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
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。受不了这些算法推荐的他，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。
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。左边的导航栏显示点赞过的视频、历史以及稍后再看等，右边的视频推荐流，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 上被反复推荐的视频拜托了。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推荐过的视频。
我们可以像使用 YouTube 一样在里面找视频，也可以按不同的主题浏览，包括艺术、软件工程、历史、思想、商业、游戏等多个不同的分类，我们甚至同样能从这 6454 个视频里面获得个性化推荐。
在视频的详情页面，我们还能看到来自 Hacker News 的评论，以及有多少位 HN 的推荐者，在什么时间内被推荐过。
一时间，社交媒体上对这个项目普遍都感到兴奋。一方面是 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 在使用上的限制非常多，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。
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。
能逆向一切的模型能力外，大家更关心的是现在我们自己也有机会用 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，抛开微信的公众号推荐流，把自己的关注的内容用更适合自己方式呈现，而不是微信推荐我们看什么就看什么。
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。
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推荐网站。
不少回复也认为，Karpathy 的HackerTube，真正的创新就是用真实用户多年来的主动推荐，替代了观看时长、点击率和缩略图 A/B 测试作为排序信号，把平台的推荐算法改成人类推荐信号。
同一套视频，但新的分发机制
在 YouTube 上，推荐系统主要根据点击、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 用户多年来的主动推荐，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。
所以我们能看到，在 AI 类别下，推荐的第一个视频就是 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」的设计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，直接不采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，除了单方面的接受，在 HackerTube 上，我们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。
因此，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 为用户而战」：当短视频和推荐系统不断争夺注意力时，个人Agent 可以站在用户一边，帮助他们过滤低质量内容。
另一部分人则从中看到了一个尚未充分开发的市场——个人推荐算法。
过去，内容筛选和推荐几乎完全由平台掌握，推荐系统几乎就是平台的基础设施。我们可以接受、屏蔽或微调，却很难真正更换整个系统。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
根据 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，主要通过对话让 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。
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适合持续的代理协作？
还有像是模型推理强度应该始终拉满，又或是采用较低强度配合反复批评和修正？
这些问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。
所以我们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」整蛊梗的常用视频等等。
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推荐层。
和 YouTube 一样，它将来同样要面临类似的局限。例如 HN 用户的兴趣不能代表所有观众，技术社区的品味也不等于普遍意义上的高质量；评论推荐会受到年代、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，最后依然需要人工挑选。
伴随而来的还有网站维护的问题，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功能复刻，到经典游戏的反编译、移植，甚至重新实现游戏玩法，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。
网友「逆向」的 Android TV 版，hackertube.tv
严格来说，HackerTube 并没有逆向 YouTube 的代码，也没有破解它的推荐算法。Karpathy 只是保留了 YouTube 的内容，用 Hacker News 多年来积累的推荐数据，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体验的部分，包括内容如何被发现、如何被排序，以及用户与推荐系统之间究竟是谁在服务谁等。
在某种程度上，我们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 这款产品所代表的产品形态、信息组织方式，以及内容与用户之间的关系。
这种变化的意义不会只停留在复制一个已有的产品上，它开始让每一个人都去探索原本那些封闭的系统，如何重新变得可理解、可修改，以及被再次创造。
对于 YouTube、Bilibili、微信公众号这样的内容平台来说，我们或许依然需要依赖它们提供的内容和基础设施，但不一定还要完全接受平台设计好的信息入口。
当逆向能力从代码层扩展到产品、信息结构乃至媒介层，我们使用互联网的方式，也可能跟着发生变化。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Karpasi made YouTube with Claude for two hours. Everyone has a connection. Network
Friday night at 11 p.m., AI Karpathy opens YouTube and wants a video.
As a result, the front page was full of exaggerating covers and numerous old topics. I can't stand these algorithms recommending him and just turn on Claude the next day and make myself another YouTube.
The whole process took only two hours to use the Medium reasoning strength of Claude Opus 5.5.
The entire video site, like the recent fire's REA reverse project, can be described as a direct regraving of the full YouTube page layout. The navigation bar on the left shows the fine video, history and later reading, and the recommended stream on the right is the same cover size, publish time and the amount of play.
He used prompt to say, "Hey, Claude, how are you? Don't make any mistakes.
Two hours after workflow ran, HackerTube went online, and all the videos were from YouTube, except Karpathy wrote a filter mechanism with Claude, the filtering rule was to find these years' discussions and recommended videos from Hacker News.
We can find videos like YouTube, and we can view them by different themes, including art, software engineering, history, thought, commerce, games, and so on, and we can even get personalized recommendations from 6454 videos.
On the video details page, we can also see comments from Hacker News and how many HN nominees were recommended in what time.
At one point, social media was generally excited about the project. On the one hand, Karpathy mentioned that he had only used Claude Code, and on the other hand, the Medium of Opus 5.5 thought about the level of reasoning, and he took back a YouTube.
But some netizens say that Claude recently has very limited access, especially since the REA fire, and that some of the reverse software often comes across the security fence.
Even the doctor of Hong Kong said that he had only asked Claude to help with the task and the running analysis, and that Claude had been labelled "training the competitor" and that Opus 5.5 downgraded to Opus 5.
Beyond the ability to reverse everything, you're more concerned that now we have the opportunity to rebuild the portal with AI.
