Chinese content近10分钟动画梳理ChatGPT之后AI演进：开放模型、推理模型与智能体
一条10分钟动画把ChatGPT之后的开放模型、推理模型到智能体串成一条线，想快速补全AI演进脉络的可以看看。
Latest
Every published item, in order · 4309 items
Sep 27
CMU Releases AI Agents Course 11-768 with 9 Lectures Online
Free access to a full graduate-level agent curriculum from two researchers behind major agent frameworks makes this a structured alternative to scattered blog-post learning. Practitioners building agent systems can use the videos and cmu-agents.com materials as a self-study path, working through the 9 published lectures before the rest of the term's recordings land.
Opus 5.5 Draws a BaoCut Promo Video Frame by Frame
The takeaway is that coding agents can double as animation pipelines: deterministic Canvas rendering plus a beat table gives Opus 5.5 a feedback loop where keyframe stills can be self-checked before a full render. Developers who want AI-made promo assets without a motion-design tool can adapt this prompt structure, grounding the model in real code and specs to avoid hallucinated features. It also shows Claude Code reusing local CLIs like Codex without API keys.
Chinese contentOpus 5.5 用 JS Canvas 生成 App Icon 实操分享
用 Opus 5.5 一句话提示词就能画 App Icon，关键是让它写 JS 画 Canvas 而不是 SVG，作者还附了真实迭代过程。
Chinese content归藏分享 Opus 5.5 游戏结算与抽卡动画提示词
归藏放出了他做游戏结算和抽卡动画的完整提示词，用 Opus 5.5 就能跑出 3D 光效动画，做前端的可以拿去直接改。
Chinese content12分钟动画串讲NLP演进：从人工规则到Transformer
用12分钟动画把NLP从规则法讲到Transformer和大模型，每代方法的优缺点都点到了，想补基础概念的可以看看。
Harrison Chase argues probability thresholds beat LLM judgment in agents
The point reframes agent reliability as an engineering problem: convert model judgment into numeric scores so thresholds, logs and eval tests become possible. Teams building production agents can adopt this pattern to make decisions auditable and testable. It contrasts with opaque end-to-end prompting where behavior is hard to regression-test.
Chinese content用 Grok 机器人给宠物生成 Apple 风格广告
有人用 Grok 做了个机器人，传几张宠物照片就能生成 Apple 发布会风格的广告片，我同事家的猫已经当上产品代言人了。
Ollaya serves open decision models locally with TypeSafe-compatible API
Decision models are a lighter, cheaper alternative to calling a full LLM for routing and classification tasks, and Ollaya makes them as easy to run locally as Ollama made general models. Developers building LLM routers, guardrails or zero-shot classifiers can drop it in behind a TypeSafe-compatible API without cloud dependencies. GLiClass and NLI-style models offer a concrete comparison point against embedding-based classifiers.
Chinese content用 Codex 的 Computer Use 监控 Claude Code 工作进度
教你让 Codex 用 Computer Use 盯着 Claude Code 干活，完事还能自动发插件商店，这个跨工具工作流挺有意思。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese content宝玉分享提示词经验：给规范不给模板，让模型自由发挥
宝玉聊了个很实用的经验：别给模型塞模板，给设计规范让它自由发挥。配 Opus 5.5 只给 TTS 和 ffmpeg 就能做视频的例子，做提示词的都该看看。
Chinese contentSimon Willison 公开 Claude 生成互动页面 kakapo-party 的提示词与完整记录
Simon Willison 把让 Claude 做互动页面的提示词和全程记录都放出来了，想学怎么写提示词的可以直接抄作业。
Chinese contentGarry Tan 分享用 Capy AI 的 GStack /autoplan 修复线上 bug 流程
Garry Tan 晒了他的修 bug 新姿势：Capy AI 的 GStack 加 /autoplan，背后跑 GPT-6，直接处理线上问题，做开发的可以看看。
Chinese content用 Claude Opus 5.5 写代码逐帧渲染《中华文明史》编年片
有人拿 Claude Opus 5.5 逐帧渲染了一支中华文明史动画，提示词里连配乐 BPM 和纹样演变都写清楚了，照着抄就能玩。
Chinese content用一段提示词让 Claude Code 把 DINOv3 讲成科普视频
宝玉晒了两个用 Claude Code 做的科普视频，Transformer 和 DINOv3 各一个，提示词原文都贴出来了，直接抄走就能让 AI 帮你做动画讲解视频。
Builders urged to learn engineering fundamentals before using agents
Agent outputs can look correct while hiding suboptimal or fragile solutions, so evaluation and engineering judgement stay essential even as coding assistants improve. Developers and technical leads adopting tools like Claude Code should invest in fundamentals — evals, architecture and concurrency — rather than delegating decisions entirely. The exchange offers a practical counterpoint to messaging that coding skills are no longer needed.
Jev Router Experiment Shows Benchmark Limits in Agent Harness Testing
Independent harness tests like this show why vendor benchmark numbers often fail to predict task-level performance for your own workflows. Builders evaluating model routers or agent harnesses can take the concrete lesson that results depend heavily on the test setup, so replicating with your own workloads matters more than leaderboard rankings. The System One versus System Two framing offers a practical lens for deciding when to route cheaply and when to pay for heavier reasoning.
