Karpathy's HackerTube rebuilds the YouTube feed from Hacker News recommendations

Andrej Karpathy has launched HackerTube, a video site that imitates YouTube's interface but stocks it exclusively with videos that Hacker News commenters have recommended over the past two decades. In a blog post dated 31 October that surfaced on Hacker News's front page, Karpathy says he built and deployed the site, hackertube.org, in roughly two hours with the help of the Opus 5.5 model.

The motivation, he writes, came the previous Friday night around 11pm, when he opened YouTube hoping to watch something and instead found exaggerated thumbnails, repetitive suggestions on the same tired topics, and what he bluntly characterises as low-grade content. His takeaway was that the recommendation algorithm may serve the median internet user reasonably well, but that it clearly does not work for him. So the next morning he handed Opus 5.5 a single task: build a YouTube clone out of whatever videos have been recommended on Hacker News.

How the catalogue was assembled

According to Karpathy, the workflow downloaded all Hacker News comments, filtered them down, and then ran Sonnet over them in an automated pipeline that stripped out spam and weak entries. That produced roughly 40,000 comments containing a YouTube video recommendation, which in turn resolve to 6,454 distinct videos spanning 2007 to 2026 across about 150 topics. The collection was wrapped in a YouTube-style interface, given a domain, and pushed to a virtual private server, which he says was secured, hardened and optimised as part of the same session.

Deliberate product choices

Karpathy lists a set of features aimed squarely at the irritations that triggered the project:

• Content quality is enforced at the source, since only videos that someone bothered to recommend on Hacker News make it into the catalogue.
• Thumbnails are taken from the 50% mark of each video rather than the uploader's chosen image, sidestepping the exaggerated-thumbnail problem.
• A For You feed generates personalised recommendations without user accounts, keeping all state in the local browser session.
• Keyboard shortcuts are built in.
• The site's full dataset is exportable, so it can be handed to an AI agent to build another version of the service with minimal effort.

He notes that the entire build takes about two hours now, and that the result has restored his ability to simply watch something worthwhile.

Why it matters

HackerTube's interesting move is not video technology but the substitution of the recommendation signal. YouTube's core product is arguably its feed, and Karpathy has replaced a watch-time-optimised machine learning system with curation accumulated through nearly two decades of Hacker News discussion. The bet is that a community's collective judgement, cleaned up by language models, produces a better feed for a technical audience than engagement metrics do.

The project is also a compact demonstration of how far AI-assisted development has moved. One person went from a Friday-night complaint to a deployed, hardened production site by Saturday, using a multi-model pipeline in which one model orchestrated another for data cleaning. The export feature extends the point: when the dataset ships alongside the site and agents can consume it directly, the cost of forking the whole idea approaches zero.

The limits are equally instructive. The videos themselves are still YouTube videos, so HackerTube is a new front door to a catalogue owned by the very platform whose feed it rejects, rather than independent infrastructure. And the curation inherits the tastes of Hacker News itself: the site works precisely for people whose interests overlap with that community, which is exactly Karpathy's argument — a feed tuned for the median user leaves everyone else underserved.