{"id":50293679,"url":"https://github.com/fancyboi999/ai-engineering-from-scratch-zh","last_synced_at":"2026-06-01T11:01:03.232Z","repository":{"id":360641953,"uuid":"1250240320","full_name":"fancyboi999/ai-engineering-from-scratch-zh","owner":"fancyboi999","description":"从零精通 AI 工程 · 20 阶段 468 课 · 中文全量翻译 + 配套站点 如何成为一名Agent 工程师  修成指南","archived":false,"fork":false,"pushed_at":"2026-05-29T06:41:35.000Z","size":13283,"stargazers_count":8,"open_issues_count":1,"forks_count":2,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-05-31T10:11:25.388Z","etag":null,"topics":["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"],"latest_commit_sha":null,"homepage":"https://aieng-zh.cn","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/fancyboi999.png","metadata":{"files":{"readme":"README.md","changelog":"CHANGELOG.md","contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":"CODE_OF_CONDUCT.md","threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":"ROADMAP.md","authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-26T12:45:07.000Z","updated_at":"2026-05-31T06:00:30.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/fancyboi999/ai-engineering-from-scratch-zh","commit_stats":null,"previous_names":["fancyboi999/ai-engineering-from-scratch-zh"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/fancyboi999/ai-engineering-from-scratch-zh","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fancyboi999%2Fai-engineering-from-scratch-zh","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fancyboi999%2Fai-engineering-from-scratch-zh/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fancyboi999%2Fai-engineering-from-scratch-zh/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fancyboi999%2Fai-engineering-from-scratch-zh/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/fancyboi999","download_url":"https://codeload.github.com/fancyboi999/ai-engineering-from-scratch-zh/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fancyboi999%2Fai-engineering-from-scratch-zh/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":33771629,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-01T02:00:06.963Z","response_time":115,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["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"],"created_at":"2026-05-28T07:31:03.890Z","updated_at":"2026-06-01T11:01:03.093Z","avatar_url":"https://github.com/fancyboi999.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003cp align=\"center\"\u003e\n  \u003cimg src=\"assets/banner.svg\" alt=\"AI Engineering from Scratch — reference manual banner\" width=\"100%\"\u003e\n\u003c/p\u003e\n\n\u003cp align=\"center\"\u003e\n  \u003ca href=\"LICENSE\"\u003e\u003cimg src=\"https://img.shields.io/badge/license-MIT-1a1a1a?style=flat-square\u0026labelColor=fafaf5\" alt=\"MIT License\"\u003e\u003c/a\u003e\n  \u003ca href=\"ROADMAP.md\"\u003e\u003cimg src=\"https://img.shields.io/badge/lessons-435-3553ff?style=flat-square\u0026labelColor=fafaf5\" alt=\"435 lessons\"\u003e\u003c/a\u003e\n  \u003ca href=\"#contents\"\u003e\u003cimg src=\"https://img.shields.io/badge/phases-20-3553ff?style=flat-square\u0026labelColor=fafaf5\" alt=\"20 phases\"\u003e\u003c/a\u003e\n  \u003ca href=\"https://github.com/fancyboi999/ai-engineering-from-scratch-zh/stargazers\"\u003e\u003cimg src=\"https://img.shields.io/github/stars/fancyboi999/ai-engineering-from-scratch-zh?style=flat-square\u0026labelColor=fafaf5\u0026color=3553ff\" alt=\"GitHub stars\"\u003e\u003c/a\u003e\n  \u003ca href=\"https://aieng-zh.cn\"\u003e\u003cimg src=\"https://img.shields.io/badge/website-live-3553ff?style=flat-square\u0026labelColor=fafaf5\" alt=\"Website\"\u003e\u003c/a\u003e\n\u003c/p\u003e\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n\u003e **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\u003e 这套课程要填的就是这道沟。\n\u003e\n\u003e 435 节课，20 个阶段，约 320 小时。Python、TypeScript、Rust、Julia。每节课都交付一件\n\u003e 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\u003e\n\u003e 你不只是学 AI，你亲手把它造出来。从头到尾，全手写。\n\n\u003e 本项目是 [AI Engineering from Scratch](https://github.com/rohitg00/ai-engineering-from-scratch) 的简体中文翻译版。感谢原作者 [Rohit Ghumare](https://github.com/rohitg00) 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，435 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 \u0026amp; Tooling\"] --\u003e P1[\"Phase 1 — Math Foundations\"]\n  P1 --\u003e P2[\"Phase 2 — ML Fundamentals\"]\n  P2 --\u003e P3[\"Phase 3 — Deep Learning Core\"]\n  P3 --\u003e P4[\"Phase 4 — Vision\"]\n  P3 --\u003e P5[\"Phase 5 — NLP\"]\n  P3 --\u003e P6[\"Phase 6 — Speech \u0026amp; Audio\"]\n  P3 --\u003e P9[\"Phase 9 — RL\"]\n  P5 --\u003e P7[\"Phase 7 — Transformers\"]\n  P7 --\u003e P8[\"Phase 8 — GenAI\"]\n  P7 --\u003e P10[\"Phase 10 — LLMs from Scratch\"]\n  P10 --\u003e P11[\"Phase 11 — LLM Engineering\"]\n  P10 --\u003e P12[\"Phase 12 — Multimodal\"]\n  P11 --\u003e P13[\"Phase 13 — Tools \u0026amp; Protocols\"]\n  P13 --\u003e P14[\"Phase 14 — Agent Engineering\"]\n  P14 --\u003e P15[\"Phase 15 — Autonomous Systems\"]\n  P15 --\u003e P16[\"Phase 16 — Multi-Agent \u0026amp; Swarms\"]\n  P14 --\u003e P17[\"Phase 17 — Infrastructure \u0026amp; Production\"]\n  P15 --\u003e P18[\"Phase 18 — Ethics \u0026amp; Alignment\"]\n  P16 --\u003e P19[\"Phase 19 — Capstone Projects\"]\n  P17 --\u003e P19\n  P18 --\u003e 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/\u003cNN\u003e-\u003cphase-name\u003e/\u003cNN\u003e-\u003clesson-name\u003e/\n├── code/      可运行的实现（Python、TypeScript、Rust、Julia）\n├── docs/\n│   └── zh.md  课程正文\n└── outputs/   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 / Use It*（动手构建 / 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/\u003e\u003csub\u003eone-line core idea\u003c/sub\u003e\"] --\u003e Pr[\"PROBLEM\u003cbr/\u003e\u003csub\u003econcrete pain\u003c/sub\u003e\"]\n  Pr --\u003e C[\"CONCEPT\u003cbr/\u003e\u003csub\u003ediagrams \u0026amp; intuition\u003c/sub\u003e\"]\n  C --\u003e B[\"BUILD IT\u003cbr/\u003e\u003csub\u003eraw math, no frameworks\u003c/sub\u003e\"]\n  B --\u003e U[\"USE IT\u003cbr/\u003e\u003csub\u003esame thing in PyTorch / sklearn\u003c/sub\u003e\"]\n  U --\u003e S[\"SHIP IT\u003cbr/\u003e\u003csub\u003eprompt · skill · agent · MCP\u003c/sub\u003e\"]\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://aieng-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://github.com/fancyboi999/ai-engineering-from-scratch-zh.git\ncd ai-engineering-from-scratch-zh\npython phases/01-math-foundations/01-linear-algebra-intuition/code/vectors.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/find-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/check-understanding 3        # 测验你对阶段 3 的掌握\nls phases/03-deep-learning-core/05-loss-functions/outputs/\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| [`/find-your-level`](.claude/skills/find-your-level/SKILL.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| [`/check-understanding \u003cphase\u003e`](.claude/skills/check-understanding/SKILL.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\u003e\n\u003ctr\u003e\n\u003cth align=\"left\" width=\"25%\"\u003e\u003cimg src=\"site/assets/figures/001-a-prompts.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.A prompts\"/\u003e\u003cbr/\u003e\u003csub\u003eFIG_001 · A\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003ePROMPTS\u003c/b\u003e\u003c/th\u003e\n\u003cth align=\"left\" width=\"25%\"\u003e\u003cimg src=\"site/assets/figures/001-b-skills.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.B skills\"/\u003e\u003cbr/\u003e\u003csub\u003eFIG_001 · B\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003eSKILLS\u003c/b\u003e\u003c/th\u003e\n\u003cth align=\"left\" width=\"25%\"\u003e\u003cimg src=\"site/assets/figures/001-c-agents.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.C agents\"/\u003e\u003cbr/\u003e\u003csub\u003eFIG_001 · C\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003eAGENTS\u003c/b\u003e\u003c/th\u003e\n\u003cth align=\"left\" width=\"25%\"\u003e\u003cimg src=\"site/assets/figures/001-d-mcp-servers.svg\" width=\"96\" height=\"96\" alt=\"FIG_001.D MCP servers\"/\u003e\u003cbr/\u003e\u003csub\u003eFIG_001 · D\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003eMCP SERVERS\u003c/b\u003e\u003c/th\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd valign=\"top\"\u003e粘进任意 AI 助手，在某个细分任务上获得专家级帮助。\u003c/td\u003e\n\u003ctd valign=\"top\"\u003e放进 Claude、Cursor、Codex、OpenClaw、Hermes，或任何能读 \u003ccode\u003eSKILL.md\u003c/code\u003e 的 agent。\u003c/td\u003e\n\u003ctd valign=\"top\"\u003e作为自主 worker 部署——那个循环你在阶段 14 自己写过。\u003c/td\u003e\n\u003ctd valign=\"top\"\u003e接入任意兼容 MCP 的客户端。在阶段 13 里从头到尾构建。\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/table\u003e\n\n\u003e 用 `python3 scripts/install_skills.py` 一次性全部安装。是真家伙，不是课后作业。\n\u003e 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\u003e\n\u003ctr\u003e\n\u003ctd valign=\"top\" width=\"50%\"\u003e\n\n**`code/agent_loop.py`** \u0026nbsp; \u003csub\u003e\u003ci\u003e动手构建\u003c/i\u003e\u003c/sub\u003e\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/td\u003e\n\u003ctd valign=\"top\" width=\"50%\"\u003e\n\n**`outputs/skill-agent-loop.md`** \u0026nbsp; \u003csub\u003e\u003ci\u003e交付\u003c/i\u003e\u003c/sub\u003e\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/prompt-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/td\u003e\n\u003c/tr\u003e\n\u003c/table\u003e\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n\u003ca id=\"contents\"\u003e\u003c/a\u003e\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\"\u003e\u003c/a\u003e\n### Phase 0: 配置与工具链 `12 lessons`\n\u003e 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/00-setup-and-tooling/01-dev-environment/) | Build | Python |\n| 02 | [Git 与协作](phases/00-setup-and-tooling/02-git-and-collaboration/) | Learn | — |\n| 03 | [GPU 配置与云端](phases/00-setup-and-tooling/03-gpu-setup-and-cloud/) | Build | Python |\n| 04 | [API 与密钥](phases/00-setup-and-tooling/04-apis-and-keys/) | Build | Python |\n| 05 | [Jupyter Notebook](phases/00-setup-and-tooling/05-jupyter-notebooks/) | Build | Python |\n| 06 | [Python 环境管理](phases/00-setup-and-tooling/06-python-environments/) | Build | Shell |\n| 07 | [面向 AI 的 Docker](phases/00-setup-and-tooling/07-docker-for-ai/) | Build | Docker |\n| 08 | [编辑器配置](phases/00-setup-and-tooling/08-editor-setup/) | Build | — |\n| 09 | [数据管理](phases/00-setup-and-tooling/09-data-management/) | Build | Python |\n| 10 | [终端与 Shell](phases/00-setup-and-tooling/10-terminal-and-shell/) | Learn | — |\n| 11 | [面向 AI 的 Linux](phases/00-setup-and-tooling/11-linux-for-ai/) | Learn | — |\n| 12 | [调试与性能分析](phases/00-setup-and-tooling/12-debugging-and-profiling/) | Build | Python |\n\n\u003cdetails id=\"phase-1\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 1 — 数学基础\u003c/b\u003e \u0026nbsp;\u003ccode\u003e22 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e每个 AI 算法背后的直觉，用代码讲清楚。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [线性代数直觉](phases/01-math-foundations/01-linear-algebra-intuition/) | Learn | Python, Julia |\n| 02 | [向量、矩阵与运算](phases/01-math-foundations/02-vectors-matrices-operations/) | Build | Python, Julia |\n| 03 | [矩阵变换与特征值](phases/01-math-foundations/03-matrix-transformations/) | Build | Python, Julia |\n| 04 | [机器学习里的微积分：导数与梯度](phases/01-math-foundations/04-calculus-for-ml/) | Learn | Python |\n| 05 | [链式法则与自动微分](phases/01-math-foundations/05-chain-rule-and-autodiff/) | Build | Python |\n| 06 | [概率与分布](phases/01-math-foundations/06-probability-and-distributions/) | Learn | Python |\n| 07 | [贝叶斯定理与统计思维](phases/01-math-foundations/07-bayes-theorem/) | Build | Python |\n| 08 | [优化：梯度下降家族](phases/01-math-foundations/08-optimization/) | Build | Python |\n| 09 | [信息论：熵与 KL 散度](phases/01-math-foundations/09-information-theory/) | Learn | Python |\n| 10 | [降维：PCA、t-SNE、UMAP](phases/01-math-foundations/10-dimensionality-reduction/) | Build | Python |\n| 11 | [奇异值分解](phases/01-math-foundations/11-singular-value-decomposition/) | Build | Python, Julia |\n| 12 | [张量运算](phases/01-math-foundations/12-tensor-operations/) | Build | Python |\n| 13 | [数值稳定性](phases/01-math-foundations/13-numerical-stability/) | Build | Python |\n| 14 | [范数与距离](phases/01-math-foundations/14-norms-and-distances/) | Build | Python |\n| 15 | [机器学习里的统计学](phases/01-math-foundations/15-statistics-for-ml/) | Build | Python |\n| 16 | [采样方法](phases/01-math-foundations/16-sampling-methods/) | Build | Python |\n| 17 | [线性方程组](phases/01-math-foundations/17-linear-systems/) | Build | Python |\n| 18 | [凸优化](phases/01-math-foundations/18-convex-optimization/) | Build | Python |\n| 19 | [面向 AI 的复数](phases/01-math-foundations/19-complex-numbers/) | Learn | Python |\n| 20 | [傅里叶变换](phases/01-math-foundations/20-fourier-transform/) | Build | Python |\n| 21 | [机器学习里的图论](phases/01-math-foundations/21-graph-theory/) | Build | Python |\n| 22 | [随机过程](phases/01-math-foundations/22-stochastic-processes/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-2\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 2 — 机器学习基础\u003c/b\u003e \u0026nbsp;\u003ccode\u003e18 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e经典机器学习——至今仍是大多数生产 AI 的骨架。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [什么是机器学习](phases/02-ml-fundamentals/01-what-is-machine-learning/) | Learn | Python |\n| 02 | [从零实现线性回归](phases/02-ml-fundamentals/02-linear-regression/) | Build | Python |\n| 03 | [逻辑回归与分类](phases/02-ml-fundamentals/03-logistic-regression/) | Build | Python |\n| 04 | [决策树与随机森林](phases/02-ml-fundamentals/04-decision-trees/) | Build | Python |\n| 05 | [支持向量机](phases/02-ml-fundamentals/05-support-vector-machines/) | Build | Python |\n| 06 | [KNN 与距离度量](phases/02-ml-fundamentals/06-knn-and-distances/) | Build | Python |\n| 07 | [无监督学习：K-Means、DBSCAN](phases/02-ml-fundamentals/07-unsupervised-learning/) | Build | Python |\n| 08 | [特征工程与特征选择](phases/02-ml-fundamentals/08-feature-engineering/) | Build | Python |\n| 09 | [模型评估：指标与交叉验证](phases/02-ml-fundamentals/09-model-evaluation/) | Build | Python |\n| 10 | [偏差、方差与学习曲线](phases/02-ml-fundamentals/10-bias-variance/) | Learn | Python |\n| 11 | [集成方法：Boosting、Bagging、Stacking](phases/02-ml-fundamentals/11-ensemble-methods/) | Build | Python |\n| 12 | [超参数调优](phases/02-ml-fundamentals/12-hyperparameter-tuning/) | Build | Python |\n| 13 | [机器学习流水线与实验追踪](phases/02-ml-fundamentals/13-ml-pipelines/) | Build | Python |\n| 14 | [朴素贝叶斯](phases/02-ml-fundamentals/14-naive-bayes/) | Build | Python |\n| 15 | [时间序列基础](phases/02-ml-fundamentals/15-time-series/) | Build | Python |\n| 16 | [异常检测](phases/02-ml-fundamentals/16-anomaly-detection/) | Build | Python |\n| 17 | [处理不平衡数据](phases/02-ml-fundamentals/17-imbalanced-data/) | Build | Python |\n| 18 | [特征选择](phases/02-ml-fundamentals/18-feature-selection/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-3\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 3 — 深度学习核心\u003c/b\u003e \u0026nbsp;\u003ccode\u003e13 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e从第一性原理出发的神经网络。先自己造一个，再碰框架。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [感知机：一切的起点](phases/03-deep-learning-core/01-the-perceptron/) | Build | Python |\n| 02 | [多层网络与前向传播](phases/03-deep-learning-core/02-multi-layer-networks/) | Build | Python |\n| 03 | [从零实现反向传播](phases/03-deep-learning-core/03-backpropagation/) | Build | Python |\n| 04 | [激活函数：ReLU、Sigmoid、GELU 及其原因](phases/03-deep-learning-core/04-activation-functions/) | Build | Python |\n| 05 | [损失函数：MSE、交叉熵、对比损失](phases/03-deep-learning-core/05-loss-functions/) | Build | Python |\n| 06 | [优化器：SGD、Momentum、Adam、AdamW](phases/03-deep-learning-core/06-optimizers/) | Build | Python |\n| 07 | [正则化：Dropout、权重衰减、BatchNorm](phases/03-deep-learning-core/07-regularization/) | Build | Python |\n| 08 | [权重初始化与训练稳定性](phases/03-deep-learning-core/08-weight-initialization/) | Build | Python |\n| 09 | [学习率调度与 Warmup](phases/03-deep-learning-core/09-learning-rate-schedules/) | Build | Python |\n| 10 | [造一个你自己的迷你框架](phases/03-deep-learning-core/10-mini-framework/) | Build | Python |\n| 11 | [PyTorch 入门](phases/03-deep-learning-core/11-intro-to-pytorch/) | Build | Python |\n| 12 | [JAX 入门](phases/03-deep-learning-core/12-intro-to-jax/) | Build | Python |\n| 13 | [调试神经网络](phases/03-deep-learning-core/13-debugging-neural-networks/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-4\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 4 — 计算机视觉\u003c/b\u003e \u0026nbsp;\u003ccode\u003e28 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e从像素到理解——图像、视频、3D、VLM 和世界模型。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [图像基础：像素、通道、色彩空间](phases/04-computer-vision/01-image-fundamentals/) | Learn | Python |\n| 02 | [从零实现卷积](phases/04-computer-vision/02-convolutions-from-scratch/) | Build | Python |\n| 03 | [CNN：从 LeNet 到 ResNet](phases/04-computer-vision/03-cnns-lenet-to-resnet/) | Build | Python |\n| 04 | [图像分类](phases/04-computer-vision/04-image-classification/) | Build | Python |\n| 05 | [迁移学习与微调](phases/04-computer-vision/05-transfer-learning/) | Build | Python |\n| 06 | [目标检测——从零实现 YOLO](phases/04-computer-vision/06-object-detection-yolo/) | Build | Python |\n| 07 | [语义分割——U-Net](phases/04-computer-vision/07-semantic-segmentation-unet/) | Build | Python |\n| 08 | [实例分割——Mask R-CNN](phases/04-computer-vision/08-instance-segmentation-mask-rcnn/) | Build | Python |\n| 09 | [图像生成——GAN](phases/04-computer-vision/09-image-generation-gans/) | Build | Python |\n| 10 | [图像生成——扩散模型](phases/04-computer-vision/10-image-generation-diffusion/) | Build | Python |\n| 11 | [Stable Diffusion——架构与微调](phases/04-computer-vision/11-stable-diffusion/) | Build | Python |\n| 12 | [视频理解——时序建模](phases/04-computer-vision/12-video-understanding/) | Build | Python |\n| 13 | [3D 视觉：点云、NeRF](phases/04-computer-vision/13-3d-vision-nerf/) | Build | Python |\n| 14 | [Vision Transformer（ViT）](phases/04-computer-vision/14-vision-transformers/) | Build | Python |\n| 15 | [实时视觉：边缘部署](phases/04-computer-vision/15-real-time-edge/) | Build | Python |\n| 16 | [构建一条完整的视觉流水线](phases/04-computer-vision/16-vision-pipeline-capstone/) | Build | Python |\n| 17 | [自监督视觉——SimCLR、DINO、MAE](phases/04-computer-vision/17-self-supervised-vision/) | Build | Python |\n| 18 | [开放词表视觉——CLIP](phases/04-computer-vision/18-open-vocab-clip/) | Build | Python |\n| 19 | [OCR 与文档理解](phases/04-computer-vision/19-ocr-document-understanding/) | Build | Python |\n| 20 | [图像检索与度量学习](phases/04-computer-vision/20-image-retrieval-metric/) | Build | Python |\n| 21 | [关键点检测与姿态估计](phases/04-computer-vision/21-keypoint-pose/) | Build | Python |\n| 22 | [从零实现 3D 高斯泼溅](phases/04-computer-vision/22-3d-gaussian-splatting/) | Build | Python |\n| 23 | [Diffusion Transformer 与 Rectified Flow](phases/04-computer-vision/23-diffusion-transformers-rectified-flow/) | Build | Python |\n| 24 | [SAM 3 与开放词表分割](phases/04-computer-vision/24-sam3-open-vocab-segmentation/) | Build | Python |\n| 25 | [视觉语言模型（ViT-MLP-LLM）](phases/04-computer-vision/25-vision-language-models/) | Build | Python |\n| 26 | [单目深度与几何估计](phases/04-computer-vision/26-monocular-depth/) | Build | Python |\n| 27 | [多目标跟踪与视频记忆](phases/04-computer-vision/27-multi-object-tracking/) | Build | Python |\n| 28 | [世界模型与视频扩散](phases/04-computer-vision/28-world-models-video-diffusion/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-5\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 5 — NLP：从基础到进阶\u003c/b\u003e \u0026nbsp;\u003ccode\u003e29 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e语言是通往智能的接口。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [文本处理：分词、词干提取、词形还原](phases/05-nlp-foundations-to-advanced/01-text-processing/) | Build | Python |\n| 02 | [词袋、TF-IDF 与文本表示](phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/) | Build | Python |\n| 03 | [词嵌入：从零实现 Word2Vec](phases/05-nlp-foundations-to-advanced/03-word-embeddings-word2vec/) | Build | Python |\n| 04 | [GloVe、FastText 与子词嵌入](phases/05-nlp-foundations-to-advanced/04-glove-fasttext-subword/) | Build | Python |\n| 05 | [情感分析](phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/) | Build | Python |\n| 06 | [命名实体识别（NER）](phases/05-nlp-foundations-to-advanced/06-named-entity-recognition/) | Build | Python |\n| 07 | [词性标注与句法分析](phases/05-nlp-foundations-to-advanced/07-pos-tagging-parsing/) | Build | Python |\n| 08 | [文本分类——用于文本的 CNN 与 RNN](phases/05-nlp-foundations-to-advanced/08-cnns-rnns-for-text/) | Build | Python |\n| 09 | [序列到序列模型](phases/05-nlp-foundations-to-advanced/09-sequence-to-sequence/) | Build | Python |\n| 10 | [注意力机制——那次突破](phases/05-nlp-foundations-to-advanced/10-attention-mechanism/) | Build | Python |\n| 11 | [机器翻译](phases/05-nlp-foundations-to-advanced/11-machine-translation/) | Build | Python |\n| 12 | [文本摘要](phases/05-nlp-foundations-to-advanced/12-text-summarization/) | Build | Python |\n| 13 | [问答系统](phases/05-nlp-foundations-to-advanced/13-question-answering/) | Build | Python |\n| 14 | [信息检索与搜索](phases/05-nlp-foundations-to-advanced/14-information-retrieval-search/) | Build | Python |\n| 15 | [主题建模：LDA、BERTopic](phases/05-nlp-foundations-to-advanced/15-topic-modeling/) | Build | Python |\n| 16 | [文本生成](phases/05-nlp-foundations-to-advanced/16-text-generation-pre-transformer/) | Build | Python |\n| 17 | [聊天机器人：从规则到神经网络](phases/05-nlp-foundations-to-advanced/17-chatbots-rule-to-neural/) | Build | Python |\n| 18 | [多语言 NLP](phases/05-nlp-foundations-to-advanced/18-multilingual-nlp/) | Build | Python |\n| 19 | [子词分词：BPE、WordPiece、Unigram、SentencePiece](phases/05-nlp-foundations-to-advanced/19-subword-tokenization/) | Learn | Python |\n| 20 | [结构化输出与约束解码](phases/05-nlp-foundations-to-advanced/20-structured-outputs-constrained-decoding/) | Build | Python |\n| 21 | [自然语言推理与文本蕴含](phases/05-nlp-foundations-to-advanced/21-nli-textual-entailment/) | Learn | Python |\n| 22 | [嵌入模型深入剖析](phases/05-nlp-foundations-to-advanced/22-embedding-models-deep-dive/) | Learn | Python |\n| 23 | [RAG 的分块策略](phases/05-nlp-foundations-to-advanced/23-chunking-strategies-rag/) | Build | Python |\n| 24 | [指代消解](phases/05-nlp-foundations-to-advanced/24-coreference-resolution/) | Learn | Python |\n| 25 | [实体链接与消歧](phases/05-nlp-foundations-to-advanced/25-entity-linking/) | Build | Python |\n| 26 | [关系抽取与知识图谱构建](phases/05-nlp-foundations-to-advanced/26-relation-extraction-kg/) | Build | Python |\n| 27 | [LLM 评估：RAGAS、DeepEval、G-Eval](phases/05-nlp-foundations-to-advanced/27-llm-evaluation-frameworks/) | Build | Python |\n| 28 | [长上下文评估：NIAH、RULER、LongBench、MRCR](phases/05-nlp-foundations-to-advanced/28-long-context-evaluation/) | Learn | Python |\n| 29 | [对话状态跟踪](phases/05-nlp-foundations-to-advanced/29-dialogue-state-tracking/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-6\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 6 — 语音与音频\u003c/b\u003e \u0026nbsp;\u003ccode\u003e17 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e听见、听懂、开口说。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [音频基础：波形、采样、FFT](phases/06-speech-and-audio/01-audio-fundamentals) | Learn | Python |\n| 02 | [频谱图、梅尔刻度与音频特征](phases/06-speech-and-audio/02-spectrograms-mel-features) | Build | Python |\n| 03 | [音频分类](phases/06-speech-and-audio/03-audio-classification) | Build | Python |\n| 04 | [语音识别（ASR）](phases/06-speech-and-audio/04-speech-recognition-asr) | Build | Python |\n| 05 | [Whisper：架构与微调](phases/06-speech-and-audio/05-whisper-architecture-finetuning) | Build | Python |\n| 06 | [说话人识别与验证](phases/06-speech-and-audio/06-speaker-recognition-verification) | Build | Python |\n| 07 | [文本转语音（TTS）](phases/06-speech-and-audio/07-text-to-speech) | Build | Python |\n| 08 | [声音克隆与音色转换](phases/06-speech-and-audio/08-voice-cloning-conversion) | Build | Python |\n| 09 | [音乐生成](phases/06-speech-and-audio/09-music-generation) | Build | Python |\n| 10 | [音频语言模型](phases/06-speech-and-audio/10-audio-language-models) | Build | Python |\n| 11 | [实时音频处理](phases/06-speech-and-audio/11-real-time-audio-processing) | Build | Python |\n| 12 | [搭一条语音助手流水线](phases/06-speech-and-audio/12-voice-assistant-pipeline) | Build | Python |\n| 13 | [神经音频编解码器——EnCodec、SNAC、Mimi、DAC](phases/06-speech-and-audio/13-neural-audio-codecs) | Learn | Python |\n| 14 | [语音活动检测与轮次切换](phases/06-speech-and-audio/14-voice-activity-detection-turn-taking) | Build | Python |\n| 15 | [流式语音到语音——Moshi、Hibiki](phases/06-speech-and-audio/15-streaming-speech-to-speech-moshi-hibiki) | Learn | Python |\n| 16 | [语音防伪与音频水印](phases/06-speech-and-audio/16-anti-spoofing-audio-watermarking) | Build | Python |\n| 17 | [音频评估——WER、MOS、MMAU、排行榜](phases/06-speech-and-audio/17-audio-evaluation-metrics) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-7\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 7 — Transformer 深入剖析\u003c/b\u003e \u0026nbsp;\u003ccode\u003e14 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e那个改变了一切的架构。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [为什么用 Transformer：RNN 的问题](phases/07-transformers-deep-dive/01-why-transformers/) | Learn | Python |\n| 02 | [从零实现自注意力](phases/07-transformers-deep-dive/02-self-attention-from-scratch/) | Build | Python |\n| 03 | [多头注意力](phases/07-transformers-deep-dive/03-multi-head-attention/) | Build | Python |\n| 04 | [位置编码：正弦、RoPE、ALiBi](phases/07-transformers-deep-dive/04-positional-encoding/) | Build | Python |\n| 05 | [完整的 Transformer：编码器 + 解码器](phases/07-transformers-deep-dive/05-full-transformer/) | Build | Python |\n| 06 | [BERT——掩码语言建模](phases/07-transformers-deep-dive/06-bert-masked-language-modeling/) | Build | Python |\n| 07 | [GPT——因果语言建模](phases/07-transformers-deep-dive/07-gpt-causal-language-modeling/) | Build | Python |\n| 08 | [T5、BART——编码器-解码器模型](phases/07-transformers-deep-dive/08-t5-bart-encoder-decoder/) | Learn | Python |\n| 09 | [Vision Transformer（ViT）](phases/07-transformers-deep-dive/09-vision-transformers/) | Build | Python |\n| 10 | [音频 Transformer——Whisper 架构](phases/07-transformers-deep-dive/10-audio-transformers-whisper/) | Learn | Python |\n| 11 | [专家混合（MoE）](phases/07-transformers-deep-dive/11-mixture-of-experts/) | Build | Python |\n| 12 | [KV Cache、Flash Attention 与推理优化](phases/07-transformers-deep-dive/12-kv-cache-flash-attention/) | Build | Python |\n| 13 | [缩放定律](phases/07-transformers-deep-dive/13-scaling-laws/) | Learn | Python |\n| 14 | [从零构建一个 Transformer](phases/07-transformers-deep-dive/14-build-a-transformer-capstone/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-8\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 8 — 生成式 AI\u003c/b\u003e \u0026nbsp;\u003ccode\u003e14 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e生成图像、视频、音频、3D，等等。