A self-defined Douyin, or a self-defined Bilibi, which contains only selected videos, or a news stream that rebuilds the portal, leaving aside the WeChat Official Account recommended stream of WeChat and presenting its concerns in a more self-appropriate way than WeChat suggesting that we look at.
It's like having a guy who's done a Hacker Papers in a similar way, and sorted out over 1,000 arXiv papers from Hacker News.
There are more than 15,000 books identified by netizens from HackerNews' post, which is a dedicated book referral site.
A number of responses also suggested that the real innovation of HackerTube in Karpathy was to replace viewing hours, hits and thumbnail A/B tests as sorting signals by using the many years of voluntary offers from real users to convert the Platform ' s recommended algorithm to a human recommended signal.
Same set of videos, but new distribution mechanisms
On YouTube, the recommended system determines what is worth seeing mainly on the basis of the indicators of click, view duration, retention and interaction.
For HackerTube, the initial filtering signal came from Hacker News, a user who thought that a video was worth sharing with another person and actually posted the link.
So we can see that under the AI category, the first video that is recommended is Karpathy's own programming class, writing a small auto-guided engine step by step, building, training a neural network, making it clear how reverse transmission works.
The design of this reduced algorithmic manipulation is also reflected in the fact that the website does not require access and that the viewing is kept in the local browser. Second, the cover that determines whether the audience wants to go in is also radically altered, not directly by the uploader ' s finely refined cover, but by choosing the frame in the middle of the video.
And as a video site, in addition to being accepted unilaterally, on HackerTube, we can export the bottom data and give it to our AI agent to recreate a version.
Thus, HackerTube ' s distribution mechanism is to bring back personal preferences from the growth of the server from the platform. After item publish, the discussion on X quickly exceeded an AI programming case.
According to a representative view, HackerTube shows "Claude fights for users": When short video and the referral system are competing for attention, the individual Agent can stand by the users and help them filter low-quality content.
Another part of the population saw a market that had not yet been fully developed — an individual recommended algorithm.
In the past, content screening and recommendations were almost entirely owned by the Platform, and the recommended systems were almost the platform ' s infrastructure. We can accept, shield or fine-tune, but it is difficult to really change the system. Now, with Claude Code, a developer can process data from trusted communities into their own information portals.
Two hours is not enough.
According to Karpathy's own share, almost a sentence was sent to Claude Code, waiting for it to be done automatically.
Karpathy uses the Claude Code desktop version on Mac. The default model is Opus 5.5, with reasoning strength in Medium. He said that he was not very good at reading and writing, running orders at the end, and that traditional engineering work, such as that of running orders at the end, was mainly about making Claude do his job through dialogue.
In this project, what he did was create Google Cloud API Key, start VPS, buy domain names. Code writing, data processing, deployment, security reinforcement and optimization were handed over to Claude.
So there's a discussion about whether the terminal Claude Code is still the best interface for this Agent project. Is desktop application now more suitable for continuous proxy collaboration?
Or should the strength of the model's reasoning always be full, or should it be accompanied by a lower intensity of criticism and correction?
It is probably difficult to have a uniform answer to these questions, but Karpathy’s workflow at least explains that the division of AI programming is moving from “model completion of several lines of code” to “human targeting, processing of a small number of external accounts and permission, acting on the remaining engineering links”.
But you're exaggerating.
HackerTube did not re-establish video hosting, uploading, copyright management, creator division and global content distribution networks. It still relies on YouTube's video and metadata, and it stands on the community language accumulated by Hacker News over the past two decades.
So we can see that the Top list of these videos is a dozen years old, the first of which is Pinkfong's Sharkbaby, a children's song and dance that repeats the "do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-do-
More precisely, it is a personalized and recommendatory layer built on existing platforms.
Like YouTube, it will face similar limitations in the future. For example, the interest of HN users does not represent all audiences, nor does the taste of the technological community amount to quality in the general sense; recommendations for comments are influenced by age, circles and community preferences.
Some of the actual testers also found that the results of the search would continue to be mixed with irrelevant content and, ultimately, would still need to be manually selected.
There was also a problem with website maintenance, and in his responses he was remembered for his previous paper tool, arxiv-sanity. He said that he also missed it, but it was too hard to maintain it at the time to turn it off; he thought it would be much easier to leave it today.
AI is turning "reverse" into a universal capability.
Indeed, almost everything can be reversed today. From the interfaces and functions of software, to the inversions, transplants and even re-enactment of the classic game, to the reverse television, more and more projects requiring extensive professional development work in the past are beginning to be tried by AI Agent.
Android TV version, hackertube.tv
Strictly speaking, HackerTube does not have a reverse YouTube code, nor does it break its recommended algorithm. Karpathy simply retained YouTube content, recasting a screening, sorting and display mechanism with recommended data accumulated over the years by Hacker News.
Although not reconstructed, the process disintegrates the most influential parts of the video media that affect individual experiences, including how content is found, how it is sequenced, and who is serving whom between the user and the recommended system.
In a way, we can also interpret this as a reverse, which is something more up the top, namely, the form of the product that YouTube represents, the way information is organized, and the relationship between content and users.
The meaning of this change is not limited to the reproduction of an existing product, but it begins to allow everyone to explore the original closed systems, how they can be understood, modified and created again.
For content platforms such as YouTube, Bilibi, WeChatWeChat Official Account, we may still need to rely on the content and infrastructure they provide, but we may not have to fully accept the platform ' s designed information portal.
The way we use the Internet can also change as reverse capabilities expand from the code layer to the product, information structure and even the media layer.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.