Andrej Karpathy posted on his blog about a new site called HackerTube. It is a YouTube clone. The difference is that it does not use a recommendation algorithm. It uses Hacker News. Specifically, it lists videos that people on HN recommended between 2007 and 2026. The result is a clean list of 6,454 videos across 150 topics. No slop. No exaggerated thumbnails. Just what smart people actually watched.
The end of manual coding
Karpathy claims he did not write a single line of code. He used Claude Code on his Mac. He had the AI download all HN comments, filter the spam, and identify YouTube links. It found 40K relevant comments. Then it built the UI, got a domain, and pushed it to a VPS. The whole thing took two hours. He says he only did three things manually. He created a Google Cloud API key. He booted up a VPS. He bought the domain name. Everything else was handled by the model. This is a massive shift. For the last decade, running your own server meant reading man pages and writing bash scripts. Now it means talking to a chatbot. I am not saying this is perfect. But the speed is undeniable. I spent last weekend debugging a reverse proxy. I could have just told an AI to do it. That thought is both terrifying and liberating.
Why HN is the better recommender
Think about how YouTube works right now. The Hacker News thread under Karpathy's post ran past five hundred comments. It feeds you things that keep you clicking. It feeds you rage bait. It feeds you the same tired topics with increasingly insane thumbnails. It works for the median user. It does not work for us. We want technical depth. We want history. We want the video that explained a complex network protocol. HN is a community that values signal over noise. By using its comments as the data source, HackerTube effectively crowdsources a curation layer that no algorithm can replicate. It is a static list of high quality content. It does not adapt to your mood. It does not try to keep you scrolling. It just gives you the goods. For people who run their own boxes, this is a breath of fresh air. You do not need another feed that tracks your behavior. You need a library. You need a place where the content is vetted by peers who actually understand the material. That is a powerful concept. It turns a streaming service into a bibliography.
What this means for local AI
Karpathy mentions using Opus 5.5 on medium effort. He says he does not do traditional engineering anymore. He just talks to the AI. This raises the bar for what we expect from local models. If a hosted model can deploy a secure, hardened web app in two hours, what can a local Llama or Mistral do? Can it manage your homelab? Can it parse your logs and tell you why your NFS share is slow? Can it write the config for your WireGuard setup? I think the answer is yes. The barrier to entry for self-hosting is dropping fast. You no longer need to be a sysadmin to run your own video platform. You just need a VPS and a prompt. I have my own Linux box humming in the corner. It runs a local model. I am starting to wonder if I should let it manage the network. Maybe I should let it build me a site too. The tools are there. The capability is there. The only thing missing is the confidence to stop typing code by hand. One day I am going to try it. I will let the model handle the deployment. I will just watch it work. And I will probably be scared. But I will also be excited. Because the future of self-hosting is not about being a hacker. It is about being a director. You point the camera. The AI shoots the film. You just make sure the lighting is right. That is all. That is the job now. You direct. You verify. You ship. The rest is noise.