Karpathy filtered 43M HN comments into a YouTube app in 2 hours
The only work he did by hand was the three steps that live behind a browser login.
Andrej Karpathy built HackerTube, a YouTube alternative whose recommendations come from 43 million Hacker News comments spanning 20 years, filtered down to 6,454 videos that keep getting recommended. He says the whole thing took roughly two hours, using Claude Code in the Mac desktop app with Opus 5.5 on medium effort as his default.
I only talk to Claude and have it do everything
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 interesting part is the division of labour. Karpathy says he only talked to Claude and had it do everything, and lists exactly three pieces of manual work he did himself, all of them in a browser. He created a Google Cloud API key so the app could query YouTube for video metadata, he booted up a VPS box for serving, and he bought the domain name. After that he handed Claude the IP address and told it to update, secure, harden, deploy and optimize the code, and reports that it did the whole thing.
What the app actually does
No logins, with state kept in the local browser session. Thumbnails are a representative frame from the middle of the video rather than the uploaded image, which Karpathy describes as "super duper A/B test hyper-optimized". He also published the underlying comment analysis data as a single zip file and suggests pointing your own Claude at it to build your own version.
For working engineers
This is one person's weekend hack, not a benchmark, and Karpathy does not report how much of the two hours was his attention versus the agent grinding. What makes it worth reading is the shape of the remaining human work. Everything that was left was gated behind a browser login and a credit card, namely a cloud console, a VPS signup and a domain purchase. The code, the data pipeline over 43 million comments, and the server hardening were all delegated.
That lines up with a pattern showing up elsewhere in agent workflows, where the bottleneck shifts from writing code to the handful of steps an agent cannot authenticate its way into. Compare it with Anthropic's own report that it rewrote its Messages API in Rust without Rust developers on the team. The claim in both cases is about who did the typing, not about how the result holds up over months of maintenance, and nobody has published that part yet.
Karpathy's own framing is the broader one. Projects that would normally have taken weeks or months, he says, can now be few-shotted for breakfast, and he ends by asking what else people are not thinking of.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.