# 路线图

每个阶段、每节课的状态追踪表。本文件里的状态符号会喂给网站（`site/build.js` 会把它们解析进 `site/data.js`），别改它们的形状。

总预计时长：约 342 小时，按自己的节奏来。

**图例：** ✅ 已完成  ·  🚧 进行中  ·  ⬚ 计划中

## Phase 0: 配置与工具 — ✅ (~14 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 开发环境 | ✅ | ~75 min |
| 02 | Git 与协作 | ✅ | ~45 min |
| 03 | GPU 配置与云端 | ✅ | ~75 min |
| 04 | API 与密钥 | ✅ | ~75 min |
| 05 | Jupyter Notebook | ✅ | ~75 min |
| 06 | Python 环境 | ✅ | ~75 min |
| 07 | 面向 AI 的 Docker | ✅ | ~75 min |
| 08 | 编辑器配置 | ✅ | ~75 min |
| 09 | 数据管理 | ✅ | ~75 min |
| 10 | 终端与 Shell | ✅ | ~45 min |
| 11 | 面向 AI 的 Linux | ✅ | ~45 min |
| 12 | 调试与性能分析 | ✅ | ~75 min |

## Phase 1: 数学基础 — ✅ (~23 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 线性代数直觉 | ✅ | ~45 min |
| 02 | 向量、矩阵与运算 | ✅ | ~75 min |
| 03 | 矩阵变换与特征值 | ✅ | ~75 min |
| 04 | 面向 ML 的微积分 — 导数与梯度 | ✅ | ~45 min |
| 05 | 链式法则与自动微分 | ✅ | ~75 min |
| 06 | 概率与分布 | ✅ | ~45 min |
| 07 | 贝叶斯定理与统计思维 | ✅ | ~75 min |
| 08 | 优化 — 梯度下降家族 | ✅ | ~75 min |
| 09 | 信息论 — 熵、KL 散度 | ✅ | ~45 min |
| 10 | 降维 — PCA、t-SNE、UMAP | ✅ | ~75 min |
| 11 | 奇异值分解 | ✅ | ~75 min |
| 12 | 张量运算 | ✅ | ~75 min |
| 13 | 数值稳定性 | ✅ | ~45 min |
| 14 | 范数与距离 | ✅ | ~45 min |
| 15 | 面向 ML 的统计学 | ✅ | ~45 min |
| 16 | 采样方法 | ✅ | ~75 min |
| 17 | 线性方程组 | ✅ | ~75 min |
| 18 | 凸优化 | ✅ | ~75 min |
| 19 | 面向 AI 的复数 | ✅ | ~45 min |
| 20 | 傅里叶变换 | ✅ | ~75 min |
| 21 | 面向 ML 的图论 | ✅ | ~45 min |
| 22 | 随机过程 | ✅ | ~45 min |

## Phase 2: 机器学习基础 — ✅ (~21 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 什么是机器学习 — 类型与分类 | ✅ | ~45 min |
| 02 | 从零实现线性回归 | ✅ | ~75 min |
| 03 | 逻辑回归与分类 | ✅ | ~75 min |
| 04 | 决策树与随机森林 | ✅ | ~75 min |
| 05 | 支持向量机 | ✅ | ~75 min |
| 06 | K 近邻与距离度量 | ✅ | ~75 min |
| 07 | 无监督学习 — K-Means、DBSCAN | ✅ | ~75 min |
| 08 | 特征工程与特征选择 | ✅ | ~75 min |
| 09 | 模型评估 — 指标、交叉验证 | ✅ | ~75 min |
| 10 | 偏差、方差与学习曲线 | ✅ | ~45 min |
| 11 | 集成方法 — Boosting、Bagging、Stacking | ✅ | ~75 min |
| 12 | 超参数调优与 AutoML | ✅ | ~75 min |
| 13 | ML 流水线与实验追踪 | ✅ | ~75 min |
| 14 | 朴素贝叶斯 — 多项式、高斯、伯努利 | ✅ | ~75 min |
| 15 | 时间序列基础 | ✅ | ~45 min |
| 16 | 异常检测 | ✅ | ~75 min |
| 17 | 处理不平衡数据 | ✅ | ~75 min |
| 18 | 特征选择 | ✅ | ~75 min |

## Phase 3: 深度学习核心 — ✅ (~15 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 感知机 — 一切的起点 | ✅ | ~45 min |
| 02 | 多层网络与前向传播 | ✅ | ~75 min |
| 03 | 从零实现反向传播 | ✅ | ~75 min |
| 04 | 激活函数 — ReLU、Sigmoid、GELU 及其原理 | ✅ | ~45 min |
| 05 | 损失函数 — MSE、交叉熵、对比损失 | ✅ | ~45 min |
| 06 | 优化器 — SGD、Momentum、Adam、AdamW | ✅ | ~75 min |
| 07 | 正则化 — Dropout、权重衰减、BatchNorm | ✅ | ~75 min |
| 08 | 权重初始化与训练稳定性 | ✅ | ~45 min |
| 09 | 学习率调度与预热 | ✅ | ~45 min |
| 10 | 打造你自己的迷你框架 | ✅ | ~120 min |
| 11 | PyTorch 入门 | ✅ | ~75 min |
| 12 | JAX 入门 | ✅ | ~75 min |
| 13 | 调试神经网络 | ✅ | ~75 min |

## Phase 4: 计算机视觉 — ✅ (~27 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 图像基础 — 像素、通道、色彩空间 | ✅ | ~45 min |
| 02 | 从零实现卷积 | ✅ | ~75 min |
| 03 | CNN — 从 LeNet 到 ResNet | ✅ | ~75 min |
| 04 | 图像分类 | ✅ | ~75 min |
| 05 | 迁移学习与微调 | ✅ | ~75 min |
| 06 | 目标检测 — 从零实现 YOLO | ✅ | ~75 min |
| 07 | 语义分割 — U-Net | ✅ | ~75 min |
| 08 | 实例分割 — Mask R-CNN | ✅ | ~75 min |
| 09 | 图像生成 — GAN | ✅ | ~75 min |
| 10 | 图像生成 — 扩散模型 | ✅ | ~75 min |
| 11 | Stable Diffusion — 架构与微调 | ✅ | ~75 min |
| 12 | 视频理解 — 时序建模 | ✅ | ~45 min |
| 13 | 3D 视觉 — 点云、NeRF | ✅ | ~45 min |
| 14 | Vision Transformer (ViT) | ✅ | ~45 min |
| 15 | 实时视觉 — 边缘部署 | ✅ | ~75 min |
| 16 | 构建完整的视觉流水线 | ✅ | ~120 min |
| 17 | 自监督视觉 — SimCLR、DINO、MAE | ✅ | ~75 min |
| 18 | 开放词表视觉 — CLIP | ✅ | ~45 min |
| 19 | OCR 与文档理解 | ✅ | ~45 min |
| 20 | 图像检索与度量学习 | ✅ | ~45 min |
| 21 | 关键点检测与姿态估计 | ✅ | ~45 min |
| 22 | 从零实现 3D 高斯泼溅 | ✅ | ~90 min |
| 23 | 扩散 Transformer 与整流流 | ✅ | ~75 min |
| 24 | SAM 3 与开放词表分割 | ✅ | ~60 min |
| 25 | 视觉语言模型 (ViT-MLP-LLM) | ✅ | ~75 min |
| 26 | 单目深度与几何估计 | ✅ | ~60 min |
| 27 | 多目标跟踪与视频记忆 | ✅ | ~60 min |
| 28 | 世界模型与视频扩散 | ✅ | ~75 min |

