Karpathy builds a YouTube clone with 6454 videos from Hacker News recommendations in two hours

Andrej Karpathy, one of the best-known AI researchers, built HackerTube in two hours. It is a YouTube clone where videos are picked not by algorithms but by users of the Hacker News aggregator. The site's database holds 6454 videos sorted into 150 topics, all taken from comments in which community members have spent years telling each other what is worth watching. The project is meant as a response to the decline of recommendations on the original video platform, which, in the author's view, serves monotonous content with clickbait thumbnails. Karpathy did not write the code: the Claude Code agent did all the work.

Where the database of 6454 videos came from

The project is built on an archive of Hacker News comments from 2007 to 2026. It is published on Hugging Face in the open-index/hacker-news dataset. From those comments, the task was to extract links to videos that users had recommended as worth watching. According to the post's description, Anthropic models built the video database.

The Sonnet model filtered out spam and assembled 6454 videos across 150 topics. The result is a curated selection rather than an endless feed: behind every video is someone's recommendation in a community used to arguing about the quality of material. Video metadata was loaded through a Google API key that Karpathy created himself.

How the site works and how it differs from YouTube

HackerTube works without registration. The service saves personal recommendations in the browser's local session, so users do not need to create a profile. This is the opposite of a model where recommendations are built on accumulated data about user behavior.

The second decision concerns thumbnails. The site deliberately ignores images uploaded by authors and shows a frame from the video itself instead. That removes one of clickbait's main levers: a bright cover with an exaggerated facial expression and large text. Viewers choose a video by its title and its actual content.

What the human actually did

The main part of the story is not the site itself but how it was made. Karpathy writes that he no longer writes code by hand, and calls the project a great example of how development has changed. He ran the whole project through Claude Code, where, by his account, Opus 5.5 completed the tasks. The agent wrote the interface, set up security and handled deployment.

The human's involvement came down to three actions: registering a domain, spinning up a VPS and creating a Google API key. All three steps require accounts, payment and access rights, which for now are issued to humans rather than agents. For working developers, this is a concrete illustration: the line between "came up with it" and "built it" has shifted so far that a finished service fits into a couple of hours.

Limitations and takeaway

The information comes from a retelling of Karpathy's post in a Telegram channel, not from independent verification. The article has no data on how the site handles load, how good the selection is in terms of quality, or how much the agent's work cost. It is also unclear how the project will update its database: the list of 6454 videos was built from an archive that ends in 2026.

Still, the example is telling. A small service with a non-trivial idea, curation through a community instead of an algorithm, was built by one person and an agent, and the person performed only administrative steps. For product managers, this is a signal: the cost of a prototype has fallen so far that a hypothesis like "what if we recommended videos based on community advice" can be tested in an evening. For developers, it is a reminder that value is shifting toward defining the task, choosing the data, and deciding what exactly the site should do and what it should not.