[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-83796":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":17,"stars90d":15,"forks30d":15,"starsTrendScore":18,"compositeScore":19,"rankGlobal":10,"rankLanguage":10,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":45,"readmeContent":46,"aiSummary":10,"trendingCount":15,"starSnapshotCount":15,"syncStatus":47,"lastSyncTime":48,"discoverSource":49},83796,"ai-engineering-from-scratch-zh","fancyboi999\u002Fai-engineering-from-scratch-zh","fancyboi999","Agent工程师最全学习路径 · 从零精通 AI 工程 · 20 阶段 503 课 · 中文全量翻译 + 配套站点 · 如何成为 AI Agent 工程师的修成指南","https:\u002F\u002Faieng-zh.cn",null,"Python",208,36,3,0,56,140,210,94.63,"MIT License",false,"main",true,[25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44],"agents","ai","ai-agents","ai-engineering","chinese","chinese-translation","computer-vision","course","deep-learning","from-scratch","generative-ai","learn-ai","llm","machine-learning","mcp","nlp","python","reinforcement-learning","transformers","tutorial","2026-06-13 04:01:37","\u003Cp align=\"center\">\n  \u003Cimg src=\"assets\u002Fbanner.svg\" alt=\"AI Engineering from Scratch — reference manual banner\" width=\"100%\">\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"LICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-MIT-1a1a1a?style=flat-square&labelColor=fafaf5\" alt=\"MIT License\">\u003C\u002Fa>\n  \u003Ca href=\"ROADMAP.md\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flessons-503-3553ff?style=flat-square&labelColor=fafaf5\" alt=\"503 lessons\">\u003C\u002Fa>\n  \u003Ca href=\"#contents\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fphases-20-3553ff?style=flat-square&labelColor=fafaf5\" alt=\"20 phases\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffancyboi999\u002Fai-engineering-from-scratch-zh\u002Fstargazers\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Ffancyboi999\u002Fai-engineering-from-scratch-zh?style=flat-square&labelColor=fafaf5&color=3553ff\" alt=\"GitHub stars\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Faieng-zh.cn\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fwebsite-live-3553ff?style=flat-square&labelColor=fafaf5\" alt=\"Website\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n> **84% 的学生已经在用 AI 工具，可只有 18% 觉得自己能在专业场景里用好它们。**\n> 这套课程要填的就是这道沟。\n>\n> 503 节课，20 个阶段，约 320 小时。Python、TypeScript、Rust、Julia。每节课都交付一件\n> 能复用的东西：一个提示词、一个技能、一个 agent、一个 MCP server。免费，开源，MIT。\n>\n> 你不只是学 AI，你亲手把它造出来。从头到尾，全手写。\n\n> 本项目是 [AI Engineering from Scratch](https:\u002F\u002Fgithub.com\u002Frohitg00\u002Fai-engineering-from-scratch) 的简体中文翻译版。感谢原作者 [Rohit Ghumare](https:\u002F\u002Fgithub.com\u002Frohitg00) 创作并开源了这套课程。\n\n## How this works\n\n大多数 AI 教材都是碎片化教学。这儿一篇论文，那儿一篇微调心得，别处再来个炫酷的 agent\ndemo。这些碎片很少能拼到一起。你做出了一个聊天机器人，却讲不清它的 loss 曲线；你给\nagent 挂了个函数，却说不出调用它的那个模型内部，attention 到底在干什么。\n\n这套课程就是那根脊椎。20 个阶段，503 节课，四种语言：Python、TypeScript、Rust、Julia。\n一头是线性代数，另一头是自主 agent 集群。每个算法都先从最原始的数学手写出来。反向传播、\n分词器、注意力、agent 循环——等 PyTorch 登场时，你已经知道它底层在做什么了。\n\n每节课都跑同一个循环：读懂问题、推导数学、写代码、跑测试、留下产物。没有五分钟速成视频，\n没有复制粘贴式部署，没有手把手喂饭。免费，开源，在你自己的笔记本上就能跑。\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## The shape of the curriculum\n\n二十个阶段层层叠起来。数学是地基，agent 和生产部署是屋顶。下层的东西你已经会了，就尽管\n往前跳；但别跳过去之后，又回头纳闷上层为什么塌了。\n\n```mermaid\n%%{init: {'theme':'base','themeVariables':{'primaryColor':'#fafaf5','primaryTextColor':'#1a1a1a','primaryBorderColor':'#3553ff','lineColor':'#3553ff','fontFamily':'JetBrains Mono','fontSize':'12px'}}}%%\nflowchart TB\n  P0[\"Phase 0 — Setup & Tooling\"] --> P1[\"Phase 1 — Math Foundations\"]\n  P1 --> P2[\"Phase 2 — ML Fundamentals\"]\n  P2 --> P3[\"Phase 3 — Deep Learning Core\"]\n  P3 --> P4[\"Phase 4 — Vision\"]\n  P3 --> P5[\"Phase 5 — NLP\"]\n  P3 --> P6[\"Phase 6 — Speech & Audio\"]\n  P3 --> P9[\"Phase 9 — RL\"]\n  P5 --> P7[\"Phase 7 — Transformers\"]\n  P7 --> P8[\"Phase 8 — GenAI\"]\n  P7 --> P10[\"Phase 10 — LLMs from Scratch\"]\n  P10 --> P11[\"Phase 11 — LLM Engineering\"]\n  P10 --> P12[\"Phase 12 — Multimodal\"]\n  P11 --> P13[\"Phase 13 — Tools & Protocols\"]\n  P13 --> P14[\"Phase 14 — Agent Engineering\"]\n  P14 --> P15[\"Phase 15 — Autonomous Systems\"]\n  P15 --> P16[\"Phase 16 — Multi-Agent & Swarms\"]\n  P14 --> P17[\"Phase 17 — Infrastructure & Production\"]\n  P15 --> P18[\"Phase 18 — Ethics & Alignment\"]\n  P16 --> P19[\"Phase 19 — Capstone Projects\"]\n  P17 --> P19\n  P18 --> P19\n```\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## The shape of a lesson\n\n每节课都待在自己的文件夹里，整套课程结构统一：\n\n```\nphases\u002F\u003CNN>-\u003Cphase-name>\u002F\u003CNN>-\u003Clesson-name>\u002F\n├── code\u002F      可运行的实现（Python、TypeScript、Rust、Julia）\n├── docs\u002F\n│   └── zh.md  课程正文\n└── outputs\u002F   本节课产出的提示词、技能、agent 或 MCP server\n```\n\n每节课都走六个节拍。其中 *Build It \u002F Use It*（动手构建 \u002F 上手使用）的拆分是整节课的脊椎——\n你先从零实现算法，再用生产级的库把同样的事跑一遍。你之所以懂框架在做什么，是因为那个更小的\n版本你自己写过。\n\n```mermaid\n%%{init: {'theme':'base','themeVariables':{'primaryColor':'#fafaf5','primaryTextColor':'#1a1a1a','primaryBorderColor':'#3553ff','lineColor':'#3553ff','fontFamily':'JetBrains Mono','fontSize':'13px'}}}%%\nflowchart LR\n  M[\"MOTTO\u003Cbr\u002F>\u003Csub>one-line core idea\u003C\u002Fsub>\"] --> Pr[\"PROBLEM\u003Cbr\u002F>\u003Csub>concrete pain\u003C\u002Fsub>\"]\n  Pr --> C[\"CONCEPT\u003Cbr\u002F>\u003Csub>diagrams & intuition\u003C\u002Fsub>\"]\n  C --> B[\"BUILD IT\u003Cbr\u002F>\u003Csub>raw math, no frameworks\u003C\u002Fsub>\"]\n  B --> U[\"USE IT\u003Cbr\u002F>\u003Csub>same thing in PyTorch \u002F sklearn\u003C\u002Fsub>\"]\n  U --> S[\"SHIP IT\u003Cbr\u002F>\u003Csub>prompt · skill · agent · MCP\u003C\u002Fsub>\"]\n```\n\n## Getting started\n\n三种入门方式。挑一个。\n\n**方式 A —— 阅读。** 在\n[aieng-zh.cn](https:\u002F\u002Faieng-zh.cn) 上打开任意一节已完成的课程，\n或展开 [目录](#contents) 里的某个阶段。无需配置，无需 clone。\n\n**方式 B —— clone 下来跑。**\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Ffancyboi999\u002Fai-engineering-from-scratch-zh.git\ncd ai-engineering-from-scratch-zh\npython phases\u002F01-math-foundations\u002F01-linear-algebra-intuition\u002Fcode\u002Fvectors.py\n```\n\n**方式 C —— 测一测你的水平 *(推荐)*。** 聪明地跳级。在 Claude、Cursor、Codex、OpenClaw、Hermes，或任何装了本课程技能的 agent 里：\n\n```bash\n\u002Ffind-your-level\n```\n\n十道题。把你的知识映射到一个起始阶段，生成一条带课时估算的个性化路径。每学完一个阶段：\n\n```bash\n\u002Fcheck-understanding 3        # 测验你对阶段 3 的掌握\nls phases\u002F03-deep-learning-core\u002F05-loss-functions\u002Foutputs\u002F\n# ├── prompt-loss-function-selector.md\n# └── prompt-loss-debugger.md\n```\n\n### 前置要求\n\n- 你会写代码（任何语言都行，会 Python 更好）。\n- 你想搞懂 AI **到底是怎么运作的**，而不只是调调 API。\n\n### 内置 agent 技能（Claude、Cursor、Codex、OpenClaw、Hermes）\n\n| 技能 | 作用 |\n|---|---|\n| [`\u002Ffind-your-level`](.claude\u002Fskills\u002Ffind-your-level\u002FSKILL.md) | 十道题的定级测验。把你的知识映射到一个起始阶段，生成带课时估算的个性化路径。 |\n| [`\u002Fcheck-understanding \u003Cphase>`](.claude\u002Fskills\u002Fcheck-understanding\u002FSKILL.md) | 按阶段测验，八道题，附反馈和需要复习的具体课程。 |\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## Every lesson ships something\n\n别的课程结尾是一句 *\"恭喜，你学会了 X。\"* 这里每节课的结尾，是一件你能直接装上、\n或粘进日常工作流的 **可复用工具**。\n\n\u003Ctable>\n\u003Ctr>\n\u003Cth align=\"left\" width=\"25%\">\u003Cimg src=\"site\u002Fassets\u002Ffigures\u002F001-a-prompts.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.A prompts\"\u002F>\u003Cbr\u002F>\u003Csub>FIG_001 · A\u003C\u002Fsub>\u003Cbr\u002F>\u003Cb>PROMPTS\u003C\u002Fb>\u003C\u002Fth>\n\u003Cth align=\"left\" width=\"25%\">\u003Cimg src=\"site\u002Fassets\u002Ffigures\u002F001-b-skills.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.B skills\"\u002F>\u003Cbr\u002F>\u003Csub>FIG_001 · B\u003C\u002Fsub>\u003Cbr\u002F>\u003Cb>SKILLS\u003C\u002Fb>\u003C\u002Fth>\n\u003Cth align=\"left\" width=\"25%\">\u003Cimg src=\"site\u002Fassets\u002Ffigures\u002F001-c-agents.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.C agents\"\u002F>\u003Cbr\u002F>\u003Csub>FIG_001 · C\u003C\u002Fsub>\u003Cbr\u002F>\u003Cb>AGENTS\u003C\u002Fb>\u003C\u002Fth>\n\u003Cth align=\"left\" width=\"25%\">\u003Cimg src=\"site\u002Fassets\u002Ffigures\u002F001-d-mcp-servers.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.D MCP servers\"\u002F>\u003Cbr\u002F>\u003Csub>FIG_001 · D\u003C\u002Fsub>\u003Cbr\u002F>\u003Cb>MCP SERVERS\u003C\u002Fb>\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd valign=\"top\">粘进任意 AI 助手，在某个细分任务上获得专家级帮助。\u003C\u002Ftd>\n\u003Ctd valign=\"top\">放进 Claude、Cursor、Codex、OpenClaw、Hermes，或任何能读 \u003Ccode>SKILL.md\u003C\u002Fcode> 的 agent。\u003C\u002Ftd>\n\u003Ctd valign=\"top\">作为自主 worker 部署——那个循环你在阶段 14 自己写过。\u003C\u002Ftd>\n\u003Ctd valign=\"top\">接入任意兼容 MCP 的客户端。在阶段 13 里从头到尾构建。\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n> 用 `python3 scripts\u002Finstall_skills.py` 一次性全部安装。是真家伙，不是课后作业。\n> 学完整套课程，你会攒下 435 件作品——你是真懂它们，因为它们都是你亲手造的。\n\n### FIG_002 · 一个实例\n\n阶段 14，第 1 课：agent 循环。约 120 行纯 Python，零依赖。\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd valign=\"top\" width=\"50%\">\n\n**`code\u002Fagent_loop.py`**   \u003Csub>\u003Ci>动手构建\u003C\u002Fi>\u003C\u002Fsub>\n\n```python\ndef run(query, tools):\n    history = [user(query)]\n    for step in range(MAX_STEPS):\n        msg = llm(history)\n        if msg.tool_calls:\n            for call in msg.tool_calls:\n                result = tools[call.name](**call.args)\n                history.append(tool_result(call.id, result))\n            continue\n        return msg.content\n    raise StepLimitExceeded\n```\n\n\u003C\u002Ftd>\n\u003Ctd valign=\"top\" width=\"50%\">\n\n**`outputs\u002Fskill-agent-loop.md`**   \u003Csub>\u003Ci>交付\u003C\u002Fi>\u003C\u002Fsub>\n\n```markdown\n---\nname: agent-loop\ndescription: ReAct-style loop for any tool list\nphase: 14\nlesson: 01\n---\n\nImplement a minimal agent loop that...