## Phase 5: NLP — 从基础到进阶 — ✅ (~30 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [文本处理 — 分词、词干提取、词形还原](phases/05-nlp-foundations-to-advanced/01-text-processing) | ✅ | ~45 min |
| 02 | [词袋、TF-IDF 与文本表示](phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf) | ✅ | ~75 min |
| 03 | [词嵌入 — 从零实现 Word2Vec](phases/05-nlp-foundations-to-advanced/03-word-embeddings-word2vec) | ✅ | ~75 min |
| 04 | [GloVe、FastText 与子词嵌入](phases/05-nlp-foundations-to-advanced/04-glove-fasttext-subword) | ✅ | ~45 min |
| 05 | [情感分析](phases/05-nlp-foundations-to-advanced/05-sentiment-analysis) | ✅ | ~75 min |
| 06 | [命名实体识别 (NER)](phases/05-nlp-foundations-to-advanced/06-named-entity-recognition) | ✅ | ~75 min |
| 07 | [词性标注与句法分析](phases/05-nlp-foundations-to-advanced/07-pos-tagging-parsing) | ✅ | ~45 min |
| 08 | [文本分类 — 用于文本的 CNN 与 RNN](phases/05-nlp-foundations-to-advanced/08-cnns-rnns-for-text) | ✅ | ~75 min |
| 09 | [序列到序列模型](phases/05-nlp-foundations-to-advanced/09-sequence-to-sequence) | ✅ | ~75 min |
| 10 | [注意力机制 — 突破性进展](phases/05-nlp-foundations-to-advanced/10-attention-mechanism) | ✅ | ~45 min |
| 11 | [机器翻译](phases/05-nlp-foundations-to-advanced/11-machine-translation) | ✅ | ~75 min |
| 12 | [文本摘要](phases/05-nlp-foundations-to-advanced/12-text-summarization) | ✅ | ~75 min |
| 13 | [问答系统](phases/05-nlp-foundations-to-advanced/13-question-answering) | ✅ | ~75 min |
| 14 | [信息检索与搜索](phases/05-nlp-foundations-to-advanced/14-information-retrieval-search) | ✅ | ~75 min |
| 15 | [主题建模 — LDA、BERTopic](phases/05-nlp-foundations-to-advanced/15-topic-modeling) | ✅ | ~45 min |
| 16 | [文本生成 — Transformer 之前的语言模型](phases/05-nlp-foundations-to-advanced/16-text-generation-pre-transformer) | ✅ | ~45 min |
| 17 | [聊天机器人 — 从规则驱动到神经网络](phases/05-nlp-foundations-to-advanced/17-chatbots-rule-to-neural) | ✅ | ~75 min |
| 18 | [多语言 NLP](phases/05-nlp-foundations-to-advanced/18-multilingual-nlp) | ✅ | ~45 min |
| 19 | [子词分词 — BPE、WordPiece、Unigram、SentencePiece](phases/05-nlp-foundations-to-advanced/19-subword-tokenization) | ✅ | ~60 min |
| 20 | [结构化输出与约束解码](phases/05-nlp-foundations-to-advanced/20-structured-outputs-constrained-decoding) | ✅ | ~60 min |
| 21 | [自然语言推理与文本蕴含](phases/05-nlp-foundations-to-advanced/21-nli-textual-entailment) | ✅ | ~60 min |
| 22 | [嵌入模型深入剖析](phases/05-nlp-foundations-to-advanced/22-embedding-models-deep-dive) | ✅ | ~60 min |
| 23 | [面向 RAG 的分块策略](phases/05-nlp-foundations-to-advanced/23-chunking-strategies-rag) | ✅ | ~60 min |
| 24 | [共指消解](phases/05-nlp-foundations-to-advanced/24-coreference-resolution) | ✅ | ~60 min |
| 25 | [实体链接与消歧](phases/05-nlp-foundations-to-advanced/25-entity-linking) | ✅ | ~60 min |
| 26 | [关系抽取与知识图谱构建](phases/05-nlp-foundations-to-advanced/26-relation-extraction-kg) | ✅ | ~60 min |
| 27 | [LLM 评估 — RAGAS、DeepEval、G-Eval](phases/05-nlp-foundations-to-advanced/27-llm-evaluation-frameworks) | ✅ | ~75 min |
| 28 | [长上下文评估 — NIAH、RULER、LongBench、MRCR](phases/05-nlp-foundations-to-advanced/28-long-context-evaluation) | ✅ | ~60 min |
| 29 | [对话状态追踪](phases/05-nlp-foundations-to-advanced/29-dialogue-state-tracking) | ✅ | ~75 min |

## Phase 6: 语音与音频 — ✅ (~18 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [音频基础 — 波形、采样、傅里叶变换](phases/06-speech-and-audio/01-audio-fundamentals) | ✅ | ~45 min |
| 02 | [频谱图、梅尔标度与音频特征](phases/06-speech-and-audio/02-spectrograms-mel-features) | ✅ | ~45 min |
| 03 | [音频分类](phases/06-speech-and-audio/03-audio-classification) | ✅ | ~75 min |
| 04 | [语音识别 (ASR)](phases/06-speech-and-audio/04-speech-recognition-asr) | ✅ | ~45 min |
| 05 | [Whisper — 架构与微调](phases/06-speech-and-audio/05-whisper-architecture-finetuning) | ✅ | ~75 min |
| 06 | [说话人识别与验证](phases/06-speech-and-audio/06-speaker-recognition-verification) | ✅ | ~45 min |
| 07 | [文本转语音 (TTS)](phases/06-speech-and-audio/07-text-to-speech) | ✅ | ~75 min |
| 08 | [声音克隆与音色转换](phases/06-speech-and-audio/08-voice-cloning-conversion) | ✅ | ~75 min |
| 09 | [音乐生成](phases/06-speech-and-audio/09-music-generation) | ✅ | ~75 min |
| 10 | [音频语言模型](phases/06-speech-and-audio/10-audio-language-models) | ✅ | ~45 min |
| 11 | [实时音频处理](phases/06-speech-and-audio/11-real-time-audio-processing) | ✅ | ~75 min |
| 12 | [构建语音助手流水线](phases/06-speech-and-audio/12-voice-assistant-pipeline) | ✅ | ~120 min |
| 13 | [神经音频编解码器 — EnCodec、SNAC、Mimi、DAC](phases/06-speech-and-audio/13-neural-audio-codecs) | ✅ | ~60 min |
| 14 | [语音活动检测与轮次切换](phases/06-speech-and-audio/14-voice-activity-detection-turn-taking) | ✅ | ~45 min |
| 15 | [流式语音到语音 — Moshi、Hibiki](phases/06-speech-and-audio/15-streaming-speech-to-speech-moshi-hibiki) | ✅ | ~75 min |
| 16 | [语音防伪与音频水印](phases/06-speech-and-audio/16-anti-spoofing-audio-watermarking) | ✅ | ~75 min |
| 17 | [音频评估 — WER、MOS、MMAU、排行榜](phases/06-speech-and-audio/17-audio-evaluation-metrics) | ✅ | ~60 min |

## Phase 7: Transformer 深入剖析 — ✅ (~14 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [为什么需要 Transformer — RNN 的问题所在](phases/07-transformers-deep-dive/01-why-transformers) | ✅ | ~45 min |
| 02 | [从零实现自注意力](phases/07-transformers-deep-dive/02-self-attention-from-scratch) | ✅ | ~75 min |
| 03 | [多头注意力](phases/07-transformers-deep-dive/03-multi-head-attention) | ✅ | ~75 min |
| 04 | [位置编码 — 正弦、RoPE、ALiBi](phases/07-transformers-deep-dive/04-positional-encoding) | ✅ | ~45 min |
| 05 | [完整的 Transformer — 编码器 + 解码器](phases/07-transformers-deep-dive/05-full-transformer) | ✅ | ~75 min |
| 06 | [BERT — 掩码语言建模](phases/07-transformers-deep-dive/06-bert-masked-language-modeling) | ✅ | ~45 min |
| 07 | [GPT — 因果语言建模](phases/07-transformers-deep-dive/07-gpt-causal-language-modeling) | ✅ | ~75 min |
| 08 | [T5、BART — 编码器-解码器模型](phases/07-transformers-deep-dive/08-t5-bart-encoder-decoder) | ✅ | ~45 min |
| 09 | [Vision Transformer (ViT)](phases/07-transformers-deep-dive/09-vision-transformers) | ✅ | ~45 min |
| 10 | [音频 Transformer — Whisper 架构](phases/07-transformers-deep-dive/10-audio-transformers-whisper) | ✅ | ~45 min |
| 11 | [混合专家 (MoE)](phases/07-transformers-deep-dive/11-mixture-of-experts) | ✅ | ~45 min |
| 12 | [KV 缓存、Flash Attention 与推理优化](phases/07-transformers-deep-dive/12-kv-cache-flash-attention) | ✅ | ~75 min |
| 13 | [缩放定律](phases/07-transformers-deep-dive/13-scaling-laws) | ✅ | ~45 min |
| 14 | [从零实现一个 Transformer — 总结项目](phases/07-transformers-deep-dive/14-build-a-transformer-capstone) | ✅ | ~120 min |
| 15 | [Attention 变体——滑动窗口、稀疏、差分](phases/07-transformers-deep-dive/15-attention-variants) | ✅ | ~90 min |
| 16 | [投机解码——草稿、验证、重复](phases/07-transformers-deep-dive/16-speculative-decoding) | ✅ | ~90 min |

