Photo: Tim Witzdam / Pexels
The former Tesla AI chief used Anthropic's models to mine nearly two decades of comments for 6,454 community-vetted videos
Andrej Karpathy has launched HackerTube, a video site that skips the algorithm and asks a simpler question: what did Hacker News commenters actually recommend?
The answer, after AI processing, is 6,454 videos sorted into 150 topics. It’s a YouTube alternative assembled from the collective memory of one of the internet’s most opinionated comment sections.
How HackerTube was built #
Karpathy, the former head of AI at Tesla, launched the site on October 31, 2026. It lives at hackertube.org.
The concept is narrow by design. Every video on the platform had to be recommended somewhere in Hacker News comments, the long-running discussion board popular with engineers, founders, and people with strong feelings about text editors.
To build it, Karpathy used Anthropic’s AI models. Claude Opus 5.5 handled management of the project, while Claude Sonnet did the filtering work.
Those models chewed through nearly 40,000 Hacker News comments over a two-hour period. The comments span from 2007 to 2026, which covers almost two decades of people telling strangers what to watch.
The bar for inclusion was not a single enthusiastic post. A video needed recommendations from multiple users, and those recommendations had to come across different months.
The compute bill was substantial in raw terms. The project consumed around 504M input tokens, with 90% of them cached, plus 7.6M output tokens.
What’s on the site, and who did the work #
HackerTube is not just a list of links. The platform ships with a few features that signal a builder’s sensibility.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Thumbnails use a 50% video-frame treatment. The site also supports keyboard shortcuts, a small touch that its likely audience will appreciate.
Recommendations work without accounts. Instead of logging in and handing over a viewing history, users get suggestions that are stored in their own browser.
There is also a full data export option. Users can download the entire dataset as a zip file and customize it however they like.
Perhaps the most telling detail is how little of the build Karpathy did by hand. His personal involvement was limited to three tasks: obtaining a Google Cloud API key, spinning up a VPS, and buying the domain name.
Background: Karpathy and the AI-assisted build #
Karpathy is one of the more influential voices in artificial intelligence. When he ships a side project, a lot of developers pay attention to how it was made, not just what it does.
The source material itself is a natural fit for that kind of work. Hacker News threads are full of links buried inside long discussions. Pulling out the video recommendations by hand from almost two decades of comments would be tedious work. A model can do it in an afternoon.
What this means #
The most interesting thing about HackerTube may be its curation philosophy. The research describes its community-vetting approach as diverging from the mainstream approach to content curation.
Requiring multiple recommenders across different months filters out one-off hype and rewards staying power.
The collection reflects the tastes of one community, so its 150 topics will skew toward whatever Hacker News users care about.
The privacy design is a second takeaway. Browser-stored recommendations and no required account mean the site does not need to know who you are to suggest something. The zip export goes further, giving users the raw material to remix the collection on their own terms.
For developers, the bigger story is the workflow. A recognizable figure in AI built and shipped a working site with 6,454 curated entries while personally handling only an API key, a server, and a domain. The token figures offer a useful reference point too. Roughly 504M input tokens sounds enormous, but the 90% cache rate shows how much efficiency comes from reusing context.
For Anthropic, the project is an unpaid showcase. Splitting duties between Claude Opus 5.5 for management and Claude Sonnet for filtering illustrates a practical pattern: use the heavier model to coordinate and a lighter one for repetitive grunt work. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.