Tool Suggestions
HackerTube: Andrej Karpathy's AI-Built YouTube Alternative With 6,454 Videos Picked From Hacker News Comments
Andrej Karpathy has built an alternative to YouTube recommendations, and he built it in a weekend with Claude. HackerTube, shared on November 1 and live at hackertube.org, surfaces 6,454 videos chosen from 43 million Hacker News comments — a human-curated video library, assembled by AI, with no engagement algorithm in sight.
The project started with frustration. Karpathy has said he was annoyed with YouTube's recommendations, which optimise for watch time and clicks rather than quality. Hacker News comments offered a different signal: when someone in a technical discussion recommends a video, that recommendation carries context and taste that an algorithm cannot fake.
The pipeline shows how the library was built. Claude Opus 5.5 scanned about 43 million Hacker News comments from 2007 to 2026. Regex and URL extraction plus spam filtering narrowed that flood to roughly 39,834 genuine recommendation comments. A Claude Sonnet screening step, run inside a Claude Code workflow, then turned those comments into 6,454 videos across 150 topics — from programming and physics to history and design.
The design choices are the product. There is no login: your state lives in your local browser session, not in an account. Thumbnails are a representative frame taken from the middle of each video, not the creator-uploaded thumbnail — so there is no clickbait image competing for your attention. A HackerTube Longs section collects the longer viewing. And the underlying comment-analysis data is downloadable as a single ZIP, so anyone can point their own Claude at it and build a variant.
The build story is almost as interesting as the site. Karpathy says the site took about two hours with Claude Code, using the Mac desktop app with Opus 5.5 at Medium effort. His only manual tasks were creating a Google Cloud API key for YouTube metadata, booting a virtual private server and buying the domain. Claude then updated, secured, hardened, deployed and optimised the code itself.
The token figures, Karpathy's own, give a sense of the scale. The pipeline used 504 million input tokens, about 90% of them cached, and 7.6 million output tokens. Building and deploying the site itself took about 1 million Opus 5.5 tokens, according to coverage of his post. Caching is doing heavy lifting here: without it, scanning 43 million comments would cost far more.
Two caveats keep this honest. The two-hour figure is Karpathy's estimate for a personal project by one of the world's most capable AI engineers working with frontier tools — it is not a typical development time, and nobody should budget a product on it. And HackerTube is a curation layer, not a new video platform: the videos themselves are still hosted on YouTube. If YouTube changes access, HackerTube feels it.
Why it matters goes beyond one weekend project. First, it shows community signal beating engagement optimisation for discovery: recommendations with a reason attached, from people whose taste you can judge. Second, it shows what one person plus a coding agent can ship: a real, usable product in an evening. Third, the open data invites variants — the same pipeline could run over any community's recommendations.
Karpathy is an unusual builder for this kind of project, which is part of why it travelled so far. The former Tesla AI director and OpenAI co-founder has a habit of publishing small, well-built tools — his nanoGPT repository taught a generation of engineers how language models work — and HackerTube follows the same pattern: a clear idea, built in public, with the workings left open for others to reuse. The downloadable comment data is the most Karpathy part of the whole project: the site is useful, but the dataset is what lets a hundred variants bloom.
Browsing it feels deliberately calm. Topics run from programming languages and machine learning to history, physics and design, and the Longs section is for the multi-hour talks that recommendation feeds rarely surface. There is no autoplay funnel and no next-video algorithm optimising for retention — the next thing you watch is something you chose, from a list somebody recommended for a reason you can read in the original comment thread.
Trying it costs nothing. Open hackertube.org, browse by topic, and watch something a Hacker News commenter thought was worth your time. No account, no feed tuned to keep you scrolling — just 6,454 videos that other curious people recommended.