卡帕西用 Claude 两小时重做了 YouTube,每个人都能拥有一个自己的互联网Karpathy rebuilt YouTube with Claude in two hours


AI free token/AI免费积分

周五晚上 11 点,AI 大神 Karpathy 打开 YouTube,想找个视频看看。

结果首页全是夸张的封面、看过无数遍的老话题。受不了这些算法建议的他,第二天干脆打开 Claude,给自己重新做了一个 YouTube。

整个过程只花了两个小时,用的还是 Claude Opus 5.5 的 Medium 推理强度。

Image

图|HackerTube 网址:https://hackertube.org/

整个视频网站和最近大火的 REA 逆向工程一样,可以说是直接复刻了完整的 YouTube 网页布局。左边的导航栏表明点赞过的视频、历史以及稍后再看等,右边的视频建议流,也是同样的封面大小、发布时间以及播放量等。

他用的提示词只有一句话,「嗨 Claude 早上好你好吗我很好谢谢帮我建一个 YouTube 风格的网站但只限于多年来在 HN 上被反复建议的视频拜托了。别犯任何错误。」

两个小时的工作流跑完,这个 HackerTube 就上线了,里面所有的视频都是来自 YouTube,只不过 Karpathy 用 Claude 写了一套过滤机制,过滤的规则就是从 Hacker News 上找到这些年讨论和建议过的视频。

Image

笔者可以像使用 YouTube 一样在里面找视频,也可以按不同的主题浏览,涵盖艺术、软件工程、历史、思想、商业、游戏等多个不同的分类,笔者甚至同样能从这 6454 个视频里面取得个性化建议。

在视频的详情页面,笔者还能看到来自 Hacker News 的评论,以及有多少位 HN 的建议者,在什么时间内被建议过。

Image

一时间,社交媒体上对这个项目普遍都感到兴奋。一维度是 Karpathy 提到自己只是用了 Claude Code,而且还是 Opus 5.5 的 Medium 思考推理程度,就把一个 YouTube 给复刻出来了。

但也有网友说,最近 Claude 在使用上的限制格外多,尤其是 REA 火了之后,逆向一些软件经常会有安全围栏的碰壁,怎么 Karpathy 就能用 Opus 来逆向 YouTube。

卡帕西用 Claude 两小时重做了 YouTube,每个人都能拥有一个自己的互联网

Image

甚至港科大的博士说自己只是让 Claude 帮忙提交任务和运行分析,就被 Claude 标记为「训练竞争对手」,Opus 5.5 降级到 Opus 5。

Image

能逆向一切的模型能力外,人们更关心的是现在笔者自己也有机会用 AI 重建信息入口。

一个自定义的抖音,或者自定义的 Bilibili,里面只包含那些经过筛选的视频;又或者是一个重建入口的新闻资讯流,抛开微信的公众号建议流,把自己的关注的内容用更合适自己方式展现,而不是微信建议笔者看什么就看什么。

就像已经有网友按类似思路做出了一个 Hacker Papers,也是从 Hacker News 上整理出 1000 多篇 arXiv 论文。

Image

还有网友从 HackerNews 的发帖里识别到了 1 万 5 多千本书,整理成了一个专门书籍建议网站。

Image

不少回复也认为,Karpathy 的HackerTube,真正的创新就是用真实使用者多年来的主动建议,替代了观看时长、点击率和缩略图 A/B 测试作为排序信号,把平台的建议算法改成人类建议信号。

同一套视频,但新的分发机制

在 YouTube 上,建议平台主要依据点击、观看时长、留存和互动等指标决定什么值得被看见。

而对 HackerTube 来说,最初的筛选信号来自 Hacker News 使用者多年来的主动建议,一个人认为某段视频值得分享给另一个人,并真的把链接贴了出来。

Image

所以笔者能看到,在 AI 类别下,建议的第一个视频就是 Karpathy 自己的编程课,一步步写出一个小型自动求导引擎,再搭建、训练神经网络,讲清楚反向传播怎么工作。

