Skip to content
by InfinyAI ES

IA en un minutoNewsAgents

Agents

Andrej Karpathy launches HackerTube, a YouTube alternative curated by Claude from Hacker News comments

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

Karpathy has put together HackerTube, a video site in YouTube's style that collects 6,454 videos recommended by Hacker News users, from 2007 to 2026, over 150 topics. According to him, Claude Opus 5.5 handled the technical work, with Sonnet weeding out the spam, and the whole thing took him roughly two hours. His only manual chores were getting a key to query YouTube, setting up the server and buying the domain.

On Saturday, October 31, Andrej Karpathy introduced HackerTube on his blog: a site that looks like YouTube but holds only videos that users of Hacker News (the tech forum) have recommended in their comments. He says the idea came on Friday night, when he opened YouTube and ran into oversized thumbnails and the same tired suggestions; the next morning he asked Claude Opus 5.5 to make his own YouTube clone out of whatever gets recommended on Hacker News. RuntimeWire picked up his account the same day.

The catalog holds 6,454 YouTube videos recommended between 2007 and 2026, sorted into 150 topics. Karpathy says Opus pulled down every Hacker News comment, filtered them and ran Sonnet over them inside a Workflow (the Claude Code feature for chaining tasks) to discard spam, ending up with about 40,000 comments that recommended a video. HackerTube's About page specifies that the figure was 39,834 comments and that a video only gets in if at least three different people linked it with some commentary, in at least three different months; roughly 7,000 videos were left, and Claude then read each comment to tell whether it recommended the video, cited it as a source, used it as a joke or criticized it. Videos are ranked by the number of separate months in which they were recommended, so mentions repeated over the years weigh more than a one-off spike.

Karpathy says the whole process took roughly two hours. He works with the Claude Code desktop app for Mac and with Opus 5.5 on medium effort, and states that he no longer reads or writes code or runs commands himself: he just talks to Claude. His manual work was limited to getting a key in Google Cloud to look up the details of the YouTube videos, starting up a server (a VPS) and buying the domain name. He then gave Claude the server's IP address and asked it to update, deploy, secure and optimize the code, and Claude did.

On the site, the home page mixes Hacker News classics, videos that are liked there but have few views, recent favorites and topics not yet explored. According to Karpathy, the "For You" recommendations work without user accounts because the state is kept in the browser's local session, and instead of the thumbnail each author uploads, the site uses a different image. Each video page shows the actual Hacker News comments, linked to their threads. There are keyboard shortcuts, and the site's data can be exported as a compressed archive with a guide (DATA.md) written so that an AI agent can build another version.

Why it matters · analysis and opinion

What changes compared with video recommendations is the criterion: here watch time does not rule, but several people have pointed to the same video again and again over the years. That the raw material is a public archive of comments explains why a single person could build it with an agent. The case shows how far an agentic workflow can go from end to end: data filtering, interface, deployment and server hardening, all requested in conversation. It should be read with caution, because the account comes from the author himself and the sources include no independent measurement of how good the recommendations are or how many people will use it. For anyone who wants something similar with data from their own sector, the example suggests that a curated selection can be put together with little manual work, as long as someone checks what the agent publishes.

Source: Andrej Karpathy · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.