\n```\n\n**`outputs\u002Fprompt-debug-agent.md`**\n\n```markdown\nYou are an agent debugger. Given the trace\nof an agent run, identify the step where\nthe agent went wrong and explain why...\n```\n\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n\u003Ca id=\"contents\">\u003C\u002Fa>\n\n## Contents\n\n二十个阶段。点开任意阶段即可展开它的课程列表。\n\n\u003Ca id=\"phase-0\">\u003C\u002Fa>\n### Phase 0: 配置与工具链 `12 lessons`\n> 把环境准备好，迎接后面所有的内容。\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [开发环境](phases\u002F00-setup-and-tooling\u002F01-dev-environment\u002F) | Build | Python |\n| 02 | [Git 与协作](phases\u002F00-setup-and-tooling\u002F02-git-and-collaboration\u002F) | Learn | — |\n| 03 | [GPU 配置与云端](phases\u002F00-setup-and-tooling\u002F03-gpu-setup-and-cloud\u002F) | Build | Python |\n| 04 | [API 与密钥](phases\u002F00-setup-and-tooling\u002F04-apis-and-keys\u002F) | Build | Python |\n| 05 | [Jupyter Notebook](phases\u002F00-setup-and-tooling\u002F05-jupyter-notebooks\u002F) | Build | Python |\n| 06 | [Python 环境管理](phases\u002F00-setup-and-tooling\u002F06-python-environments\u002F) | Build | Shell |\n| 07 | [面向 AI 的 Docker](phases\u002F00-setup-and-tooling\u002F07-docker-for-ai\u002F) | Build | Docker |\n| 08 | [编辑器配置](phases\u002F00-setup-and-tooling\u002F08-editor-setup\u002F) | Build | — |\n| 09 | [数据管理](phases\u002F00-setup-and-tooling\u002F09-data-management\u002F) | Build | Python |\n| 10 | [终端与 Shell](phases\u002F00-setup-and-tooling\u002F10-terminal-and-shell\u002F) | Learn | — |\n| 11 | [面向 AI 的 Linux](phases\u002F00-setup-and-tooling\u002F11-linux-for-ai\u002F) | Learn | — |\n| 12 | [调试与性能分析](phases\u002F00-setup-and-tooling\u002F12-debugging-and-profiling\u002F) | Build | Python |\n\n\u003Cdetails id=\"phase-1\">\n\u003Csummary>\u003Cb>Phase 1 — 数学基础\u003C\u002Fb>  \u003Ccode>22 lessons\u003C\u002Fcode>  \u003Cem>每个 AI 算法背后的直觉，用代码讲清楚。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [线性代数直觉](phases\u002F01-math-foundations\u002F01-linear-algebra-intuition\u002F) | Learn | Python, Julia |\n| 02 | [向量、矩阵与运算](phases\u002F01-math-foundations\u002F02-vectors-matrices-operations\u002F) | Build | Python, Julia |\n| 03 | [矩阵变换与特征值](phases\u002F01-math-foundations\u002F03-matrix-transformations\u002F) | Build | Python, Julia |\n| 04 | [机器学习里的微积分：导数与梯度](phases\u002F01-math-foundations\u002F04-calculus-for-ml\u002F) | Learn | Python |\n| 05 | [链式法则与自动微分](phases\u002F01-math-foundations\u002F05-chain-rule-and-autodiff\u002F) | Build | Python |\n| 06 | [概率与分布](phases\u002F01-math-foundations\u002F06-probability-and-distributions\u002F) | Learn | Python |\n| 07 | [贝叶斯定理与统计思维](phases\u002F01-math-foundations\u002F07-bayes-theorem\u002F) | Build | Python |\n| 08 | [优化：梯度下降家族](phases\u002F01-math-foundations\u002F08-optimization\u002F) | Build | Python |\n| 09 | [信息论：熵与 KL 散度](phases\u002F01-math-foundations\u002F09-information-theory\u002F) | Learn | Python |\n| 10 | [降维：PCA、t-SNE、UMAP](phases\u002F01-math-foundations\u002F10-dimensionality-reduction\u002F) | Build | Python |\n| 11 | [奇异值分解](phases\u002F01-math-foundations\u002F11-singular-value-decomposition\u002F) | Build | Python, Julia |\n| 12 | [张量运算](phases\u002F01-math-foundations\u002F12-tensor-operations\u002F) | Build | Python |\n| 13 | [数值稳定性](phases\u002F01-math-foundations\u002F13-numerical-stability\u002F) | Build | Python |\n| 14 | [范数与距离](phases\u002F01-math-foundations\u002F14-norms-and-distances\u002F) | Build | Python |\n| 15 | [机器学习里的统计学](phases\u002F01-math-foundations\u002F15-statistics-for-ml\u002F) | Build | Python |\n| 16 | [采样方法](phases\u002F01-math-foundations\u002F16-sampling-methods\u002F) | Build | Python |\n| 17 | [线性方程组](phases\u002F01-math-foundations\u002F17-linear-systems\u002F) | Build | Python |\n| 18 | [凸优化](phases\u002F01-math-foundations\u002F18-convex-optimization\u002F) | Build | Python |\n| 19 | [面向 AI 的复数](phases\u002F01-math-foundations\u002F19-complex-numbers\u002F) | Learn | Python |\n| 20 | [傅里叶变换](phases\u002F01-math-foundations\u002F20-fourier-transform\u002F) | Build | Python |\n| 21 | [机器学习里的图论](phases\u002F01-math-foundations\u002F21-graph-theory\u002F) | Build | Python |\n| 22 | [随机过程](phases\u002F01-math-foundations\u002F22-stochastic-processes\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-2\">\n\u003Csummary>\u003Cb>Phase 2 — 机器学习基础\u003C\u002Fb>  \u003Ccode>18 lessons\u003C\u002Fcode>  \u003Cem>经典机器学习——至今仍是大多数生产 AI 的骨架。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [什么是机器学习](phases\u002F02-ml-fundamentals\u002F01-what-is-machine-learning\u002F) | Learn | Python |\n| 02 | [从零实现线性回归](phases\u002F02-ml-fundamentals\u002F02-linear-regression\u002F) | Build | Python |\n| 03 | [逻辑回归与分类](phases\u002F02-ml-fundamentals\u002F03-logistic-regression\u002F) | Build | Python |\n| 04 | [决策树与随机森林](phases\u002F02-ml-fundamentals\u002F04-decision-trees\u002F) | Build | Python |\n| 05 | [支持向量机](phases\u002F02-ml-fundamentals\u002F05-support-vector-machines\u002F) | Build | Python |\n| 06 | [KNN 与距离度量](phases\u002F02-ml-fundamentals\u002F06-knn-and-distances\u002F) | Build | Python |\n| 07 | [无监督学习：K-Means、DBSCAN](phases\u002F02-ml-fundamentals\u002F07-unsupervised-learning\u002F) | Build | Python |\n| 08 | [特征工程与特征选择](phases\u002F02-ml-fundamentals\u002F08-feature-engineering\u002F) | Build | Python |\n| 09 | [模型评估：指标与交叉验证](phases\u002F02-ml-fundamentals\u002F09-model-evaluation\u002F) | Build | Python |\n| 10 | [偏差、方差与学习曲线](phases\u002F02-ml-fundamentals\u002F10-bias-variance\u002F) | Learn | Python |\n| 11 | [集成方法：Boosting、Bagging、Stacking](phases\u002F02-ml-fundamentals\u002F11-ensemble-methods\u002F) | Build | Python |\n| 12 | [超参数调优](phases\u002F02-ml-fundamentals\u002F12-hyperparameter-tuning\u002F) | Build | Python |\n| 13 | [机器学习流水线与实验追踪](phases\u002F02-ml-fundamentals\u002F13-ml-pipelines\u002F) | Build | Python |\n| 14 | [朴素贝叶斯](phases\u002F02-ml-fundamentals\u002F14-naive-bayes\u002F) | Build | Python |\n| 15 | [时间序列基础](phases\u002F02-ml-fundamentals\u002F15-time-series\u002F) | Build | Python |\n| 16 | [异常检测](phases\u002F02-ml-fundamentals\u002F16-anomaly-detection\u002F) | Build | Python |\n| 17 | [处理不平衡数据](phases\u002F02-ml-fundamentals\u002F17-imbalanced-data\u002F) | Build | Python |\n| 18 | [特征选择](phases\u002F02-ml-fundamentals\u002F18-feature-selection\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-3\">\n\u003Csummary>\u003Cb>Phase 3 — 深度学习核心\u003C\u002Fb>  \u003Ccode>13 lessons\u003C\u002Fcode>  \u003Cem>从第一性原理出发的神经网络。先自己造一个，再碰框架。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [感知机：一切的起点](phases\u002F03-deep-learning-core\u002F01-the-perceptron\u002F) | Build | Python |\n| 02 | [多层网络与前向传播](phases\u002F03-deep-learning-core\u002F02-multi-layer-networks\u002F) | Build | Python |\n| 03 | [从零实现反向传播](phases\u002F03-deep-learning-core\u002F03-backpropagation\u002F) | Build | Python |\n| 04 | [激活函数：ReLU、Sigmoid、GELU 及其原因](phases\u002F03-deep-learning-core\u002F04-activation-functions\u002F) | Build | Python |\n| 05 | [损失函数：MSE、交叉熵、对比损失](phases\u002F03-deep-learning-core\u002F05-loss-functions\u002F) | Build | Python |\n| 06 | [优化器：SGD、Momentum、Adam、AdamW](phases\u002F03-deep-learning-core\u002F06-optimizers\u002F) | Build | Python |\n| 07 | [正则化：Dropout、权重衰减、BatchNorm](phases\u002F03-deep-learning-core\u002F07-regularization\u002F) | Build | Python |\n| 08 | [权重初始化与训练稳定性](phases\u002F03-deep-learning-core\u002F08-weight-initialization\u002F) | Build | Python |\n| 09 | [学习率调度与 Warmup](phases\u002F03-deep-learning-core\u002F09-learning-rate-schedules\u002F) | Build | Python |\n| 10 | [造一个你自己的迷你框架](phases\u002F03-deep-learning-core\u002F10-mini-framework\u002F) | Build | Python |\n| 11 | [PyTorch 入门](phases\u002F03-deep-learning-core\u002F11-intro-to-pytorch\u002F) | Build | Python |\n| 12 | [JAX 入门](phases\u002F03-deep-learning-core\u002F12-intro-to-jax\u002F) | Build | Python |\n| 13 | [调试神经网络](phases\u002F03-deep-learning-core\u002F13-debugging-neural-networks\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-4\">\n\u003Csummary>\u003Cb>Phase 4 — 计算机视觉\u003C\u002Fb>  \u003Ccode>28 lessons\u003C\u002Fcode>  \u003Cem>从像素到理解——图像、视频、3D、VLM 和世界模型。