## Phase 8: 生成式 AI — ✅ (~14 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [生成模型 — 分类与历史](phases/08-generative-ai/01-generative-models-taxonomy-history/) | ✅ | ~45 min |
| 02 | [自编码器与 VAE](phases/08-generative-ai/02-autoencoders-vae/) | ✅ | ~75 min |
| 03 | [GAN — 生成器对判别器](phases/08-generative-ai/03-gans-generator-discriminator/) | ✅ | ~75 min |
| 04 | [条件 GAN 与 Pix2Pix](phases/08-generative-ai/04-conditional-gans-pix2pix/) | ✅ | ~75 min |
| 05 | [StyleGAN](phases/08-generative-ai/05-stylegan/) | ✅ | ~45 min |
| 06 | [扩散模型 — 从零实现 DDPM](phases/08-generative-ai/06-diffusion-ddpm-from-scratch/) | ✅ | ~75 min |
| 07 | [潜在扩散与 Stable Diffusion](phases/08-generative-ai/07-latent-diffusion-stable-diffusion/) | ✅ | ~75 min |
| 08 | [ControlNet、LoRA 与图像条件控制](phases/08-generative-ai/08-controlnet-lora-conditioning/) | ✅ | ~75 min |
| 09 | [图像修复、扩展与编辑](phases/08-generative-ai/09-inpainting-outpainting-editing/) | ✅ | ~75 min |
| 10 | [视频生成](phases/08-generative-ai/10-video-generation/) | ✅ | ~45 min |
| 11 | [音频生成](phases/08-generative-ai/11-audio-generation/) | ✅ | ~45 min |
| 12 | [3D 生成](phases/08-generative-ai/12-3d-generation/) | ✅ | ~45 min |
| 13 | [流匹配与整流流](phases/08-generative-ai/13-flow-matching-rectified-flows/) | ✅ | ~45 min |
| 14 | [评估 — FID、CLIP Score、人类偏好](phases/08-generative-ai/14-evaluation-fid-clip-score/) | ✅ | ~45 min |
| 19 | [视觉自回归建模（VAR）：下一尺度预测](phases/08-generative-ai/19-visual-autoregressive-var) | ✅ | ~90 min |

## Phase 9: 强化学习 — ✅ (~13 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | MDP、状态、动作与奖励 | ✅ | ~45 min |
| 02 | 动态规划 | ✅ | ~75 min |
| 03 | 蒙特卡洛方法 | ✅ | ~75 min |
| 04 | 时序差分 — Q-Learning、SARSA | ✅ | ~75 min |
| 05 | 深度 Q 网络 (DQN) | ✅ | ~75 min |
| 06 | 策略梯度方法 — REINFORCE | ✅ | ~75 min |
| 07 | Actor-Critic — A2C、A3C | ✅ | ~75 min |
| 08 | 近端策略优化 (PPO) | ✅ | ~75 min |
| 09 | 奖励建模与 RLHF | ✅ | ~45 min |
| 10 | 多智能体强化学习 | ✅ | ~45 min |
| 11 | 从仿真到现实的迁移 | ✅ | ~45 min |
| 12 | 游戏中的强化学习 | ✅ | ~75 min |

## Phase 10: 从零实现 LLM — ✅ (~26 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [分词器 — BPE、WordPiece、SentencePiece](phases/10-llms-from-scratch/01-tokenizers) | ✅ | ~45 min |
| 02 | [从零实现一个分词器](phases/10-llms-from-scratch/02-building-a-tokenizer) | ✅ | ~75 min |
| 03 | [预训练数据流水线](phases/10-llms-from-scratch/03-data-pipelines) | ✅ | ~75 min |
| 04 | [预训练一个迷你 GPT (124M)](phases/10-llms-from-scratch/04-pre-training-mini-gpt) | ✅ | ~120 min |
| 05 | [扩展 — 分布式训练、FSDP、DeepSpeed](phases/10-llms-from-scratch/05-scaling-distributed) | ✅ | ~75 min |
| 06 | [指令微调 — SFT](phases/10-llms-from-scratch/06-instruction-tuning-sft) | ✅ | ~75 min |
| 07 | [RLHF — 奖励模型 + PPO 训练](phases/10-llms-from-scratch/07-rlhf) | ✅ | ~75 min |
| 08 | [DPO — 直接偏好优化](phases/10-llms-from-scratch/08-dpo) | ✅ | ~75 min |
| 09 | [Constitutional AI 与自我改进](phases/10-llms-from-scratch/09-constitutional-ai-self-improvement) | ✅ | ~45 min |
| 10 | [评估 — 基准、Eval、LM Harness](phases/10-llms-from-scratch/10-evaluation) | ✅ | ~75 min |
| 11 | [量化 — INT8、GPTQ、AWQ、GGUF](phases/10-llms-from-scratch/11-quantization) | ✅ | ~75 min |
| 12 | [推理优化](phases/10-llms-from-scratch/12-inference-optimization) | ✅ | ~75 min |
| 13 | [构建完整的 LLM 流水线](phases/10-llms-from-scratch/13-building-complete-llm-pipeline) | ✅ | ~120 min |
| 14 | [开源模型 — 架构逐步剖析](phases/10-llms-from-scratch/14-open-models-architecture-walkthroughs) | ✅ | ~45 min |
| 15 | [推测解码与 EAGLE-3](phases/10-llms-from-scratch/15-speculative-decoding-eagle3) | ✅ | ~75 min |
| 16 | [差分注意力 (V2)](phases/10-llms-from-scratch/16-differential-attention-v2) | ✅ | ~60 min |
| 17 | [原生稀疏注意力 (DeepSeek NSA)](phases/10-llms-from-scratch/17-native-sparse-attention) | ✅ | ~60 min |
| 18 | [多 token 预测 (MTP)](phases/10-llms-from-scratch/18-multi-token-prediction) | ✅ | ~60 min |
| 19 | [DualPipe 并行](phases/10-llms-from-scratch/19-dualpipe-parallelism) | ✅ | ~60 min |
| 20 | [DeepSeek-V3 架构逐步剖析](phases/10-llms-from-scratch/20-deepseek-v3-walkthrough) | ✅ | ~75 min |
| 21 | [Jamba — 混合 SSM-Transformer](phases/10-llms-from-scratch/21-jamba-hybrid-ssm-transformer) | ✅ | ~60 min |
| 22 | [异步与 Hogwild! 推理](phases/10-llms-from-scratch/22-async-hogwild-inference) | ✅ | ~60 min |
| 25 | [推测解码与 EAGLE](phases/10-llms-from-scratch/25-speculative-decoding) | ✅ | ~90 min |
| 34 | [梯度检查点与激活重算](phases/10-llms-from-scratch/34-gradient-checkpointing) | ✅ | ~90 min |

