Karpathy Built the HackerTube Video Site with Claude Code in Two Hours

At 11 p.m. on Friday, AI guru Karpathy opened YouTube, wanting to find a video to watch.

But the homepage was full of exaggerated thumbnails and old topics he'd seen countless times. Unable to stand these algorithmic recommendations, the next day he simply opened Claude and built himself a new YouTube.

The whole process took only two hours, and he used Claude Opus 5.5 at Medium reasoning effort.

The entire video site, like the recently viral REA reverse engineering, can be said to directly replicate the complete YouTube web layout. The left navigation bar shows liked videos, history, and Watch Later, while the video recommendation feed on the right has the same thumbnail size, publish time, and view counts.

His prompt was just one sentence: "Hi Claude, good morning, how are you? I'm fine, thanks. Help me build a YouTube-style website, but limited to videos that have been repeatedly recommended on HN over the years. Please. Don't make any mistakes."

After a two-hour workflow run, HackerTube went live. All the videos in it come from YouTube, except that Karpathy used Claude to write a filtering mechanism, and the filtering rule is to find videos that have been discussed and recommended on Hacker News over the years.

We can search for videos in it just like using YouTube, and we can also browse by different topics, including art, software engineering, history, ideas, business, games, and many other categories. We can even get personalized recommendations from these 6,454 videos.

On the video detail page, we can also see comments from Hacker News, as well as how many HN recommenders there were and when it was recommended.

For a time, social media was broadly excited about this project. On one hand, Karpathy mentioned that he had only used Claude Code, and at Opus 5.5's Medium thinking/reasoning level, to replicate a YouTube.

But some netizens also said that Claude has recently had many usage restrictions, especially after REA became popular, reverse engineering some software often hits safety guardrails. How could Karpathy use Opus to reverse engineer YouTube?

A PhD student at HKUST even said that he only asked Claude to help submit tasks and run analyses, and Claude flagged him as a "training competitor," downgrading Opus 5.5 to Opus 5.

A custom Douyin, or a custom Bilibili, containing only those filtered videos; or a news feed with a rebuilt entry point, setting aside WeChat's official account recommendation feed and presenting the content you follow in a way that suits you better, rather than watching whatever WeChat recommends.

Just as some netizens have already made a Hacker Papers using a similar approach, also organizing more than 1,000 arXiv papers from Hacker News.

Other netizens identified more than 15,000 books from Hacker News posts and organized them into a dedicated book recommendation website.

Many replies also believe that the real innovation of Karpathy's HackerTube is using real users' active recommendations over many years to replace watch time, click-through rate, and thumbnail A/B testing as ranking signals, changing the platform's recommendation algorithm into human recommendation signals.

The same set of videos, but a new distribution mechanism

On YouTube, the recommendation system mainly decides what deserves to be seen based on metrics such as clicks, watch time, retention, and engagement.

For HackerTube, the initial filtering signal comes from Hacker News users' active recommendations over the years—one person thinks a certain video is worth sharing with another person, and actually posts the link.

So we can see that under the AI category, the first recommended video is Karpathy's own programming course, step by step writing a small automatic differentiation engine, then building and training a neural network, and clearly explaining how backpropagation works.

This design to reduce "being manipulated by algorithms" is also reflected in the fact that the website does not require login, and viewing status is saved in the local browser. Second, the thumbnail that determines whether viewers click in has also been drastically changed: it directly does not use the carefully optimized thumbnail from the uploader, but only selects a frame from the middle of the video.

And as a video website, besides one-way reception, on HackerTube we can also export the underlying data and hand it to our own AI agent to remake a version.

Therefore, HackerTube's distribution mechanism pulls the algorithm's service target from platform growth back to personal preference. After the project was released, discussion on X quickly went beyond an AI coding case.

A representative view holds that HackerTube demonstrates "Claude fighting for users": when short videos and recommendation systems constantly compete for attention, a personal Agent can stand on the user's side and help them filter out low-quality content.

Another group of people saw in it a market that has not yet been fully developed—personal recommendation algorithms.

