Karpathy's HackerTube: YouTube Curated by Hacker News

Andrej Karpathy, the AI researcher behind some of the most watched machine learning lectures on the internet, got tired of YouTube's recommendation algorithm. On a Friday night at 11 PM, he opened YouTube looking for something to watch and found what he called "an enshittified wall of slop" [1]. Exaggerated thumbnails, the same tired topics, idiocracy-grade content. So he did what any engineer with access to a powerful AI model would do: he built his own YouTube clone in two hours.

The result is HackerTube, live at hackertube.org. It is a YouTube frontend where every video has been recommended by someone on Hacker News. Karpathy gave Claude Opus 5.5 the task, and the model did the rest: it downloaded all HN comments, filtered them for YouTube links, ran Sonnet over the results to remove spam, and narrowed everything down to roughly 40'000 comments pointing to 6'454 distinct videos across 150 topics, spanning 2007 to 2026 [2].

What makes it different

HackerTube is not just a list of links. It is a full YouTube-like experience with some genuinely thoughtful design choices:

• Quality over clicks. Every video on the platform was recommended by at least one person on Hacker News. No engagement metrics, no watch-time optimization, no algorithmic rabbit holes.
• 50% thumbnails. Instead of the uploader's clickbait thumbnail, HackerTube grabs the midpoint frame from the video. No shocked faces with red arrows.
• No accounts needed. The "For You" recommendations use local browser session storage. No login, no tracking, no data leaving your machine.
• Keyboard shortcuts. Because Karpathy is the kind of person who uses keyboard shortcuts.
• Open data. The site's data is exportable so you can hand it to your own AI agent and build your own version.

The process

What is striking is how Karpathy describes his workflow. He does not do traditional engineering anymore. No reading code, no running commands in a terminal. He talks to Claude and has it do everything [3]. The only manual work for this project was: creating a Google Cloud API key for YouTube metadata, booting a VPS, and buying the domain name. Claude Code handled the rest: writing the code, deploying it, securing and hardening the server, and optimizing performance.

This is the same person who wrote the original PyTorch tutorial that most ML engineers learned from. If Karpathy is not writing code by hand anymore, that says something about where AI-assisted development has landed. Not for prototypes or toys, but for shipping a real product to a real VPS with SSL and production hardening, in a single Saturday morning session.

Why it matters

HackerTube sits at the intersection of two trends that have been accelerating this year. First, the enshittification of mainstream platforms. YouTube's recommendation algorithm is optimized for engagement, which means it optimizes for the median viewer, which means it serves slop to anyone with niche interests. Second, the collapse of the barrier between idea and shipped product. A working, deployed, hardened web app took two hours and one human conversation. The gap between "I have an idea" and "it is live on the internet" has never been smaller.

The HN community noticed. The post hit 1'384 points and 379 comments within hours [4]. The top comments are a mix of people trying it, people suggesting improvements, and people reflecting on what it means that a former Tesla AI director can ship a YouTube competitor from his couch on a Saturday morning.

The answer, judging by hackertube.org, is that he can. And he did.