## Phase 11: LLM 工程 — ✅ (~17 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [提示工程 — 技巧与模式](phases/11-llm-engineering/01-prompt-engineering) | ✅ | ~45 min |
| 02 | [少样本、思维链、思维树](phases/11-llm-engineering/02-few-shot-cot) | ✅ | ~45 min |
| 03 | [结构化输出](phases/11-llm-engineering/03-structured-outputs) | ✅ | ~75 min |
| 04 | [嵌入与向量表示](phases/11-llm-engineering/04-embeddings) | ✅ | ~75 min |
| 05 | [上下文工程](phases/11-llm-engineering/05-context-engineering) | ✅ | ~75 min |
| 06 | [RAG — 检索增强生成](phases/11-llm-engineering/06-rag) | ✅ | ~75 min |
| 07 | [进阶 RAG](phases/11-llm-engineering/07-advanced-rag) | ✅ | ~75 min |
| 08 | [用 LoRA 与 QLoRA 微调](phases/11-llm-engineering/08-fine-tuning-lora) | ✅ | ~75 min |
| 09 | [函数调用与工具使用](phases/11-llm-engineering/09-function-calling) | ✅ | ~75 min |
| 10 | [评估与测试 LLM 应用](phases/11-llm-engineering/10-evaluation) | ✅ | ~45 min |
| 11 | [缓存、限流与成本优化](phases/11-llm-engineering/11-caching-cost) | ✅ | ~45 min |
| 12 | [护栏、安全与内容过滤](phases/11-llm-engineering/12-guardrails) | ✅ | ~45 min |
| 13 | [构建生产级 LLM 应用](phases/11-llm-engineering/13-production-app) | ✅ | ~120 min |
| 14 | [模型上下文协议 (MCP)](phases/11-llm-engineering/14-model-context-protocol) | ✅ | ~75 min |
| 15 | [提示缓存与上下文缓存](phases/11-llm-engineering/15-prompt-caching) | ✅ | ~60 min |

## Phase 12: 多模态 AI — ✅ (~65 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [Vision Transformer 与图块-token 原语](phases/12-multimodal-ai/01-vision-transformer-patch-tokens) | ✅ | ~120 min |
| 02 | [CLIP 与对比式视觉语言预训练](phases/12-multimodal-ai/02-clip-contrastive-pretraining) | ✅ | ~180 min |
| 03 | [BLIP-2 与作为模态桥梁的 Q-Former](phases/12-multimodal-ai/03-blip2-qformer-bridge) | ✅ | ~180 min |
| 04 | [Flamingo 与门控交叉注意力](phases/12-multimodal-ai/04-flamingo-gated-cross-attention) | ✅ | ~120 min |
| 05 | [LLaVA 与视觉指令微调](phases/12-multimodal-ai/05-llava-visual-instruction-tuning) | ✅ | ~180 min |
| 06 | [任意分辨率视觉：Patch-n'-Pack 与 NaFlex](phases/12-multimodal-ai/06-any-resolution-patch-n-pack) | ✅ | ~120 min |
| 07 | [开源权重 VLM 配方：真正重要的是什么](phases/12-multimodal-ai/07-open-weight-vlm-recipes) | ✅ | ~180 min |
| 08 | [LLaVA-OneVision：单图、多图、视频](phases/12-multimodal-ai/08-llava-onevision-single-multi-video) | ✅ | ~180 min |
| 09 | [Qwen-VL 家族与动态帧率视频](phases/12-multimodal-ai/09-qwen-vl-family-dynamic-fps) | ✅ | ~120 min |
| 10 | [InternVL3 原生多模态预训练](phases/12-multimodal-ai/10-internvl3-native-multimodal) | ✅ | ~120 min |
| 11 | [Chameleon 与早融合纯 token](phases/12-multimodal-ai/11-chameleon-early-fusion-tokens) | ✅ | ~180 min |
| 12 | [Emu3 用下一 token 预测做生成](phases/12-multimodal-ai/12-emu3-next-token-for-generation) | ✅ | ~120 min |
| 13 | [Transfusion 自回归 + 扩散](phases/12-multimodal-ai/13-transfusion-autoregressive-diffusion) | ✅ | ~180 min |
| 14 | [Show-o 与离散扩散统一](phases/12-multimodal-ai/14-show-o-discrete-diffusion-unified) | ✅ | ~120 min |
| 15 | [Janus-Pro 解耦编码器](phases/12-multimodal-ai/15-janus-pro-decoupled-encoders) | ✅ | ~120 min |
| 16 | [MIO 任意到任意流式处理](phases/12-multimodal-ai/16-mio-any-to-any-streaming) | ✅ | ~120 min |
| 17 | [视频语言时序定位](phases/12-multimodal-ai/17-video-language-temporal-grounding) | ✅ | ~180 min |
| 18 | [百万 token 上下文下的长视频理解](phases/12-multimodal-ai/18-long-video-million-token) | ✅ | ~180 min |
| 19 | [音频语言模型：从 Whisper 到 AF3](phases/12-multimodal-ai/19-audio-language-whisper-to-af3) | ✅ | ~180 min |
| 20 | [全模态模型：Thinker-Talker](phases/12-multimodal-ai/20-omni-models-thinker-talker) | ✅ | ~180 min |
| 21 | [具身 VLA：RT-2、OpenVLA、π0、GR00T](phases/12-multimodal-ai/21-embodied-vlas-openvla-pi0-groot) | ✅ | ~180 min |
| 22 | [文档与图表理解](phases/12-multimodal-ai/22-document-diagram-understanding) | ✅ | ~180 min |
| 23 | [ColPali 视觉原生文档 RAG](phases/12-multimodal-ai/23-colpali-vision-native-rag) | ✅ | ~180 min |
| 24 | [多模态 RAG 与跨模态检索](phases/12-multimodal-ai/24-multimodal-rag-cross-modal) | ✅ | ~180 min |
| 25 | [多模态智能体与计算机操作 (总结项目)](phases/12-multimodal-ai/25-multimodal-agents-computer-use) | ✅ | ~240 min |

## Phase 13: 工具与协议 — ✅ (~40 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [工具接口](phases/13-tools-and-protocols/01-the-tool-interface/) | ✅ | ~45 min |
| 02 | [函数调用深入剖析](phases/13-tools-and-protocols/02-function-calling-deep-dive/) | ✅ | ~75 min |
| 03 | [并行与流式工具调用](phases/13-tools-and-protocols/03-parallel-and-streaming-tool-calls/) | ✅ | ~75 min |
| 04 | [结构化输出](phases/13-tools-and-protocols/04-structured-output/) | ✅ | ~75 min |
| 05 | [工具 Schema 设计](phases/13-tools-and-protocols/05-tool-schema-design/) | ✅ | ~45 min |
| 06 | [MCP 基础](phases/13-tools-and-protocols/06-mcp-fundamentals/) | ✅ | ~45 min |
| 07 | [构建一个 MCP Server](phases/13-tools-and-protocols/07-building-an-mcp-server/) | ✅ | ~75 min |
| 08 | [构建一个 MCP Client](phases/13-tools-and-protocols/08-building-an-mcp-client/) | ✅ | ~75 min |
| 09 | [MCP 传输层](phases/13-tools-and-protocols/09-mcp-transports/) | ✅ | ~45 min |
| 10 | [MCP 资源与 Prompt](phases/13-tools-and-protocols/10-mcp-resources-and-prompts/) | ✅ | ~45 min |
| 11 | [MCP Sampling](phases/13-tools-and-protocols/11-mcp-sampling/) | ✅ | ~75 min |
| 12 | [MCP Roots 与 Elicitation](phases/13-tools-and-protocols/12-mcp-roots-and-elicitation/) | ✅ | ~45 min |
| 13 | [MCP 异步任务](phases/13-tools-and-protocols/13-mcp-async-tasks/) | ✅ | ~75 min |
| 14 | [MCP Apps](phases/13-tools-and-protocols/14-mcp-apps/) | ✅ | ~75 min |
| 15 | [MCP 安全 I — 工具投毒](phases/13-tools-and-protocols/15-mcp-security-tool-poisoning/) | ✅ | ~45 min |
| 16 | [MCP 安全 II — OAuth 2.1](phases/13-tools-and-protocols/16-mcp-security-oauth-2-1/) | ✅ | ~75 min |
| 17 | [MCP 网关与注册中心](phases/13-tools-and-protocols/17-mcp-gateways-and-registries/) | ✅ | ~45 min |
| 18 | [生产环境中的 MCP 认证 — iii 上的 DCR + JWKS](phases/13-tools-and-protocols/18-mcp-auth-production/) | ✅ | ~90 min |
| 19 | [A2A 协议](phases/13-tools-and-protocols/19-a2a-protocol/) | ✅ | ~75 min |
| 20 | [OpenTelemetry GenAI](phases/13-tools-and-protocols/20-opentelemetry-genai/) | ✅ | ~75 min |
| 21 | [LLM 路由层](phases/13-tools-and-protocols/21-llm-routing-layer/) | ✅ | ~45 min |
| 22 | [Skill 与 Agent SDK](phases/13-tools-and-protocols/22-skills-and-agent-sdks/) | ✅ | ~45 min |
| 23 | [总结项目 — 工具生态系统](phases/13-tools-and-protocols/23-capstone-tool-ecosystem/) | ✅ | ~120 min |
| 24 | [Skill 发现与渐进式披露](phases/13-tools-and-protocols/24-skill-discovery-and-progressive-disclosure/) | ✅ | ~105 min |
| 25 | [Skill 调用与路由](phases/13-tools-and-protocols/25-skill-invocation-and-routing/) | ✅ | ~105 min |
| 26 | [Skill 权限、沙箱与信任](phases/13-tools-and-protocols/26-skill-permissions-sandboxes-and-trust/) | ✅ | ~120 min |
| 27 | [Skill 评测、打包与可移植性](phases/13-tools-and-protocols/27-skill-evals-packaging-and-portability/) | ✅ | ~150 min |
| 28 | [MCP 工具契约与内容](phases/13-tools-and-protocols/28-mcp-tool-contracts-and-content/) | ✅ | ~120 min |
| 29 | [MCP 可靠性、取消与流量控制](phases/13-tools-and-protocols/29-mcp-reliability-cancellation-and-flow-control/) | ✅ | ~120 min |
| 30 | [MCP Registry 供应链：准入、漂移与回滚](phases/13-tools-and-protocols/30-mcp-registry-supply-chain-and-drift/) | ✅ | ~90 min |
| 31 | [MCP 一致性工程：版本、证据与运维](phases/13-tools-and-protocols/31-mcp-conformance-versioning-and-operations/) | ✅ | ~100 min |

