Andrej Karpathy用Opus 5.5构建HackerTube
HackerTube: YouTube built from videos recommended by HN
Karpathy亲自示范如何用AI Agent快速完成从数据采集、清洗到全栈部署的完整闭环，工程思路清晰且代码开源，非常适合想探索AI辅助开发的开发者参考实践。
HackerTube
31 Oct, 2026
2026年10月31日
It was Friday night at 11pm, I was tired, I went to YouTube to watch something and realized that I was looking at an enshittified wall of slop. Insanely exaggerated thumbnails, dumb recommendations of the same tired topics, truly terrible idiocracy-grade content. I'm sure this algorithm works well for the median person on the internet, but it definitely does not work for me. So this morning (Saturday) I took a hold of my Opus 5.5 and gave it the task: Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News. And Opus came back with: hackertube.org!
那是周五晚上11点，我很累，去YouTube想看点什么，结果发现映入眼帘的是一堵满是低质垃圾内容的墙。缩略图夸张得离谱，推荐的都是些老生常谈的无聊话题，内容质量简直糟糕透顶，达到了《蠢蛋进化论》级别的水平。我确信这个算法对互联网上的普通人很有效，但对我肯定没用。于是今天早上（周六），我拿起我的 Opus 5.5，给它布置了这样一个任务：让我们根据 Hacker News 上被推荐过的视频，写一个我们自己的 YouTube 克隆版。Opus 给出的答案是：hackertube.org！
Opus downloaded all comments of HackerNews, filtered them down, ran Sonnet over them in a workflow, got rid of spam and low grade content, and narrowed them down to just 40K comments recommending a YouTube video. In total, we get 6,454 YouTube videos that have been recommended by a few people on HN from 2007-2026 across 150 topics. Then we wrap it up in a YouTube UI, get a domain, push to a VPS, secure, harden, optimize and voila!
Opus 下载了 HackerNews 的所有评论，进行过滤，在 workflow 中用 Sonnet 处理这些评论，剔除垃圾和低质内容，最终将范围缩小到仅包含 40K 条推荐 YouTube 视频的评论。总共，我们获得了 6,454 个 YouTube 视频，这些视频由 HN 用户在 2007-2026 年间针对 150 个主题推荐过。然后我们将它包装成 YouTube 的 UI，获取域名，部署到 VPS，进行安全加固、优化，搞定！
HackerTube:
HackerTube：
- Contains quality content
- Uses the 50% thumbnail instead of the uploaded thumbnail
- Custom and good ForYou recommendations without any user accounts, storing state in local browser session
- Keyboard shortcuts
- Export of the site's data available to give to your agent to easily create your own version
- 包含优质内容
- 使用 50% 缩略图而非上传的缩略图
- 无需任何用户账户即可提供自定义且优质的“为你推荐”推荐，状态存储在本地浏览器会话中
- 键盘快捷键
- 提供网站数据的导出功能，以便交给你的 Agent 轻松创建你自己的版本
This takes ~2 hours now. I can breathe again. Thank you Opus 5.5 <3.
现在这只需要大约 2 小时。我终于能喘口气了。感谢 Opus 5.5 <3。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力