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [图像基础：像素、通道、色彩空间](phases\u002F04-computer-vision\u002F01-image-fundamentals\u002F) | Learn | Python |\n| 02 | [从零实现卷积](phases\u002F04-computer-vision\u002F02-convolutions-from-scratch\u002F) | Build | Python |\n| 03 | [CNN：从 LeNet 到 ResNet](phases\u002F04-computer-vision\u002F03-cnns-lenet-to-resnet\u002F) | Build | Python |\n| 04 | [图像分类](phases\u002F04-computer-vision\u002F04-image-classification\u002F) | Build | Python |\n| 05 | [迁移学习与微调](phases\u002F04-computer-vision\u002F05-transfer-learning\u002F) | Build | Python |\n| 06 | [目标检测——从零实现 YOLO](phases\u002F04-computer-vision\u002F06-object-detection-yolo\u002F) | Build | Python |\n| 07 | [语义分割——U-Net](phases\u002F04-computer-vision\u002F07-semantic-segmentation-unet\u002F) | Build | Python |\n| 08 | [实例分割——Mask R-CNN](phases\u002F04-computer-vision\u002F08-instance-segmentation-mask-rcnn\u002F) | Build | Python |\n| 09 | [图像生成——GAN](phases\u002F04-computer-vision\u002F09-image-generation-gans\u002F) | Build | Python |\n| 10 | [图像生成——扩散模型](phases\u002F04-computer-vision\u002F10-image-generation-diffusion\u002F) | Build | Python |\n| 11 | [Stable Diffusion——架构与微调](phases\u002F04-computer-vision\u002F11-stable-diffusion\u002F) | Build | Python |\n| 12 | [视频理解——时序建模](phases\u002F04-computer-vision\u002F12-video-understanding\u002F) | Build | Python |\n| 13 | [3D 视觉：点云、NeRF](phases\u002F04-computer-vision\u002F13-3d-vision-nerf\u002F) | Build | Python |\n| 14 | [Vision Transformer（ViT）](phases\u002F04-computer-vision\u002F14-vision-transformers\u002F) | Build | Python |\n| 15 | [实时视觉：边缘部署](phases\u002F04-computer-vision\u002F15-real-time-edge\u002F) | Build | Python |\n| 16 | [构建一条完整的视觉流水线](phases\u002F04-computer-vision\u002F16-vision-pipeline-capstone\u002F) | Build | Python |\n| 17 | [自监督视觉——SimCLR、DINO、MAE](phases\u002F04-computer-vision\u002F17-self-supervised-vision\u002F) | Build | Python |\n| 18 | [开放词表视觉——CLIP](phases\u002F04-computer-vision\u002F18-open-vocab-clip\u002F) | Build | Python |\n| 19 | [OCR 与文档理解](phases\u002F04-computer-vision\u002F19-ocr-document-understanding\u002F) | Build | Python |\n| 20 | [图像检索与度量学习](phases\u002F04-computer-vision\u002F20-image-retrieval-metric\u002F) | Build | Python |\n| 21 | [关键点检测与姿态估计](phases\u002F04-computer-vision\u002F21-keypoint-pose\u002F) | Build | Python |\n| 22 | [从零实现 3D 高斯泼溅](phases\u002F04-computer-vision\u002F22-3d-gaussian-splatting\u002F) | Build | Python |\n| 23 | [Diffusion Transformer 与 Rectified Flow](phases\u002F04-computer-vision\u002F23-diffusion-transformers-rectified-flow\u002F) | Build | Python |\n| 24 | [SAM 3 与开放词表分割](phases\u002F04-computer-vision\u002F24-sam3-open-vocab-segmentation\u002F) | Build | Python |\n| 25 | [视觉语言模型（ViT-MLP-LLM）](phases\u002F04-computer-vision\u002F25-vision-language-models\u002F) | Build | Python |\n| 26 | [单目深度与几何估计](phases\u002F04-computer-vision\u002F26-monocular-depth\u002F) | Build | Python |\n| 27 | [多目标跟踪与视频记忆](phases\u002F04-computer-vision\u002F27-multi-object-tracking\u002F) | Build | Python |\n| 28 | [世界模型与视频扩散](phases\u002F04-computer-vision\u002F28-world-models-video-diffusion\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-5\">\n\u003Csummary>\u003Cb>Phase 5 — NLP：从基础到进阶\u003C\u002Fb>  \u003Ccode>29 lessons\u003C\u002Fcode>  \u003Cem>语言是通往智能的接口。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [文本处理：分词、词干提取、词形还原](phases\u002F05-nlp-foundations-to-advanced\u002F01-text-processing\u002F) | Build | Python |\n| 02 | [词袋、TF-IDF 与文本表示](phases\u002F05-nlp-foundations-to-advanced\u002F02-bag-of-words-tfidf\u002F) | Build | Python |\n| 03 | [词嵌入：从零实现 Word2Vec](phases\u002F05-nlp-foundations-to-advanced\u002F03-word-embeddings-word2vec\u002F) | Build | Python |\n| 04 | [GloVe、FastText 与子词嵌入](phases\u002F05-nlp-foundations-to-advanced\u002F04-glove-fasttext-subword\u002F) | Build | Python |\n| 05 | [情感分析](phases\u002F05-nlp-foundations-to-advanced\u002F05-sentiment-analysis\u002F) | Build | Python |\n| 06 | [命名实体识别（NER）](phases\u002F05-nlp-foundations-to-advanced\u002F06-named-entity-recognition\u002F) | Build | Python |\n| 07 | [词性标注与句法分析](phases\u002F05-nlp-foundations-to-advanced\u002F07-pos-tagging-parsing\u002F) | Build | Python |\n| 08 | [文本分类——用于文本的 CNN 与 RNN](phases\u002F05-nlp-foundations-to-advanced\u002F08-cnns-rnns-for-text\u002F) | Build | Python |\n| 09 | [序列到序列模型](phases\u002F05-nlp-foundations-to-advanced\u002F09-sequence-to-sequence\u002F) | Build | Python |\n| 10 | [注意力机制——那次突破](phases\u002F05-nlp-foundations-to-advanced\u002F10-attention-mechanism\u002F) | Build | Python |\n| 11 | [机器翻译](phases\u002F05-nlp-foundations-to-advanced\u002F11-machine-translation\u002F) | Build | Python |\n| 12 | [文本摘要](phases\u002F05-nlp-foundations-to-advanced\u002F12-text-summarization\u002F) | Build | Python |\n| 13 | [问答系统](phases\u002F05-nlp-foundations-to-advanced\u002F13-question-answering\u002F) | Build | Python |\n| 14 | [信息检索与搜索](phases\u002F05-nlp-foundations-to-advanced\u002F14-information-retrieval-search\u002F) | Build | Python |\n| 15 | [主题建模：LDA、BERTopic](phases\u002F05-nlp-foundations-to-advanced\u002F15-topic-modeling\u002F) | Build | Python |\n| 16 | [文本生成](phases\u002F05-nlp-foundations-to-advanced\u002F16-text-generation-pre-transformer\u002F) | Build | Python |\n| 17 | [聊天机器人：从规则到神经网络](phases\u002F05-nlp-foundations-to-advanced\u002F17-chatbots-rule-to-neural\u002F) | Build | Python |\n| 18 | [多语言 NLP](phases\u002F05-nlp-foundations-to-advanced\u002F18-multilingual-nlp\u002F) | Build | Python |\n| 19 | [子词分词：BPE、WordPiece、Unigram、SentencePiece](phases\u002F05-nlp-foundations-to-advanced\u002F19-subword-tokenization\u002F) | Learn | Python |\n| 20 | [结构化输出与约束解码](phases\u002F05-nlp-foundations-to-advanced\u002F20-structured-outputs-constrained-decoding\u002F) | Build | Python |\n| 21 | [自然语言推理与文本蕴含](phases\u002F05-nlp-foundations-to-advanced\u002F21-nli-textual-entailment\u002F) | Learn | Python |\n| 22 | [嵌入模型深入剖析](phases\u002F05-nlp-foundations-to-advanced\u002F22-embedding-models-deep-dive\u002F) | Learn | Python |\n| 23 | [RAG 的分块策略](phases\u002F05-nlp-foundations-to-advanced\u002F23-chunking-strategies-rag\u002F) | Build | Python |\n| 24 | [指代消解](phases\u002F05-nlp-foundations-to-advanced\u002F24-coreference-resolution\u002F) | Learn | Python |\n| 25 | [实体链接与消歧](phases\u002F05-nlp-foundations-to-advanced\u002F25-entity-linking\u002F) | Build | Python |\n| 26 | [关系抽取与知识图谱构建](phases\u002F05-nlp-foundations-to-advanced\u002F26-relation-extraction-kg\u002F) | Build | Python |\n| 27 | [LLM 评估：RAGAS、DeepEval、G-Eval](phases\u002F05-nlp-foundations-to-advanced\u002F27-llm-evaluation-frameworks\u002F) | Build | Python |\n| 28 | [长上下文评估：NIAH、RULER、LongBench、MRCR](phases\u002F05-nlp-foundations-to-advanced\u002F28-long-context-evaluation\u002F) | Learn | Python |\n| 29 | [对话状态跟踪](phases\u002F05-nlp-foundations-to-advanced\u002F29-dialogue-state-tracking\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-6\">\n\u003Csummary>\u003Cb>Phase 6 — 语音与音频\u003C\u002Fb>  \u003Ccode>17 lessons\u003C\u002Fcode>  \u003Cem>听见、听懂、开口说。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [音频基础：波形、采样、FFT](phases\u002F06-speech-and-audio\u002F01-audio-fundamentals) | Learn | Python |\n| 02 | [频谱图、梅尔刻度与音频特征](phases\u002F06-speech-and-audio\u002F02-spectrograms-mel-features) | Build | Python |\n| 03 | [音频分类](phases\u002F06-speech-and-audio\u002F03-audio-classification) | Build | Python |\n| 04 | [语音识别（ASR）](phases\u002F06-speech-and-audio\u002F04-speech-recognition-asr) | Build | Python |\n| 05 | [Whisper：架构与微调](phases\u002F06-speech-and-audio\u002F05-whisper-architecture-finetuning) | Build | Python |\n| 06 | [说话人识别与验证](phases\u002F06-speech-and-audio\u002F06-speaker-recognition-verification) | Build | Python |\n| 07 | [文本转语音（TTS）](phases\u002F06-speech-and-audio\u002F07-text-to-speech) | Build | Python |\n| 08 | [声音克隆与音色转换](phases\u002F06-speech-and-audio\u002F08-voice-cloning-conversion) | Build | Python |\n| 09 | [音乐生成](phases\u002F06-speech-and-audio\u002F09-music-generation) | Build | Python |\n| 10 | [音频语言模型](phases\u002F06-speech-and-audio\u002F10-audio-language-models) | Build | Python |\n| 11 | [实时音频处理](phases\u002F06-speech-and-audio\u002F11-real-time-audio-processing) | Build | Python |\n| 12 | [搭一条语音助手流水线](phases\u002F06-speech-and-audio\u002F12-voice-assistant-pipeline) | Build | Python |\n| 13 | [神经音频编解码器——EnCodec、SNAC、Mimi、DAC](phases\u002F06-speech-and-audio\u002F13-neural-audio-codecs) | Learn | Python |\n| 14 | [语音活动检测与轮次切换](phases\u002F06-speech-and-audio\u002F14-voice-activity-detection-turn-taking) | Build | Python |\n| 15 | [流式语音到语音——Moshi、Hibiki](phases\u002F06-speech-and-audio\u002F15-streaming-speech-to-speech-moshi-hibiki) | Learn | Python |\n| 16 | [语音防伪与音频水印](phases\u002F06-speech-and-audio\u002F16-anti-spoofing-audio-watermarking) | Build | Python |\n| 17 | [音频评估——WER、MOS、MMAU、排行榜](phases\u002F06-speech-and-audio\u002F17-audio-evaluation-metrics) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-7\">\n\u003Csummary>\u003Cb>Phase 7 — Transformer 深入剖析\u003C\u002Fb>  \u003Ccode>14 lessons\u003C\u002Fcode>  \u003Cem>那个改变了一切的架构。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [为什么用 Transformer：RNN 的问题](phases\u002F07-transformers-deep-dive\u002F01-why-transformers\u002F) | Learn | Python |\n| 02 | [从零实现自注意力](phases\u002F07-transformers-deep-dive\u002F02-self-attention-from-scratch\u002F) | Build | Python |\n| 03 | [多头注意力](phases\u002F07-transformers-deep-dive\u002F03-multi-head-attention\u002F) | Build | Python |\n| 04 | [位置编码：正弦、RoPE、ALiBi](phases\u002F07-transformers-deep-dive\u002F04-positional-encoding\u002F) | Build | Python |\n| 05 | [完整的 Transformer：编码器 + 解码器](phases\u002F07-transformers-deep-dive\u002F05-full-transformer\u002F) | Build | Python |\n| 06 | [BERT——掩码语言建模](phases\u002F07-transformers-deep-dive\u002F06-bert-masked-language-modeling\u002F) | Build | Python |\n| 07 | [GPT——因果语言建模](phases\u002F07-transformers-deep-dive\u002F07-gpt-causal-language-modeling\u002F) | Build | Python |\n| 08 | [T5、BART——编码器-解码器模型](phases\u002F07-transformers-deep-dive\u002F08-t5-bart-encoder-decoder\u002F) | Learn | Python |\n| 09 | [Vision Transformer（ViT）](phases\u002F07-transformers-deep-dive\u002F09-vision-transformers\u002F) | Build | Python |\n| 10 | [音频 Transformer——Whisper 架构](phases\u002F07-transformers-deep-dive\u002F10-audio-transformers-whisper\u002F) | Learn | Python |\n| 11 | [专家混合（MoE）](phases\u002F07-transformers-deep-dive\u002F11-mixture-of-experts\u002F) | Build | Python |\n| 12 | [KV Cache、Flash Attention 与推理优化](phases\u002F07-transformers-deep-dive\u002F12-kv-cache-flash-attention\u002F) | Build | Python |\n| 13 | [缩放定律](phases\u002F07-transformers-deep-dive\u002F13-scaling-laws\u002F) | Learn | Python |\n| 14 | [从零构建一个 Transformer](phases\u002F07-transformers-deep-dive\u002F14-build-a-transformer-capstone\u002F) | Build | Python |\n| 15 | [Attention 变体——滑动窗口、稀疏、差分](phases\u002F07-transformers-deep-dive\u002F15-attention-variants\u002F) | Build | Python |\n| 16 | [投机解码——草稿、验证、重复](phases\u002F07-transformers-deep-dive\u002F16-speculative-decoding\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-8\">\n\u003Csummary>\u003Cb>Phase 8 — 生成式 AI\u003C\u002Fb>  \u003Ccode>14 lessons\u003C\u002Fcode>  \u003Cem>生成图像、视频、音频、3D，等等。