⏱️ Reading time: 9 min
Karpathy couldn’t stand YouTube’s recommendation algorithm anymore: on a Friday at 11 PM he found exaggerated thumbnails and repetitive content, and by the next day he already had Hacker News HackerTube running, a YouTube clone built with Claude Opus 5.5 that catalogs 6,454 videos recommended in the forum’s comments between 2007 and 2026.
📑 En este artículo
The site, published at hackertube.org, solves one specific problem: watching only the videos that the Hacker News community considered good enough to mention in a comment. No YouTube algorithm, no accounts, no watch history to sell to advertisers.
TL;DR
- Karpathy launched HackerTube, a YouTube clone with 6,454 videos recommended on HN between 2007 and 2026.
- The site filtered 40,000 HN comments using a Claude Code Workflow and Sonnet to discard spam.
- The 6,454 videos cover 150 topics and use the thumbnail taken at the 50% mark of the video instead of the original thumbnail.
- The entire process, scraping, filtering, deployment, and hardening, took about two hours with Claude Opus 5.5.
- HackerTube doesn’t require accounts: it stores recommendations in the user’s local browser session.
What HackerTube Is
Hacker News HackerTube is the platform built by Andrej Karpathy that gathers, with an interface copied straight from YouTube, the videos that different users recommended in Hacker News comments between 2007 and 2026; it groups 6,454 unique videos into 150 topics and doesn’t require registration.
Karpathy is well known in the artificial intelligence community for his time at OpenAI and Tesla, and for later founding Eureka Labs, an AI-focused education project; in 2026 he spends a good part of his time experimenting with Anthropic’s coding agents to build personal tools. HackerTube is, by his own account, the result of one Saturday’s work with Claude Opus 5.5 as the only programmer.
What Happened
It all started on a Friday at 11 PM: Karpathy went to YouTube to find something to watch and ran into, in his own words, a wall of degraded content, with exaggerated thumbnails and repeated recommendations of the same topics. The next day, Saturday, he decided to fix it with his own tool instead of adjusting settings.
He gave Claude the task: build a YouTube clone based on the videos the Hacker News community had recommended over time. The result, described in his original post, was hackertube.org.
The process, described step by step by Karpathy, was: download all of Hacker News’s historical comments, filter them, run them through Sonnet inside a Claude Code Workflow to discard spam, and keep only the real video recommendations. That produced 40,000 comments that mentioned a YouTube video, narrowed down to 6,454 distinct videos spread across 150 topics.
According to Karpathy himself, the manual work was limited to three steps: creating an API key in Google Cloud to query YouTube metadata, setting up a VPS, and buying the domain. Everything else (scraping, filtering, deployment, hardening, and server optimization) was handled by Claude. The entire process took him about two hours.
Context and History
Hacker News has existed since 2007, when Paul Graham launched it as Y Combinator’s forum for startup and technology news. Almost two decades of accumulated comments mean that any YouTube video ever mentioned in a technical thread is, in theory, traceable.
What’s new isn’t the dataset, it’s the time it took to build something useful with it. By 2026 it’s already common to see individual programmers announce finished products (scraping, backend, frontend, deployment, security) built in a single session with a coding agent, without writing a single line by hand. Karpathy sums it up in his post: I don’t do anything resembling traditional engineering, I just talk to Claude and let it do everything.
That shift in role, from programmer to agent director, has been discussed in the community for months, but few examples are as concrete as this one: a curated dataset, a working interface, and a domain of its own, all in one Saturday afternoon.
Technical Details: The Mechanism Behind HackerTube
The bottleneck in a project like this isn’t displaying YouTube videos, any embedded player handles that; it’s separating genuine video mentions from all the noise around them. With 40,000 candidate comments and only 6,454 final videos, the filter discarded about 84% of the initial mentions: broken links, repeated videos, and comments where the link appeared without being an actual recommendation.
To solve this, Karpathy used Workflows, the Claude Code feature that lets you orchestrate several agents in parallel, each with a fixed role. In this case, Sonnet ran over the downloaded comments with a single task: deciding whether a comment with a YouTube link was a genuine recommendation or noise. Those that passed the filter were deduplicated by video ID and fed into the final catalog.