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [生成模型：分类与历史](phases/08-generative-ai/01-generative-models-taxonomy-history/) | Learn | Python |\n| 02 | [自编码器与 VAE](phases/08-generative-ai/02-autoencoders-vae/) | Build | Python |\n| 03 | [GAN：生成器 vs 判别器](phases/08-generative-ai/03-gans-generator-discriminator/) | Build | Python |\n| 04 | [条件 GAN 与 Pix2Pix](phases/08-generative-ai/04-conditional-gans-pix2pix/) | Build | Python |\n| 05 | [StyleGAN](phases/08-generative-ai/05-stylegan/) | Build | Python |\n| 06 | [扩散模型——从零实现 DDPM](phases/08-generative-ai/06-diffusion-ddpm-from-scratch/) | Build | Python |\n| 07 | [潜在扩散与 Stable Diffusion](phases/08-generative-ai/07-latent-diffusion-stable-diffusion/) | Build | Python |\n| 08 | [ControlNet、LoRA 与条件控制](phases/08-generative-ai/08-controlnet-lora-conditioning/) | Build | Python |\n| 09 | [图像修复、扩展与编辑](phases/08-generative-ai/09-inpainting-outpainting-editing/) | Build | Python |\n| 10 | [视频生成](phases/08-generative-ai/10-video-generation/) | Build | Python |\n| 11 | [音频生成](phases/08-generative-ai/11-audio-generation/) | Build | Python |\n| 12 | [3D 生成](phases/08-generative-ai/12-3d-generation/) | Build | Python |\n| 13 | [Flow Matching 与 Rectified Flow](phases/08-generative-ai/13-flow-matching-rectified-flows/) | Build | Python |\n| 14 | [评估：FID、CLIP Score](phases/08-generative-ai/14-evaluation-fid-clip-score/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-9\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 9 — 强化学习\u003c/b\u003e \u0026nbsp;\u003ccode\u003e12 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003eRLHF 和会玩游戏的 AI 的基石。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [MDP、状态、动作与奖励](phases/09-reinforcement-learning/01-mdps-states-actions-rewards/) | Learn | Python |\n| 02 | [动态规划](phases/09-reinforcement-learning/02-dynamic-programming/) | Build | Python |\n| 03 | [蒙特卡洛方法](phases/09-reinforcement-learning/03-monte-carlo-methods/) | Build | Python |\n| 04 | [Q-Learning、SARSA](phases/09-reinforcement-learning/04-q-learning-sarsa/) | Build | Python |\n| 05 | [深度 Q 网络（DQN）](phases/09-reinforcement-learning/05-dqn/) | Build | Python |\n| 06 | [策略梯度——REINFORCE](phases/09-reinforcement-learning/06-policy-gradients-reinforce/) | Build | Python |\n| 07 | [Actor-Critic——A2C、A3C](phases/09-reinforcement-learning/07-actor-critic-a2c-a3c/) | Build | Python |\n| 08 | [PPO](phases/09-reinforcement-learning/08-ppo/) | Build | Python |\n| 09 | [奖励建模与 RLHF](phases/09-reinforcement-learning/09-reward-modeling-rlhf/) | Build | Python |\n| 10 | [多智能体强化学习](phases/09-reinforcement-learning/10-multi-agent-rl/) | Build | Python |\n| 11 | [仿真到现实的迁移](phases/09-reinforcement-learning/11-sim-to-real-transfer/) | Build | Python |\n| 12 | [游戏中的强化学习](phases/09-reinforcement-learning/12-rl-for-games/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-10\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 10 — 从零实现 LLM\u003c/b\u003e \u0026nbsp;\u003ccode\u003e22 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e构建、训练并真正理解大语言模型。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [分词器：BPE、WordPiece、SentencePiece](phases/10-llms-from-scratch/01-tokenizers/) | Build | Python, Rust |\n| 02 | [从零实现一个分词器](phases/10-llms-from-scratch/02-building-a-tokenizer/) | Build | Python |\n| 03 | [预训练的数据流水线](phases/10-llms-from-scratch/03-data-pipelines/) | Build | Python |\n| 04 | [预训练一个迷你 GPT（124M）](phases/10-llms-from-scratch/04-pre-training-mini-gpt/) | Build | Python |\n| 05 | [分布式训练、FSDP、DeepSpeed](phases/10-llms-from-scratch/05-scaling-distributed/) | Build | Python |\n| 06 | [指令微调——SFT](phases/10-llms-from-scratch/06-instruction-tuning-sft/) | Build | Python |\n| 07 | [RLHF——奖励模型 + PPO](phases/10-llms-from-scratch/07-rlhf/) | Build | Python |\n| 08 | [DPO——直接偏好优化](phases/10-llms-from-scratch/08-dpo/) | Build | Python |\n| 09 | [Constitutional AI 与自我改进](phases/10-llms-from-scratch/09-constitutional-ai-self-improvement/) | Build | Python |\n| 10 | [评估——基准与 evals](phases/10-llms-from-scratch/10-evaluation/) | Build | Python |\n| 11 | [量化：INT8、GPTQ、AWQ、GGUF](phases/10-llms-from-scratch/11-quantization/) | Build | Python |\n| 12 | [推理优化](phases/10-llms-from-scratch/12-inference-optimization/) | Build | Python |\n| 13 | [搭一条完整的 LLM 流水线](phases/10-llms-from-scratch/13-building-complete-llm-pipeline/) | Build | Python |\n| 14 | [开源模型：架构逐一拆解](phases/10-llms-from-scratch/14-open-models-architecture-walkthroughs/) | Learn | Python |\n| 15 | [投机解码与 EAGLE-3](phases/10-llms-from-scratch/15-speculative-decoding-eagle3/) | Build | Python |\n| 16 | [差分注意力（V2）](phases/10-llms-from-scratch/16-differential-attention-v2/) | Build | Python |\n| 17 | [原生稀疏注意力（DeepSeek NSA）](phases/10-llms-from-scratch/17-native-sparse-attention/) | Build | Python |\n| 18 | [多 token 预测（MTP）](phases/10-llms-from-scratch/18-multi-token-prediction/) | Build | Python |\n| 19 | [DualPipe 并行](phases/10-llms-from-scratch/19-dualpipe-parallelism/) | Learn | Python |\n| 20 | [DeepSeek-V3 架构拆解](phases/10-llms-from-scratch/20-deepseek-v3-walkthrough/) | Learn | Python |\n| 21 | [Jamba——SSM-Transformer 混合架构](phases/10-llms-from-scratch/21-jamba-hybrid-ssm-transformer/) | Learn | Python |\n| 22 | [异步与 Hogwild! 推理](phases/10-llms-from-scratch/22-async-hogwild-inference/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-11\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 11 — LLM 工程\u003c/b\u003e \u0026nbsp;\u003ccode\u003e17 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e让 LLM 在生产环境里干活。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [提示工程：技巧与套路](phases/11-llm-engineering/01-prompt-engineering/) | Build | Python |\n| 02 | [Few-Shot、CoT、Tree-of-Thought](phases/11-llm-engineering/02-few-shot-cot/) | Build | Python |\n| 03 | [结构化输出](phases/11-llm-engineering/03-structured-outputs/) | Build | Python |\n| 04 | [嵌入与向量表示](phases/11-llm-engineering/04-embeddings/) | Build | Python |\n| 05 | [上下文工程](phases/11-llm-engineering/05-context-engineering/) | Build | Python |\n| 06 | [RAG：检索增强生成](phases/11-llm-engineering/06-rag/) | Build | Python |\n| 07 | [进阶 RAG：分块、重排](phases/11-llm-engineering/07-advanced-rag/) | Build | Python |\n| 08 | [用 LoRA 与 QLoRA 微调](phases/11-llm-engineering/08-fine-tuning-lora/) | Build | Python |\n| 09 | [函数调用与工具使用](phases/11-llm-engineering/09-function-calling/) | Build | Python |\n| 10 | [评估与测试](phases/11-llm-engineering/10-evaluation/) | Build | Python |\n| 11 | [缓存、限流与成本](phases/11-llm-engineering/11-caching-cost/) | Build | Python |\n| 12 | [护栏与安全](phases/11-llm-engineering/12-guardrails/) | Build | Python |\n| 13 | [构建一个生产级 LLM 应用](phases/11-llm-engineering/13-production-app/) | Build | Python |\n| 14 | [模型上下文协议（MCP）](phases/11-llm-engineering/14-model-context-protocol/) | Build | Python |\n| 15 | [提示缓存与上下文缓存](phases/11-llm-engineering/15-prompt-caching/) | Build | Python |\n| 16 | [LangGraph：面向 agent 的状态机](phases/11-llm-engineering/16-langgraph-state-machines/) | Build | Python |\n| 17 | [agent 框架的取舍](phases/11-llm-engineering/17-agent-framework-tradeoffs/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-12\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 12 — 多模态 AI\u003c/b\u003e \u0026nbsp;\u003ccode\u003e25 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e跨模态地看、听、读、推理——从 ViT 的图块到操作电脑的 agent。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [Vision Transformer 与图块-token 原语](phases/12-multimodal-ai/01-vision-transformer-patch-tokens/) | Learn | Python |\n| 02 | [CLIP 与对比式视觉语言预训练](phases/12-multimodal-ai/02-clip-contrastive-pretraining/) | Build | Python |\n| 03 | [BLIP-2 Q-Former 作为模态桥梁](phases/12-multimodal-ai/03-blip2-qformer-bridge/) | Build | Python |\n| 04 | [Flamingo 与门控交叉注意力](phases/12-multimodal-ai/04-flamingo-gated-cross-attention/) | Learn | Python |\n| 05 | [LLaVA 与视觉指令微调](phases/12-multimodal-ai/05-llava-visual-instruction-tuning/) | Build | Python |\n| 06 | [任意分辨率视觉——Patch-n'-Pack 与 NaFlex](phases/12-multimodal-ai/06-any-resolution-patch-n-pack/) | Build | Python |\n| 07 | [开源权重 VLM 配方：真正要紧的是什么](phases/12-multimodal-ai/07-open-weight-vlm-recipes/) | Learn | Python |\n| 08 | [LLaVA-OneVision：单图、多图、视频](phases/12-multimodal-ai/08-llava-onevision-single-multi-video/) | Build | Python |\n| 09 | [Qwen-VL 家族与动态 FPS 视频](phases/12-multimodal-ai/09-qwen-vl-family-dynamic-fps/) | Learn | Python |\n| 10 | [InternVL3 原生多模态预训练](phases/12-multimodal-ai/10-internvl3-native-multimodal/) | Learn | Python |\n| 11 | [Chameleon 早融合纯 token](phases/12-multimodal-ai/11-chameleon-early-fusion-tokens/) | Build | Python |\n| 12 | [Emu3 用下一 token 预测做生成](phases/12-multimodal-ai/12-emu3-next-token-for-generation/) | Learn | Python |\n| 13 | [Transfusion：自回归 + 扩散](phases/12-multimodal-ai/13-transfusion-autoregressive-diffusion/) | Build | Python |\n| 14 | [Show-o 离散扩散统一架构](phases/12-multimodal-ai/14-show-o-discrete-diffusion-unified/) | Learn | Python |\n| 15 | [Janus-Pro 解耦编码器](phases/12-multimodal-ai/15-janus-pro-decoupled-encoders/) | Build | Python |\n| 16 | [MIO 任意到任意流式](phases/12-multimodal-ai/16-mio-any-to-any-streaming/) | Learn | Python |\n| 17 | [视频语言时序定位](phases/12-multimodal-ai/17-video-language-temporal-grounding/) | Build | Python |\n| 18 | [百万 token 上下文下的长视频](phases/12-multimodal-ai/18-long-video-million-token/) | Build | Python |\n| 19 | [音频语言模型：从 Whisper 到 AF3](phases/12-multimodal-ai/19-audio-language-whisper-to-af3/) | Build | Python |\n| 20 | [Omni 模型：Thinker-Talker 流式](phases/12-multimodal-ai/20-omni-models-thinker-talker/) | Build | Python |\n| 21 | [具身 VLA：RT-2、OpenVLA、π0、GR00T](phases/12-multimodal-ai/21-embodied-vlas-openvla-pi0-groot/) | Learn | Python |\n| 22 | [文档与图表理解](phases/12-multimodal-ai/22-document-diagram-understanding/) | Build | Python |\n| 23 | [ColPali 视觉原生文档 RAG](phases/12-multimodal-ai/23-colpali-vision-native-rag/) | Build | Python |\n| 24 | [多模态 RAG 与跨模态检索](phases/12-multimodal-ai/24-multimodal-rag-cross-modal/) | Build | Python |\n| 25 | [多模态 agent 与操作电脑（综合项目）](phases/12-multimodal-ai/25-multimodal-agents-computer-use/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-13\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 13 — 工具与协议\u003c/b\u003e \u0026nbsp;\u003ccode\u003e23 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003eAI 与真实世界之间的接口。