这种降低「被算法操控」的外观还体现在网站不要求登录,观看状态保存在本地浏览器。其次,决定观众是否要点进去的封面也是大刀阔斧地改,直接不使用上传者精心优化的封面,只是选取视频中间的一帧。

而作为视频网站,除了单维度的接受,在 HackerTube 上,笔者还可以导出底层数据,交给自己的 AI 代理重新制作一个版本。

因而,HackerTube 的分发机制,就是把算法的服务对象从平台增长重新拉回了个人偏好。项目发布后,X 上的讨论很快超出了一个 AI 编程案例。

一种代表性观点认为,HackerTube 展示的是「Claude 为使用者而战」:当短视频和建议平台不断争夺注意力时,个人Agent 可以站在使用者一边,协助他们过滤低质量内容。

Image

另一部分人则从中看到了一个尚未充分开发的市面——个人建议算法。

过去,内容筛选和建议几乎完全由平台掌握,建议平台几乎就是平台的基础设施。笔者可以接受、屏蔽或微调,却很难真正更换整个平台。现在,一个开发者借助 Claude Code,就能把可信社区的数据加工成自己的信息入口。

两个小时还是不够夸张

依据 Karpathy 自己的分享,几乎是发送一句话给 Claude Code,就等着它自动完成了。

Karpathy 使用的是 Mac 上的 Claude Code 桌面版,默认模型为 Opus 5.5,推理强度设在 Medium。他说,自己已经不太做读写代码、在终端运行命令这类传统工程工作,主要借助对话让 Claude 执行任务。

在这个项目里,他自己动手完成的事情只有创建用于查询视频元数据的 Google Cloud API Key、启动 VPS、购买域名,这三件事。代码编写、数据处理、部署、安全加固和优化,都交给了 Claude。

Image

于是也有网友开始讨论,在做这种 Agent 项目时,终端式的 Claude Code 还能不能说是最好的接口。桌面应用现在看来是否又更合适持续的代理协作?

还有像是模型推理强度应该始终拉满,又或是使用较低强度配合反复批评和修正?

这些难题大概很难有一个统一答案,但 Karpathy 的工作流至少说明,AI 编程的分工正在从「模型补全几行代码」,转向「人提出目标、处理少量外部账号和权限,代理完成其余工程链路」。

但「重做 YouTube」夸张了

HackerTube 并没有重新建立视频托管、上传、版权管理、创作者分成和全球内容分发网络。它仍然依赖 YouTube 的视频与元数据,也站在 Hacker News 近二十年积累的社区语料之上。

Image

所以笔者能看到在这些视频的播放量 Top 排行榜单里,都是一些十几年的视频,第一个就是 Pinkfong 的《鲨鱼宝宝》,那首反复唱「嘟嘟嘟」的儿童歌舞,还有鸟叔的《江南 Style》、「瑞克摇」整蛊梗的常用视频之类。

更准确地说,它是一个建立在现有平台之上的个性化策展与建议层。

和 YouTube 一样,它将来同样要面临类似的局限。例如 HN 使用者的兴趣不能代表所有观众,工艺社区的品味也不等于普遍意义上的高质量;评论建议会受到年代、圈层和社区偏好的影响。

一些实际试用者也发现,搜索结果仍会混入不相关内容,最后仍旧需要人工挑选。

Image

伴随而来的还有网站维护的难题,有人在回复里怀念他以前做的论文工具 arxiv-sanity。他说,自己也很怀念,但当时维护起来实在太费工夫,只能把它关掉;放在今天,他觉得这件事会容易得多。

AI 正在把「逆向」变成一种通用能力

确实如此,放在今天,几乎什么都可以逆向。从各种软件的界面和能力复刻,到经典游戏的反编译、移植,甚至重新达成游戏玩法,还有逆向电视,越来越多过去需要大量专业开发工作的项目,开始有人尝试交给 AI Agent 完成。

Image

网友「逆向」的 Android TV 版,hackertube.tv

严格来说,HackerTube 并没有逆向 YouTube 的代码,也没有破解它的建议算法。Karpathy 只是保留了 YouTube 的内容,用 Hacker News 多年来积累的建议数据,重新做了一套筛选、排序和展示机制。