Once the video is identified, the site requests its metadata (title, channel, duration) from the YouTube Data API, the one point where Karpathy had to intervene by hand: generating an API key in Google Cloud. The rest of the personalization, the accountless ForYou feed, lives entirely in the browser session: there’s no user database or login, just local state that gets lost if the cache is cleared.
That flow (downloading, classifying with a model, deduplicating, querying an external API, and deploying) is replicable, but not free: every comment filtered by Sonnet means a call to the model, so a batch of 40,000 comments means tens of thousands of calls before reaching the final catalog of 6,454 videos.
flowchart TD
A["Hacker News comments 2007-2026"] --> B["Filter with Sonnet in a Workflow"]
B --> C["40,000 video recommendations"]
C --> D["6,454 unique videos across 150 topics"]
D --> E["YouTube-style interface"]
E --> F[("Secured and optimized VPS")]
A minor detail Karpathy highlights: the site doesn’t use the thumbnail the channel uploaded, but the automatic thumbnail YouTube generates at the halfway point of the video’s duration. As he describes in his post, that thumbnail tends to represent the actual content better than the thumbnail designed to generate clicks.
📌 Note: HackerTube doesn’t retrain any model or store usage data: the filtering with Sonnet ran once, over a historical dataset, to build the catalog. Every subsequent visit is read-only.
Hacker News HackerTube’s Impact on the Dev Community
What makes this case interesting isn’t the result, a curated video catalog, but the cost of building it. A few years ago, putting together a comment scraper, a spam classifier, a YouTube-style interface, and deploying it all on a secure VPS was weeks of work for a small team. Karpathy cut that down to one afternoon, with three specific manual interventions.
That doesn’t mean anyone can replicate it frictionlessly: HackerTube depends on the YouTube Data API, which has limited daily quotas and could stall a project if traffic grows. The catalog’s 6,454 videos are also a fixed snapshot up to the publication date; if Hacker News keeps generating comments with YouTube links, the site won’t add them on its own, since it doesn’t run a continuous scraping process.
There’s also no real editorial curation beyond the spam filter: the catalog reflects whatever the Hacker News community commented on, with its bias toward programming, science, and technology topics, and little general entertainment. Anyone looking for a different kind of content won’t find it there, and that’s partly the point.
What’s Next
Karpathy didn’t announce a maintenance plan or a date for adding new Hacker News comments to the catalog. What he did make available is a full export of the site’s data, explicitly meant for another developer, or their own coding agent, to take the dataset and build a different version without repeating the scraping and filtering work from scratch.
It’s reasonable to expect forks: the combination of a public dataset, a pipeline documented in a public post, and a replicable pattern (downloading forum comments, classifying them with a model, building an interface) is exactly the kind of weekend project that proliferates once someone influential like Karpathy shows it working.
If you want to see which videos almost two decades of Hacker News comments marked as must-watch, hackertube.org is already up right now, and it doesn’t ask for an email or a password.
Frequently Asked Questions
What is hackertube.org and who built it?
It’s a YouTube clone with 6,454 videos recommended in Hacker News comments between 2007 and 2026, built by Andrej Karpathy with Claude Opus 5.5 in one Saturday work session.
How does HackerTube filter spam from Hacker News comments?
A Claude Code Workflow ran the Sonnet model over 40,000 comments with YouTube links, discarding spam, broken links, and mentions that weren’t real recommendations, before deduplicating by video.
Do I need an account to use hackertube.org?
No. The site has no login or user database: personalized recommendations, the ForYou feed, are calculated and stored in the local browser session.
How long did it take Karpathy to build the Hacker News YouTube clone?
According to his own post, the entire process (scraping, filtering, interface, deployment, security, and optimization) took about two hours, with only three manual steps: creating a YouTube API key, setting up a VPS, and buying the domain.
Can I download Hacker News HackerTube’s data to make my own version?
Yes. The site offers a full export of the catalog’s data, meant for another developer or coding agent to build their own version without repeating the scraping process from scratch.
References
- Karpathy, HackerTube: the original post where Andrej Karpathy describes how he built the site with Claude Opus 5.5.
- Wikipedia: Hacker News: history and origin of the Y Combinator forum, founded in 2007.
- YouTube Data API: official documentation for the API HackerTube uses to query video metadata.
- Anthropic: developer of Claude Code and the Opus and Sonnet models used to build the site.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de Zulfugar Karimov en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments