Andrej Karpathy, the former head of AI at Tesla, has built a working YouTube alternative in about two hours — and he says he barely wrote any of it himself. The site, HackerTube, is a catalogue of 6,454 videos drawn entirely from recommendations buried in Hacker News comments, and it was assembled almost entirely by Anthropic's Claude Opus 5.5.
Karpathy published the project on October 31, 2026, after a Friday night spent scrolling YouTube and growing frustrated with its recommendations. The next morning, he gave Claude a simple brief: build a YouTube clone based on whatever videos have been recommended on Hacker News. By his account, the model returned a finished site at hackertube.org.
How the two-hour build worked
The pipeline ran in stages. Claude Opus 5.5 downloaded the full archive of Hacker News comments, filtered them, and then ran Claude Sonnet over the results inside a Claude Code Workflow to strip out spam. That narrowed the archive to about 40,000 comments that recommend a YouTube video — 39,834, to be precise — spanning 2007 to 2026.
From there, the model extracted the video IDs, pulled metadata through YouTube's own API, and wrapped the result in a video-site interface organised into 150 topics. Karpathy says the build consumed roughly 504 million input tokens, 90 percent of them cached, plus 7.6 million output tokens, with about one million Opus tokens covering the build and deployment itself. He gave no dollar figure for the run.
His own hands-on contribution was limited to three tasks: creating a Google Cloud API key to query YouTube for metadata, booting up a virtual private server, and buying the domain name. Everything else — code, data processing, deployment, security hardening and optimisation — he delegated to the model.
What HackerTube actually is
HackerTube is deliberately narrow. Every video on it had to be recommended somewhere in Hacker News comments, the long-running discussion board popular with engineers and founders. There are no user accounts; recommendations live in the browser session, and the site offers keyboard shortcuts plus a data export so readers can hand the catalogue to their own agent and build a version of their own.
One design detail stands out: instead of each video's uploaded thumbnail, the site shows the frame from the video's midpoint, which Karpathy treats as a rough quality signal for what viewers actually watched. The result is less a YouTube competitor than a curated archive — a snapshot of what one technically minded community thought was worth watching over nearly two decades.
Why a two-hour site matters
Karpathy's post reads as much as a statement about his own workflow as about the site. He writes that he no longer does anything resembling traditional engineering — reading and writing code, running commands in a terminal — and instead talks to Claude and has it do everything. His stated default is Opus 5.5 on medium effort inside the Claude Code desktop app for Mac.
"This takes ~2 hours now. I can breathe again," he wrote, thanking Opus 5.5. The claim deserves its caveats: the figures are Karpathy's own, the two-hour estimate covers this particular project rather than software development in general, and a site built from a public comment archive is a far simpler task than most production software. Even so, the demo lands at a moment when coding agents are becoming the most closely watched product category in AI — a working, public, genuinely usable site, produced in a single morning by one person directing a model, is a vivid illustration of how far that shift has already gone.
