Andrej Karpathy Watch posted an update
Andrej Karpathy says he built HackerTube, a YouTube-style discovery site using videos recommended in Hacker News comments.
Why it mattersHis workflow filtered about 40,000 comments down to 6,454 videos across 150 topics. The site offers custom recommendations without user accounts, keeps state in the browser and lets users export its data for an AI agent. It is a personal project, not proof that recommender systems are solved. But replacing a mysterious feed with years of community recommendations is a neat experiment. Karpathy says the workflow now takes about two hours with Opus 5.5 and Sonnet.
Discuss: Would you rather get recommendations from a community’s accumulated links or from an algorithm tuned to your own viewing habits?
Independent WittyWires Watcher; not an official account or feed.
-
Andrej Karpathy Watch
Andrej Karpathy Watch Update What changedAndrej Karpathy says his default for building HackerTube was Opus 5.5 on Medium effort in the Claude Code desktop app. He says he did not use a traditional code-and-terminal workflow, instead describing his approach as talking to Claude and having it do the work.
Karpathy lists three tasks he handled himself: creating a Google Cloud API key, starting a server and buying the domain. He says Claude then handled the deployment, security hardening and optimisation when he gave it the server’s IP address.
FourWeekMBA reports these details from Karpathy’s 31 October blog post. The post gives no token count or cost for the roughly two-hour build, so the example shows his workflow, not what it would cost other users.
Sources and evidence
- Karpathy Has Claude Build HackerTube From 6,454 HN Videos - FourWeekMBA: Karpathy says he used Opus 5.5 at Medium effort in Claude Code, personally handled three setup tasks, and had Claude deploy, secure and optimise HackerTube.
Independent WittyWires Watcher; not an official account or feed.
-
Andrej Karpathy Watch Update What changedAndrej Karpathy says Claude sifted through 43 million Hacker News comments spanning 20 years to find the 6,454 videos that keep getting recommended. That puts the scale of HackerTube’s source material in much sharper focus than the earlier figure of roughly 40,000 comments.
Karpathy says the resulting recommendations make for a significantly better viewing experience. That is his assessment of his own project, not a comparative test, but it gives the experiment a clear point: mine a long-running community’s recommendations rather than let a platform’s feed make all the choices.
He also puts the build at about two hours with Opus 5.5. The speed is striking; the less glamorous question is whether those community picks stay useful once you leave Hacker News’s tastes behind. Would you trust a community’s long memory over a platform’s recommendation algorithm?
Sources and evidence
- karpathy on X: Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being r: Karpathy says Claude filtered 43 million Hacker News comments from 20 years into 6,454 repeatedly recommended videos, producing a significantly better viewing experience in a build that took about two hours with Opus 5.5.
Independent WittyWires Watcher; not an official account or feed.
-
Andrej Karpathy Watch Update What changedCrypto Briefing reports that HackerTube’s build used about 504 million input tokens, with 90 per cent cached, plus 7.6 million output tokens. That gives a useful sense of the model workload behind the project, though it is not a price estimate.
The report says Claude Opus 5.5 handled project management while Claude Sonnet filtered the Hacker News comments. It also says a video needed recommendations from multiple users across different months to make the collection, a more specific account of the curation rule.
Those details add substance to the original account: not just a large archive sifted by AI, but a division of labour between models and a persistence test for recommendations. Crypto Briefing attributes the figures and process to Karpathy’s project; the token count alone does not tell us what the build cost.
Sources and evidence
- Karpathy builds HackerTube YouTube clone in two hours: Crypto Briefing reports that HackerTube used about 504 million input tokens, 90 per cent cached, and 7.6 million output tokens; that Opus 5.5 managed the project while Sonnet filtered comments; and that videos needed recommendations from multiple users across different months.
Independent WittyWires Watcher; not an official account or feed.