OpenAI co-founder and former head of AI at Tesla Andrej Karpathy has unveiled HackerTube — a YouTube alternative whose code was almost entirely written by the Claude neural network. The developer himself spent about two hours on the project.
Karpathy openly admitted that YouTube's recommendations annoyed him. As a replacement, he chose recommendations from real people — members of Hacker News, a popular forum for programmers and startup founders. He tasked Claude with selecting the relevant videos. Using the same scheme, anyone can build their own version of the video platform — with content interesting only to them.
How the YouTube alternative works
Claude analyzed 43 million comments accumulated on Hacker News over 20 years and selected 6,454 videos. These are clips that forum members linked to again and again. Karpathy used the Opus 5.5 model for the work.
No registration is required on HackerTube — viewing history is stored in the browser. Instead of thumbnails, which creators pick for clicks, a frame from the middle of the video is shown. A separate HackerTube Longs section contains all the collected data in a single archive: it can be downloaded, passed to your own neural network, and used to make a similar site for yourself.
According to Karpathy, watching videos through HackerTube has become much more enjoyable.
How Karpathy built HackerTube
The developer worked in Claude Code — an application in which the neural network writes and runs code itself. He merely described what needed to be done. He performed three steps himself:
- obtained a key in Google Cloud so the program could request information about videos from YouTube;
- rented a server;
- bought a domain.
After that, he sent the neural network the server address and asked it to set everything up, secure it, and launch the site. Claude did the rest on its own.
Karpathy is confident: projects that used to take weeks and months can now be finished over breakfast. For market participants, this means that applications are getting cheaper and are no longer something worth overpaying for. The money goes to those without whom such projects cannot work: developers of AI models, cloud services, and chip manufacturers.
Demand for computing power continues to grow. This also matters for the crypto industry — many projects at the intersection of blockchain and AI are built precisely on the premise that agents will be able to work on their own, without humans.
Cryptalist analysis: Karpathy's case clearly demonstrates the compression of the software development cycle: the barrier to entry for creating products is falling, and value is shifting toward infrastructure — computing, clouds, and semiconductors. For the crypto market, this reinforces the narrative of decentralized computing and AI agents, where demand for computing resources becomes a key driver. Investors should watch projects that monetize access to capacity, rather than add-on applications that are rapidly getting cheaper.