Andrej Karpathy builds HackerTube, an AI-sorted video site from Hacker News
By Anika SharmaAI & Innovations Editor, Peepals Tech
Representative image: subarasikiai / Pixabay
Andrej Karpathy launched HackerTube around October 31, 2026: 6,454 videos recommended on Hacker News, sorted into 150 topics with AI help.
What Happened
Andrej Karpathy, the artificial intelligence researcher known for his work at OpenAI and Tesla, has launched HackerTube, a website that lets people browse videos recommended by users of the technology forum Hacker News instead of relying on YouTube's own suggestions. The site went live around October 31, 2026 and holds 6,454 videos sorted into 150 topics. Karpathy said he built it in roughly two hours by directing AI models from Anthropic's Claude family to do most of the work, and that he was prompted by frustration with YouTube's homepage, which he called an "enshittified wall of slop". His name is trending in India as the project spreads among developers.
Key Facts
- HackerTube collects videos that Hacker News users have recommended in comments over the years: 6,454 videos across 150 topics.
- The catalogue was narrowed down from about 40,000 comments that mentioned videos, according to most reports.
- Karpathy says Claude Opus 5.5 handled the data work and Claude Sonnet screened comments for spam and labelled recommendations, inside a Claude Code workflow.
- He reported about 504 million input tokens, around 90% of them cached, and 7.6 million output tokens; the comment-labelling calls made up more than 90% of usage.
- His own manual steps were creating a Google Cloud key for YouTube data, starting a rented server and buying the web domain.
- The site gives recommendations without user accounts, keeping its memory of what a visitor likes in the browser.
- The two-hour build time is Karpathy's own estimate.
Why It Matters
HackerTube is a small project with a big message. Here Karpathy showed an AI system doing the bulk of a job that would normally take a developer days: reading tens of thousands of forum comments, pulling out video links, filtering spam, fetching video details and organising them into a browsable site. For students and young developers in India, the project is a practical example of "directing" AI models rather than writing every line of code.
It is also a pointed comment on recommendation algorithms. Karpathy built it because he was unhappy with what YouTube's homepage showed him. HackerTube takes a different approach: it trusts the judgement of a community of engineers and enthusiasts who recommended videos in discussion threads, and it does so without accounts or personal data stored on a server. That approach will not replace a platform of YouTube's size, but it shows how anyone with an archive of human recommendations can now build a curated alternative quickly.
Some caution is needed. The build time and token counts come from Karpathy himself and have not been independently measured, and a two-hour build by an experienced AI researcher does not tell us how long the same job would take an ordinary developer. Reports also differ on some details, such as the exact number of comments processed.
Peepals calculation: with 6,454 videos in 150 topics, each topic holds about 43 videos on average (6,454 ÷ 150 ≈ 43). If about 90% of the 504 million input tokens were cached, roughly 50 million were fresh input (504 × 0.10 ≈ 50), and the AI read about 66 tokens for every token it wrote (504 ÷ 7.6 ≈ 66). Our calculation from the reported figures.
| Item | Figure reported |
|---|---|
| Videos | 6,454 |
| Topics | 150 |
| Input tokens | about 504 million (about 90% cached) |
| Output tokens | 7.6 million |
| Build time | about two hours (Karpathy's estimate) |
Impact
Short-term: Developers are trying the site and studying how it was built, and the project adds to debate about how much of everyday software work AI agents can take over.
Long-term: If such quick builds become common, more communities may turn their own archives of recommendations into curated sites, offering alternatives to the algorithm-driven feeds of large platforms.
Who is affected: Developers and students learning to build with AI tools, Hacker News readers, viewers looking for alternatives to algorithmic feeds, and AI companies whose models power such projects.
Key Takeaway
Andrej Karpathy's HackerTube turns years of Hacker News video recommendations into a 6,454-video site, built in about two hours with AI doing most of the work, according to him.
Questions and Answers
What is Andrej Karpathy's HackerTube?
HackerTube is a website launched around October 31, 2026 that collects 6,454 videos recommended in Hacker News comments and sorts them into 150 topics, offering an alternative to YouTube's algorithmic suggestions.
How did Andrej Karpathy build HackerTube?
He says he directed Anthropic's Claude Opus 5.5 and Claude Sonnet models through a Claude Code workflow to process the comments, filter spam and organise the videos, doing only a few manual steps such as buying the domain.
How long did it take to build HackerTube?
About two hours, by Karpathy's own estimate. That figure has not been independently checked.
Do I need an account to use HackerTube?
No. The site gives recommendations without accounts and keeps a visitor's preferences in their own browser.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
By Anika Sharma, AI & Innovations EditorCovers AI & Innovation and more for Peepals Tech. All stories by Anika Sharma
CVE-2026-98432: critical Linux TCP flaw still unpatched on some major distros
Nothing OS 5.0 update brings Android 17 to Phone 4a, 3a series
Get the day's stories in your inbox
Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.
Anika Sharma is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.
