Andrej Karpathy, the AI researcher, opened YouTube late one night, disliked what it showed him, and asked Claude to build him a replacement. The result, HackerTube, holds 6,454 videos that people on the Hacker News discussion site recommended in comments between 2007 and 2026. It took about two hours, and his only hands-on jobs were a YouTube access key, a rented server and the web address.
The model judges, and plain code does the heavy lifting
Nobody has an AI read 43 million comments one by one. Simple code first picks out the comments that contain a video link and have enough text of their own. A smaller Claude model then reads those and removes spam, which left about 40,000 comments that truly recommend a video. Only then does the site-building start.
The rule that decides what gets in is the part to steal: a video stays only if at least three different people recommended it in at least three different months. One excited thread cannot get a video in. Years of different people returning to it can.
Karpathy then needed about three minutes to run the same pattern on another person's published list of favourite channels, producing a 47,000-video viewer. Once the pattern exists, the second site costs almost nothing.
It works on any pile of recommendations
The book copy shows what travels and what breaks. It used a cheap Google model, not Claude, and a different coding tool, so the recipe does not belong to one vendor. Its builder checked 83 recommendations against the original comments, and the authors were right every time. But 470 of the quotes the model pulled out were not word for word what the commenter wrote, so he blanked them. Let the model decide whether a comment is a recommendation, and send readers back to the original comment for the words.
For example, a school department head with ten years of staff emails, in which teachers pass around videos and worksheets that worked, could ask for every resource named by at least three teachers in at least three different terms. Code could cut perhaps 30,000 emails to the few thousand with links, a model reads those, and a new teacher would get a shelf of materials that colleagues kept using, not what was shared once in September. My estimate is an afternoon of work and a few dollars of model time. The step that needs a person is permission: staff wrote those emails to each other, so ask them before you mine them.
What this changes for feeds
A feed from a big platform can see everyone's clicks, which is a real advantage, but it ranks by attention. A group that keeps its own records can rank by something slower: what people were still recommending a year later. Building that list used to mean a developer and weeks. Now it is an evening and a few dollars.
So expect many small, specific recommendation lists: what plumbers keep recommending to each other, what a hospital's nurses use to learn a new system, what a neighbourhood group trusts. For sellers, being recommended again and again by real customers over months will matter more than a spike of attention, because that is the signal these lists will use.