Karpathy 用 Claude 重造 YouTube 推荐
Karpathy 用 Claude Code + Opus 5.5 分析 4300 万条 HN 评论，筛选出 6454 个高质视频做成 HackerTube，无需登录、无算法茧房。AI 工程师可借鉴这种"用 LLM 替代平台算法"的个人化工作流。
这件事的真正信号不是"又出了个新工具"，而是顶级 AI 研究者正在用 LLM 大规模替代互联网平台的原生服务。Karpathy 花 2 小时用 Claude 做出比 YouTube 算法更懂他的推荐系统，成本几乎为零。
这指向一个趋势：平台算法护城河在 LLM 时代被快速侵蚀。做内容聚合、推荐系统的团队需要重新思考产品壁垒在哪里——不再是数据规模，而是用户意图的理解深度和实时性。
对读者的直接行动建议：如果你还在忍受某个平台的垃圾推荐，现在就可以用 Claude Code + 公开数据集（Reddit、HN、Twitter API）搭一个个人版，这比等平台优化快得多。
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: https://t.co/JIhXN7OSh1 - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
查看原文 →RT @heyshrutimishra: Yann LeCun was right & nobody in this industry wants to say it out loud. 109 million parameters beat 72 billion. He…
查看原文 →For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. So we built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on where they most want help from a personal assistant—from finding childcare to accessing government services to managing personal finances. Over the past year, we’ve used LifeVal to measure and improve how well our models handle these everyday tasks. https://t.co/GlUjixzEM0
查看原文 →RT @askalphaxiv: Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve i…
查看原文 →RT @haider1: Yann LeCun says: People in AI have made the same mistake for 70 years: if a machine can do a task humans see as intellectuall…
查看原文 →RT @SkinnerPm: A truly disgraceful public official. A blatant racist and dishonest appeal
查看原文 →RT @JeffDean: Election Tuesday is just two days away. For those eligible to vote in the U.S., please make time to go to the polls to vote…
查看原文 →RT @ylecun: @iamsupersocks Tout cela est correct. Merci. Sauf une chose : j'ai bien créé FAIR en 2013, y compris la branche parisienne en…
查看原文 →RT @KenRoth: The Trump administration is openly meddling in European democracies to promote far-right political parties that support Putin…
查看原文 →Nothing will ever be the same for developers. It's just incredible that you can build a complete application and deploy it in under 5 minutes. @vercel added Python support, and I'm in heaven! https://t.co/ZVDo8lVmTu
查看原文 →This is an open-source tool that will let you carry knowledge between different coding assistants. This is super cool for moving between assistants when you hit a usage limit. Basically, this will record your coding sessions and turn useful fixes, corrections, and workflows into lessons. These lessons will then become available to other coding assistants.
查看原文 →RT @tonyzzhao: This is a great model. Congratulations @DrJimFan and @yunzeman for the awesome work!
查看原文 →This is a new decision model for audio input: • You provide the audio • You provide a list of questions • Model outputs answers + confidence scores The model analyzes the audio directly (no transcript needed). Something really cool is the model's ability to answer questions based on tone, hesitation, laughter, and other audio clues that wouldn't be in a transcript.
查看原文 →I'm adding every account that replies with AI-generated slop to a list, so I can mass-mute and block them all. This is probably an uphill battle, but I'll try to fight back because the replies on this platform are useless right now. I'm using Grok to identify slop, and based on my tests, it does a decent job. To avoid false positives, I'm only blocking accounts that consistently show up as bots. Even if a human reply gets caught up once or twice, the bot is unlikely to classify it incorrectly consistently. I'll share the list later, so you can ask your bot to mute everyone there as well.
查看原文 →