虽然没有重建,但这个过程却拆解了视频媒介中最影响个人感受的部分,涵盖内容如何被发现、如何被排序,以及使用者与建议平台之间究竟是谁在服务谁等。

Image

在某种程度上,笔者也可以把这理解为是一种逆向,它逆向的是更上层的东西,即 YouTube 这款机型所代表的机型形态、信息组织方式,以及内容与使用者之间的关系。

这种变化的意义不会只停留在复制一个已有的机型上,它开始让每一个人都去探索原本那些封闭的平台,如何重新变得可理解、可修改,以及被再次创造。

Image

对于 YouTube、Bilibili、微信公众号这样的内容平台来说,笔者或许仍旧需要依赖它们给出的内容和基础设施,但不一定还要完全接受平台外观好的信息入口。

当逆向能力从代码层扩展到机型、信息结构乃至媒介层,笔者使用互联网的方式,也可能跟着发生变化。

本文来源:https://www.ifanr.com/1687094
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

On Friday night at 11 p.m., AI heavyweight 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. Fed up with these algorithmic recommendations, he opened Claude the next day and simply rebuilt himself a YouTube.

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

Image

Image | HackerTube URL: https://hackertube.org/

Like REA (reverse engineering) that has blown up recently, the entire video site essentially directly replicates the full YouTube webpage layout. The left sidebar shows liked videos, history, watch later, etc.; the right-hand recommendation stream has the same thumbnail sizes, upload times, view counts, and so on.

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

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

Image

I can search for videos inside it just like using YouTube, and I can also browse by different themes, covering multiple categories such as art, software engineering, history, ideas, business, games, etc. I can even get personalized recommendations from these 6454 videos.

On the video detail page, I can also see comments from Hacker News, how many HN recommenders there were, and during what time period they were recommended.

Image

For a time, people on social media were widely excited about this project. One dimension is that Karpathy mentioned he only used Claude Code—and at Medium thinking/reasoning level on Opus 5.5—to replicate a YouTube.

But some netizens also said that Claude has recently had especially many usage restrictions; especially after REA became popular, reverse engineering some software often hits safety guardrails—so how was Karpathy able to use Opus to reverse engineer YouTube?

Karpathy rebuilt YouTube with Claude in two hours

Image

Even a PhD student from HKUST said that he only asked Claude to help submit tasks and run analyses, yet he was flagged by Claude as a “training competitor,” and Opus 5.5 was downgraded to Opus 5.

Image

Beyond the model’s ability to reverse engineer almost anything, what people care about more is that now I also have the chance to rebuild information gateways with AI.

A custom Douyin, or a custom Bilibili, containing only curated videos; or a rebuilt news feed gateway that sets aside WeChat’s official-account recommendation stream, presenting the content I follow in a way that suits me better—instead of just watching whatever WeChat recommends I watch.

Just like one netizen has already built something called Hacker Papers along similar lines, also compiling over 1000 arXiv papers from Hacker News.

Image

Another netizen identified more than 15,000 books from Hacker News posts and put them together into a dedicated book-recommendation website.

Image

Many replies also believe that the real innovation of Karpathy’s HackerTube is replacing watch time, click-through rate, and thumbnail A/B testing as ranking signals with real users’ long-standing, voluntary recommendations—turning the platform’s recommendation algorithm into a human-recommendation signal.

The same set of videos, but a new distribution mechanism

On YouTube, the recommendation platform 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’ voluntary recommendations over the years: one person believes a video is worth sharing with another person, and actually posts the link.

Image

So I can see that under the AI category, the first recommended video is Karpathy’s own programming course, which walks through writing a small autograd engine step by step, then building and training a neural network, and explains clearly how backpropagation works.

This reduced appearance of being manipulated by algorithms is also reflected in the fact that the website does not require login, and viewing status is saved locally in the browser. Furthermore, the cover images that determine whether viewers click in have been drastically changed: instead of using the carefully optimized covers by uploaders, it simply selects a single frame from the middle of the video.

And as a video website, beyond just one-dimensional consumption, on HackerTube, the author can also export the underlying data and hand it over to their own AI agent to create a new version.