Chinese content用 Astra 加 Seedance 2.5 不到 1 小时生成角色视频
有人用 Astra 加 Seedance 2.5，一小时做出角色视频，分镜板都公开了，想搞 AI 视频的可以抄这个流程。
Chinese contentChamath 分享他定制的 AI agent 用法：谈判、人生决策
Chamath 自己讲他怎么用定制 AI agent，比如谈判前算纳什均衡，还把基础模型包成了 microservices，思路挺具体。
Chinese content给智能体写行为指令，通过 adapter 持续优化工作方式
svpino 分享了个实用小技巧：直接用自然语言告诉智能体该怎么干活，配合 adapter 用在哪类 agent 框架上都行，还能越调越贴合你的习惯。
Chinese contentAndrew Ng：AI 项目不同阶段应采用不同的工程与测试策略
Andrew Ng 谈 AI 工程里最容易被搞砸的一环：早期项目别急着上 rigid 测试，怎么按阶段调整评估和架构，写得挺实操。
Chinese contentNotebookLM 接入 Gemini 3.8 TTS，中文语音自然，还能快速产出视频
NotebookLM 升级了 Gemini 3.8 TTS，中文终于自然了，推文还顺带整理了四套 AI 做视频的完整路线，做知识类视频的可以直接抄作业。
Sep 26
Chinese contentHamel Husain 分享合成数据生成方法：先组合场景再生成查询
Hamel Husain 讲了个实用的合成数据技巧：先列场景组合再生成查询，比直接让模型瞎编靠谱多了。
Chinese content用 Claude Opus 5.5 一次性生成 1000 张 matplotlib 图表
goodside 让 Claude Opus 5.5 一口气画了 1000 张 matplotlib 图，想看模型批量出图长什么样可以点进来看看。
Chinese contentGemini 3.8 中文写作体验如何？作者求助是否该试试 Antigravity
有人用 Gemini 3.8 写中文稿感觉比 agent 味淡，纠结要不要换 Google 的 Antigravity，评论区有真实使用反馈
Chinese content《Plan Mode 已死》：强模型时代不再需要先写规格文档
一篇 HN 热文说 plan 模式过时了，强模型边做边改就行，作者自己都卸了一堆 skills，写代码的可以看看
Chinese content用 GPT-6 调用 Hyper3D Rodin MCP，零代码做赛车游戏
有人用 GPT-6 加 Hyper3D Rodin 的 MCP，两句提示词就做出 3D 赛车游戏，一行代码没写，教程步骤都列出来了。
Sebastian Raschka's Reasoning From Scratch Video 5 Covers Log-Probability Scoring
Log-probability scoring underlies loss functions like cross-entropy in pre-training and distillation, so the material transfers well beyond answer ranking. Practitioners building inference-time scaling or self-refinement pipelines can follow along with the timestamped PyTorch implementations and the MATH-500 evaluation setup. It is a hands-on complement to reading papers on reasoning methods.
Chinese contentRaycast AI 登顶 Product Hunt 榜一，附生成中文信息图的提示词
作者把一段现成的 Raycast 提示词贴出来了，选中文字就能自动生成中文信息图，拿来就能用。
Chinese contentMatt Pocock 开源一套 Claude Skill 工作流，从对齐需求到 code review 全覆盖
Matt Pocock 把 grill-me 背后的整套 Skill 开源了，从聊需求到 code review 一步一个，靠文档驱动换 Agent 也不怕，直接抄这套工作流。
Chinese content开发者分享用满 Claude 和 Codex 订阅额度的真实体验
一个重度开发者晒出每月上千刀 AI 订阅的用法：Opus 5.5 写产品、GPT-6 做评审、Grok 4.7 清 issue，附带了真实的工具搭配经验。
Chinese content爆火 AI 创业公司的 7 个增长模式拆解
一篇实操向增长拆解，把 Higgsfield 和 ElevenLabs 的打法归纳成 7 条，做 AI 产品投放和获客的可以照着对照自查。
Chinese contentGergely Orosz：AI 让 hack 式开发更容易，但那不等于工程
Gergely Orosz 发了条推，说 vibe coding 就是从前抄 StackOverflow 那套，AI 只是让它更快，算不上工程，但有些活儿也够用了。
Chinese content突然爆发的 AI 创业公司往往在重复同一套增长模式
总结了爆发型 AI 创业公司的发布节奏，比如 Higgsfield 每周发 2-3 次，想做增长的人可以对照自己公司的频率看看。
Chinese contentMuse 连接器体验：B 站、Steam、DeepSeek 一键接入
作者实测了 Muse 接 B 站、Steam、DeepSeek 的流程，扫码、填 token 就搞定，还把 DeepSeek Harness 装进它的虚拟机查国内内容，玩法可以抄。
Open-source Skill teaches Claude Code APK reverse engineering
The interesting part is the workflow design: instead of shipping binaries, the Skill encodes expert decision-making and lets the agent provision its own toolchain, which is a reusable pattern for other domain skills. Reverse engineers and mobile security researchers can plug it into Claude Code or Codex and get a working APK analysis pipeline. Being MIT-licensed, it is free to fork and adapt for internal triage work.
Only the latest 2000 items are listed; for older ones see the daily archive