Karpathy's YouTube Alternative Reportedly Built in Two Hours

TVS News AI, editorially supervised, presents Karpathy's YouTube Alternative Reportedly Built in Two Hours.

On Friday evening at 11 p.m., Andrej Karpathy opened YouTube, according to his own account, and encountered what he called "algorithmic mush." Within less than two days, it led to the HackerTube project. The website has been online since October 31, 2026. It does adopt some YouTube patterns, but uses a different mechanism for selecting content: videos that were repeatedly recommended in 43 million comments from 20 years of Hacker News.

The numbers give the project an unusually concrete scale. According to Karpathy's project description, the comments yielded approximately 40,000 posts explicitly recommending a video. From those, Claude filtered out 6,454 videos published between 2007 and 2026. Karpathy organized them into 150 topics.

43 Million Comments Became 6,454 Videos

The community provides the selection criteria

On HackerTube, an invisible feed algorithm does not determine which videos appear on the home page. The foundation is recommendations from comments on the Hacker News forum. This is the central difference from the original YouTube: The platform filters through a vast collection of discussions spanning decades and retains videos that repeatedly gained attention there.

Claude did more than simply perform a keyword search. The system reviewed 43 million posts, removed spam and uninteresting mentions, and isolated around 40,000 comments in which a specific video was recommended. These formed the curated selection of 6,454 titles.

For you, this filter is the main practical benefit. You do not get a general mixture of entertainment, advertising, and popular videos, but a technically curated collection. However, the source cannot substantiate how representative this selection is: It provides no figures on user numbers, views, or a better performance compared with YouTube. Karpathy's phrase "a significantly better viewing experience" is therefore primarily his own assessment.

The content stretches back to 2007

The range encompasses 200 years of videos from 20 years — more precisely: 6,454 publications from 2007 to 2026, distributed across 150 topics. Karpathy has also created his own playlists. One is called "The HN Canon" and brings together videos that, according to the account, belong to the well-known classics of the tech world.

This means HackerTube is not merely a single filter for the latest posts. The platform collects material from two decades and turns it into a kind of technically curated archive. Although the detailed report from Numerama does not name any individual topic categories, it does explicitly mention the 150 topics and the "The HN Canon" playlist.

Claude Handled the Code, Design, and Publication

By his own account, Karpathy did not write a single line of code

Karpathy, co-founder of OpenAI and former head of AI at Tesla, worked with Claude Code in the Mac app. He used Opus 5.5 with a medium effort level as the model. He gave Claude the task of updating the website, securing it, and deploying it on the server.

The manual work remained manageable. Karpathy created a Google Cloud API key, rented a server, and bought the domain. Once the server was running, Claude was supposed to operate the finished application there. According to Karpathy, this process took about two hours.

Numerama summarizes the entire development process as a weekend hack, writing that the website was entirely conceived and put online by Claude. The source does not explain the precise way the time estimates relate to one another. Both are therefore documented: the weekend as the overall framework and Karpathy's statement that the concrete implementation took about two hours with Opus 5.5.

No account and no centralized video history

HackerTube works without registration. Recommendations also work without logging in. If you use your viewing history, it is stored in your browser, rather than with a central user account. This is a significant difference from the original YouTube, where the watch history is tied to the signed-in account.

The interface is also modeled on YouTube. HackerTube includes watch history, the "Watch Later" feature, and playlists. For thumbnails, however, the platform takes a different approach: Instead of the image chosen by the creator, a still from the middle of the video appears.

The Curated Selection Does Not End with YouTube

HackerTube Longs replaces vertical clips

"HackerTube Longs" provides a dedicated section for short-form content. It follows the principle of YouTube Shorts without being limited to vertical videos. The site therefore takes in not only the classic features of a video archive, but also the format through which YouTube captures attention particularly quickly.

The open approach to the data is especially interesting. Karpathy makes the data derived from the comment analysis available in a simple archive. He invites people to feed it into their own AI assistant and build a personal version from it. The method is therefore not limited to HackerTube.

According to Numerama, others have already picked up on the idea. There is now a variant for scientific publications on arXiv and an Android app. The source does not state how many projects have been created or what features they offer.

What the Two Hours Really Mean

The impressive number applies to the work with Claude

The strongest claim in the report lies not only in the 6,454 videos, but in the combination of a complete platform and the processed dataset. Karpathy gave Claude access to 43 million comments, had it build a recommendation system from them, connected that system to a YouTube-like interface, and deployed the result on his own server.

The project also shows what the statement "without writing a single line of code" means — and what it does not mean. Karpathy himself provided the technical access: API key, server, and domain. Claude created and published the website. Without those three manual steps, the two-hour implementation would not have been possible.

The strongest counterargument to the claimed efficiency is the lack of an independent assessment. The source cites neither user numbers nor views or a comparative study with YouTube. The 43 million comments, 6,454 videos, 150 topics, and roughly two hours with Opus 5.5 are firmly documented. Whether the new feed works better over the long term will have to be demonstrated by the actual recommendations.