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [工具接口](phases/13-tools-and-protocols/01-the-tool-interface/) | Learn | Python |\n| 02 | [函数调用深入剖析](phases/13-tools-and-protocols/02-function-calling-deep-dive/) | Build | Python |\n| 03 | [并行与流式工具调用](phases/13-tools-and-protocols/03-parallel-and-streaming-tool-calls/) | Build | Python |\n| 04 | [结构化输出](phases/13-tools-and-protocols/04-structured-output/) | Build | Python |\n| 05 | [工具 Schema 设计](phases/13-tools-and-protocols/05-tool-schema-design/) | Learn | Python |\n| 06 | [MCP 基础](phases/13-tools-and-protocols/06-mcp-fundamentals/) | Learn | Python |\n| 07 | [构建一个 MCP server](phases/13-tools-and-protocols/07-building-an-mcp-server/) | Build | Python |\n| 08 | [构建一个 MCP client](phases/13-tools-and-protocols/08-building-an-mcp-client/) | Build | Python |\n| 09 | [MCP 传输层](phases/13-tools-and-protocols/09-mcp-transports/) | Learn | Python |\n| 10 | [MCP 资源与提示](phases/13-tools-and-protocols/10-mcp-resources-and-prompts/) | Build | Python |\n| 11 | [MCP Sampling](phases/13-tools-and-protocols/11-mcp-sampling/) | Build | Python |\n| 12 | [MCP Roots 与 Elicitation](phases/13-tools-and-protocols/12-mcp-roots-and-elicitation/) | Build | Python |\n| 13 | [MCP 异步任务](phases/13-tools-and-protocols/13-mcp-async-tasks/) | Build | Python |\n| 14 | [MCP Apps](phases/13-tools-and-protocols/14-mcp-apps/) | Build | Python |\n| 15 | [MCP 安全 I——工具投毒](phases/13-tools-and-protocols/15-mcp-security-tool-poisoning/) | Learn | Python |\n| 16 | [MCP 安全 II——OAuth 2.1](phases/13-tools-and-protocols/16-mcp-security-oauth-2-1/) | Build | Python |\n| 17 | [MCP 网关与注册表](phases/13-tools-and-protocols/17-mcp-gateways-and-registries/) | Learn | Python |\n| 18 | [生产环境的 MCP 认证——iii 上的 DCR + JWKS](phases/13-tools-and-protocols/18-mcp-auth-production/) | Build | Python |\n| 19 | [A2A 协议](phases/13-tools-and-protocols/19-a2a-protocol/) | Build | Python |\n| 20 | [OpenTelemetry GenAI](phases/13-tools-and-protocols/20-opentelemetry-genai/) | Build | Python |\n| 21 | [LLM 路由层](phases/13-tools-and-protocols/21-llm-routing-layer/) | Learn | Python |\n| 22 | [Skills 与 Agent SDK](phases/13-tools-and-protocols/22-skills-and-agent-sdks/) | Learn | Python |\n| 23 | [综合项目——工具生态](phases/13-tools-and-protocols/23-capstone-tool-ecosystem/) | Build | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-14\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 14 — Agent 工程\u003c/b\u003e \u0026nbsp;\u003ccode\u003e42 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e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/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [Agent 循环](phases/14-agent-engineering/01-the-agent-loop/) | Build | Python |\n| 02 | [ReWOO 与 Plan-and-Execute](phases/14-agent-engineering/02-rewoo-plan-and-execute/) | Build | Python |\n| 03 | [Reflexion 与言语强化学习](phases/14-agent-engineering/03-reflexion-verbal-rl/) | Build | Python |\n| 04 | [Tree of Thoughts 与 LATS](phases/14-agent-engineering/04-tree-of-thoughts-lats/) | Build | Python |\n| 05 | [Self-Refine 与 CRITIC](phases/14-agent-engineering/05-self-refine-and-critic/) | Build | Python |\n| 06 | [工具使用与函数调用](phases/14-agent-engineering/06-tool-use-and-function-calling/) | Build | Python |\n| 07 | [记忆——虚拟上下文与 MemGPT](phases/14-agent-engineering/07-memory-virtual-context-memgpt/) | Build | Python |\n| 08 | [记忆块与睡眠时计算](phases/14-agent-engineering/08-memory-blocks-sleep-time-compute/) | Build | Python |\n| 09 | [混合记忆——Mem0 向量 + 图 + KV](phases/14-agent-engineering/09-hybrid-memory-mem0/) | Build | Python |\n| 10 | [技能库与终身学习——Voyager](phases/14-agent-engineering/10-skill-libraries-voyager/) | Build | Python |\n| 11 | [用 HTN 与进化搜索做规划](phases/14-agent-engineering/11-planning-htn-and-evolutionary/) | Build | Python |\n| 12 | [Anthropic 的工作流模式](phases/14-agent-engineering/12-anthropic-workflow-patterns/) | Build | Python |\n| 13 | [LangGraph——有状态图与持久化执行](phases/14-agent-engineering/13-langgraph-stateful-graphs/) | Build | Python |\n| 14 | [AutoGen v0.4——Actor 模型](phases/14-agent-engineering/14-autogen-actor-model/) | Build | Python |\n| 15 | [CrewAI——基于角色的团队与流程](phases/14-agent-engineering/15-crewai-role-based-crews/) | Build | Python |\n| 16 | [OpenAI Agents SDK——交接、护栏、追踪](phases/14-agent-engineering/16-openai-agents-sdk/) | Build | Python |\n| 17 | [Claude Agent SDK——子 agent 与会话存储](phases/14-agent-engineering/17-claude-agent-sdk/) | Build | Python |\n| 18 | [Agno 与 Mastra——生产级运行时](phases/14-agent-engineering/18-agno-and-mastra-runtimes/) | Learn | Python |\n| 19 | [基准——SWE-bench、GAIA、AgentBench](phases/14-agent-engineering/19-benchmarks-swebench-gaia/) | Learn | Python |\n| 20 | [基准——WebArena 与 OSWorld](phases/14-agent-engineering/20-benchmarks-webarena-osworld/) | Learn | Python |\n| 21 | [操作电脑——Claude、OpenAI CUA、Gemini](phases/14-agent-engineering/21-computer-use-agents/) | Build | Python |\n| 22 | [语音 agent——Pipecat 与 LiveKit](phases/14-agent-engineering/22-voice-agents-pipecat-livekit/) | Build | Python |\n| 23 | [OpenTelemetry GenAI 语义约定](phases/14-agent-engineering/23-otel-genai-conventions/) | Build | Python |\n| 24 | [Agent 可观测性——Langfuse、Phoenix、Opik](phases/14-agent-engineering/24-agent-observability-platforms/) | Learn | Python |\n| 25 | [多 agent 辩论与协作](phases/14-agent-engineering/25-multi-agent-debate/) | Build | Python |\n| 26 | [失败模式——agent 为什么会崩](phases/14-agent-engineering/26-failure-modes-agentic/) | Build | Python |\n| 27 | [提示注入与 PVE 防御](phases/14-agent-engineering/27-prompt-injection-defense/) | Build | Python |\n| 28 | [编排模式——Supervisor、Swarm、分层](phases/14-agent-engineering/28-orchestration-patterns/) | Build | Python |\n| 29 | [生产级运行时——队列、事件、Cron](phases/14-agent-engineering/29-production-runtimes/) | Learn | Python |\n| 30 | [Eval 驱动的 agent 开发](phases/14-agent-engineering/30-eval-driven-agent-development/) | Build | Python |\n| 31 | [Agent 工作台：能力强的模型为什么仍会失败](phases/14-agent-engineering/31-agent-workbench-why-models-fail/) | Learn | Python |\n| 32 | [最小化 agent 工作台](phases/14-agent-engineering/32-minimal-agent-workbench/) | Build | Python |\n| 33 | [把 agent 指令写成可执行约束](phases/14-agent-engineering/33-instructions-as-executable-constraints/) | Build | Python |\n| 34 | [仓库记忆与持久化状态](phases/14-agent-engineering/34-repo-memory-and-state/) | Build | Python |\n| 35 | [给 agent 的初始化脚本](phases/14-agent-engineering/35-initialization-scripts/) | Build | Python |\n| 36 | [范围契约与任务边界](phases/14-agent-engineering/36-scope-contracts/) | Build | Python |\n| 37 | [运行时反馈回路](phases/14-agent-engineering/37-runtime-feedback-loops/) | Build | Python |\n| 38 | [验证关卡](phases/14-agent-engineering/38-verification-gates/) | Build | Python |\n| 39 | [审查 agent：把构建者和评判者分开](phases/14-agent-engineering/39-reviewer-agent/) | Build | Python |\n| 40 | [多会话交接](phases/14-agent-engineering/40-multi-session-handoff/) | Build | Python |\n| 41 | [在真实仓库上跑工作台](phases/14-agent-engineering/41-workbench-for-real-repos/) | Build | Python |\n| 42 | [综合项目：交付一套可复用的 agent 工作台包](phases/14-agent-engineering/42-agent-workbench-capstone/) | 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/details\u003e\n\n\u003cdetails id=\"phase-15\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 15 — 自主系统\u003c/b\u003e \u0026nbsp;\u003ccode\u003e22 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e长程 agent、自我改进，以及 2026 年的安全技术栈。