Karpathy finds YouTube recommendations so poor that he had Claude build HackerTube in two hours.

Beating AI News Flash: OpenAI co-founder and current Anthropic pretraining team member Andrej Karpathy, fed up with YouTube's monotonous recommendations and exaggerated thumbnails, simply had Claude build him a video recommendation site called HackerTube—taking only about two hours from start to finish.

HackerTube catalogs 6,454 YouTube videos across 150 topics. The videos all come from nearly 20 years of recommendation records in the Hacker News tech community, rather than YouTube's own recommendation algorithm. The site requires no login—just open it and browse—and personal browsing state is saved locally. Video thumbnails no longer use the exaggerated thumbnails uploaded by creators; instead, a frame from the middle of each video is uniformly captured.

To find these videos, Claude processed approximately 43 million historical comments from Hacker News. It first narrowed the scope through link matching, regular expressions, and other methods, then called Sonnet to identify comments that genuinely recommended videos and filter out spam, ultimately compiling this batch of videos from about 40,000 valid comments.

The entire project was completed by Claude Code, with Opus 5.5 handling the data processing pipeline, website development, and deployment, and Sonnet handling large-scale comment classification. Karpathy said he did not hand-write code or operate a terminal—he only manually applied for a YouTube API Key, purchased a domain, launched a server, and configured DNS. The remaining work, including server security hardening and performance optimization, was left to Claude.

Karpathy disclosed that the project consumed a cumulative 504 million input tokens (about 90% cache hits) and 7.6 million output tokens. More than 90% of the usage went to Sonnet's comment classification, while website development and deployment themselves consumed about 1 million Opus 5.5 tokens.

HackerTube is now live, and the compiled recommendation data is also available for download. Others can hand the data to their own Agent to build video sites tailored to their personal interests.