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [生成模型：分类与历史](phases\u002F08-generative-ai\u002F01-generative-models-taxonomy-history\u002F) | Learn | Python |\n| 02 | [自编码器与 VAE](phases\u002F08-generative-ai\u002F02-autoencoders-vae\u002F) | Build | Python |\n| 03 | [GAN：生成器 vs 判别器](phases\u002F08-generative-ai\u002F03-gans-generator-discriminator\u002F) | Build | Python |\n| 04 | [条件 GAN 与 Pix2Pix](phases\u002F08-generative-ai\u002F04-conditional-gans-pix2pix\u002F) | Build | Python |\n| 05 | [StyleGAN](phases\u002F08-generative-ai\u002F05-stylegan\u002F) | Build | Python |\n| 06 | [扩散模型——从零实现 DDPM](phases\u002F08-generative-ai\u002F06-diffusion-ddpm-from-scratch\u002F) | Build | Python |\n| 07 | [潜在扩散与 Stable Diffusion](phases\u002F08-generative-ai\u002F07-latent-diffusion-stable-diffusion\u002F) | Build | Python |\n| 08 | [ControlNet、LoRA 与条件控制](phases\u002F08-generative-ai\u002F08-controlnet-lora-conditioning\u002F) | Build | Python |\n| 09 | [图像修复、扩展与编辑](phases\u002F08-generative-ai\u002F09-inpainting-outpainting-editing\u002F) | Build | Python |\n| 10 | [视频生成](phases\u002F08-generative-ai\u002F10-video-generation\u002F) | Build | Python |\n| 11 | [音频生成](phases\u002F08-generative-ai\u002F11-audio-generation\u002F) | Build | Python |\n| 12 | [3D 生成](phases\u002F08-generative-ai\u002F12-3d-generation\u002F) | Build | Python |\n| 13 | [Flow Matching 与 Rectified Flow](phases\u002F08-generative-ai\u002F13-flow-matching-rectified-flows\u002F) | Build | Python |\n| 14 | [评估：FID、CLIP Score](phases\u002F08-generative-ai\u002F14-evaluation-fid-clip-score\u002F) | Build | Python |\n| 19 | [视觉自回归建模（VAR）：下一尺度预测](phases\u002F08-generative-ai\u002F19-visual-autoregressive-var\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-9\">\n\u003Csummary>\u003Cb>Phase 9 — 强化学习\u003C\u002Fb>  \u003Ccode>12 lessons\u003C\u002Fcode>  \u003Cem>RLHF 和会玩游戏的 AI 的基石。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [MDP、状态、动作与奖励](phases\u002F09-reinforcement-learning\u002F01-mdps-states-actions-rewards\u002F) | Learn | Python |\n| 02 | [动态规划](phases\u002F09-reinforcement-learning\u002F02-dynamic-programming\u002F) | Build | Python |\n| 03 | [蒙特卡洛方法](phases\u002F09-reinforcement-learning\u002F03-monte-carlo-methods\u002F) | Build | Python |\n| 04 | [Q-Learning、SARSA](phases\u002F09-reinforcement-learning\u002F04-q-learning-sarsa\u002F) | Build | Python |\n| 05 | [深度 Q 网络（DQN）](phases\u002F09-reinforcement-learning\u002F05-dqn\u002F) | Build | Python |\n| 06 | [策略梯度——REINFORCE](phases\u002F09-reinforcement-learning\u002F06-policy-gradients-reinforce\u002F) | Build | Python |\n| 07 | [Actor-Critic——A2C、A3C](phases\u002F09-reinforcement-learning\u002F07-actor-critic-a2c-a3c\u002F) | Build | Python |\n| 08 | [PPO](phases\u002F09-reinforcement-learning\u002F08-ppo\u002F) | Build | Python |\n| 09 | [奖励建模与 RLHF](phases\u002F09-reinforcement-learning\u002F09-reward-modeling-rlhf\u002F) | Build | Python |\n| 10 | [多智能体强化学习](phases\u002F09-reinforcement-learning\u002F10-multi-agent-rl\u002F) | Build | Python |\n| 11 | [仿真到现实的迁移](phases\u002F09-reinforcement-learning\u002F11-sim-to-real-transfer\u002F) | Build | Python |\n| 12 | [游戏中的强化学习](phases\u002F09-reinforcement-learning\u002F12-rl-for-games\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-10\">\n\u003Csummary>\u003Cb>Phase 10 — 从零实现 LLM\u003C\u002Fb>  \u003Ccode>22 lessons\u003C\u002Fcode>  \u003Cem>构建、训练并真正理解大语言模型。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [分词器：BPE、WordPiece、SentencePiece](phases\u002F10-llms-from-scratch\u002F01-tokenizers\u002F) | Build | Python, Rust |\n| 02 | [从零实现一个分词器](phases\u002F10-llms-from-scratch\u002F02-building-a-tokenizer\u002F) | Build | Python |\n| 03 | [预训练的数据流水线](phases\u002F10-llms-from-scratch\u002F03-data-pipelines\u002F) | Build | Python |\n| 04 | [预训练一个迷你 GPT（124M）](phases\u002F10-llms-from-scratch\u002F04-pre-training-mini-gpt\u002F) | Build | Python |\n| 05 | [分布式训练、FSDP、DeepSpeed](phases\u002F10-llms-from-scratch\u002F05-scaling-distributed\u002F) | Build | Python |\n| 06 | [指令微调——SFT](phases\u002F10-llms-from-scratch\u002F06-instruction-tuning-sft\u002F) | Build | Python |\n| 07 | [RLHF——奖励模型 + PPO](phases\u002F10-llms-from-scratch\u002F07-rlhf\u002F) | Build | Python |\n| 08 | [DPO——直接偏好优化](phases\u002F10-llms-from-scratch\u002F08-dpo\u002F) | Build | Python |\n| 09 | [Constitutional AI 与自我改进](phases\u002F10-llms-from-scratch\u002F09-constitutional-ai-self-improvement\u002F) | Build | Python |\n| 10 | [评估——基准与 evals](phases\u002F10-llms-from-scratch\u002F10-evaluation\u002F) | Build | Python |\n| 11 | [量化：INT8、GPTQ、AWQ、GGUF](phases\u002F10-llms-from-scratch\u002F11-quantization\u002F) | Build | Python |\n| 12 | [推理优化](phases\u002F10-llms-from-scratch\u002F12-inference-optimization\u002F) | Build | Python |\n| 13 | [搭一条完整的 LLM 流水线](phases\u002F10-llms-from-scratch\u002F13-building-complete-llm-pipeline\u002F) | Build | Python |\n| 14 | [开源模型：架构逐一拆解](phases\u002F10-llms-from-scratch\u002F14-open-models-architecture-walkthroughs\u002F) | Learn | Python |\n| 15 | [投机解码与 EAGLE-3](phases\u002F10-llms-from-scratch\u002F15-speculative-decoding-eagle3\u002F) | Build | Python |\n| 16 | [差分注意力（V2）](phases\u002F10-llms-from-scratch\u002F16-differential-attention-v2\u002F) | Build | Python |\n| 17 | [原生稀疏注意力（DeepSeek NSA）](phases\u002F10-llms-from-scratch\u002F17-native-sparse-attention\u002F) | Build | Python |\n| 18 | [多 token 预测（MTP）](phases\u002F10-llms-from-scratch\u002F18-multi-token-prediction\u002F) | Build | Python |\n| 19 | [DualPipe 并行](phases\u002F10-llms-from-scratch\u002F19-dualpipe-parallelism\u002F) | Learn | Python |\n| 20 | [DeepSeek-V3 架构拆解](phases\u002F10-llms-from-scratch\u002F20-deepseek-v3-walkthrough\u002F) | Learn | Python |\n| 21 | [Jamba——SSM-Transformer 混合架构](phases\u002F10-llms-from-scratch\u002F21-jamba-hybrid-ssm-transformer\u002F) | Learn | Python |\n| 22 | [异步与 Hogwild! 推理](phases\u002F10-llms-from-scratch\u002F22-async-hogwild-inference\u002F) | Build | Python |\n| 25 | [推测解码与 EAGLE](phases\u002F10-llms-from-scratch\u002F25-speculative-decoding\u002F) | Build | Python |\n| 34 | [梯度检查点与激活重算](phases\u002F10-llms-from-scratch\u002F34-gradient-checkpointing\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-11\">\n\u003Csummary>\u003Cb>Phase 11 — LLM 工程\u003C\u002Fb>  \u003Ccode>17 lessons\u003C\u002Fcode>  \u003Cem>让 LLM 在生产环境里干活。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [提示工程：技巧与套路](phases\u002F11-llm-engineering\u002F01-prompt-engineering\u002F) | Build | Python |\n| 02 | [Few-Shot、CoT、Tree-of-Thought](phases\u002F11-llm-engineering\u002F02-few-shot-cot\u002F) | Build | Python |\n| 03 | [结构化输出](phases\u002F11-llm-engineering\u002F03-structured-outputs\u002F) | Build | Python |\n| 04 | [嵌入与向量表示](phases\u002F11-llm-engineering\u002F04-embeddings\u002F) | Build | Python |\n| 05 | [上下文工程](phases\u002F11-llm-engineering\u002F05-context-engineering\u002F) | Build | Python |\n| 06 | [RAG：检索增强生成](phases\u002F11-llm-engineering\u002F06-rag\u002F) | Build | Python |\n| 07 | [进阶 RAG：分块、重排](phases\u002F11-llm-engineering\u002F07-advanced-rag\u002F) | Build | Python |\n| 08 | [用 LoRA 与 QLoRA 微调](phases\u002F11-llm-engineering\u002F08-fine-tuning-lora\u002F) | Build | Python |\n| 09 | [函数调用与工具使用](phases\u002F11-llm-engineering\u002F09-function-calling\u002F) | Build | Python |\n| 10 | [评估与测试](phases\u002F11-llm-engineering\u002F10-evaluation\u002F) | Build | Python |\n| 11 | [缓存、限流与成本](phases\u002F11-llm-engineering\u002F11-caching-cost\u002F) | Build | Python |\n| 12 | [护栏与安全](phases\u002F11-llm-engineering\u002F12-guardrails\u002F) | Build | Python |\n| 13 | [构建一个生产级 LLM 应用](phases\u002F11-llm-engineering\u002F13-production-app\u002F) | Build | Python |\n| 14 | [模型上下文协议（MCP）](phases\u002F11-llm-engineering\u002F14-model-context-protocol\u002F) | Build | Python |\n| 15 | [提示缓存与上下文缓存](phases\u002F11-llm-engineering\u002F15-prompt-caching\u002F) | Build | Python |\n| 16 | [LangGraph：面向 agent 的状态机](phases\u002F11-llm-engineering\u002F16-langgraph-state-machines\u002F) | Build | Python |\n| 17 | [agent 框架的取舍](phases\u002F11-llm-engineering\u002F17-agent-framework-tradeoffs\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-12\">\n\u003Csummary>\u003Cb>Phase 12 — 多模态 AI\u003C\u002Fb>  \u003Ccode>25 lessons\u003C\u002Fcode>  \u003Cem>跨模态地看、听、读、推理——从 ViT 的图块到操作电脑的 agent。