TL;DR
HackerTube is a YouTube-style site assembled from videos recommended in Hacker News comments. Its creator says the project filtered the forum’s comments to identify 6,454 videos across 150 topics, then used an AI coding assistant to build and deploy the site. The collection and workflow are described by the creator; independent verification and details about ongoing maintenance are not provided.
Developer Andrej Karpathy says he built and launched HackerTube, a YouTube-style site that organizes videos recommended in Hacker News comments, after asking an AI coding assistant to create an alternative feed. The project’s stated dataset contains 6,454 YouTube videos drawn from comments posted between 2007 and 2026, giving users a way to browse recommendations from the forum rather than rely on YouTube’s usual recommendation feed.
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
In a post dated October 31, 2026, Karpathy said he began the project after finding YouTube’s recommendations unhelpful for his own viewing. He described the feed as dominated by exaggerated thumbnails and repetitive topics; that is his assessment, not an independently measured finding about YouTube’s recommendations. He said he asked Claude Opus 5.5 to build a clone using videos recommended on Hacker News, and the resulting site was named HackerTube.
According to Karpathy, the process started with downloading Hacker News comments and filtering them for references to YouTube videos. He said an AI-assisted workflow removed spam and narrowed the material to about 40,000 comments recommending a video. Those comments yielded 6,454 videos associated with 150 topics. The post does not detail the filtering criteria or explain how repeated recommendations were counted.
Karpathy said the site presents the videos in a YouTube-style interface, uses what he called a “50% thumbnail” rather than each video’s uploaded thumbnail, and provides personalized recommendations without user accounts by keeping state in the browser session. He also said users can export the site’s data to use with their own AI agent. The post describes the site as available at hackertube.org; its stated features and the dataset figures come from the creator’s account.
A Different Route to Video Discovery
HackerTube offers a small-scale example of how a community’s past links can be repackaged as a discovery tool. Instead of building a profile from an individual’s watch history, the project draws on recommendations made in public forum discussions. That changes the source of the feed: users browse videos that commenters chose to mention on Hacker News, rather than videos selected for them by YouTube’s recommendation system.
The site may interest viewers who want a curated starting point or who prefer recommendations that can be traced to a community conversation. But the dataset is not necessarily a representative sample of YouTube, Hacker News users, or video quality. It reflects what people posted in one forum, what the filtering process retained, and what was available as a YouTube link. Karpathy’s description of the result as quality content is a personal characterization, not an independent evaluation of the videos.
The project also draws attention to how AI coding tools can compress a prototype’s development process. Karpathy said the work took about two hours and that the assistant handled coding, deployment, security hardening and optimization. That is his account of this project, not a general measure of how long comparable services take to build or how much human review they require.
Products Worth Considering

OMRON Complete 2-in-1 Upper Arm Blood Pressure Monitor with EKG

Withings BPM Connect Wi-Fi and Bluetooth Blood Pressure Monitor

Levoit Core 300S-P Smart WiFi Air Purifier for Home and Pets
As an affiliate, we earn on qualifying purchases.
From Forum Comments to a Site
Hacker News is a discussion forum where users frequently share links, including videos. Karpathy’s idea was to treat those historical comments as a recommendation archive and put the resulting links into a browsable interface. The data described in the post spans 2007–2026, making HackerTube a retrospective collection rather than a feed limited to newly posted recommendations.
Karpathy said he used Claude Code, Anthropic’s coding tool, with Opus 5.5 set to medium effort. He described his role as directing the assistant rather than writing code or running terminal commands himself. His stated manual tasks included creating a Google Cloud API key for YouTube metadata, starting a virtual private server and buying the domain. He said he then supplied the server’s IP address and asked the assistant to update, deploy, secure, harden and optimize the software.
The post is a first-person account from the project’s creator. It does not include an independent technical review, a detailed methodology for classifying videos by topic, or a comparison with YouTube’s own recommendation system. The account supports the reported development process and the project’s stated design, but does not by itself verify how the site performs for users.
How the Collection Was Filtered
The creator’s post does not publish the full dataset, the filtering rules, or a breakdown of how the 6,454 videos were identified and grouped into 150 topics. It is not clear how the process handled duplicate links, deleted or unavailable videos, ambiguous references, or comments that mentioned a video without recommending it. The reported 40,000 comments and resulting video count should therefore be treated as figures supplied by Karpathy, rather than independently audited totals.
Other details also remain unreported, including how often the site’s collection will be refreshed, whether users can submit corrections, and how its personalized recommendations are generated from browser-session state. The post gives no independent assessment of the site’s security, uptime, or user numbers. It is also unclear whether the two-hour estimate includes time spent on the assistant’s work or only the creator’s active involvement.
Site Availability and Future Updates
Karpathy’s post presents hackertube.org as the deployed site and says its data can be exported for people who want to create their own version. The post does not announce a schedule for new features, a regular refresh of the Hacker News archive, or a process for community contributions.
Readers seeking further confirmation can check the live site and any subsequent updates from Karpathy. Until the creator publishes more technical documentation or a fuller account of the dataset, the collection’s scope and the site’s continued operation remain based on the launch description.
Key Questions
What is HackerTube?
HackerTube is a YouTube-style site that presents videos recommended in Hacker News comments, according to its creator, Andrej Karpathy.
How many videos does the site include?
Karpathy says the collection contains 6,454 videos, identified from about 40,000 comments and grouped across 150 topics. Those figures have not been independently verified in the supplied account.
How were the videos selected?
Karpathy says Hacker News comments were downloaded and filtered, with an AI-assisted workflow used to remove spam and identify comments recommending YouTube videos. The post does not provide the full filtering method.
Did an AI assistant build the site?
Karpathy says he used Claude Code with Opus 5.5 to build and deploy the project. He also reports handling several setup tasks himself, including obtaining an API key, starting a server and purchasing the domain.
Is HackerTube an official YouTube service?
The launch post describes HackerTube as a separate site built around YouTube video links. It does not say that YouTube operates or endorses the project.
Source: karpathy.bearblog.dev
