The project and the method
Andrej Karpathy, co-founder of OpenAI and now a researcher in the field of AI, built HackerTube, a curated video platform, in a weekend. As The Decoder reports, the work was based on Claude's Opus 5.5 model, used to sift through 43 million comments collected over twenty years of Hacker News and derive 6,454 videos that the community continues to recommend.
According to the source, Karpathy left almost the entire execution to Claude Code: from data analysis to server configuration. The manual steps were reportedly limited to three elements: a Google Cloud API key, a rented server, and the purchase of the domain. No costs, exact times, or technical details about the architecture or content selection are given.
Design choices and comparison with YouTube
The platform does not require login and stores state locally in the browser. Instead of optimized thumbnails, it shows a single frame taken from the middle of the video. These are elements that, together with the catalog built from comments, outline an approach different from that of traditional video platforms, more oriented toward curation than maximizing engagement.
Karpathy writes on X that the result offers a "significantly better" viewing experience than YouTube's recommendations. However, this is a statement by the author, reported by the source, and not an independent assessment: no metrics, comparative tests, or user behavior data are published that would make it possible to verify it. Even the comparison with YouTube's suggestions remains, for now, a subjective judgment.
The second experiment and the emerging picture
Immediately afterward, according to The Decoder, Karpathy built TristanTube, described as produced "in a single shot" with Claude Code in three minutes, with 47,000 videos derived from a list of YouTube channels provided by Tristan Hume. The sequence of the two projects suggests that the same tool can be reapplied to different catalogs by changing the data source, but the source does not clarify how much preparation or verification work was required in each case.
The news should be read above all as an indication of a change in the nature of personal projects: Karpathy argues that tasks that once required weeks can now be completed in a few hours, thanks to models that handle analysis and infrastructure. This is the author's thesis, not a measured result. Still to be verified are the quality of the video selection, the platform's scalability, possible copyright limits on aggregated content, and its actual superiority over established recommendation systems. The source also does not specify the exact dates on which the two projects were built.
