Andrej Karpathy built HackerTube from 6,454 videos in two hours

YouTube knows what holds people's attention, but that does not always match what they actually want to watch. Andrej Karpathy ran into this familiar problem and built HackerTube over a weekend—a video catalog whose recommendations come not from an algorithmic feed, but from Hacker News comments.

The result is a website featuring 6,454 videos that HN members recommended to one another over 20 years. According to Karpathy, Claude completed the core work in about two hours. The developer himself bought a domain, created a YouTube API key, and spun up a VPS.

It sounds like yet another story in the “an agent wrote an app over breakfast” genre. But the useful part is not the speed at which the code was produced. Karpathy showed how to turn a large, unstructured archive into a focused product with a clear value proposition: identify recurring recommendations, connect them to videos, and give users a proper interface for searching and watching them.

How it works, using a simple example

Imagine that people leave links to lectures, interviews, and technical deep dives across thousands of discussions. One link may appear once and mean nothing. Another keeps resurfacing for years in different threads, accompanied by phrases such as “the best explanation of the subject” or “this is worth watching in full.”

HackerTube uses the collective trail of these recommendations as its source signal. Instead of a personalized model trying to predict the next click, the service shows videos that have already passed a human filter within a technical community. Karpathy says that 43 million Hacker News comments were processed to build the catalog, yielding 6,454 videos.

This is an important distinction. The agent does not invent recommendations from scratch; it helps extract structure from existing data and turn the result into a product. The separately published analysis archive can also be reused—for example, users can ask their own agents to select material on compilers, startup history, or machine learning.

What the agent actually accelerated

In the past, an experiment like this would have required separately implementing data ingestion, link processing, YouTube integration, a server-side component, and an interface. A modern coding agent can connect these pieces in a single session when the task is constrained, the sources are known, and a human makes infrastructure decisions quickly.

The practical takeaway for teams is simple: a prototype should now begin not with a lengthy technical specification, but with a testable question. For example: “Will developers search for videos through HN recommendations instead of the YouTube feed?” An agent can build a minimal product, and real users can provide the answer faster than presentations and weeks of planning.

That said, the phrase “built in two hours” does not mean the product is ready for heavy traffic or commercial operation. Query costs, YouTube API limits, data refreshes, moderation, security, backups, and the legal terms governing content use still remain. There are also not yet enough public measurements of the selection's quality: it is unknown how many useful links were missed or how strongly the catalog is skewed toward the interests of the Hacker News audience.

What to apply in your own work

Developers should look for problems where a valuable dataset already exists but lacks a convenient interface. Product teams should measure not the amount of generated code, but the time from an idea to the first user signal. Founders should distinguish in advance between a rapid prototype and a system that can be maintained safely.

HackerTube is compelling precisely as an example of this approach. The agent dramatically reduced the cost of building the wrapper, but the idea, the data source, and the usefulness criterion remained human work. It appears that the main constraint is no longer the ability to build another interface quickly, but the ability to notice a good dataset and ask the right question of it.