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [Vision Transformer 与图块-token 原语](phases\u002F12-multimodal-ai\u002F01-vision-transformer-patch-tokens\u002F) | Learn | Python |\n| 02 | [CLIP 与对比式视觉语言预训练](phases\u002F12-multimodal-ai\u002F02-clip-contrastive-pretraining\u002F) | Build | Python |\n| 03 | [BLIP-2 Q-Former 作为模态桥梁](phases\u002F12-multimodal-ai\u002F03-blip2-qformer-bridge\u002F) | Build | Python |\n| 04 | [Flamingo 与门控交叉注意力](phases\u002F12-multimodal-ai\u002F04-flamingo-gated-cross-attention\u002F) | Learn | Python |\n| 05 | [LLaVA 与视觉指令微调](phases\u002F12-multimodal-ai\u002F05-llava-visual-instruction-tuning\u002F) | Build | Python |\n| 06 | [任意分辨率视觉——Patch-n'-Pack 与 NaFlex](phases\u002F12-multimodal-ai\u002F06-any-resolution-patch-n-pack\u002F) | Build | Python |\n| 07 | [开源权重 VLM 配方：真正要紧的是什么](phases\u002F12-multimodal-ai\u002F07-open-weight-vlm-recipes\u002F) | Learn | Python |\n| 08 | [LLaVA-OneVision：单图、多图、视频](phases\u002F12-multimodal-ai\u002F08-llava-onevision-single-multi-video\u002F) | Build | Python |\n| 09 | [Qwen-VL 家族与动态 FPS 视频](phases\u002F12-multimodal-ai\u002F09-qwen-vl-family-dynamic-fps\u002F) | Learn | Python |\n| 10 | [InternVL3 原生多模态预训练](phases\u002F12-multimodal-ai\u002F10-internvl3-native-multimodal\u002F) | Learn | Python |\n| 11 | [Chameleon 早融合纯 token](phases\u002F12-multimodal-ai\u002F11-chameleon-early-fusion-tokens\u002F) | Build | Python |\n| 12 | [Emu3 用下一 token 预测做生成](phases\u002F12-multimodal-ai\u002F12-emu3-next-token-for-generation\u002F) | Learn | Python |\n| 13 | [Transfusion：自回归 + 扩散](phases\u002F12-multimodal-ai\u002F13-transfusion-autoregressive-diffusion\u002F) | Build | Python |\n| 14 | [Show-o 离散扩散统一架构](phases\u002F12-multimodal-ai\u002F14-show-o-discrete-diffusion-unified\u002F) | Learn | Python |\n| 15 | [Janus-Pro 解耦编码器](phases\u002F12-multimodal-ai\u002F15-janus-pro-decoupled-encoders\u002F) | Build | Python |\n| 16 | [MIO 任意到任意流式](phases\u002F12-multimodal-ai\u002F16-mio-any-to-any-streaming\u002F) | Learn | Python |\n| 17 | [视频语言时序定位](phases\u002F12-multimodal-ai\u002F17-video-language-temporal-grounding\u002F) | Build | Python |\n| 18 | [百万 token 上下文下的长视频](phases\u002F12-multimodal-ai\u002F18-long-video-million-token\u002F) | Build | Python |\n| 19 | [音频语言模型：从 Whisper 到 AF3](phases\u002F12-multimodal-ai\u002F19-audio-language-whisper-to-af3\u002F) | Build | Python |\n| 20 | [Omni 模型：Thinker-Talker 流式](phases\u002F12-multimodal-ai\u002F20-omni-models-thinker-talker\u002F) | Build | Python |\n| 21 | [具身 VLA：RT-2、OpenVLA、π0、GR00T](phases\u002F12-multimodal-ai\u002F21-embodied-vlas-openvla-pi0-groot\u002F) | Learn | Python |\n| 22 | [文档与图表理解](phases\u002F12-multimodal-ai\u002F22-document-diagram-understanding\u002F) | Build | Python |\n| 23 | [ColPali 视觉原生文档 RAG](phases\u002F12-multimodal-ai\u002F23-colpali-vision-native-rag\u002F) | Build | Python |\n| 24 | [多模态 RAG 与跨模态检索](phases\u002F12-multimodal-ai\u002F24-multimodal-rag-cross-modal\u002F) | Build | Python |\n| 25 | [多模态 agent 与操作电脑（综合项目）](phases\u002F12-multimodal-ai\u002F25-multimodal-agents-computer-use\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-13\">\n\u003Csummary>\u003Cb>Phase 13 — 工具与协议\u003C\u002Fb>  \u003Ccode>23 lessons\u003C\u002Fcode>  \u003Cem>AI 与真实世界之间的接口。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [工具接口](phases\u002F13-tools-and-protocols\u002F01-the-tool-interface\u002F) | Learn | Python |\n| 02 | [函数调用深入剖析](phases\u002F13-tools-and-protocols\u002F02-function-calling-deep-dive\u002F) | Build | Python |\n| 03 | [并行与流式工具调用](phases\u002F13-tools-and-protocols\u002F03-parallel-and-streaming-tool-calls\u002F) | Build | Python |\n| 04 | [结构化输出](phases\u002F13-tools-and-protocols\u002F04-structured-output\u002F) | Build | Python |\n| 05 | [工具 Schema 设计](phases\u002F13-tools-and-protocols\u002F05-tool-schema-design\u002F) | Learn | Python |\n| 06 | [MCP 基础](phases\u002F13-tools-and-protocols\u002F06-mcp-fundamentals\u002F) | Learn | Python |\n| 07 | [构建一个 MCP server](phases\u002F13-tools-and-protocols\u002F07-building-an-mcp-server\u002F) | Build | Python |\n| 08 | [构建一个 MCP client](phases\u002F13-tools-and-protocols\u002F08-building-an-mcp-client\u002F) | Build | Python |\n| 09 | [MCP 传输层](phases\u002F13-tools-and-protocols\u002F09-mcp-transports\u002F) | Learn | Python |\n| 10 | [MCP 资源与提示](phases\u002F13-tools-and-protocols\u002F10-mcp-resources-and-prompts\u002F) | Build | Python |\n| 11 | [MCP Sampling](phases\u002F13-tools-and-protocols\u002F11-mcp-sampling\u002F) | Build | Python |\n| 12 | [MCP Roots 与 Elicitation](phases\u002F13-tools-and-protocols\u002F12-mcp-roots-and-elicitation\u002F) | Build | Python |\n| 13 | [MCP 异步任务](phases\u002F13-tools-and-protocols\u002F13-mcp-async-tasks\u002F) | Build | Python |\n| 14 | [MCP Apps](phases\u002F13-tools-and-protocols\u002F14-mcp-apps\u002F) | Build | Python |\n| 15 | [MCP 安全 I——工具投毒](phases\u002F13-tools-and-protocols\u002F15-mcp-security-tool-poisoning\u002F) | Learn | Python |\n| 16 | [MCP 安全 II——OAuth 2.1](phases\u002F13-tools-and-protocols\u002F16-mcp-security-oauth-2-1\u002F) | Build | Python |\n| 17 | [MCP 网关与注册表](phases\u002F13-tools-and-protocols\u002F17-mcp-gateways-and-registries\u002F) | Learn | Python |\n| 18 | [生产环境的 MCP 认证——iii 上的 DCR + JWKS](phases\u002F13-tools-and-protocols\u002F18-mcp-auth-production\u002F) | Build | Python |\n| 19 | [A2A 协议](phases\u002F13-tools-and-protocols\u002F19-a2a-protocol\u002F) | Build | Python |\n| 20 | [OpenTelemetry GenAI](phases\u002F13-tools-and-protocols\u002F20-opentelemetry-genai\u002F) | Build | Python |\n| 21 | [LLM 路由层](phases\u002F13-tools-and-protocols\u002F21-llm-routing-layer\u002F) | Learn | Python |\n| 22 | [Skills 与 Agent SDK](phases\u002F13-tools-and-protocols\u002F22-skills-and-agent-sdks\u002F) | Learn | Python |\n| 23 | [综合项目——工具生态](phases\u002F13-tools-and-protocols\u002F23-capstone-tool-ecosystem\u002F) | Build | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-14\">\n\u003Csummary>\u003Cb>Phase 14 — Agent 工程\u003C\u002Fb>  \u003Ccode>42 lessons\u003C\u002Fcode>  \u003Cem>从第一性原理构建 agent——循环、记忆、规划、框架、基准、生产、工作台。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [Agent 循环](phases\u002F14-agent-engineering\u002F01-the-agent-loop\u002F) | Build | Python |\n| 02 | [ReWOO 与 Plan-and-Execute](phases\u002F14-agent-engineering\u002F02-rewoo-plan-and-execute\u002F) | Build | Python |\n| 03 | [Reflexion 与言语强化学习](phases\u002F14-agent-engineering\u002F03-reflexion-verbal-rl\u002F) | Build | Python |\n| 04 | [Tree of Thoughts 与 LATS](phases\u002F14-agent-engineering\u002F04-tree-of-thoughts-lats\u002F) | Build | Python |\n| 05 | [Self-Refine 与 CRITIC](phases\u002F14-agent-engineering\u002F05-self-refine-and-critic\u002F) | Build | Python |\n| 06 | [工具使用与函数调用](phases\u002F14-agent-engineering\u002F06-tool-use-and-function-calling\u002F) | Build | Python |\n| 07 | [记忆——虚拟上下文与 MemGPT](phases\u002F14-agent-engineering\u002F07-memory-virtual-context-memgpt\u002F) | Build | Python |\n| 08 | [记忆块与睡眠时计算](phases\u002F14-agent-engineering\u002F08-memory-blocks-sleep-time-compute\u002F) | Build | Python |\n| 09 | [混合记忆——Mem0 向量 + 图 + KV](phases\u002F14-agent-engineering\u002F09-hybrid-memory-mem0\u002F) | Build | Python |\n| 10 | [技能库与终身学习——Voyager](phases\u002F14-agent-engineering\u002F10-skill-libraries-voyager\u002F) | Build | Python |\n| 11 | [用 HTN 与进化搜索做规划](phases\u002F14-agent-engineering\u002F11-planning-htn-and-evolutionary\u002F) | Build | Python |\n| 12 | [Anthropic 的工作流模式](phases\u002F14-agent-engineering\u002F12-anthropic-workflow-patterns\u002F) | Build | Python |\n| 13 | [LangGraph——有状态图与持久化执行](phases\u002F14-agent-engineering\u002F13-langgraph-stateful-graphs\u002F) | Build | Python |\n| 14 | [AutoGen v0.4——Actor 模型](phases\u002F14-agent-engineering\u002F14-autogen-actor-model\u002F) | Build | Python |\n| 15 | [CrewAI——基于角色的团队与流程](phases\u002F14-agent-engineering\u002F15-crewai-role-based-crews\u002F) | Build | Python |\n| 16 | [OpenAI Agents SDK——交接、护栏、追踪](phases\u002F14-agent-engineering\u002F16-openai-agents-sdk\u002F) | Build | Python |\n| 17 | [Claude Agent SDK——子 agent 与会话存储](phases\u002F14-agent-engineering\u002F17-claude-agent-sdk\u002F) | Build | Python |\n| 18 | [Agno 与 Mastra——生产级运行时](phases\u002F14-agent-engineering\u002F18-agno-and-mastra-runtimes\u002F) | Learn | Python |\n| 19 | [基准——SWE-bench、GAIA、AgentBench](phases\u002F14-agent-engineering\u002F19-benchmarks-swebench-gaia\u002F) | Learn | Python |\n| 20 | [基准——WebArena 与 OSWorld](phases\u002F14-agent-engineering\u002F20-benchmarks-webarena-osworld\u002F) | Learn | Python |\n| 21 | [操作电脑——Claude、OpenAI CUA、Gemini](phases\u002F14-agent-engineering\u002F21-computer-use-agents\u002F) | Build | Python |\n| 22 | [语音 agent——Pipecat 与 LiveKit](phases\u002F14-agent-engineering\u002F22-voice-agents-pipecat-livekit\u002F) | Build | Python |\n| 23 | [OpenTelemetry GenAI 语义约定](phases\u002F14-agent-engineering\u002F23-otel-genai-conventions\u002F) | Build | Python |\n| 24 | [Agent 可观测性——Langfuse、Phoenix、Opik](phases\u002F14-agent-engineering\u002F24-agent-observability-platforms\u002F) | Learn | Python |\n| 25 | [多 agent 辩论与协作](phases\u002F14-agent-engineering\u002F25-multi-agent-debate\u002F) | Build | Python |\n| 26 | [失败模式——agent 为什么会崩](phases\u002F14-agent-engineering\u002F26-failure-modes-agentic\u002F) | Build | Python |\n| 27 | [提示注入与 PVE 防御](phases\u002F14-agent-engineering\u002F27-prompt-injection-defense\u002F) | Build | Python |\n| 28 | [编排模式——Supervisor、Swarm、分层](phases\u002F14-agent-engineering\u002F28-orchestration-patterns\u002F) | Build | Python |\n| 29 | [生产级运行时——队列、事件、Cron](phases\u002F14-agent-engineering\u002F29-production-runtimes\u002F) | Learn | Python |\n| 30 | [Eval 驱动的 agent 开发](phases\u002F14-agent-engineering\u002F30-eval-driven-agent-development\u002F) | Build | Python |\n| 31 | [Agent 工作台：能力强的模型为什么仍会失败](phases\u002F14-agent-engineering\u002F31-agent-workbench-why-models-fail\u002F) | Learn | Python |\n| 32 | [最小化 agent 工作台](phases\u002F14-agent-engineering\u002F32-minimal-agent-workbench\u002F) | Build | Python |\n| 33 | [把 agent 指令写成可执行约束](phases\u002F14-agent-engineering\u002F33-instructions-as-executable-constraints\u002F) | Build | Python |\n| 34 | [仓库记忆与持久化状态](phases\u002F14-agent-engineering\u002F34-repo-memory-and-state\u002F) | Build | Python |\n| 35 | [给 agent 的初始化脚本](phases\u002F14-agent-engineering\u002F35-initialization-scripts\u002F) | Build | Python |\n| 36 | [范围契约与任务边界](phases\u002F14-agent-engineering\u002F36-scope-contracts\u002F) | Build | Python |\n| 37 | [运行时反馈回路](phases\u002F14-agent-engineering\u002F37-runtime-feedback-loops\u002F) | Build | Python |\n| 38 | [验证关卡](phases\u002F14-agent-engineering\u002F38-verification-gates\u002F) | Build | Python |\n| 39 | [审查 agent：把构建者和评判者分开](phases\u002F14-agent-engineering\u002F39-reviewer-agent\u002F) | Build | Python |\n| 40 | [多会话交接](phases\u002F14-agent-engineering\u002F40-multi-session-handoff\u002F) | Build | Python |\n| 41 | [在真实仓库上跑工作台](phases\u002F14-agent-engineering\u002F41-workbench-for-real-repos\u002F) | Build | Python |\n| 42 | [综合项目：交付一套可复用的 agent 工作台包](phases\u002F14-agent-engineering\u002F42-agent-workbench-capstone\u002F) | Build | Python |\n\n阶段 14 里每节工作台课程（31-42）都附带一份 `mission.md`，在 agent 打开完整课程文档前先给它做简报。