With Claude, Karpathy Builds a Custom YouTube in a Weekend
Andrej Karpathy built HackerTube in a weekend with Claude Opus 5.5. The project filters 43 million Hacker News comments and generates 6,454 recommended videos. A very concrete demonstration of what agentic AI is already accelerating.
Andrej Karpathy, co-founder of OpenAI and researcher at Anthropic, built HackerTube in a weekend, a curated video platform relying on Claude Opus 5.5. The site filters 43 million comments from 20 years of Hacker News and extracts 6,454 videos, then recommended on the platform. Karpathy states on X that this provides a “significantly better viewing experience” than YouTube recommendations, bringing an old promise back to the center of the debate: AI can now assemble a useful product almost in one stroke.
A minimalist site, but a fairly heavy mechanism
HackerTube does not aim to impress with its interface. No login, no big account system, no hunt for optimized thumbnails: the site displays a fixed image taken from the middle of the video and keeps state locally in the browser. This choice offers a simple-to-use service but above all very quick to set up when the background work is entrusted to AI.
According to Karpathy, Claude Code handled almost the entire chain, from data analysis to server configuration. The only manual steps mentioned are a Google Cloud API key, a rented server, and domain name purchase. The message is clear, and a bit embarrassing for devs who like to do everything by hand: projects that used to take weeks can now be completed in a few hours.
Two projects, one method: fit everything into a weekend
Following this, Karpathy says he also created TristanTube, “one-shotted with Claude Code in 3 minutes,” from a list of YouTube channels provided by Tristan Hume. This second version aggregates 47,000 videos. The sequence is interesting because it is not just an isolated prototype: it shows that the same tooling can digest very different corpora and quickly produce a usable application.
The use case remains modest, but the impact is broader. For product teams, independents, and startups, the boundary between idea and demonstrator is shifting fast. The question is no longer just whether a model writes code, but whether it can orchestrate entire tasks without continuous supervision. That is where the shift becomes economic, not just technical.
What HackerTube Really Shows
The project illustrates a clear trend: agentic AI is no longer limited to completing a line or suggesting a commit. It begins to link data sources, configure infrastructure, and deliver a finished product. However, the result still depends on very concrete elements, such as API access, hosting, and dataset quality. The magic still needs a server, which makes it much less mystical.
In HackerTube’s case, the interest goes beyond personal showmanship. The project tests another way to recommend content, based on a community archive rather than a classic advertising algorithm. If this logic spreads, it could interest media, specialized communities, or niche platforms seeking finer recommendations without necessarily expanding their technical teams.
Key points
- HackerTube was built in a weekend.
- Claude filtered 43 million comments.
- The site displays 6,454 recommended videos.
- TristanTube aggregates 47,000 videos in 3 minutes.
- Karpathy cites Claude Opus 5.5 and Claude Code.
By the numbers
- 43 million comments — from 20 years of Hacker News, according to Karpathy.
- 6,454 videos — selected for HackerTube, per the X post.
- 47,000 videos — included in TristanTube, according to Karpathy.
- 3 minutes — time announced for TristanTube with Claude Code.
- 1 weekend — claimed time to build HackerTube.