrated with major platforms' recommendation algorithms, a focused selection with no login or tracking can be a product in its own right. Keep in mind that the bottleneck is shifting: code and data processing are getting cheaper, while a strong source of signals and good judgment about what counts as "good" are becoming more valuable.

Finally, pay attention to the boundary of automation. People are still needed where payments and access are involved: keys, a server, a domain. Those steps remain human responsibilities for now, and they are precisely what you should prepare in advance if you want to match this pace.