# Karpathy Launches HackerTube: YouTube Recommendations Built With Claude Code

HackerTube turns years of Hacker News recommendations into a searchable video library, with a reusable dataset behind the interface.

By Paras Madan | 2026-11-01


Andrej Karpathy introduced HackerTube on October 31, an independent site that uses AI to organize YouTube videos recommended on Hacker News. It launches with 6,454 videos across 150 topics. The useful change is how it finds recommendations: repeated endorsements from a community, rather than a fresh feed of uploads.

Karpathy says he built it in roughly two hours using Claude Opus 5.5. That is his account of one project, not a benchmark for how quickly anyone can build and operate a production service. The more interesting part is the work behind the video cards.

## How HackerTube chooses videos

According to the [HackerTube project documentation](https://hackertube.org/about), the pipeline scanned a public archive of Hacker News comments from 2007 through 2026 for YouTube links. Its initial filter required a video to appear with substantive commentary from at least three different people across at least three different months.

Claude classified whether comments actually endorsed the videos. Treating every mention as a recommendation would have produced a very different library.

Ranking favors recommendations that recur across months rather than one burst of attention.

This gives the product a clear editorial preference. It favors material with a history of being useful to Hacker News readers. That can help someone find an older technical talk, but it also makes the site a poor substitute for a complete view of what people are publishing now.

## What Claude built, and what people can use

In his [launch post](https://karpathy.bearblog.dev/hacker-tube/), Karpathy describes Opus coordinating the data work and the website, with Sonnet processing comments. The result includes topic browsing, personalized recommendations without accounts and thumbnails taken from the middle of videos instead of their uploaded promotional images.

The site identifies Claude Code as the development environment and links video pages back to the original comments.

Those links matter because the AI-generated explanation is only an interpretation. Readers can check whether the underlying discussion supports the recommendation instead of having to trust the card's summary.

## The reusable dataset may matter more than the interface

HackerTube offers downloadable video records, labeled comments, topics and search embeddings, with documentation for agents. The release requires attribution and excludes some YouTube metadata. Check its terms before reusing it.

That makes a different interface a plausible next step. A developer could use the same curated material for a reading-group queue or a narrow collection of systems talks, without rebuilding the entire discovery pipeline.

The constraint would still be the source material. A different interface cannot remove the community's selection bias, and AI classification can misunderstand a joke or an ambiguous endorsement. An archive also needs a refresh process if it is supposed to remain useful for new material.

## What this says about small AI-built products

HackerTube is a concrete example of a familiar dataset becoming a useful product through classification, organization and an interface. The original human judgments were already there, scattered through discussions. AI made them easier to retrieve and use together.

The lesson for builders is to look for valuable decisions people have already made but cannot easily find again. A focused product around those decisions can be more useful than another general-purpose chat box.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product decisions and launches into clear stories that developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min)