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [从聊天机器人到长程 agent（METR）](phases/15-autonomous-systems/01-long-horizon-agents/) | Learn | Python |\n| 02 | [STaR、V-STaR、Quiet-STaR：自学推理](phases/15-autonomous-systems/02-star-family-reasoning/) | Learn | Python |\n| 03 | [AlphaEvolve：进化式编码 agent](phases/15-autonomous-systems/03-alphaevolve-evolutionary-coding/) | Learn | Python |\n| 04 | [Darwin Gödel Machine：自我修改的 agent](phases/15-autonomous-systems/04-darwin-godel-machine/) | Learn | Python |\n| 05 | [AI Scientist v2：研讨会级别的科研](phases/15-autonomous-systems/05-ai-scientist-v2/) | Learn | Python |\n| 06 | [自动化对齐研究（Anthropic AAR）](phases/15-autonomous-systems/06-automated-alignment-research/) | Learn | Python |\n| 07 | [递归式自我改进：能力 vs 对齐](phases/15-autonomous-systems/07-recursive-self-improvement/) | Learn | Python |\n| 08 | [有界自我改进的设计](phases/15-autonomous-systems/08-bounded-self-improvement/) | Learn | Python |\n| 09 | [自主编码 agent 全景（SWE-bench、CodeAct）](phases/15-autonomous-systems/09-coding-agent-landscape/) | Learn | Python |\n| 10 | [Claude Code 的权限模式与 Auto 模式](phases/15-autonomous-systems/10-claude-code-permission-modes/) | Learn | Python |\n| 11 | [浏览器 agent 与间接提示注入](phases/15-autonomous-systems/11-browser-agents/) | Learn | Python |\n| 12 | [长时运行 agent 的持久化执行](phases/15-autonomous-systems/12-durable-execution/) | Learn | Python |\n| 13 | [动作预算、迭代上限、成本管控](phases/15-autonomous-systems/13-cost-governors/) | Learn | Python |\n| 14 | [急停开关、熔断器、金丝雀 token](phases/15-autonomous-systems/14-kill-switches-canaries/) | Learn | Python |\n| 15 | [人在回路：先提议后提交](phases/15-autonomous-systems/15-propose-then-commit/) | Learn | Python |\n| 16 | [检查点与回滚](phases/15-autonomous-systems/16-checkpoints-rollback/) | Learn | Python |\n| 17 | [Constitutional AI 与规则覆盖](phases/15-autonomous-systems/17-constitutional-ai/) | Learn | Python |\n| 18 | [Llama Guard 与输入/输出分类](phases/15-autonomous-systems/18-llama-guard/) | Learn | Python |\n| 19 | [Anthropic 负责任扩展政策 v3.0](phases/15-autonomous-systems/19-anthropic-rsp/) | Learn | Python |\n| 20 | [OpenAI Preparedness 框架与 DeepMind FSF](phases/15-autonomous-systems/20-openai-preparedness-deepmind-fsf/) | Learn | Python |\n| 21 | [METR 时间跨度与外部评估](phases/15-autonomous-systems/21-metr-external-evaluation/) | Learn | Python |\n| 22 | [CAIS、CAISI 与社会规模风险](phases/15-autonomous-systems/22-cais-caisi-societal-risk/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-16\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 16 — 多 agent 与集群\u003c/b\u003e \u0026nbsp;\u003ccode\u003e25 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e协调、涌现，以及集体智能。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [为什么要多 agent](phases/16-multi-agent-and-swarms/01-why-multi-agent/) | Learn | TypeScript |\n| 02 | [FIPA-ACL 传承与言语行为](phases/16-multi-agent-and-swarms/02-fipa-acl-heritage/) | Learn | Python |\n| 03 | [通信协议](phases/16-multi-agent-and-swarms/03-communication-protocols/) | Build | TypeScript |\n| 04 | [多 agent 原语模型](phases/16-multi-agent-and-swarms/04-primitive-model/) | Learn | Python |\n| 05 | [Supervisor / 编排者-worker 模式](phases/16-multi-agent-and-swarms/05-supervisor-orchestrator-pattern/) | Build | Python |\n| 06 | [分层架构与分解漂移](phases/16-multi-agent-and-swarms/06-hierarchical-architecture/) | Learn | Python |\n| 07 | [心智社会与多 agent 辩论](phases/16-multi-agent-and-swarms/07-society-of-mind-debate/) | Build | Python |\n| 08 | [角色专精——规划者 / 批评者 / 执行者 / 验证者](phases/16-multi-agent-and-swarms/08-role-specialization/) | Build | Python |\n| 09 | [并行集群与网络化架构](phases/16-multi-agent-and-swarms/09-parallel-swarm-networks/) | Build | Python |\n| 10 | [群聊与发言人选择](phases/16-multi-agent-and-swarms/10-group-chat-speaker-selection/) | Build | Python |\n| 11 | [交接与例程（无状态编排）](phases/16-multi-agent-and-swarms/11-handoffs-and-routines/) | Build | Python |\n| 12 | [A2A——Agent 到 Agent 协议](phases/16-multi-agent-and-swarms/12-a2a-protocol/) | Build | Python |\n| 13 | [共享记忆与黑板模式](phases/16-multi-agent-and-swarms/13-shared-memory-blackboard/) | Build | Python |\n| 14 | [共识与拜占庭容错](phases/16-multi-agent-and-swarms/14-consensus-and-bft/) | Build | Python |\n| 15 | [投票、自洽性与辩论拓扑](phases/16-multi-agent-and-swarms/15-voting-debate-topology/) | Build | Python |\n| 16 | [协商与议价](phases/16-multi-agent-and-swarms/16-negotiation-bargaining/) | Build | Python |\n| 17 | [生成式 agent 与涌现式仿真](phases/16-multi-agent-and-swarms/17-generative-agents-simulation/) | Build | Python |\n| 18 | [心智理论与涌现式协调](phases/16-multi-agent-and-swarms/18-theory-of-mind-coordination/) | Build | Python |\n| 19 | [群体优化（PSO、ACO）](phases/16-multi-agent-and-swarms/19-swarm-optimization-pso-aco/) | Build | Python |\n| 20 | [MARL——MADDPG、QMIX、MAPPO](phases/16-multi-agent-and-swarms/20-marl-maddpg-qmix-mappo/) | Learn | Python |\n| 21 | [Agent 经济、token 激励、声誉](phases/16-multi-agent-and-swarms/21-agent-economies/) | Learn | Python |\n| 22 | [生产级扩展——队列、检查点、持久性](phases/16-multi-agent-and-swarms/22-production-scaling-queues-checkpoints/) | Build | Python |\n| 23 | [失败模式——MAST、群体思维、单一文化](phases/16-multi-agent-and-swarms/23-failure-modes-mast-groupthink/) | Learn | Python |\n| 24 | [评估与协调基准](phases/16-multi-agent-and-swarms/24-evaluation-coordination-benchmarks/) | Learn | Python |\n| 25 | [案例研究与 2026 最新进展](phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-17\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 17 — 基础设施与生产\u003c/b\u003e \u0026nbsp;\u003ccode\u003e28 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e把 AI 交付到真实世界。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [托管 LLM 平台 — Bedrock、Azure OpenAI、Vertex AI](phases/17-infrastructure-and-production/01-managed-llm-platforms/) | Learn | Python |\n| 02 | [推理平台经济学 — Fireworks、Together、Baseten、Modal](phases/17-infrastructure-and-production/02-inference-platform-economics/) | Learn | Python |\n| 03 | [Kubernetes 上的 GPU 自动扩缩 — Karpenter、KAI Scheduler](phases/17-infrastructure-and-production/03-gpu-autoscaling-kubernetes/) | Learn | Python |\n| 04 | [vLLM 服务内部机制 — PagedAttention、连续批处理、分块预填充](phases/17-infrastructure-and-production/04-vllm-serving-internals/) | Learn | Python |\n| 05 | [生产环境中的 EAGLE-3 推测解码](phases/17-infrastructure-and-production/05-eagle3-speculative-decoding/) | Learn | Python |\n| 06 | [面向前缀密集型负载的 SGLang 与 RadixAttention](phases/17-infrastructure-and-production/06-sglang-radixattention/) | Learn | Python |\n| 07 | [Blackwell 上用 FP8 与 NVFP4 的 TensorRT-LLM](phases/17-infrastructure-and-production/07-tensorrt-llm-blackwell/) | Learn | Python |\n| 08 | [推理指标 — TTFT、TPOT、ITL、Goodput、P99](phases/17-infrastructure-and-production/08-inference-metrics-goodput/) | Learn | Python |\n| 09 | [生产级量化 — AWQ、GPTQ、GGUF、FP8、NVFP4](phases/17-infrastructure-and-production/09-production-quantization/) | Learn | Python |\n| 10 | [无服务器 LLM 的冷启动缓解](phases/17-infrastructure-and-production/10-cold-start-mitigation/) | Learn | Python |\n| 11 | [多区域 LLM 服务与 KV 缓存局部性](phases/17-infrastructure-and-production/11-multi-region-kv-locality/) | Learn | Python |\n| 12 | [边缘推理 — ANE、Hexagon、WebGPU、Jetson](phases/17-infrastructure-and-production/12-edge-inference/) | Learn | Python |\n| 13 | [LLM 可观测性技术栈选型](phases/17-infrastructure-and-production/13-llm-observability/) | Learn | Python |\n| 14 | [提示缓存与语义缓存的经济学](phases/17-infrastructure-and-production/14-prompt-semantic-caching/) | Learn | Python |\n| 15 | [批处理 API — 50% 折扣作为行业标准](phases/17-infrastructure-and-production/15-batch-apis/) | Learn | Python |\n| 16 | [把模型路由作为降本原语](phases/17-infrastructure-and-production/16-model-routing/) | Learn | Python |\n| 17 | [预填充/解码分离 — NVIDIA Dynamo 与 llm-d](phases/17-infrastructure-and-production/17-disaggregated-prefill-decode/) | Learn | Python |\n| 18 | [带 LMCache KV 卸载的 vLLM 生产栈](phases/17-infrastructure-and-production/18-vllm-production-stack-lmcache/) | Learn | Python |\n| 19 | [AI 网关 — LiteLLM、Portkey、Kong、Bifrost](phases/17-infrastructure-and-production/19-ai-gateways/) | Learn | Python |\n| 20 | [影子、金丝雀与渐进式部署](phases/17-infrastructure-and-production/20-shadow-canary-progressive/) | Learn | Python |\n| 21 | [LLM 功能的 A/B 测试 — GrowthBook 与 Statsig](phases/17-infrastructure-and-production/21-ab-testing-llm-features/) | Learn | Python |\n| 22 | [LLM API 的负载测试 — k6、LLMPerf、GenAI-Perf](phases/17-infrastructure-and-production/22-load-testing-llm-apis/) | Build | Python |\n| 23 | [面向 AI 的 SRE — 多智能体事件响应](phases/17-infrastructure-and-production/23-sre-for-ai/) | Learn | Python |\n| 24 | [面向 LLM 生产的混沌工程](phases/17-infrastructure-and-production/24-chaos-engineering-llm/) | Learn | Python |\n| 25 | [安全 — 密钥、PII 脱敏、审计日志](phases/17-infrastructure-and-production/25-security-secrets-audit/) | Learn | Python |\n| 26 | [合规 — SOC 2、HIPAA、GDPR、EU AI Act、ISO 42001](phases/17-infrastructure-and-production/26-compliance-frameworks/) | Learn | Python |\n| 27 | [面向 LLM 的 FinOps — 单位经济与多租户归因](phases/17-infrastructure-and-production/27-finops-llms/) | Learn | Python |\n| 28 | [自托管服务选型 — llama.cpp、Ollama、TGI、vLLM、SGLang](phases/17-infrastructure-and-production/28-self-hosted-serving-selection/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-18\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 18 — 伦理、安全与对齐\u003c/b\u003e \u0026nbsp;\u003ccode\u003e30 lessons\u003c/code\u003e\u0026nbsp; \u003cem\u003e构建对人类有益的 AI。这不是选修。