In the past, content filtering and recommendation were almost entirely controlled by platforms, and recommendation systems were essentially the infrastructure of platforms. We could accept, block, or fine-tune them, but it was very difficult to truly replace the entire system. Now, with Claude Code, a developer can process data from trusted communities into their own information gateway.

Two hours is still not exaggerated enough

According to Karpathy's own account, it was almost a matter of sending one sentence to Claude Code and waiting for it to complete the task automatically.

Karpathy used the Claude Code desktop version on Mac, with the default model Opus 5.5 and reasoning effort set to Medium. He said he no longer really does traditional engineering work such as reading and writing code or running commands in the terminal, and mainly has Claude execute tasks through conversation.

In this project, the only things he did himself were creating a Google Cloud API Key for querying video metadata, starting a VPS, and purchasing a domain name—these three things. Code writing, data processing, deployment, security hardening, and optimization were all handed over to Claude.

As a result, some netizens also began discussing whether, when doing this kind of Agent project, terminal-style Claude Code can still be called the best interface. Does the desktop application now seem more suitable for sustained agent collaboration?

There are also questions such as whether model reasoning effort should always be maxed out, or whether a lower effort should be used together with repeated criticism and correction.

These questions probably have no unified answer, but Karpathy's workflow at least shows that the division of labor in AI programming is shifting from "the model completes a few lines of code" to "a person proposes the goal, handles a small number of external accounts and permissions, and the agent completes the rest of the engineering chain."

But "remaking YouTube" is an exaggeration

HackerTube has not rebuilt video hosting, uploading, copyright management, creator revenue sharing, or a global content delivery network. It still relies on YouTube's videos and metadata, and it also stands on the community corpus accumulated by Hacker News over nearly twenty years.

So we can see that in the Top rankings by view count for these videos, they are all videos from over a decade ago. The first is Pinkfong's "Baby Shark," that children's song and dance that repeatedly sings "doo doo doo," along with Psy's "Gangnam Style," the commonly used video for the "Rickroll" prank meme, and so on.

More accurately, it is a personalized curation and recommendation layer built on top of existing platforms.

Like YouTube, it will likewise face similar limitations in the future. For example, the interests of HN users cannot represent all viewers, and the taste of a technical community does not equal high quality in the general sense; comment recommendations will be affected by era, circle, and community preferences.

Along with this comes the issue of website maintenance. Someone in the replies reminisced about the paper tool arxiv-sanity that he had made before. He said that he also missed it very much, but maintaining it at the time was really too laborious, so he had no choice but to shut it down; today, he thinks this matter would be much easier.

Indeed, today almost anything can be reverse-engineered. From replicating the interfaces and functions of various software, to decompiling and porting classic games, even reimplementing gameplay, and reverse-engineering TVs, more and more projects that previously required a large amount of professional development work are beginning to be attempted by people handing them to AI Agents.

A netizen's "reverse-engineered" Android TV version, hackertube.tv

Strictly speaking, HackerTube did not reverse-engineer YouTube's code, nor did it crack its recommendation algorithm. Karpathy merely retained YouTube's content and used the recommendation data accumulated by Hacker News over the years to rebuild a set of filtering, ranking, and display mechanisms.

Although it did not rebuild, this process dismantled the parts of video media that most affect personal experience, including how content is discovered, how it is ranked, and who is actually serving whom between users and recommendation systems.

To some extent, we can also understand this as a kind of reverse engineering. What it reverse-engineers is something at a higher level, namely the product form represented by YouTube as a product, the way information is organized, and the relationship between content and users.

The significance of this change will not stop at copying an existing product. It begins to let everyone explore how those originally closed systems can become understandable, modifiable, and created anew.

For content platforms such as YouTube, Bilibili, and WeChat Official Accounts, we may still need to rely on the content and infrastructure they provide, but we do not necessarily have to fully accept the information gateways designed by the platforms.

When reverse-engineering capability expands from the code layer to the product, information structure, and even media layers, the way we use the internet may also change accordingly.

This article comes from the WeChat Official Account "APPSO," author: APPSO, which discovers tomorrow's products, published by 36Kr with authorization.