## Phase 14: 智能体工程 — ✅ (~55 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 智能体循环 | ✅ | ~60 min |
| 02 | ReWOO 与规划-执行 | ✅ | ~60 min |
| 03 | Reflexion 与言语强化学习 | ✅ | ~60 min |
| 04 | 思维树与 LATS | ✅ | ~75 min |
| 05 | Self-Refine 与 CRITIC | ✅ | ~60 min |
| 06 | 工具使用与函数调用 | ✅ | ~60 min |
| 07 | Agent 记忆 — 虚拟上下文与记忆分页 | ✅ | ~75 min |
| 08 | 记忆块与休眠时计算 | ✅ | ~75 min |
| 09 | 混合记忆 — 向量 + 图 + KV | ✅ | ~75 min |
| 10 | 技能库与终身学习 (Voyager) | ✅ | ~75 min |
| 11 | 用 HTN 与进化搜索做规划 | ✅ | ~75 min |
| 12 | Anthropic 的工作流模式 | ✅ | ~60 min |
| 13 | 有状态图编排 — 持久执行与检查点 | ✅ | ~75 min |
| 14 | Agent 的 Actor 模型 — 异步消息与类型化运行时 | ✅ | ~75 min |
| 15 | 基于角色的 Agent 团队 — 角色、任务与流程 | ✅ | ~60 min |
| 16 | OpenAI Agents SDK — 移交、护栏、追踪 | ✅ | ~75 min |
| 17 | Harness 即库 — 子 agent 与会话存储 | ✅ | ~75 min |
| 18 | 生产级 Agent 运行时 — 快速实例化与类型化工作流 | ✅ | ~45 min |
| 19 | 基准 — SWE-bench、GAIA、AgentBench | ✅ | ~60 min |
| 20 | 基准 — WebArena 与 OSWorld | ✅ | ~60 min |
| 21 | 计算机操作 — Claude、OpenAI CUA、Gemini | ✅ | ~60 min |
| 22 | 语音智能体 — Pipecat 与 LiveKit | ✅ | ~60 min |
| 23 | OpenTelemetry GenAI 语义约定 | ✅ | ~60 min |
| 24 | 智能体可观测性 — Langfuse、Phoenix、Opik | ✅ | ~45 min |
| 25 | 多智能体辩论与协作 | ✅ | ~60 min |
| 26 | 故障模式 — 智能体为何崩溃 | ✅ | ~60 min |
| 27 | 提示注入与 PVE 防御 | ✅ | ~75 min |
| 28 | 编排模式 — 主管、蜂群、分层 | ✅ | ~60 min |
| 29 | 生产级运行时 — 队列、事件、定时任务 | ✅ | ~60 min |
| 30 | 评估驱动的智能体开发 | ✅ | ~60 min |
| 31 | 智能体工作台：能力强的模型为何仍会失败 | ✅ | ~45 min |
| 32 | 最小智能体工作台 | ✅ | ~45 min |
| 33 | 把智能体指令写成可执行的约束 | ✅ | ~50 min |
| 34 | 仓库记忆与持久化状态 | ✅ | ~60 min |
| 35 | 智能体的初始化脚本 | ✅ | ~45 min |
| 36 | 范围契约与任务边界 | ✅ | ~50 min |
| 37 | 运行时反馈循环 | ✅ | ~50 min |
| 38 | 验证关卡 | ✅ | ~55 min |
| 39 | 审查智能体：把构建者和评分者分开 | ✅ | ~55 min |
| 40 | 多会话交接 | ✅ | ~50 min |
| 41 | 在真实仓库上跑工作台 | ✅ | ~60 min |
| 42 | 总结项目：交付一个可复用的智能体工作台包 | ✅ | ~75 min |
| 43 | 在 agent 写代码前框定任务 | ✅ | ~60 min |
| 44 | 制定由证据支撑的执行计划 | ✅ | ~65 min |
| 45 | 通过隔离与合并契约委派 agent 工作 | ✅ | ~70 min |
| 46 | 把每次 agent 纠正变成系统改进 | ✅ | ~65 min |
| 47 | 在选择产出前定义结果 | ✅ | ~60 min |
| 48 | 发现人们实际执行的工作流 | ✅ | ~70 min |
| 49 | 绘制假设地图，先化解风险最高的一项 | ✅ | ~65 min |
| 50 | 选择能改变决定的最小可测试切片 | ✅ | ~65 min |
| 51 | 编写保留判断空间的规格说明 | ✅ | ~75 min |
| 52 | 在结果出现前设计成功指标 | ✅ | ~70 min |
| 53 | 有意识地选择原型、试点或生产 | ✅ | ~70 min |
| 54 | 构建带所有权与淘汰机制的反馈棘轮 | ✅ | ~75 min |

## Phase 15: 自主系统 — ✅ (~20 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 从聊天机器人到长时程智能体 (METR) | ✅ | ~45 min |
| 02 | STaR、V-STaR、Quiet-STaR — 自学推理 | ✅ | ~60 min |
| 03 | AlphaEvolve — 进化式编码智能体 | ✅ | ~60 min |
| 04 | Darwin Gödel Machine — 自我修改智能体 | ✅ | ~60 min |
| 05 | AI Scientist v2 — 研讨会级研究 | ✅ | ~60 min |
| 06 | 自动化对齐研究 (Anthropic AAR) | ✅ | ~60 min |
| 07 | 递归自我改进 — 能力对齐之争 | ✅ | ~60 min |
| 08 | 有界自我改进设计 | ✅ | ~60 min |
| 09 | 自主编码智能体全景 (SWE-bench、CodeAct) | ✅ | ~45 min |
| 10 | 自主 Agent 的权限模式 | ✅ | ~45 min |
| 11 | 浏览器智能体与间接提示注入 | ✅ | ~45 min |
| 12 | 长时运行智能体的持久化执行 | ✅ | ~60 min |
| 13 | 动作预算、迭代上限、成本调控器 | ✅ | ~60 min |
| 14 | 急停开关、熔断器、金丝雀 token | ✅ | ~60 min |
| 15 | HITL — 先提议后提交 | ✅ | ~60 min |
| 16 | 检查点与回滚 | ✅ | ~60 min |
| 17 | Constitutional AI 与规则覆盖 | ✅ | ~60 min |
| 18 | Llama Guard 与输入/输出分类 | ✅ | ~45 min |
| 19 | Anthropic 负责任扩展政策 v3.0 | ✅ | ~45 min |
| 20 | OpenAI 准备框架与 DeepMind FSF | ✅ | ~45 min |
| 21 | METR 时间跨度与外部评估 | ✅ | ~60 min |
| 22 | CAIS、CAISI 与社会尺度风险 | ✅ | ~45 min |

## Phase 16: 多智能体与蜂群 — ✅ (~28 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [为什么需要多智能体](phases/16-multi-agent-and-swarms/01-why-multi-agent/) | ✅ | ~45 min |
| 02 | [FIPA-ACL 传承与言语行为](phases/16-multi-agent-and-swarms/02-fipa-acl-heritage/) | ✅ | ~60 min |
| 03 | [通信协议](phases/16-multi-agent-and-swarms/03-communication-protocols/) | ✅ | ~45 min |
| 04 | [多智能体原语模型](phases/16-multi-agent-and-swarms/04-primitive-model/) | ✅ | ~60 min |
| 05 | [主管 / 编排器-工作者模式](phases/16-multi-agent-and-swarms/05-supervisor-orchestrator-pattern/) | ✅ | ~75 min |
| 06 | [分层架构与分解漂移](phases/16-multi-agent-and-swarms/06-hierarchical-architecture/) | ✅ | ~60 min |
| 07 | [心智社会与多智能体辩论](phases/16-multi-agent-and-swarms/07-society-of-mind-debate/) | ✅ | ~75 min |
| 08 | [角色专精 — 规划者 / 评论者 / 执行者 / 验证者](phases/16-multi-agent-and-swarms/08-role-specialization/) | ✅ | ~75 min |
| 09 | [并行蜂群与网络化架构](phases/16-multi-agent-and-swarms/09-parallel-swarm-networks/) | ✅ | ~60 min |
| 10 | [群聊与发言人选择](phases/16-multi-agent-and-swarms/10-group-chat-speaker-selection/) | ✅ | ~60 min |
| 11 | [移交与例程 (无状态编排)](phases/16-multi-agent-and-swarms/11-handoffs-and-routines/) | ✅ | ~60 min |
| 12 | [A2A — 智能体到智能体协议](phases/16-multi-agent-and-swarms/12-a2a-protocol/) | ✅ | ~75 min |
| 13 | [共享记忆与黑板模式](phases/16-multi-agent-and-swarms/13-shared-memory-blackboard/) | ✅ | ~75 min |
| 14 | [面向智能体的共识与拜占庭容错](phases/16-multi-agent-and-swarms/14-consensus-and-bft/) | ✅ | ~75 min |
| 15 | [投票、自洽性与辩论拓扑](phases/16-multi-agent-and-swarms/15-voting-debate-topology/) | ✅ | ~75 min |
| 16 | [协商与议价](phases/16-multi-agent-and-swarms/16-negotiation-bargaining/) | ✅ | ~75 min |
| 17 | [生成式智能体与涌现式仿真](phases/16-multi-agent-and-swarms/17-generative-agents-simulation/) | ✅ | ~75 min |
| 18 | [心智理论与涌现式协调](phases/16-multi-agent-and-swarms/18-theory-of-mind-coordination/) | ✅ | ~75 min |
| 19 | [面向 LLM 的群体优化 (PSO、ACO)](phases/16-multi-agent-and-swarms/19-swarm-optimization-pso-aco/) | ✅ | ~75 min |
| 20 | [MARL — MADDPG、QMIX、MAPPO](phases/16-multi-agent-and-swarms/20-marl-maddpg-qmix-mappo/) | ✅ | ~90 min |
| 21 | [智能体经济、token 激励、声誉](phases/16-multi-agent-and-swarms/21-agent-economies/) | ✅ | ~75 min |
| 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/) | ✅ | ~75 min |
| 23 | [故障模式 — MAST、群体思维、单一文化、级联](phases/16-multi-agent-and-swarms/23-failure-modes-mast-groupthink/) | ✅ | ~75 min |
| 24 | [评估与协调基准](phases/16-multi-agent-and-swarms/24-evaluation-coordination-benchmarks/) | ✅ | ~75 min |
| 25 | [案例研究与 2026 年最新进展](phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/) | ✅ | ~90 min |

## Phase 17: 基础设施与生产 — ✅ (~32 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | 托管 LLM 平台 — Bedrock、Azure OpenAI、Vertex AI | ✅ | ~60 min |
| 02 | 推理平台经济学 — Fireworks、Together、Baseten、Modal | ✅ | ~60 min |
| 03 | Kubernetes 上的 GPU 自动扩缩 — Karpenter、KAI Scheduler | ✅ | ~75 min |
| 04 | 推理服务引擎内部机制 — PagedAttention、连续批处理与分块预填充 | ✅ | ~75 min |
| 05 | 生产环境中的 EAGLE-3 推测解码 | ✅ | ~60 min |
| 06 | 前缀缓存推理服务 — RadixAttention 与 KV 复用 | ✅ | ~60 min |
| 07 | 硬件专用推理编译 — Blackwell 上的 FP8 与 NVFP4 | ✅ | ~75 min |
| 08 | 推理指标 — TTFT、TPOT、ITL、Goodput、P99 | ✅ | ~60 min |
| 09 | 生产级量化 — AWQ、GPTQ、GGUF、FP8、NVFP4 | ✅ | ~75 min |
| 10 | 无服务器 LLM 的冷启动缓解 | ✅ | ~60 min |
| 11 | 多区域 LLM 服务与 KV 缓存局部性 | ✅ | ~60 min |
| 12 | 边缘推理 — ANE、Hexagon、WebGPU、Jetson | ✅ | ~60 min |
| 13 | LLM 可观测性技术栈选型 | ✅ | ~60 min |
| 14 | 提示缓存与语义缓存的经济学 | ✅ | ~60 min |
| 15 | 批处理 API — 50% 折扣作为行业标准 | ✅ | ~45 min |
| 16 | 把模型路由作为降本原语 | ✅ | ~60 min |
| 17 | 预填充/解码分离 — NVIDIA Dynamo 与 llm-d | ✅ | ~75 min |
| 18 | 生产级推理服务栈 — KV 卸载与缓存感知路由 | ✅ | ~60 min |
| 19 | AI 网关 — LiteLLM、Portkey、Kong、Bifrost | ✅ | ~60 min |
| 20 | 影子、金丝雀与渐进式部署 | ✅ | ~60 min |
| 21 | LLM 功能的 A/B 测试 — GrowthBook 与 Statsig | ✅ | ~60 min |
| 22 | LLM API 的负载测试 — k6、LLMPerf、GenAI-Perf | ✅ | ~75 min |
| 23 | 面向 AI 的 SRE — 多智能体事件响应 | ✅ | ~60 min |
| 24 | 面向 LLM 生产的混沌工程 | ✅ | ~60 min |
| 25 | 安全 — 密钥、PII 脱敏、审计日志 | ✅ | ~60 min |
| 26 | 合规 — SOC 2、HIPAA、GDPR、EU AI Act、ISO 42001 | ✅ | ~60 min |
| 27 | 面向 LLM 的 FinOps — 单位经济与多租户归因 | ✅ | ~60 min |
| 28 | 自托管推理服务选型 — 让引擎匹配硬件与规模 | ✅ | ~45 min |

## Phase 18: 伦理、安全与对齐 — ✅ (~31 hours)

| # | Lesson | Status | Est. |
|---|--------|--------|------|
| 01 | [指令遵循作为对齐信号](phases/18-ethics-safety-alignment/01-instruction-following-alignment-signal) | ✅ | ~45 min |
| 02 | [奖励黑客与古德哈特定律](phases/18-ethics-safety-alignment/02-reward-hacking-goodhart) | ✅ | ~60 min |
| 03 | [直接偏好优化家族](phases/18-ethics-safety-alignment/03-direct-preference-optimization-family) | ✅ | ~60 min |
| 04 | [谄媚作为 RLHF 的放大效应](phases/18-ethics-safety-alignment/04-sycophancy-rlhf-amplification) | ✅ | ~45 min |
| 05 | [Constitutional AI 与 RLAIF](phases/18-ethics-safety-alignment/05-constitutional-ai-rlaif) | ✅ | ~60 min |
| 06 | [Mesa 优化与欺骗性对齐](phases/18-ethics-safety-alignment/06-mesa-optimization-deceptive-alignment) | ✅ | ~75 min |
| 07 | [潜伏智能体 — 持久性欺骗](phases/18-ethics-safety-alignment/07-sleeper-agents-persistent-deception) | ✅ | ~60 min |
| 08 | [前沿模型中的上下文阴谋](phases/18-ethics-safety-alignment/08-in-context-scheming-frontier-models) | ✅ | ~60 min |
| 09 | [对齐伪装](phases/18-ethics-safety-alignment/09-alignment-faking) | ✅ | ~60 min |
| 10 | [AI 控制 — 颠覆之下仍保安全](phases/18-ethics-safety-alignment/10-ai-control-subversion) | ✅ | ~75 min |
| 11 | [可扩展监督与弱到强泛化](phases/18-ethics-safety-alignment/11-scalable-oversight-weak-to-strong) | ✅ | ~60 min |
| 12 | [红队 — PAIR 与自动化攻击](phases/18-ethics-safety-alignment/12-red-teaming-pair-automated-attacks) | ✅ | ~75 min |
| 13 | [多样本越狱](phases/18-ethics-safety-alignment/13-many-shot-jailbreaking) | ✅ | ~45 min |
| 14 | [ASCII 艺术与视觉越狱](phases/18-ethics-safety-alignment/14-ascii-art-visual-jailbreaks) | ✅ | ~60 min |
| 15 | [间接提示注入](phases/18-ethics-safety-alignment/15-indirect-prompt-injection) | ✅ | ~75 min |
| 16 | [红队工具 — Garak、Llama Guard、PyRIT](phases/18-ethics-safety-alignment/16-red-team-tooling-garak-llamaguard-pyrit) | ✅ | ~75 min |
| 17 | [WMDP 与两用能力评估](phases/18-ethics-safety-alignment/17-wmdp-dual-use-evaluation) | ✅ | ~60 min |
| 18 | [前沿安全框架 — RSP、PF、FSF](phases/18-ethics-safety-alignment/18-frontier-safety-frameworks-rsp-pf-fsf) | ✅ | ~75 min |
| 19 | [模型福祉研究](phases/18-ethics-safety-alignment/19-model-welfare-research) | ✅ | ~45 min |
| 20 | [偏见与表征伤害](phases/18-ethics-safety-alignment/20-bias-representational-harm) | ✅ | ~60 min |
| 21 | [公平性准则 — 群体、个体、反事实](phases/18-ethics-safety-alignment/21-fairness-criteria-group-individual-counterfactual) | ✅ | ~60 min |
| 22 | [面向 LLM 的差分隐私](phases/18-ethics-safety-alignment/22-differential-privacy-for-llms) | ✅ | ~60 min |
| 23 | [水印 — SynthID、Stable Signature、C2PA](phases/18-ethics-safety-alignment/23-watermarking-synthid-stable-signature-c2pa) | ✅ | ~75 min |
| 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) | ✅ | ~75 min |
| 25 | [EchoLeak 与面向 AI 的 CVE](phases/18-ethics-safety-alignment/25-echoleak-cves-for-ai) | ✅ | ~45 min |
| 26 | [模型卡、系统卡与数据集卡](phases/18-ethics-safety-alignment/26-model-system-dataset-cards) | ✅ | ~60 min |
| 27 | [数据溯源与训练数据治理](phases/18-ethics-safety-alignment/27-data-provenance-training-governance) | ✅ | ~60 min |
| 28 | [对齐研究生态 — MATS、Redwood、Apollo、METR](phases/18-ethics-safety-alignment/28-alignment-research-ecosystem) | ✅ | ~45 min |
| 29 | [审核系统 — OpenAI、Perspective、Llama Guard](phases/18-ethics-safety-alignment/29-moderation-systems-openai-perspective-llamaguard) | ✅ | ~60 min |
| 30 | [两用风险 — 网络、生物、化学、核](phases/18-ethics-safety-alignment/30-dual-use-risk-cyber-bio-chem-nuclear) | ✅ | ~75 min |

## Phase 19: 总结项目 — ✅ (~800 hours)

| # | Project | Status | Est. |
|---|---------|--------|------|
| 01 | [终端原生编码智能体](phases/19-capstone-projects/01-terminal-native-coding-agent) | ✅ | ~35 hr |
| 02 | [面向代码库的 RAG (跨仓库语义搜索)](phases/19-capstone-projects/02-rag-over-codebase) | ✅ | ~30 hr |
| 03 | [实时语音助手 (ASR 到 LLM 到 TTS)](phases/19-capstone-projects/03-realtime-voice-assistant) | ✅ | ~30 hr |
| 04 | [多模态文档问答 (视觉优先)](phases/19-capstone-projects/04-multimodal-document-qa) | ✅ | ~30 hr |
| 05 | [自主研究智能体 (AI-Scientist 级)](phases/19-capstone-projects/05-autonomous-research-agent) | ✅ | ~40 hr |
| 06 | [面向 Kubernetes 的 DevOps 排障智能体](phases/19-capstone-projects/06-devops-troubleshooting-agent) | ✅ | ~30 hr |
| 07 | [端到端微调流水线](phases/19-capstone-projects/07-end-to-end-fine-tuning-pipeline) | ✅ | ~35 hr |
| 08 | [生产级 RAG 聊天机器人 (受监管垂直领域)](phases/19-capstone-projects/08-production-rag-chatbot) | ✅ | ~30 hr |
| 09 | [代码迁移智能体 (仓库级升级)](phases/19-capstone-projects/09-code-migration-agent) | ✅ | ~30 hr |
| 10 | [多智能体软件工程团队](phases/19-capstone-projects/10-multi-agent-software-team) | ✅ | ~40 hr |
| 11 | [LLM 可观测性与评估仪表盘](phases/19-capstone-projects/11-llm-observability-dashboard) | ✅ | ~25 hr |
| 12 | [视频理解流水线 (场景到问答)](phases/19-capstone-projects/12-video-understanding-pipeline) | ✅ | ~30 hr |
| 13 | [带注册中心与治理的 MCP Server](phases/19-capstone-projects/13-mcp-server-with-registry) | ✅ | ~25 hr |
| 14 | [推测解码推理服务器](phases/19-capstone-projects/14-speculative-decoding-server) | ✅ | ~30 hr |
| 15 | [Constitutional 安全护栏 + 红队靶场](phases/19-capstone-projects/15-constitutional-safety-harness) | ✅ | ~25 hr |
| 16 | [GitHub Issue 到 PR 的自主智能体](phases/19-capstone-projects/16-github-issue-to-pr-agent) | ✅ | ~30 hr |
| 17 | [个人 AI 导师 (自适应、多模态)](phases/19-capstone-projects/17-personal-ai-tutor) | ✅ | ~30 hr |
| 20 | [Agent Harness Loop 契约](phases/19-capstone-projects/20-agent-harness-loop-contract) | ✅ | ~90 min |
| 21 | [带 Schema 校验的 Tool Registry](phases/19-capstone-projects/21-tool-registry-schema-validation) | ✅ | ~90 min |
| 22 | [基于换行分隔 stdio 的 JSON-RPC 2.0](phases/19-capstone-projects/22-jsonrpc-stdio-transport) | ✅ | ~90 min |
| 23 | [Function Call Dispatcher](phases/19-capstone-projects/23-function-call-dispatcher) | ✅ | ~90 min |
| 24 | [Plan-Execute 控制流](phases/19-capstone-projects/24-plan-execute-control-flow) | ✅ | ~90 min |
| 25 | [Verification Gate 与 Observation Budget](phases/19-capstone-projects/25-verification-gates-observation-budget) | ✅ | ~90 min |
| 26 | [带 Denylist 与 Path Jail 的 Sandbox Runner](phases/19-capstone-projects/26-sandbox-runner-denylist) | ✅ | ~90 min |
| 27 | [带 Fixture Tasks 的 Eval Harness](phases/19-capstone-projects/27-eval-harness-fixture-tasks) | ✅ | ~90 min |
| 28 | [用 OTel GenAI Span 与 Prometheus 做 Observability](phases/19-capstone-projects/28-observability-otel-traces) | ✅ | ~90 min |
| 29 | [端到端 Coding Agent Demo](phases/19-capstone-projects/29-end-to-end-coding-task-demo) | ✅ | ~90 min |
| 30 | [从零实现 BPE Tokenizer](phases/19-capstone-projects/30-bpe-tokenizer-from-scratch) | ✅ | ~90 min |
| 31 | [带 Sliding Window 的 Tokenized Dataset](phases/19-capstone-projects/31-tokenized-dataset-sliding-window) | ✅ | ~90 min |
| 32 | [Token Embedding 与 Positional Embedding](phases/19-capstone-projects/32-token-positional-embeddings) | ✅ | ~90 min |
| 33 | [Multi-Head Self-Attention](phases/19-capstone-projects/33-multihead-self-attention) | ✅ | ~90 min |
| 34 | [从零实现 Transformer Block](phases/19-capstone-projects/34-transformer-block) | ✅ | ~90 min |
| 35 | [GPT 模型组装](phases/19-capstone-projects/35-gpt-model-assembly) | ✅ | ~90 min |
| 36 | [训练循环与评估](phases/19-capstone-projects/36-training-loop-eval) | ✅ | ~90 min |
| 37 | [加载预训练权重](phases/19-capstone-projects/37-loading-pretrained-weights) | ✅ | ~90 min |
| 38 | [通过换 Head 做分类微调](phases/19-capstone-projects/38-classifier-finetuning) | ✅ | ~90 min |
| 39 | [通过 SFT 做 Instruction Tuning](phases/19-capstone-projects/39-instruction-tuning-sft) | ✅ | ~90 min |
| 40 | [从零实现 DPO](phases/19-capstone-projects/40-dpo-from-scratch) | ✅ | ~90 min |
| 41 | [完整评估流水线](phases/19-capstone-projects/41-eval-pipeline) | ✅ | ~90 min |
| 42 | [大规模语料下载器](phases/19-capstone-projects/42-large-corpus-downloader) | ✅ | ~90 min |
| 43 | [HDF5 Tokenized Corpus](phases/19-capstone-projects/43-hdf5-tokenized-corpus) | ✅ | ~90 min |
| 44 | [Cosine 学习率 + 线性 Warmup](phases/19-capstone-projects/44-cosine-lr-warmup) | ✅ | ~90 min |
| 45 | [Gradient Clipping 与混合精度训练](phases/19-capstone-projects/45-gradient-clipping-amp) | ✅ | ~90 min |
| 46 | [梯度累积](phases/19-capstone-projects/46-gradient-accumulation) | ✅ | ~90 min |
| 47 | [Checkpoint 保存与恢复](phases/19-capstone-projects/47-checkpoint-save-resume) | ✅ | ~90 min |
| 48 | [从零实现分布式数据并行与 FSDP](phases/19-capstone-projects/48-distributed-fsdp-ddp) | ✅ | ~90 min |
| 49 | [语言模型评测框架](phases/19-capstone-projects/49-lm-eval-harness) | ✅ | ~90 min |
| 50 | [假设生成器](phases/19-capstone-projects/50-hypothesis-generator) | ✅ | ~90 min |
| 51 | [文献检索](phases/19-capstone-projects/51-literature-retrieval) | ✅ | ~90 min |
| 52 | [实验执行器](phases/19-capstone-projects/52-experiment-runner) | ✅ | ~90 min |
| 53 | [结果评估器](phases/19-capstone-projects/53-result-evaluator) | ✅ | ~90 min |
| 54 | [论文生成器](phases/19-capstone-projects/54-paper-writer) | ✅ | ~90 min |
| 55 | [评审循环](phases/19-capstone-projects/55-critic-loop) | ✅ | ~90 min |
| 56 | [迭代调度器](phases/19-capstone-projects/56-iteration-scheduler) | ✅ | ~90 min |
| 57 | [端到端研究 Demo](phases/19-capstone-projects/57-end-to-end-research-demo) | ✅ | ~90 min |
| 58 | [Vision Encoder 的 Patch 切分](phases/19-capstone-projects/58-vision-encoder-patches) | ✅ | ~90 min |
| 59 | [Vision Transformer Encoder（ViT）](phases/19-capstone-projects/59-vit-transformer) | ✅ | ~90 min |
| 60 | [用 Projection Layer 做模态对齐](phases/19-capstone-projects/60-projection-layer-modality-align) | ✅ | ~90 min |
| 61 | [Cross-Attention 融合](phases/19-capstone-projects/61-cross-attention-fusion) | ✅ | ~90 min |
| 62 | [Vision-Language 预训练](phases/19-capstone-projects/62-vision-language-pretraining) | ✅ | ~90 min |
| 63 | [多模态评测](phases/19-capstone-projects/63-multimodal-eval) | ✅ | ~90 min |
| 64 | [Chunking 策略横向对比](phases/19-capstone-projects/64-chunking-strategies-advanced) | ✅ | ~90 min |
| 65 | [用 BM25 与 Dense Embedding 做 Hybrid Retrieval](phases/19-capstone-projects/65-hybrid-retrieval-bm25-dense) | ✅ | ~90 min |
| 66 | [Cross-Encoder Reranker](phases/19-capstone-projects/66-reranker-cross-encoder) | ✅ | ~90 min |
| 67 | [Query 改写：HyDE、Multi-Query 与 Decomposition](phases/19-capstone-projects/67-query-rewriting-hyde) | ✅ | ~90 min |
| 68 | [RAG 评测：Precision、Recall、MRR、nDCG 等](phases/19-capstone-projects/68-rag-eval-precision-recall) | ✅ | ~90 min |
| 69 | [端到端 RAG 系统](phases/19-capstone-projects/69-end-to-end-rag-system) | ✅ | ~90 min |
| 70 | [任务规格格式](phases/19-capstone-projects/70-task-spec-format) | ✅ | ~90 min |
| 71 | [经典评测指标](phases/19-capstone-projects/71-classical-metrics) | ✅ | ~90 min |
| 72 | [代码执行评测指标](phases/19-capstone-projects/72-code-exec-metric) | ✅ | ~90 min |
| 73 | [perplexity 与 calibration](phases/19-capstone-projects/73-perplexity-calibration) | ✅ | ~90 min |
| 74 | [leaderboard 聚合](phases/19-capstone-projects/74-leaderboard-aggregation) | ✅ | ~90 min |
| 75 | [端到端 eval runner](phases/19-capstone-projects/75-end-to-end-eval-runner) | ✅ | ~90 min |
| 76 | [从零实现集合通信](phases/19-capstone-projects/76-collective-ops-from-scratch) | ✅ | ~90 min |
| 77 | [数据并行 DDP](phases/19-capstone-projects/77-data-parallel-ddp) | ✅ | ~90 min |
| 78 | [ZeRO Optimizer State 分片](phases/19-capstone-projects/78-zero-parameter-sharding) | ✅ | ~90 min |
| 79 | [Pipeline Parallel 与 Bubble 分析](phases/19-capstone-projects/79-pipeline-parallel) | ✅ | ~90 min |
| 80 | [分片 Checkpoint 与原子化恢复](phases/19-capstone-projects/80-checkpoint-sharded-resume) | ✅ | ~90 min |
| 81 | [端到端分布式训练](phases/19-capstone-projects/81-end-to-end-distributed-train) | ✅ | ~90 min |
| 82 | [越狱分类法](phases/19-capstone-projects/82-jailbreak-taxonomy) | ✅ | ~90 min |
| 83 | [Prompt 注入检测器](phases/19-capstone-projects/83-prompt-injection-detector) | ✅ | ~90 min |
| 84 | [拒答评估](phases/19-capstone-projects/84-refusal-evaluation) | ✅ | ~90 min |
| 85 | [内容分类器集成](phases/19-capstone-projects/85-content-classifier-integration) | ✅ | ~90 min |
| 86 | [Constitutional 规则引擎](phases/19-capstone-projects/86-constitutional-rules-engine) | ✅ | ~90 min |
| 87 | [端到端 safety gate](phases/19-capstone-projects/87-end-to-end-safety-gate) | ✅ | ~90 min |

---

**总计：20 个阶段，523 节课 | 523 已完成 | 预计约 1128 小时**

想出一份力？挑任意一节 ⬚ 课提交 PR。详见 [CONTRIBUTING.md](CONTRIBUTING.md)。