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

A representative view suggests that what HackerTube demonstrates is "Claude fighting for the user": while short-form videos and recommendation platforms continuously compete for attention, personal agents can stand on the user's side, helping them filter out low-quality content.

Image

Another group of people sees an underdeveloped market in this—personal recommendation algorithms.

In the past, content filtering and recommendations were almost entirely controlled by platforms; recommendation platforms were virtually the infrastructure of the platforms themselves. Users could accept, block, or fine-tune them, but it was difficult to truly switch the entire platform. Now, with the help of Claude Code, a developer can process data from trusted communities into their own information gateway.

Two hours is still not exaggerated.

According to Karpathy's own sharing, it was almost just sending a single sentence to Claude Code and waiting for it to automatically complete the task.

Karpathy used the Claude Code desktop version on a Mac, with the default model set to Opus 5.5 and reasoning effort set to Medium. He mentioned that he no longer does traditional engineering work like reading/writing code or running commands in the terminal; instead, he mainly relies on conversations to let Claude execute tasks.

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

Image

Consequently, some netizens began discussing whether the terminal-based Claude Code can still be considered the best interface when working on such Agent projects. Does the desktop application now seem more suitable for continuous agent collaboration?

There are also questions like whether the model's reasoning effort should always be maxed out, or if a lower effort paired with repeated criticism and correction is better?

It is probably difficult to have a unified answer for these challenges, but Karpathy's workflow at least illustrates that the division of labor in AI programming is shifting from "the model completing a few lines of code" to "humans proposing goals, handling a small number of external accounts and permissions, and agents completing the rest of the engineering pipeline."

But "rebuilding YouTube" is an exaggeration.

HackerTube did not rebuild video hosting, uploading, copyright management, creator revenue sharing, or a global content distribution network. It still relies on YouTube's videos and metadata, and it stands on the community corpus accumulated by Hacker News over nearly two decades.

Image

Therefore, the author noticed that in the top view count rankings for these videos, they are all videos from over a decade ago. The first one is Pinkfong's "Baby Shark," the children's song and dance that repeatedly sings "doo-doo-doo," as well as PSY's "Gangnam Style," and videos commonly used for the "Rickroll" internet meme, among others.

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

Like YouTube, it will face similar limitations in the future. For example, the interests of HN users cannot represent all viewers; the tastes of the craft community are not equivalent to high quality in a universal sense; and recommendation algorithms will be influenced by the era, social circles, and community preferences.

Some actual users also found that search results still mix in irrelevant content, ultimately still requiring manual selection.

Image

Along with this comes the challenge of website maintenance. Someone replied nostalgically about the paper tool arxiv-sanity he had built previously. He said he also misses it, but maintaining it back then took too much effort, forcing him to shut it down; he feels that doing this today would be much easier.

AI is turning "reverse engineering" into a general capability.

Indeed, looking at today, almost anything can be reverse-engineered. From replicating interfaces and capabilities of various software, to decompiling, porting, and even recreating the gameplay of classic games, as well as reverse-engineering televisions, more and more projects that previously required extensive professional development work are now being attempted by handing them over to AI agents.

Image

The 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 new filtering, sorting, and display mechanism.

Although it was not rebuilt, this process did dissect the parts of the video medium that most affect personal experience, covering how content is discovered, how it is ranked, and exactly who is serving whom between the user and the recommendation platform.

Image

To some extent, the author can also interpret this as a kind of reverse engineering—one that works against the higher-level structures, namely the device form represented by YouTube, its method of organizing information, and the relationship between content and users.

The significance of this change does not remain confined to replicating an already existing device; rather, it begins to prompt everyone to explore how once-closed platforms can be rendered intelligible, modifiable, and ripe for re-creation.

Image

For content platforms such as YouTube, Bilibili, and WeChat Official Accounts, the author may still depend on the content and infrastructure they provide, but it is not necessary to wholly embrace the platforms’ visually polished information entry points.

As the capacity for reverse engineering expands from the code layer to the device, information architecture, and even the medium layer, the way the author uses the Internet may also change accordingly.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry