HackerTube · AI Agent Build · October 2026
HackerTube: How Karpathy Built a YouTube Clone from Hacker News Recommendations in ~2 Hours
On Friday night, October 30, 2026, Andrej Karpathy opened YouTube and found what he called an “enshittified wall of slop”: exaggerated thumbnails, repeated topics, and recommendations tuned for the median internet user — not for him. On Saturday morning he gave Claude Opus 5.5 a different assignment: build a YouTube clone from whatever videos Hacker News commenters had recommended. The result, HackerTube, went live the same day at hackertube.org, topped the Hacker News front page, and — more usefully than the headline — documented a repeatable method for replacing an engagement feed with a feed curated by a community you trust. This page breaks down the pipeline, what makes it different from YouTube, and how to build your own version.
Last updated: November 1, 2026. Facts checked against Karpathy's own post (karpathy.bearblog.dev, October 31, 2026) and the Hacker News discussion it started; the “about two hours” figure is Karpathy's account of his own build, including deployment hardening, not a promised reproduction time.
At a Glance
| Product | HackerTube — a YouTube-style interface over HN-recommended videos (hackertube.org) |
|---|---|
| Built by / with | Andrej Karpathy, with Claude Opus 5.5 (and a Sonnet workflow for filtering), on October 31, 2026 |
| Corpus | Hacker News comments, 2007–2026, filtered to about 40,000 comments recommending a YouTube video |
| Output | 6,454 videos across 150 topics, in a YouTube-style UI |
| Accounts / state | No accounts; recommendations and state live in the local browser session |
| Reusability | The site's data is exportable, so you can hand it to your own agent and build your own version |
| Key limitation | A one-off snapshot as described — no update cadence is stated in the original post |
Why This Exists: A Feed Tuned for Someone Else
Karpathy's complaint is specific, and that specificity is the whole idea. YouTube's recommendations are optimized for the median user and for engagement — which is why the homepage converges on exaggerated thumbnails and the same tired topics. Hacker News recommendations are optimized for nothing except that a technically literate commenter thought a video was worth sharing. HackerTube is what you get when you swap the optimization target: from “what keeps the most people watching” to “what did people I respect bother to recommend.”
That swap is the transferable insight. Every feed you dislike is a curation source you can replace. The question is never “can an agent build a feed UI” — of course it can. The question is whether you have a corpus whose taste you trust more than the algorithm's, and a quality filter that keeps it clean.
The 3-Step Pipeline
1. Mine a public corpus. Opus downloaded all Hacker News comments — nearly two decades of them, 2007 to 2026. The corpus matters more than the model here: HN comments are public, timestamped, threaded, and already carry a rough quality signal (someone chose to post a link in a technical discussion). A private Slack history or a paywalled forum would not have worked the same way.
2. Filter with a cheaper model in a workflow. Raw comments are mostly not recommendations. A Sonnet workflow removed spam and low-grade content and kept only comments that actually recommend a YouTube video — about 40,000 of them. This is the step most replications will underestimate: the pipeline is a classifier with a taste threshold, and where you set that threshold decides whether your feed feels curated or feels like a dump. Deduplication matters too — the same popular video recommended by many people should count as one strong signal, not many entries.
3. Wrap it in a familiar UI and ship it properly. The 6,454 surviving videos went into a YouTube-style interface with 150 topics, custom ForYou recommendations, keyboard shortcuts, and no accounts — state stays in the local browser session. Then came the unglamorous part Karpathy lists explicitly: get a domain, push to a VPS, secure, harden, and optimize. A demo that only runs on localhost is not the product; the hardening is part of the two hours, and it is the part first-time builders most often skip.
Three Design Choices Worth Stealing
The 50% thumbnail. HackerTube ignores the uploader's thumbnail and shows a frame from the middle of each video instead. Thumbnails are advertising; a mid-video frame is evidence. If you build any video feed, this one change removes an entire class of clickbait at zero cost.
No accounts, local state. Recommendations work without sign-in because state lives in the browser session. That is a privacy win and a scope win: no auth system, no user database, no password resets — the build stays a weekend project instead of becoming a SaaS. The trade-off is real too: no sync across devices, and clearing the browser resets your feed.
Exportable data. HackerTube ships an export of its data so anyone can feed it to their own agent and build their own version. This is the opposite of a walled garden, and it is why the project travels: the artifact is not just a site, it is a starter dataset for the next hundred niche feeds.
Build Your Own Version: A Practical Recipe
Pick a corpus with taste you can defend. Good candidates: your team's shared links over a year, a subreddit or forum whose moderation you trust, a newsletter archive, or HackerTube's own export as a starting point. Bad candidates: raw trending feeds (that just rebuilds the algorithm you were escaping) and any corpus you cannot legally or ethically reuse.
Set the quality bar before you automate. Write down what counts as a recommendation versus a mention, what counts as spam, and how you handle duplicates — then encode that as the filter prompt. Spot-check at least 50 filtered items by hand before you trust the pipeline; a filter you never audited is a feed you do not control.
Ship the boring parts. Domain, VPS, HTTPS, basic hardening, and a refresh plan. HackerTube's post states no update cadence, so its library is a snapshot; decide yours up front — re-run the mining pipeline weekly, or label the feed as a snapshot with its coverage dates. A feed that silently goes stale is worse than a feed that honestly says when it stops.
The Honest Caveats
First, “about two hours” is Karpathy's time for his build, with his judgment calls and his tolerance for rough edges — not a benchmark. Your corpus will have its own spam patterns, its own duplicates, and its own deployment surprises. Second, 6,454 videos is a library, not a universe: it inherits Hacker News taste, including its blind spots, and it will feel thin outside technical topics. Third, recommendations without accounts are private but forgetful; if cross-device sync matters to your users, the no-account design is a cost, not just a feature.
Official sources: HackerTube — Karpathy's original post (October 31, 2026), HackerTube (hackertube.org), and the Hacker News discussion.
FAQ
What is HackerTube?
HackerTube (hackertube.org) is a YouTube-style interface built by Andrej Karpathy on October 31, 2026. Instead of YouTube's recommendation algorithm, it only shows videos that Hacker News commenters recommended between 2007 and 2026: 6,454 videos, drawn from about 40,000 comments, organized into 150 topics. There are no accounts; recommendations and state live in the local browser session.
How was HackerTube built?
Karpathy describes a three-step pipeline in his own post: download all Hacker News comments, filter them down with a Sonnet workflow that removes spam and low-grade content until only about 40,000 comments recommending a YouTube video remain, then wrap the resulting 6,454 videos in a YouTube-style UI, register a domain, and deploy to a VPS with security hardening and optimization. He reports the whole thing took about two hours — his account of his own build, not a promised reproduction time.
Why does HackerTube use the 50% thumbnail?
Instead of the uploader-chosen thumbnail — which on YouTube is often an exaggerated, clickbait-style image — HackerTube shows a frame taken from the middle (50%) of each video. The point is to judge a video by what it actually contains rather than by its packaging, which is the same reason the site exists at all: recommendations chosen by a trusted community instead of an engagement algorithm.
Can I build my own version of HackerTube?
Yes, and Karpathy explicitly supports it: HackerTube offers an export of the site's data so you can hand it to your own agent and build your own version. The reusable method is the pipeline, not the video list — pick a corpus you trust (your team's shared links, a community's bookmarks, a niche forum's recommendations), filter it with a human-set quality bar, and wrap it in a simple feed UI. Do not just clone the video library and call it a product; the value is the curation source.
Is HackerTube a live, updating YouTube alternative?
No — at least not as described in Karpathy's post. The post describes a one-off build from comments up to 2026 and states no update cadence, so treat the library as a snapshot, not a live feed. Hacker News commenters have also raised freshness questions about the snapshot; Karpathy's post itself does not promise ongoing updates. If you build your own version, plan a refresh step (re-run the mining pipeline on a schedule) or say clearly that your feed is a snapshot.
What does HackerTube not prove about AI agents?
It proves that a strong agent can compress a weekend-project pipeline — data mining, filtering, UI, deployment, hardening — into about two hours in skilled hands. It does not prove that any agent reproduces that result on your first try, on your corpus, or to production quality. Deployment hardening, spam filtering quality, and recommendation taste are the parts most likely to take longer than the headline time suggests.
Turn ideas into AI video with HAILUOAI
Explore HAILUOAI's text-to-video and image-to-video tools to create your own AI-generated videos.
HAILUOAI