\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-15\">\n\u003Csummary>\u003Cb>Phase 15 — 自主系统\u003C\u002Fb>  \u003Ccode>22 lessons\u003C\u002Fcode>  \u003Cem>长程 agent、自我改进，以及 2026 年的安全技术栈。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [从聊天机器人到长程 agent（METR）](phases\u002F15-autonomous-systems\u002F01-long-horizon-agents\u002F) | Learn | Python |\n| 02 | [STaR、V-STaR、Quiet-STaR：自学推理](phases\u002F15-autonomous-systems\u002F02-star-family-reasoning\u002F) | Learn | Python |\n| 03 | [AlphaEvolve：进化式编码 agent](phases\u002F15-autonomous-systems\u002F03-alphaevolve-evolutionary-coding\u002F) | Learn | Python |\n| 04 | [Darwin Gödel Machine：自我修改的 agent](phases\u002F15-autonomous-systems\u002F04-darwin-godel-machine\u002F) | Learn | Python |\n| 05 | [AI Scientist v2：研讨会级别的科研](phases\u002F15-autonomous-systems\u002F05-ai-scientist-v2\u002F) | Learn | Python |\n| 06 | [自动化对齐研究（Anthropic AAR）](phases\u002F15-autonomous-systems\u002F06-automated-alignment-research\u002F) | Learn | Python |\n| 07 | [递归式自我改进：能力 vs 对齐](phases\u002F15-autonomous-systems\u002F07-recursive-self-improvement\u002F) | Learn | Python |\n| 08 | [有界自我改进的设计](phases\u002F15-autonomous-systems\u002F08-bounded-self-improvement\u002F) | Learn | Python |\n| 09 | [自主编码 agent 全景（SWE-bench、CodeAct）](phases\u002F15-autonomous-systems\u002F09-coding-agent-landscape\u002F) | Learn | Python |\n| 10 | [Claude Code 的权限模式与 Auto 模式](phases\u002F15-autonomous-systems\u002F10-claude-code-permission-modes\u002F) | Learn | Python |\n| 11 | [浏览器 agent 与间接提示注入](phases\u002F15-autonomous-systems\u002F11-browser-agents\u002F) | Learn | Python |\n| 12 | [长时运行 agent 的持久化执行](phases\u002F15-autonomous-systems\u002F12-durable-execution\u002F) | Learn | Python |\n| 13 | [动作预算、迭代上限、成本管控](phases\u002F15-autonomous-systems\u002F13-cost-governors\u002F) | Learn | Python |\n| 14 | [急停开关、熔断器、金丝雀 token](phases\u002F15-autonomous-systems\u002F14-kill-switches-canaries\u002F) | Learn | Python |\n| 15 | [人在回路：先提议后提交](phases\u002F15-autonomous-systems\u002F15-propose-then-commit\u002F) | Learn | Python |\n| 16 | [检查点与回滚](phases\u002F15-autonomous-systems\u002F16-checkpoints-rollback\u002F) | Learn | Python |\n| 17 | [Constitutional AI 与规则覆盖](phases\u002F15-autonomous-systems\u002F17-constitutional-ai\u002F) | Learn | Python |\n| 18 | [Llama Guard 与输入\u002F输出分类](phases\u002F15-autonomous-systems\u002F18-llama-guard\u002F) | Learn | Python |\n| 19 | [Anthropic 负责任扩展政策 v3.0](phases\u002F15-autonomous-systems\u002F19-anthropic-rsp\u002F) | Learn | Python |\n| 20 | [OpenAI Preparedness 框架与 DeepMind FSF](phases\u002F15-autonomous-systems\u002F20-openai-preparedness-deepmind-fsf\u002F) | Learn | Python |\n| 21 | [METR 时间跨度与外部评估](phases\u002F15-autonomous-systems\u002F21-metr-external-evaluation\u002F) | Learn | Python |\n| 22 | [CAIS、CAISI 与社会规模风险](phases\u002F15-autonomous-systems\u002F22-cais-caisi-societal-risk\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-16\">\n\u003Csummary>\u003Cb>Phase 16 — 多 agent 与集群\u003C\u002Fb>  \u003Ccode>25 lessons\u003C\u002Fcode>  \u003Cem>协调、涌现，以及集体智能。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [为什么要多 agent](phases\u002F16-multi-agent-and-swarms\u002F01-why-multi-agent\u002F) | Learn | TypeScript |\n| 02 | [FIPA-ACL 传承与言语行为](phases\u002F16-multi-agent-and-swarms\u002F02-fipa-acl-heritage\u002F) | Learn | Python |\n| 03 | [通信协议](phases\u002F16-multi-agent-and-swarms\u002F03-communication-protocols\u002F) | Build | TypeScript |\n| 04 | [多 agent 原语模型](phases\u002F16-multi-agent-and-swarms\u002F04-primitive-model\u002F) | Learn | Python |\n| 05 | [Supervisor \u002F 编排者-worker 模式](phases\u002F16-multi-agent-and-swarms\u002F05-supervisor-orchestrator-pattern\u002F) | Build | Python |\n| 06 | [分层架构与分解漂移](phases\u002F16-multi-agent-and-swarms\u002F06-hierarchical-architecture\u002F) | Learn | Python |\n| 07 | [心智社会与多 agent 辩论](phases\u002F16-multi-agent-and-swarms\u002F07-society-of-mind-debate\u002F) | Build | Python |\n| 08 | [角色专精——规划者 \u002F 批评者 \u002F 执行者 \u002F 验证者](phases\u002F16-multi-agent-and-swarms\u002F08-role-specialization\u002F) | Build | Python |\n| 09 | [并行集群与网络化架构](phases\u002F16-multi-agent-and-swarms\u002F09-parallel-swarm-networks\u002F) | Build | Python |\n| 10 | [群聊与发言人选择](phases\u002F16-multi-agent-and-swarms\u002F10-group-chat-speaker-selection\u002F) | Build | Python |\n| 11 | [交接与例程（无状态编排）](phases\u002F16-multi-agent-and-swarms\u002F11-handoffs-and-routines\u002F) | Build | Python |\n| 12 | [A2A——Agent 到 Agent 协议](phases\u002F16-multi-agent-and-swarms\u002F12-a2a-protocol\u002F) | Build | Python |\n| 13 | [共享记忆与黑板模式](phases\u002F16-multi-agent-and-swarms\u002F13-shared-memory-blackboard\u002F) | Build | Python |\n| 14 | [共识与拜占庭容错](phases\u002F16-multi-agent-and-swarms\u002F14-consensus-and-bft\u002F) | Build | Python |\n| 15 | [投票、自洽性与辩论拓扑](phases\u002F16-multi-agent-and-swarms\u002F15-voting-debate-topology\u002F) | Build | Python |\n| 16 | [协商与议价](phases\u002F16-multi-agent-and-swarms\u002F16-negotiation-bargaining\u002F) | Build | Python |\n| 17 | [生成式 agent 与涌现式仿真](phases\u002F16-multi-agent-and-swarms\u002F17-generative-agents-simulation\u002F) | Build | Python |\n| 18 | [心智理论与涌现式协调](phases\u002F16-multi-agent-and-swarms\u002F18-theory-of-mind-coordination\u002F) | Build | Python |\n| 19 | [群体优化（PSO、ACO）](phases\u002F16-multi-agent-and-swarms\u002F19-swarm-optimization-pso-aco\u002F) | Build | Python |\n| 20 | [MARL——MADDPG、QMIX、MAPPO](phases\u002F16-multi-agent-and-swarms\u002F20-marl-maddpg-qmix-mappo\u002F) | Learn | Python |\n| 21 | [Agent 经济、token 激励、声誉](phases\u002F16-multi-agent-and-swarms\u002F21-agent-economies\u002F) | Learn | Python |\n| 22 | [生产级扩展——队列、检查点、持久性](phases\u002F16-multi-agent-and-swarms\u002F22-production-scaling-queues-checkpoints\u002F) | Build | Python |\n| 23 | [失败模式——MAST、群体思维、单一文化](phases\u002F16-multi-agent-and-swarms\u002F23-failure-modes-mast-groupthink\u002F) | Learn | Python |\n| 24 | [评估与协调基准](phases\u002F16-multi-agent-and-swarms\u002F24-evaluation-coordination-benchmarks\u002F) | Learn | Python |\n| 25 | [案例研究与 2026 最新进展](phases\u002F16-multi-agent-and-swarms\u002F25-case-studies-2026-sota\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-17\">\n\u003Csummary>\u003Cb>Phase 17 — 基础设施与生产\u003C\u002Fb>  \u003Ccode>28 lessons\u003C\u002Fcode>  \u003Cem>把 AI 交付到真实世界。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [托管 LLM 平台 — Bedrock、Azure OpenAI、Vertex AI](phases\u002F17-infrastructure-and-production\u002F01-managed-llm-platforms\u002F) | Learn | Python |\n| 02 | [推理平台经济学 — Fireworks、Together、Baseten、Modal](phases\u002F17-infrastructure-and-production\u002F02-inference-platform-economics\u002F) | Learn | Python |\n| 03 | [Kubernetes 上的 GPU 自动扩缩 — Karpenter、KAI Scheduler](phases\u002F17-infrastructure-and-production\u002F03-gpu-autoscaling-kubernetes\u002F) | Learn | Python |\n| 04 | [vLLM 服务内部机制 — PagedAttention、连续批处理、分块预填充](phases\u002F17-infrastructure-and-production\u002F04-vllm-serving-internals\u002F) | Learn | Python |\n| 05 | [生产环境中的 EAGLE-3 推测解码](phases\u002F17-infrastructure-and-production\u002F05-eagle3-speculative-decoding\u002F) | Learn | Python |\n| 06 | [面向前缀密集型负载的 SGLang 与 RadixAttention](phases\u002F17-infrastructure-and-production\u002F06-sglang-radixattention\u002F) | Learn | Python |\n| 07 | [Blackwell 上用 FP8 与 NVFP4 的 TensorRT-LLM](phases\u002F17-infrastructure-and-production\u002F07-tensorrt-llm-blackwell\u002F) | Learn | Python |\n| 08 | [推理指标 — TTFT、TPOT、ITL、Goodput、P99](phases\u002F17-infrastructure-and-production\u002F08-inference-metrics-goodput\u002F) | Learn | Python |\n| 09 | [生产级量化 — AWQ、GPTQ、GGUF、FP8、NVFP4](phases\u002F17-infrastructure-and-production\u002F09-production-quantization\u002F) | Learn | Python |\n| 10 | [无服务器 LLM 的冷启动缓解](phases\u002F17-infrastructure-and-production\u002F10-cold-start-mitigation\u002F) | Learn | Python |\n| 11 | [多区域 LLM 服务与 KV 缓存局部性](phases\u002F17-infrastructure-and-production\u002F11-multi-region-kv-locality\u002F) | Learn | Python |\n| 12 | [边缘推理 — ANE、Hexagon、WebGPU、Jetson](phases\u002F17-infrastructure-and-production\u002F12-edge-inference\u002F) | Learn | Python |\n| 13 | [LLM 可观测性技术栈选型](phases\u002F17-infrastructure-and-production\u002F13-llm-observability\u002F) | Learn | Python |\n| 14 | [提示缓存与语义缓存的经济学](phases\u002F17-infrastructure-and-production\u002F14-prompt-semantic-caching\u002F) | Learn | Python |\n| 15 | [批处理 API — 50% 折扣作为行业标准](phases\u002F17-infrastructure-and-production\u002F15-batch-apis\u002F) | Learn | Python |\n| 16 | [把模型路由作为降本原语](phases\u002F17-infrastructure-and-production\u002F16-model-routing\u002F) | Learn | Python |\n| 17 | [预填充\u002F解码分离 — NVIDIA Dynamo 与 llm-d](phases\u002F17-infrastructure-and-production\u002F17-disaggregated-prefill-decode\u002F) | Learn | Python |\n| 18 | [带 LMCache KV 卸载的 vLLM 生产栈](phases\u002F17-infrastructure-and-production\u002F18-vllm-production-stack-lmcache\u002F) | Learn | Python |\n| 19 | [AI 网关 — LiteLLM、Portkey、Kong、Bifrost](phases\u002F17-infrastructure-and-production\u002F19-ai-gateways\u002F) | Learn | Python |\n| 20 | [影子、金丝雀与渐进式部署](phases\u002F17-infrastructure-and-production\u002F20-shadow-canary-progressive\u002F) | Learn | Python |\n| 21 | [LLM 功能的 A\u002FB 测试 — GrowthBook 与 Statsig](phases\u002F17-infrastructure-and-production\u002F21-ab-testing-llm-features\u002F) | Learn | Python |\n| 22 | [LLM API 的负载测试 — k6、LLMPerf、GenAI-Perf](phases\u002F17-infrastructure-and-production\u002F22-load-testing-llm-apis\u002F) | Build | Python |\n| 23 | [面向 AI 的 SRE — 多智能体事件响应](phases\u002F17-infrastructure-and-production\u002F23-sre-for-ai\u002F) | Learn | Python |\n| 24 | [面向 LLM 生产的混沌工程](phases\u002F17-infrastructure-and-production\u002F24-chaos-engineering-llm\u002F) | Learn | Python |\n| 25 | [安全 — 密钥、PII 脱敏、审计日志](phases\u002F17-infrastructure-and-production\u002F25-security-secrets-audit\u002F) | Learn | Python |\n| 26 | [合规 — SOC 2、HIPAA、GDPR、EU AI Act、ISO 42001](phases\u002F17-infrastructure-and-production\u002F26-compliance-frameworks\u002F) | Learn | Python |\n| 27 | [面向 LLM 的 FinOps — 单位经济与多租户归因](phases\u002F17-infrastructure-and-production\u002F27-finops-llms\u002F) | Learn | Python |\n| 28 | [自托管服务选型 — llama.cpp、Ollama、TGI、vLLM、SGLang](phases\u002F17-infrastructure-and-production\u002F28-self-hosted-serving-selection\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-18\">\n\u003Csummary>\u003Cb>Phase 18 — 伦理、安全与对齐\u003C\u002Fb>  \u003Ccode>30 lessons\u003C\u002Fcode>  \u003Cem>构建对人类有益的 AI。这不是选修。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [把遵循指令当作对齐信号](phases\u002F18-ethics-safety-alignment\u002F01-instruction-following-alignment-signal\u002F) | Learn | Python |\n| 02 | [奖励黑客与古德哈特定律](phases\u002F18-ethics-safety-alignment\u002F02-reward-hacking-goodhart\u002F) | Learn | Python |\n| 03 | [直接偏好优化家族](phases\u002F18-ethics-safety-alignment\u002F03-direct-preference-optimization-family\u002F) | Learn | Python |\n| 04 | [阿谀奉承：RLHF 的放大效应](phases\u002F18-ethics-safety-alignment\u002F04-sycophancy-rlhf-amplification\u002F) | Learn | Python |\n| 05 | [Constitutional AI 与 RLAIF](phases\u002F18-ethics-safety-alignment\u002F05-constitutional-ai-rlaif\u002F) | Learn | Python |\n| 06 | [Mesa 优化与欺骗性对齐](phases\u002F18-ethics-safety-alignment\u002F06-mesa-optimization-deceptive-alignment\u002F) | Learn | Python |\n| 07 | [潜伏 agent——持续性欺骗](phases\u002F18-ethics-safety-alignment\u002F07-sleeper-agents-persistent-deception\u002F) | Learn | Python |\n| 08 | [前沿模型中的上下文内谋划](phases\u002F18-ethics-safety-alignment\u002F08-in-context-scheming-frontier-models\u002F) | Learn | Python |\n| 09 | [对齐造假](phases\u002F18-ethics-safety-alignment\u002F09-alignment-faking\u002F) | Learn | Python |\n| 10 | [AI Control——即便被颠覆也保安全](phases\u002F18-ethics-safety-alignment\u002F10-ai-control-subversion\u002F) | Learn | Python |\n| 11 | [可扩展监督与弱到强](phases\u002F18-ethics-safety-alignment\u002F11-scalable-oversight-weak-to-strong\u002F) | Learn | Python |\n| 12 | [红队：PAIR 与自动化攻击](phases\u002F18-ethics-safety-alignment\u002F12-red-teaming-pair-automated-attacks\u002F) | Build | Python |\n| 13 | [多样本越狱](phases\u002F18-ethics-safety-alignment\u002F13-many-shot-jailbreaking\u002F) | Learn | Python |\n| 14 | [ASCII 字符画与视觉越狱](phases\u002F18-ethics-safety-alignment\u002F14-ascii-art-visual-jailbreaks\u002F) | Build | Python |\n| 15 | [间接提示注入](phases\u002F18-ethics-safety-alignment\u002F15-indirect-prompt-injection\u002F) | Build | Python |\n| 16 | [红队工具：Garak、Llama Guard、PyRIT](phases\u002F18-ethics-safety-alignment\u002F16-red-team-tooling-garak-llamaguard-pyrit\u002F) | Build | Python |\n| 17 | [WMDP 与双用途能力评估](phases\u002F18-ethics-safety-alignment\u002F17-wmdp-dual-use-evaluation\u002F) | Learn | Python |\n| 18 | [前沿安全框架——RSP、PF、FSF](phases\u002F18-ethics-safety-alignment\u002F18-frontier-safety-frameworks-rsp-pf-fsf\u002F) | Learn | Python |\n| 19 | [模型福祉研究](phases\u002F18-ethics-safety-alignment\u002F19-model-welfare-research\u002F) | Learn | Python |\n| 20 | [偏见与表征伤害](phases\u002F18-ethics-safety-alignment\u002F20-bias-representational-harm\u002F) | Build | Python |\n| 21 | [公平性准则：群体、个体、反事实](phases\u002F18-ethics-safety-alignment\u002F21-fairness-criteria-group-individual-counterfactual\u002F) | Learn | Python |\n| 22 | [面向 LLM 的差分隐私](phases\u002F18-ethics-safety-alignment\u002F22-differential-privacy-for-llms\u002F) | Build | Python |\n| 23 | [水印：SynthID、Stable Signature、C2PA](phases\u002F18-ethics-safety-alignment\u002F23-watermarking-synthid-stable-signature-c2pa\u002F) | Build | Python |\n| 24 | [监管框架：欧盟、美国、英国、韩国](phases\u002F18-ethics-safety-alignment\u002F24-regulatory-frameworks-eu-us-uk-korea\u002F) | Learn | Python |\n| 25 | [EchoLeak 与 AI 的 CVE](phases\u002F18-ethics-safety-alignment\u002F25-echoleak-cves-for-ai\u002F) | Learn | Python |\n| 26 | [模型卡、系统卡与数据集卡](phases\u002F18-ethics-safety-alignment\u002F26-model-system-dataset-cards\u002F) | Build | Python |\n| 27 | [数据溯源与训练数据治理](phases\u002F18-ethics-safety-alignment\u002F27-data-provenance-training-governance\u002F) | Learn | Python |\n| 28 | [对齐研究生态：MATS、Redwood、Apollo、METR](phases\u002F18-ethics-safety-alignment\u002F28-alignment-research-ecosystem\u002F) | Learn | Python |\n| 29 | [内容审核系统：OpenAI、Perspective、Llama Guard](phases\u002F18-ethics-safety-alignment\u002F29-moderation-systems-openai-perspective-llamaguard\u002F) | Build | Python |\n| 30 | [双用途风险：网络、生物、化学、核](phases\u002F18-ethics-safety-alignment\u002F30-dual-use-risk-cyber-bio-chem-nuclear\u002F) | Learn | Python |\n\n\u003C\u002Fdetails>\n\n\u003Cdetails id=\"phase-19\">\n\u003Csummary>\u003Cb>Phase 19 — 综合项目\u003C\u002Fb>  \u003Ccode>85 projects\u003C\u002Fcode>  \u003Cem>2026 年的端到端可交付产品，每个 20-40 小时。\u003C\u002Fem>\u003C\u002Fsummary>\n\u003Cbr\u002F>\n\n| # | Project | Combines | Lang |\n|:---:|---------|----------|------|\n| 01 | [终端原生编码 agent](phases\u002F19-capstone-projects\u002F01-terminal-native-coding-agent\u002F) | P0 P5 P7 P10 P11 P13 P14 P15 P17 P18 | Python |\n| 02 | [代码库 RAG（跨仓库语义搜索）](phases\u002F19-capstone-projects\u002F02-rag-over-codebase\u002F) | P5 P7 P11 P13 P17 | Python |\n| 03 | [实时语音助手（ASR → LLM → TTS）](phases\u002F19-capstone-projects\u002F03-realtime-voice-assistant\u002F) | P6 P7 P11 P13 P14 P17 | Python |\n| 04 | [多模态文档问答（视觉优先）](phases\u002F19-capstone-projects\u002F04-multimodal-document-qa\u002F) | P4 P5 P7 P11 P12 P17 | Python |\n| 05 | [自主科研 agent（AI-Scientist 级别）](phases\u002F19-capstone-projects\u002F05-autonomous-research-agent\u002F) | P0 P2 P3 P7 P10 P14 P15 P16 P18 | Python |\n| 06 | [面向 Kubernetes 的 DevOps 排障 agent](phases\u002F19-capstone-projects\u002F06-devops-troubleshooting-agent\u002F) | P11 P13 P14 P15 P17 P18 | Python |\n| 07 | [端到端微调流水线](phases\u002F19-capstone-projects\u002F07-end-to-end-fine-tuning-pipeline\u002F) | P2 P3 P7 P10 P11 P17 P18 | Python |\n| 08 | [生产级 RAG 聊天机器人（受监管垂直行业）](phases\u002F19-capstone-projects\u002F08-production-rag-chatbot\u002F) | P5 P7 P11 P12 P17 P18 | Python |\n| 09 | [代码迁移 agent（仓库级升级）](phases\u002F19-capstone-projects\u002F09-code-migration-agent\u002F) | P5 P7 P11 P13 P14 P15 P17 | Python |\n| 10 | [多 agent 软件工程团队](phases\u002F19-capstone-projects\u002F10-multi-agent-software-team\u002F) | P11 P13 P14 P15 P16 P17 | Python |\n| 11 | [LLM 可观测性与 Eval 仪表盘](phases\u002F19-capstone-projects\u002F11-llm-observability-dashboard\u002F) | P11 P13 P17 P18 | Python |\n| 12 | [视频理解流水线（场景 → 问答）](phases\u002F19-capstone-projects\u002F12-video-understanding-pipeline\u002F) | P4 P6 P7 P11 P12 P17 | Python |\n| 13 | [带注册表与治理的 MCP server](phases\u002F19-capstone-projects\u002F13-mcp-server-with-registry\u002F) | P11 P13 P14 P17 P18 | Python |\n| 14 | [投机解码推理服务器](phases\u002F19-capstone-projects\u002F14-speculative-decoding-server\u002F) | P3 P7 P10 P17 | Python |\n| 15 | [Constitutional 安全测试架 + 红队靶场](phases\u002F19-capstone-projects\u002F15-constitutional-safety-harness\u002F) | P10 P11 P13 P14 P18 | Python |\n| 16 | [GitHub Issue 到 PR 的自主 agent](phases\u002F19-capstone-projects\u002F16-github-issue-to-pr-agent\u002F) | P11 P13 P14 P15 P17 | Python |\n| 17 | [个人 AI 导师（自适应、多模态）](phases\u002F19-capstone-projects\u002F17-personal-ai-tutor\u002F) | P5 P6 P11 P12 P14 P17 P18 | Python |\n| 20 | [Agent Harness Loop 契约](phases\u002F19-capstone-projects\u002F20-agent-harness-loop-contract\u002F) | A. Agent harness | Python |\n| 21 | [带 Schema 校验的 Tool Registry](phases\u002F19-capstone-projects\u002F21-tool-registry-schema-validation\u002F) | A. Agent harness | Python |\n| 22 | [基于换行分隔 stdio 的 JSON-RPC 2.0](phases\u002F19-capstone-projects\u002F22-jsonrpc-stdio-transport\u002F) | A. Agent harness | Python |\n| 23 | [Function Call Dispatcher](phases\u002F19-capstone-projects\u002F23-function-call-dispatcher\u002F) | A. Agent harness | Python |\n| 24 | [Plan-Execute 控制流](phases\u002F19-capstone-projects\u002F24-plan-execute-control-flow\u002F) | A. Agent harness | Python |\n| 25 | [Verification Gate 与 Observation Budget](phases\u002F19-capstone-projects\u002F25-verification-gates-observation-budget\u002F) | A. Agent harness | Python |\n| 26 | [带 Denylist 与 Path Jail 的 Sandbox Runner](phases\u002F19-capstone-projects\u002F26-sandbox-runner-denylist\u002F) | A. Agent harness | Python |\n| 27 | [带 Fixture Tasks 的 Eval Harness](phases\u002F19-capstone-projects\u002F27-eval-harness-fixture-tasks\u002F) | A. Agent harness | Python |\n| 28 | [用 OTel GenAI Span 与 Prometheus 做 Observability](phases\u002F19-capstone-projects\u002F28-observability-otel-traces\u002F) | A. Agent harness | Python |\n| 29 | [端到端 Coding Agent Demo](phases\u002F19-capstone-projects\u002F29-end-to-end-coding-task-demo\u002F) | A. Agent harness | Python |\n| 30 | [从零实现 BPE Tokenizer](phases\u002F19-capstone-projects\u002F30-bpe-tokenizer-from-scratch\u002F) | B. NLP LLM | Python |\n| 31 | [带 Sliding Window 的 Tokenized Dataset](phases\u002F19-capstone-projects\u002F31-tokenized-dataset-sliding-window\u002F) | B. NLP LLM | Python |\n| 32 | [Token Embedding 与 Positional Embedding](phases\u002F19-capstone-projects\u002F32-token-positional-embeddings\u002F) | B. NLP LLM | Python |\n| 33 | [Multi-Head Self-Attention](phases\u002F19-capstone-projects\u002F33-multihead-self-attention\u002F) | B. NLP LLM | Python |\n| 34 | [从零实现 Transformer Block](phases\u002F19-capstone-projects\u002F34-transformer-block\u002F) | B. NLP LLM | Python |\n| 35 | [GPT 模型组装](phases\u002F19-capstone-projects\u002F35-gpt-model-assembly\u002F) | B. NLP LLM | Python |\n| 36 | [训练循环与评估](phases\u002F19-capstone-projects\u002F36-training-loop-eval\u002F) | B. NLP LLM | Python |\n| 37 | [加载预训练权重](phases\u002F19-capstone-projects\u002F37-loading-pretrained-weights\u002F) | B. NLP LLM | Python |\n| 38 | [通过换 Head 做分类微调](phases\u002F19-capstone-projects\u002F38-classifier-finetuning\u002F) | B. NLP LLM | Python |\n| 39 | [通过 SFT 做 Instruction Tuning](phases\u002F19-capstone-projects\u002F39-instruction-tuning-sft\u002F) | B. NLP LLM | Python |\n| 40 | [从零实现 DPO](phases\u002F19-capstone-projects\u002F40-dpo-from-scratch\u002F) | B. NLP LLM | Python |\n| 41 | [完整评估流水线](phases\u002F19-capstone-projects\u002F41-eval-pipeline\u002F) | B. NLP LLM | Python |\n| 42 | [大规模语料下载器](phases\u002F19-capstone-projects\u002F42-large-corpus-downloader\u002F) | C. 端到端训练 | Python |\n| 43 | [HDF5 Tokenized Corpus](phases\u002F19-capstone-projects\u002F43-hdf5-tokenized-corpus\u002F) | C. 端到端训练 | Python |\n| 44 | [Cosine 学习率 + 线性 Warmup](phases\u002F19-capstone-projects\u002F44-cosine-lr-warmup\u002F) | C. 端到端训练 | Python |\n| 45 | [Gradient Clipping 与混合精度训练](phases\u002F19-capstone-projects\u002F45-gradient-clipping-amp\u002F) | C. 端到端训练 | Python |\n| 46 | [梯度累积](phases\u002F19-capstone-projects\u002F46-gradient-accumulation\u002F) | C. 端到端训练 | Python |\n| 47 | [Checkpoint 保存与恢复](phases\u002F19-capstone-projects\u002F47-checkpoint-save-resume\u002F) | C. 端到端训练 | Python |\n| 48 | [从零实现分布式数据并行与 FSDP](phases\u002F19-capstone-projects\u002F48-distributed-fsdp-ddp\u002F) | C. 端到端训练 | Python |\n| 49 | [语言模型评测框架](phases\u002F19-capstone-projects\u002F49-lm-eval-harness\u002F) | C. 端到端训练 | Python |\n| 50 | [假设生成器](phases\u002F19-capstone-projects\u002F50-hypothesis-generator\u002F) | D. 自动研究 | Python |\n| 51 | [文献检索](phases\u002F19-capstone-projects\u002F51-literature-retrieval\u002F) | D. 自动研究 | Python |\n| 52 | [实验执行器](phases\u002F19-capstone-projects\u002F52-experiment-runner\u002F) | D. 自动研究 | Python |\n| 53 | [结果评估器](phases\u002F19-capstone-projects\u002F53-result-evaluator\u002F) | D. 自动研究 | Python |\n| 54 | [论文生成器](phases\u002F19-capstone-projects\u002F54-paper-writer\u002F) | D. 自动研究 | Python |\n| 55 | [评审循环](phases\u002F19-capstone-projects\u002F55-critic-loop\u002F) | D. 自动研究 | Python |\n| 56 | [迭代调度器](phases\u002F19-capstone-projects\u002F56-iteration-scheduler\u002F) | D. 自动研究 | Python |\n| 57 | [端到端研究 Demo](phases\u002F19-capstone-projects\u002F57-end-to-end-research-demo\u002F) | D. 自动研究 | Python |\n| 58 | [Vision Encoder 的 Patch 切分](phases\u002F19-capstone-projects\u002F58-vision-encoder-patches\u002F) | E. 多模态 | Python |\n| 59 | [Vision Transformer Encoder（ViT）](phases\u002F19-capstone-projects\u002F59-vit-transformer\u002F) | E. 多模态 | Python |\n| 60 | [用 Projection Layer 做模态对齐](phases\u002F19-capstone-projects\u002F60-projection-layer-modality-align\u002F) | E. 多模态 | Python |\n| 61 | [Cross-Attention 融合](phases\u002F19-capstone-projects\u002F61-cross-attention-fusion\u002F) | E. 多模态 | Python |\n| 62 | [Vision-Language 预训练](phases\u002F19-capstone-projects\u002F62-vision-language-pretraining\u002F) | E. 多模态 | Python |\n| 63 | [多模态评测](phases\u002F19-capstone-projects\u002F63-multimodal-eval\u002F) | E. 多模态 | Python |\n| 64 | [Chunking 策略横向对比](phases\u002F19-capstone-projects\u002F64-chunking-strategies-advanced\u002F) | F. 高级 RAG | Python |\n| 65 | [用 BM25 与 Dense Embedding 做 Hybrid Retrieval](phases\u002F19-capstone-projects\u002F65-hybrid-retrieval-bm25-dense\u002F) | F. 高级 RAG | Python |\n| 66 | [Cross-Encoder Reranker](phases\u002F19-capstone-projects\u002F66-reranker-cross-encoder\u002F) | F. 高级 RAG | Python |\n| 67 | [Query 改写：HyDE、Multi-Query 与 Decomposition](phases\u002F19-capstone-projects\u002F67-query-rewriting-hyde\u002F) | F. 高级 RAG | Python |\n| 68 | [RAG 评测：Precision、Recall、MRR、nDCG 等](phases\u002F19-capstone-projects\u002F68-rag-eval-precision-recall\u002F) | F. 高级 RAG | Python |\n| 69 | [端到端 RAG 系统](phases\u002F19-capstone-projects\u002F69-end-to-end-rag-system\u002F) | F. 高级 RAG | Python |\n| 70 | [任务规格格式](phases\u002F19-capstone-projects\u002F70-task-spec-format\u002F) | G. 评测体系 | Python |\n| 71 | [经典评测指标](phases\u002F19-capstone-projects\u002F71-classical-metrics\u002F) | G. 评测体系 | Python |\n| 72 | [代码执行评测指标](phases\u002F19-capstone-projects\u002F72-code-exec-metric\u002F) | G. 评测体系 | Python |\n| 73 | [perplexity 与 calibration](phases\u002F19-capstone-projects\u002F73-perplexity-calibration\u002F) | G. 评测体系 | Python |\n| 74 | [leaderboard 聚合](phases\u002F19-capstone-projects\u002F74-leaderboard-aggregation\u002F) | G. 评测体系 | Python |\n| 75 | [端到端 eval runner](phases\u002F19-capstone-projects\u002F75-end-to-end-eval-runner\u002F) | G. 评测体系 | Python |\n| 76 | [从零实现集合通信](phases\u002F19-capstone-projects\u002F76-collective-ops-from-scratch\u002F) | H. 分布式训练 | Python |\n| 77 | [数据并行 DDP](phases\u002F19-capstone-projects\u002F77-data-parallel-ddp\u002F) | H. 分布式训练 | Python |\n| 78 | [ZeRO Optimizer State 分片](phases\u002F19-capstone-projects\u002F78-zero-parameter-sharding\u002F) | H. 分布式训练 | Python |\n| 79 | [Pipeline Parallel 与 Bubble 分析](phases\u002F19-capstone-projects\u002F79-pipeline-parallel\u002F) | H. 分布式训练 | Python |\n| 80 | [分片 Checkpoint 与原子化恢复](phases\u002F19-capstone-projects\u002F80-checkpoint-sharded-resume\u002F) | H. 分布式训练 | Python |\n| 81 | [端到端分布式训练](phases\u002F19-capstone-projects\u002F81-end-to-end-distributed-train\u002F) | H. 分布式训练 | Python |\n| 82 | [越狱分类法](phases\u002F19-capstone-projects\u002F82-jailbreak-taxonomy\u002F) | I. 安全护栏 | Python |\n| 83 | [Prompt 注入检测器](phases\u002F19-capstone-projects\u002F83-prompt-injection-detector\u002F) | I. 安全护栏 | Python |\n| 84 | [拒答评估](phases\u002F19-capstone-projects\u002F84-refusal-evaluation\u002F) | I. 安全护栏 | Python |\n| 85 | [内容分类器集成](phases\u002F19-capstone-projects\u002F85-content-classifier-integration\u002F) | I. 安全护栏 | Python |\n| 86 | [Constitutional 规则引擎](phases\u002F19-capstone-projects\u002F86-constitutional-rules-engine\u002F) | I. 安全护栏 | Python |\n| 87 | [端到端 safety gate](phases\u002F19-capstone-projects\u002F87-end-to-end-safety-gate\u002F) | I. 安全护栏 | Python |\n\n\u003C\u002Fdetails>\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## 工具箱\n\n每节课都会产出一件可复用的产物。学完之后，你手里会有：\n\n```\noutputs\u002F\n├── prompts\u002F      覆盖每类 AI 任务的提示词模板\n└── skills\u002F       给 AI 编码 agent 用的 SKILL.md 文件\n```\n\n用 `npx skills add` 安装。把它们接进 Claude、Cursor、Codex、OpenClaw、Hermes，\n或任何能读 SKILL.md \u002F AGENTS.md 目录的 agent。都是真家伙，不是课后作业。\n\n### 把所有课程技能装进你的 agent\n\n仓库在 `phases\u002F**\u002Foutputs\u002F` 下交付了 378 个技能和 99 个提示词。\n\n**推荐：通过 [skills.sh](https:\u002F\u002Fskills.sh) 安装。** 不用 clone，不用 Python，\n自动识别你 agent 的技能目录：\n\n```bash\nnpx skills add fancyboi999\u002Fai-engineering-from-scratch-zh                       # 所有技能\nnpx skills add fancyboi999\u002Fai-engineering-from-scratch-zh --skill agent-loop    # 单个技能\nnpx skills add fancyboi999\u002Fai-engineering-from-scratch-zh --phase 14            # 单个阶段\n```\n\n`skills` 会写到你 agent 实际读取的那个目录：`.claude\u002Fskills\u002F`、`.cursor\u002Fskills\u002F`、\n`.codex\u002Fskills\u002F`、OpenClaw 的技能文件夹、Hermes 的 bundle 路径，或任何识别 SKILL.md\n的工具。一条命令，覆盖所有 agent。\n\n**进阶：用 `scripts\u002Finstall_skills.py` 做离线 \u002F 自定义布局。** 需要先 clone 仓库。\n当你需要按标签过滤、dry-run，或非默认布局时很有用：\n\n```bash\npython3 scripts\u002Finstall_skills.py \u003Ctarget>                                 # 所有技能，默认 --layout skills（嵌套）\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --layout skills                 # 同上，显式写出\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --type all                      # 技能 + 提示词 + agent\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --phase 14                      # 只装一个阶段\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --tag rag                       # 按标签过滤\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --layout flat                   # 扁平文件\npython3 scripts\u002Finstall_skills.py \u003Ctarget> --",2,"2026-06-11 04:11:29","CREATED_QUERY"]