HackerTube replaces YouTube thumbnails with frames from halfway through the video
An AI-built viewing catalog puts discovery back in the audience’s hands, using an exportable record of what people actually recommended.

The thumbnail has to come from the video. Imagine that. In his account of building HackerTube, Andrej Karpathy lists the halfway-frame feature. His own site identifies him as an AI educator who publishes videos himself. He has a reason to care about what happens between making a useful video and getting it in front of someone.
His coding agent filtered a comment dataset down to 40,000 video recommendations, then to 6,454 videos spanning 2007 through 2026 and 150 topics. Those are the maker’s figures for the catalog.
Karpathy describes arriving at the idea late on a Friday, tired of exaggerated images and repetitive recommendations. “This takes ~2 hours now,” he writes of the resulting project. By Saturday, he had asked his agent to assemble an alternative. The work included filtering comments, removing spam and using a second model to interpret recommendations. A pasted video link alone doesn’t explain whether someone endorsed it, mocked it or warned everybody away from it.
The development account includes a useful limit on the two-hour figure: Karpathy still obtained a Google API key for video metadata, started a server and bought a domain. That leaves some real-world errands between asking for a website and having one people can reach. The agent handled deployment tasks, according to his account; the time estimate is his report, rather than a repeatable benchmark.
The site keeps recommendation state in the browser session, requires no user account, and offers a data export. The halfway image can certainly be boring. It also makes a video creator’s exaggerated reaction face much harder to use as bait.
The Open Index dataset he links makes the raw material inspectable. Its records include item identifiers, authors, timestamps, comment text and parent relationships. The documentation shows how to query selected months or years without first treating the entire history as one enormous download. Keeping those relationships matters: a recommendation acquires meaning from the conversation around it.
The dataset also separates stories from comments and describes comments as the overwhelming majority of its contents. That makes the mining problem more interesting than copying a list of front-page links. Useful recommendations can live several replies deep, the online equivalent of somebody finally mentioning the good movie after everybody has finished arguing about the trailer.
A source catalog can support a very different viewing experience without hosting the videos itself. The selection still reflects the people who posted recommendations and the model that filtered them. Calling a feed community-sourced doesn’t make it universal. Someone looking for a cooking demonstration and someone looking for a compiler lecture can want completely different evenings, and both can be right.
The export makes that question something another builder can explore. A person could preserve the recommended links while changing the topics, presentation or way the next item is chosen. The attractive possibility is an audience keeping useful discovery data when it changes interfaces. The harder question is how that personal selection gets refreshed after the initial collection has been watched, skipped or quietly forgotten.