实践哥MinLi (@MinLiBuilds)

Yang Zhilin has a few mentors worth naming: undergrad with Jie Tang — founder of Zhipu AI; PhD at CMU, co-advised by Russ Salakhutdinov (a Hinton student) and William W. Cohen. Yang is Chinese researchers in China. Stack his academic path against the Chinese researchers at OpenAI and Anthropic in Silicon Valley, and he’s in the same tier: elite undergrad, CMU-level PhD, hard pretraining-era papers, plus big-lab research experience. With Kimi K3 just out, this is a good moment to share Andrew Ng’s short interview with Russ. It was recorded around 2017–2018, while Yang was still a PhD student at CMU. Follow the thread: what kind of advisor produces a researcher like this? You can feel Russ’s energy the whole way — especially worth watching if you care about deep learning. It’s an episode from Andrew Ng’s Heroes of Deep Learning series. At the time, Russ was still Apple’s AI Research Director and a CMU professor. About 15 minutes, fast-paced: how he got into the field, how he thinks about deep learning, and how he chooses between academia and industry. Timestamps: 1:49 — His story with Hinton on the eve of the deep learning revival 7:34 — How his view of deep learning changed, plus supervised vs. unsupervised — still sharp against today’s pretraining 10:46 — Advice for newcomers: PhD or industry for AI, and how he’s moved between both 15:05 — The research directions he was most excited about then ─── Another thread: world models. Yang has also named a unified multimodal world model as a key milestone. Kimi leans more toward long context, knowledge work, and agents operating in the digital world. The piece below is a more spatial line: Playable World — from watching a video to walking into a world. Watch the video and then read the article

1 一个人走进 AI 生成的巷子，说了句召唤——天上真飞下来一只鹰。🦅 这不就是《盗梦空间》么？ Alaya World 干的就是这种事：一句话、一张图丢进去，实时长出一个能逛的世界。你挪视角，街巷往前展开；走到一半打一句提示词，里面真会跟着变。不是后期特效，是当场生成的。 说个暴论：未来游戏开发可能要变天了——先生成一个可玩的 demo，然后就直接上线了。🎮 我们现在用的 AI，多半还是打字进、文字出，最多再给你一张图、一段视频。看完就结束，你始终站在屏幕外面。 可以这么理解： • 💬 聊天模型陪你说话 • 🎬 视频模型给你放片 • 🌍 世界模型开始给你一个能走进去晃两下的地方 它还不完美，偏研究向，家用电脑也跑不动——别当成明天就能下载开玩的游戏。但代码开源了，主页上 Demo 很多，光看就能明白：AI 下一步可能不只是更会聊天，而是开始撑起一个空间。 想直观感受，直接刷 @alayastd 的 Demo，比看任何介绍都快。 1.🐦‍ 召唤火鸟 冷却好了，点火鸟技能。烟尘里冲出一只燃烧的凤凰，双翼一展，脚下黑烟烈焰

天才团队不会被埋没，马斯克是真伯乐。 25 年初 cursor 和 claude 珠联璧合，一个做编程 harness，一个做模型。 25 年初，奥特曼和梁文峰都不喜欢做 coding 场景，一个搞 sora，一个搞有竞争力低成本推理。 25 年初 cursor $20 每月订阅，copilot $10。 当时 cursor 搭配的 opus 3.5-3.7，用的人都反馈贼好，硅谷炸子鸡，一度估值 70 亿美元。 转机发生在 1 季度中旬，claude 大量用了 cursor 的业务数据，蒸馏出了 3.7 以及后面版本，推出 claude code。 cursor 的业务节节败退，增长的导数变成负数，claude 的增长导数是指数级。 25 年真是魔幻之年，下半场，claude gemini codex 都意识到了编程和 long agentic 场景，grok 紧随其后。 前沿 AI 公司的名单上没了 cursor，他们没有自己的模型，没有算力，人们嘲笑他们为 claude 做嫁衣。微调一个 kimi都会被喷。 伟大牛逼人类先知@elonmusk 出手了，他为 cursor 提供了一切支持，就像当年支持 openai 一样。 天才 cursor 团队也交了完美的答卷，恭喜 cursor，恭喜 SpaceXAI

给 Fable 跪下了。 我问他如何查看 lidar 采集的数据，他直接给我搓了一个工具，可以查看全貌可以测量。 放以前，不得搞 2 个月 Fable 还没停，我的戒断反应已经来了

突然发现，之前免费用户为了做点 deep research 各种瞎凑合Gemini、 GPT，问几句就被限速，全浪费时间。 这才是终极的Deep Research工具，登录就能用，旗舰模型，sota，免费。目前体验下来最舒服的产品，一次开出三十几个Agent，又深、又快、又准。 下文我会分享几个常用做 deep research 的场景。 这次我用它来调研十几个港股的打新策略，报告附在后面 👇

On a recent podcast, Anthropic's CEO was asked if you should still learn to code. His answer: "Coding is going away first." The man building the AI just told a generation their plan is obsolete. Watch the whole thing — what he says to do instead is the real point. Timestamps: 00:47 Which careers still have runway in the AI era? 03:28 The most important skill: critical thinking 04:25 Which human skills will AI weaken? 06:38 How should beginners learn Claude Code? I subtitled it hands-free with a free Claude skill I built — beats CapCut, runs local: Me doing the 5%, letting AI do the 95%.