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Lesson | Type | Lang |\n|:---:|--------|:----:|------|\n| 01 | [把遵循指令当作对齐信号](phases/18-ethics-safety-alignment/01-instruction-following-alignment-signal/) | Learn | Python |\n| 02 | [奖励黑客与古德哈特定律](phases/18-ethics-safety-alignment/02-reward-hacking-goodhart/) | Learn | Python |\n| 03 | [直接偏好优化家族](phases/18-ethics-safety-alignment/03-direct-preference-optimization-family/) | Learn | Python |\n| 04 | [阿谀奉承：RLHF 的放大效应](phases/18-ethics-safety-alignment/04-sycophancy-rlhf-amplification/) | Learn | Python |\n| 05 | [Constitutional AI 与 RLAIF](phases/18-ethics-safety-alignment/05-constitutional-ai-rlaif/) | Learn | Python |\n| 06 | [Mesa 优化与欺骗性对齐](phases/18-ethics-safety-alignment/06-mesa-optimization-deceptive-alignment/) | Learn | Python |\n| 07 | [潜伏 agent——持续性欺骗](phases/18-ethics-safety-alignment/07-sleeper-agents-persistent-deception/) | Learn | Python |\n| 08 | [前沿模型中的上下文内谋划](phases/18-ethics-safety-alignment/08-in-context-scheming-frontier-models/) | Learn | Python |\n| 09 | [对齐造假](phases/18-ethics-safety-alignment/09-alignment-faking/) | Learn | Python |\n| 10 | [AI Control——即便被颠覆也保安全](phases/18-ethics-safety-alignment/10-ai-control-subversion/) | Learn | Python |\n| 11 | [可扩展监督与弱到强](phases/18-ethics-safety-alignment/11-scalable-oversight-weak-to-strong/) | Learn | Python |\n| 12 | [红队：PAIR 与自动化攻击](phases/18-ethics-safety-alignment/12-red-teaming-pair-automated-attacks/) | Build | Python |\n| 13 | [多样本越狱](phases/18-ethics-safety-alignment/13-many-shot-jailbreaking/) | Learn | Python |\n| 14 | [ASCII 字符画与视觉越狱](phases/18-ethics-safety-alignment/14-ascii-art-visual-jailbreaks/) | Build | Python |\n| 15 | [间接提示注入](phases/18-ethics-safety-alignment/15-indirect-prompt-injection/) | Build | Python |\n| 16 | [红队工具：Garak、Llama Guard、PyRIT](phases/18-ethics-safety-alignment/16-red-team-tooling-garak-llamaguard-pyrit/) | Build | Python |\n| 17 | [WMDP 与双用途能力评估](phases/18-ethics-safety-alignment/17-wmdp-dual-use-evaluation/) | Learn | Python |\n| 18 | [前沿安全框架——RSP、PF、FSF](phases/18-ethics-safety-alignment/18-frontier-safety-frameworks-rsp-pf-fsf/) | Learn | Python |\n| 19 | [模型福祉研究](phases/18-ethics-safety-alignment/19-model-welfare-research/) | Learn | Python |\n| 20 | [偏见与表征伤害](phases/18-ethics-safety-alignment/20-bias-representational-harm/) | Build | Python |\n| 21 | [公平性准则：群体、个体、反事实](phases/18-ethics-safety-alignment/21-fairness-criteria-group-individual-counterfactual/) | Learn | Python |\n| 22 | [面向 LLM 的差分隐私](phases/18-ethics-safety-alignment/22-differential-privacy-for-llms/) | Build | Python |\n| 23 | [水印：SynthID、Stable Signature、C2PA](phases/18-ethics-safety-alignment/23-watermarking-synthid-stable-signature-c2pa/) | Build | Python |\n| 24 | [监管框架：欧盟、美国、英国、韩国](phases/18-ethics-safety-alignment/24-regulatory-frameworks-eu-us-uk-korea/) | Learn | Python |\n| 25 | [EchoLeak 与 AI 的 CVE](phases/18-ethics-safety-alignment/25-echoleak-cves-for-ai/) | Learn | Python |\n| 26 | [模型卡、系统卡与数据集卡](phases/18-ethics-safety-alignment/26-model-system-dataset-cards/) | Build | Python |\n| 27 | [数据溯源与训练数据治理](phases/18-ethics-safety-alignment/27-data-provenance-training-governance/) | Learn | Python |\n| 28 | [对齐研究生态：MATS、Redwood、Apollo、METR](phases/18-ethics-safety-alignment/28-alignment-research-ecosystem/) | Learn | Python |\n| 29 | [内容审核系统：OpenAI、Perspective、Llama Guard](phases/18-ethics-safety-alignment/29-moderation-systems-openai-perspective-llamaguard/) | Build | Python |\n| 30 | [双用途风险：网络、生物、化学、核](phases/18-ethics-safety-alignment/30-dual-use-risk-cyber-bio-chem-nuclear/) | Learn | Python |\n\n\u003c/details\u003e\n\n\u003cdetails id=\"phase-19\"\u003e\n\u003csummary\u003e\u003cb\u003ePhase 19 — 综合项目\u003c/b\u003e \u0026nbsp;\u003ccode\u003e55 projects\u003c/code\u003e\u0026nbsp; \u003cem\u003e2026 年的端到端可交付产品，每个 20-40 小时。\u003c/em\u003e\u003c/summary\u003e\n\u003cbr/\u003e\n\n| # | Project | Combines | Lang |\n|:---:|---------|----------|------|\n| 01 | [终端原生编码 agent](phases/19-capstone-projects/01-terminal-native-coding-agent/) | 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/19-capstone-projects/02-rag-over-codebase/) | P5 P7 P11 P13 P17 | Python |\n| 03 | [实时语音助手（ASR → LLM → TTS）](phases/19-capstone-projects/03-realtime-voice-assistant/) | P6 P7 P11 P13 P14 P17 | Python |\n| 04 | [多模态文档问答（视觉优先）](phases/19-capstone-projects/04-multimodal-document-qa/) | P4 P5 P7 P11 P12 P17 | Python |\n| 05 | [自主科研 agent（AI-Scientist 级别）](phases/19-capstone-projects/05-autonomous-research-agent/) | P0 P2 P3 P7 P10 P14 P15 P16 P18 | Python |\n| 06 | [面向 Kubernetes 的 DevOps 排障 agent](phases/19-capstone-projects/06-devops-troubleshooting-agent/) | P11 P13 P14 P15 P17 P18 | Python |\n| 07 | [端到端微调流水线](phases/19-capstone-projects/07-end-to-end-fine-tuning-pipeline/) | 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/19-capstone-projects/08-production-rag-chatbot/) | P5 P7 P11 P12 P17 P18 | Python |\n| 09 | [代码迁移 agent（仓库级升级）](phases/19-capstone-projects/09-code-migration-agent/) | P5 P7 P11 P13 P14 P15 P17 | Python |\n| 10 | [多 agent 软件工程团队](phases/19-capstone-projects/10-multi-agent-software-team/) | P11 P13 P14 P15 P16 P17 | Python |\n| 11 | [LLM 可观测性与 Eval 仪表盘](phases/19-capstone-projects/11-llm-observability-dashboard/) | P11 P13 P17 P18 | Python |\n| 12 | [视频理解流水线（场景 → 问答）](phases/19-capstone-projects/12-video-understanding-pipeline/) | P4 P6 P7 P11 P12 P17 | Python |\n| 13 | [带注册表与治理的 MCP server](phases/19-capstone-projects/13-mcp-server-with-registry/) | P11 P13 P14 P17 P18 | Python |\n| 14 | [投机解码推理服务器](phases/19-capstone-projects/14-speculative-decoding-server/) | P3 P7 P10 P17 | Python |\n| 15 | [Constitutional 安全测试架 + 红队靶场](phases/19-capstone-projects/15-constitutional-safety-harness/) | P10 P11 P13 P14 P18 | Python |\n| 16 | [GitHub Issue 到 PR 的自主 agent](phases/19-capstone-projects/16-github-issue-to-pr-agent/) | P11 P13 P14 P15 P17 | Python |\n| 17 | [个人 AI 导师（自适应、多模态）](phases/19-capstone-projects/17-personal-ai-tutor/) | P5 P6 P11 P12 P14 P17 P18 | Python |\n| 20 | [Agent Harness Loop 契约](phases/19-capstone-projects/20-agent-harness-loop-contract/) | A. Agent harness | Python |\n| 21 | [带 Schema 校验的 Tool Registry](phases/19-capstone-projects/21-tool-registry-schema-validation/) | A. Agent harness | Python |\n| 22 | [基于换行分隔 stdio 的 JSON-RPC 2.0](phases/19-capstone-projects/22-jsonrpc-stdio-transport/) | A. Agent harness | Python |\n| 23 | [Function Call Dispatcher](phases/19-capstone-projects/23-function-call-dispatcher/) | A. Agent harness | Python |\n| 24 | [Plan-Execute 控制流](phases/19-capstone-projects/24-plan-execute-control-flow/) | A. Agent harness | Python |\n| 25 | [Verification Gate 与 Observation Budget](phases/19-capstone-projects/25-verification-gates-observation-budget/) | A. Agent harness | Python |\n| 26 | [带 Denylist 与 Path Jail 的 Sandbox Runner](phases/19-capstone-projects/26-sandbox-runner-denylist/) | A. Agent harness | Python |\n| 27 | [带 Fixture Tasks 的 Eval Harness](phases/19-capstone-projects/27-eval-harness-fixture-tasks/) | A. Agent harness | Python |\n| 28 | [用 OTel GenAI Span 与 Prometheus 做 Observability](phases/19-capstone-projects/28-observability-otel-traces/) | A. Agent harness | Python |\n| 29 | [端到端 Coding Agent Demo](phases/19-capstone-projects/29-end-to-end-coding-task-demo/) | A. Agent harness | Python |\n| 30 | [从零实现 BPE Tokenizer](phases/19-capstone-projects/30-bpe-tokenizer-from-scratch/) | B. NLP LLM | Python |\n| 31 | [带 Sliding Window 的 Tokenized Dataset](phases/19-capstone-projects/31-tokenized-dataset-sliding-window/) | B. NLP LLM | Python |\n| 32 | [Token Embedding 与 Positional Embedding](phases/19-capstone-projects/32-token-positional-embeddings/) | B. NLP LLM | Python |\n| 33 | [Multi-Head Self-Attention](phases/19-capstone-projects/33-multihead-self-attention/) | B. NLP LLM | Python |\n| 34 | [从零实现 Transformer Block](phases/19-capstone-projects/34-transformer-block/) | B. NLP LLM | Python |\n| 35 | [GPT 模型组装](phases/19-capstone-projects/35-gpt-model-assembly/) | B. NLP LLM | Python |\n| 36 | [训练循环与评估](phases/19-capstone-projects/36-training-loop-eval/) | B. NLP LLM | Python |\n| 37 | [加载预训练权重](phases/19-capstone-projects/37-loading-pretrained-weights/) | B. NLP LLM | Python |\n| 38 | [通过换 Head 做分类微调](phases/19-capstone-projects/38-classifier-finetuning/) | B. NLP LLM | Python |\n| 39 | [通过 SFT 做 Instruction Tuning](phases/19-capstone-projects/39-instruction-tuning-sft/) | B. NLP LLM | Python |\n| 40 | [从零实现 DPO](phases/19-capstone-projects/40-dpo-from-scratch/) | B. NLP LLM | Python |\n| 41 | [完整评估流水线](phases/19-capstone-projects/41-eval-pipeline/) | B. NLP LLM | Python |\n| 42 | [大规模语料下载器](phases/19-capstone-projects/42-large-corpus-downloader/) | C. 端到端训练 | Python |\n| 43 | [HDF5 Tokenized Corpus](phases/19-capstone-projects/43-hdf5-tokenized-corpus/) | C. 端到端训练 | Python |\n| 44 | [Cosine 学习率 + 线性 Warmup](phases/19-capstone-projects/44-cosine-lr-warmup/) | C. 端到端训练 | Python |\n| 45 | [Gradient Clipping 与混合精度训练](phases/19-capstone-projects/45-gradient-clipping-amp/) | C. 端到端训练 | Python |\n| 46 | [梯度累积](phases/19-capstone-projects/46-gradient-accumulation/) | C. 端到端训练 | Python |\n| 47 | [Checkpoint 保存与恢复](phases/19-capstone-projects/47-checkpoint-save-resume/) | C. 端到端训练 | Python |\n| 48 | [从零实现分布式数据并行与 FSDP](phases/19-capstone-projects/48-distributed-fsdp-ddp/) | C. 端到端训练 | Python |\n| 49 | [语言模型评测框架](phases/19-capstone-projects/49-lm-eval-harness/) | C. 端到端训练 | Python |\n| 50 | [假设生成器](phases/19-capstone-projects/50-hypothesis-generator/) | D. 自动研究 | Python |\n| 51 | [文献检索](phases/19-capstone-projects/51-literature-retrieval/) | D. 自动研究 | Python |\n| 52 | [实验执行器](phases/19-capstone-projects/52-experiment-runner/) | D. 自动研究 | Python |\n| 53 | [结果评估器](phases/19-capstone-projects/53-result-evaluator/) | D. 自动研究 | Python |\n| 54 | [论文生成器](phases/19-capstone-projects/54-paper-writer/) | D. 自动研究 | Python |\n| 55 | [评审循环](phases/19-capstone-projects/55-critic-loop/) | D. 自动研究 | Python |\n| 56 | [迭代调度器](phases/19-capstone-projects/56-iteration-scheduler/) | D. 自动研究 | Python |\n| 57 | [端到端研究 Demo](phases/19-capstone-projects/57-end-to-end-research-demo/) | D. 自动研究 | Python |\n\n\u003c/details\u003e\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/\n├── prompts/      覆盖每类 AI 任务的提示词模板\n└── skills/       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 / 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/**/outputs/` 下交付了 378 个技能和 99 个提示词。\n\n**推荐：通过 [skills.sh](https://skills.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/ai-engineering-from-scratch-zh                       # 所有技能\nnpx skills add fancyboi999/ai-engineering-from-scratch-zh --skill agent-loop    # 单个技能\nnpx skills add fancyboi999/ai-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/skills/`、`.cursor/skills/`、\n`.codex/skills/`、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/install_skills.py` 做离线 / 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/install_skills.py \u003ctarget\u003e                                 # 所有技能，默认 --layout skills（嵌套）\npython3 scripts/install_skills.py \u003ctarget\u003e --layout skills                 # 同上，显式写出\npython3 scripts/install_skills.py \u003ctarget\u003e --type all                      # 技能 + 提示词 + agent\npython3 scripts/install_skills.py \u003ctarget\u003e --phase 14                      # 只装一个阶段\npython3 scripts/install_skills.py \u003ctarget\u003e --tag rag                       # 按标签过滤\npython3 scripts/install_skills.py \u003ctarget\u003e --layout flat                   # 扁平文件\npython3 scripts/install_skills.py \u003ctarget\u003e --dry-run                       # 只预览，不写入\npython3 scripts/install_skills.py \u003ctarget\u003e --force                         # 覆盖已有文件\n```\n\n`\u003ctarget\u003e` 是你 agent 的技能目录（例如：\n`~/.claude/skills/`、`~/.cursor/skills/`、`~/.config/openclaw/skills/`、\n`.skills/`，或任何你 agent 读取的路径）。\n\n默认情况下，脚本拒绝覆盖已存在的目标，会列出每个冲突路径后以退出码 1 退出。\n用 `--dry-run` 预览冲突，或用 `--force` 覆盖。每次非 dry-run 的运行都会在目标里\n写一份 `manifest.json`，按类型和阶段分组列出完整清单。挑你 agent 读取的那种布局：\n\n| `--layout`  | Path written |\n|---|---|\n| `skills`    | `\u003ctarget\u003e/\u003cname\u003e/SKILL.md`（嵌套约定，Claude / Cursor / Codex / OpenClaw / Hermes 都支持） |\n| `by-phase`  | `\u003ctarget\u003e/phase-NN/\u003cname\u003e.md` |\n| `flat`      | `\u003ctarget\u003e/\u003cname\u003e.md` |\n\n### 把 agent 工作台塞进你自己的仓库\n\n阶段 14 的综合项目交付了一套可复用的 Agent Workbench 包（AGENTS.md、schema、\ninit / verify / handoff 脚本）。用下面的命令把它脚手架进任意仓库：\n\n```bash\npython3 scripts/scaffold_workbench.py path/to/your-repo            # 完整包 + 种子文件\npython3 scripts/scaffold_workbench.py path/to/your-repo --minimal  # 跳过 docs/\npython3 scripts/scaffold_workbench.py path/to/your-repo --dry-run  # 只预览\npython3 scripts/scaffold_workbench.py path/to/your-repo --force    # 覆盖\n```\n\n你会得到接好线的七个工作台界面、一份起步用的 `task_board.json`，\n以及一份 `schema_version: 1` 的全新 `agent_state.json`。从这里开始：编辑任务、\n编辑 `AGENTS.md`、运行 `scripts/init_agent.py`，把契约交给你的 agent。包的源码在\n`phases/14-agent-engineering/42-agent-workbench-capstone/outputs/agent-workbench-pack/`。\n\n### 把整套课程当 JSON 来浏览\n\n`scripts/build_catalog.py` 会遍历磁盘上的每个阶段、每节课、每件产物，\n在仓库根目录写出 `catalog.json`。一个文件，囊括课程的全部真相。\n\n```bash\npython3 scripts/build_catalog.py               # 写到 \u003crepo\u003e/catalog.json\npython3 scripts/build_catalog.py --stdout      # 输出到 stdout，不动仓库\npython3 scripts/build_catalog.py --out path/to/file.json\n```\n\n这份目录是从文件系统派生的，不是从 README 派生的，所以计数永远和磁盘上的实际内容一致。\n可以用它做站点构建、下游工具，或核对 README 的计数有没有漂移。Schema 在脚本顶部有说明。\n\n一个 GitHub Action（`.github/workflows/curriculum.yml`）会在每个 PR 上重建\n`catalog.json`，如果提交的文件过期就让构建失败。改完任意课程后，跑一遍\n`python3 scripts/build_catalog.py` 并提交结果，否则 CI 会拒掉 PR。同一个工作流还会以\nwarn-only 模式运行 `audit_lessons.py`（这样已有的漂移不会卡住贡献者）。\n\n### 给每节课的 Python 代码做冒烟检查\n\n`scripts/lesson_run.py` 会逐字节编译每节课 `code/` 目录下的每个 `.py` 文件。\n默认模式只做语法检查——不执行、不需要 API key、不需要重型 ML 依赖。专抓贡献者最常\n引入的回归（缩进错误、f-string 写坏、误改）。\n\n```bash\npython3 scripts/lesson_run.py                  # 语法检查整套课程\npython3 scripts/lesson_run.py --phase 14       # 只查一个阶段\npython3 scripts/lesson_run.py --json           # 在 stdout 输出 JSON 报告\npython3 scripts/lesson_run.py --strict         # 任一课程失败就 exit 1\npython3 scripts/lesson_run.py --execute        # 真正运行，每节课 10 秒超时\n```\n\n`--execute` 会运行每节课的 `code/main.py`（或第一个 `.py` 文件），10 秒超时。\n入口文件开头带 `# requires: pkg1, pkg2` 注释（列出非标准库依赖）的课程会被跳过，\n原因标为 `needs \u003cdeps\u003e`。这个脚本是可选的，没接入 CI。\n\n仅用标准库，Python 3.10+。设置 `LINK_CHECK_SKIP=domain1,domain2` 可以覆盖默认跳过名单\n（`twitter.com`、`x.com`、`linkedin.com`、`instagram.com`、`medium.com`——这些域名\n会主动屏蔽自动化的 HEAD/GET）。\n\n## 从哪里开始\n\n| Background | Start at | Estimated time |\n|---|---|---|\n| 编程和 AI 都是新手 | Phase 0 — 配置 | ~306 hours |\n| 会 Python，刚接触 ML | Phase 1 — 数学基础 | ~270 hours |\n| 懂 ML，刚接触深度学习 | Phase 3 — 深度学习核心 | ~200 hours |\n| 懂深度学习，想学 LLM 和 agent | Phase 10 — 从零实现 LLM | ~100 hours |\n| 资深工程师，只想要 agent 工程 | Phase 14 — Agent 工程 | ~60 hours |\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## 为什么这件事现在很重要\n\n\u003ctable\u003e\n\u003ctr\u003e\n\u003cth align=\"left\" width=\"50%\"\u003e\u003csub\u003eFIG_003 · A\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003eTHE INDUSTRY SIGNAL\u003c/b\u003e\u003c/th\u003e\n\u003cth align=\"left\" width=\"50%\"\u003e\u003csub\u003eFIG_003 · B\u003c/sub\u003e\u003cbr/\u003e\u003cb\u003eFOUNDATIONAL PAPERS COVERED\u003c/b\u003e\u003c/th\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd valign=\"top\"\u003e\n\n\u003e *\"最热门的新编程语言，是英语。\"*\u003cbr/\u003e\n\u003e — **Andrej Karpathy** ([tweet](https://x.com/karpathy/status/1617979122625712128))\n\n\u003e *\"软件工程正在我们眼前被重塑。\"*\u003cbr/\u003e\n\u003e — **Boris Cherny**，Claude Code 的作者\n\n\u003e *\"模型只会越来越强。真正会复利增长的技能，是**知道该造什么**。\"*\u003cbr/\u003e\n\u003e — 行业共识，2026\n\n\u003c/td\u003e\n\u003ctd valign=\"top\"\u003e\n\n- *Attention Is All You Need* — Vaswani et al., 2017 → [Phase 7](#phase-7)\n- *Language Models are Few-Shot Learners* (GPT-3) → [Phase 10](#phase-10)\n- *Denoising Diffusion Probabilistic Models* → [Phase 8](#phase-8)\n- *InstructGPT / RLHF* → [Phase 10](#phase-10)\n- *Direct Preference Optimization* → [Phase 10](#phase-10)\n- *Chain-of-Thought Prompting* → [Phase 11](#phase-11)\n- *ReAct: Reasoning + Acting in LLMs* → [Phase 14](#phase-14)\n- *Model Context Protocol* — Anthropic → [Phase 13](#phase-13)\n\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/table\u003e\n\n```\n░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒░░░▒▒▒\n```\n\n## 参与贡献\n\n| Goal | Read |\n|---|---|\n| 贡献一节课或一处修复 | [CONTRIBUTING.md](CONTRIBUTING.md) |\n| 为你的团队或学校 fork | [FORKING.md](FORKING.md) |\n| 课程模板 | [LESSON_TEMPLATE.md](LESSON_TEMPLATE.md) |\n| 跟踪进度 | [ROADMAP.md](ROADMAP.md) |\n| 术语表 | [glossary/terms.md](glossary/terms.md) |\n| 行为准则 | [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md) |\n\n提交一节课之前，先跑一遍不变量检查：\n\n```bash\npython3 scripts/audit_lessons.py           # 整套课程\npython3 scripts/audit_lessons.py --phase 14  # 单个阶段\npython3 scripts/audit_lessons.py --json    # 适合 CI 的输出\n```\n\n任一规则失败时退出码非零。规则（L001–L010）会校验目录结构、`docs/zh.md` 是否存在\n及是否有 H1、`code/` 是否非空、`quiz.json` 的 schema（拒绝引发 issue #102 的旧版\n`q/choices/answer` 键），以及课程文档里的相对链接。\n\n## Star 历史\n\n\u003ca href=\"https://star-history.com/#fancyboi999/ai-engineering-from-scratch-zh\u0026Date\"\u003e\n  \u003cpicture\u003e\n    \u003csource media=\"(prefers-color-scheme: dark)\" srcset=\"https://api.star-history.com/svg?repos=fancyboi999/ai-engineering-from-scratch-zh\u0026type=Date\u0026theme=dark\"\u003e\n    \u003cimg alt=\"Star history\" src=\"https://api.star-history.com/svg?repos=fancyboi999/ai-engineering-from-scratch-zh\u0026type=Date\" width=\"100%\"\u003e\n  \u003c/picture\u003e\n\u003c/a\u003e\n\n如果这份手册帮到了你，给仓库点个 star。这能让项目活下去。\n\n## 许可\n\nMIT。随你怎么用——fork、拿去教学、拿去卖、拿去交付。欢迎署名，但不强制。\n\n由 [Rohit Ghumare](https://github.com/rohitg00) 和社区共同维护。\n\n\u003csub\u003e\n  \u003ca href=\"https://x.com/ghumare64\"\u003e@ghumare64\u003c/a\u003e \u0026nbsp;·\u0026nbsp;\n  \u003ca href=\"https://aieng-zh.cn\"\u003eaieng-zh.cn\u003c/a\u003e \u0026nbsp;·\u0026nbsp;\n  \u003ca href=\"https://github.com/fancyboi999/ai-engineering-from-scratch-zh/issues/new/choose\"\u003eReport / Suggest\u003c/a\u003e\n\u003c/sub\u003e\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffancyboi999%2Fai-engineering-from-scratch-zh","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ffancyboi999%2Fai-engineering-from-scratch-zh","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffancyboi999%2Fai-engineering-from-scratch-zh/lists"}