The Map
The series as a dependency graph The whole premise of the series is that the LLM stack is a dependency graph, and that nothing should get used before the intuition for it has been earned. This page is that claim, drawn. Every edge comes from the posts themselves — a line means one post actually links back to another as prior material. Find a concept you need, then walk its chain backwards until you hit ground you already know. That's your reading list.
ARC 01 Math & Computation ARC 02 Before Transformers ARC 03 Tokenization ARC 04 Transformers ARC 05 Decoding & Inference ARC 06 Engines & Serving ARC 07 Training ARC 08 Evaluation ARC 09 Retrieval & Context ARC 10 Tools & Agents ARC 11 Provider Ecosystems ARC 12 Governance & Trajectory 1.1 Vectors, Matrices, and … 1.2 Norms, Dot Products, an… 1.3 Distributions, Softmax,… 1.4 Cross-Entropy, KL Diver… 1.5 Gradients and How Machi… 1.6 Optimizers: Momentum, A… 1.7 GPUs, Floating Point, a… 1.8 A Short Prehistory of S… 2.1 Language Modeling as Ne… 2.2 N-gram Models and the C… 2.3 Word2vec and the Embedd… 2.4 GloVe, FastText, and th… 2.5 Recurrent Neural Networ… 2.6 LSTMs, GRUs, and Gated … 2.7 Seq2seq, Bahdanau Atten… 3.1 Unicode, Bytes, and Wha… 3.2 Byte-Pair Encoding from… 3.3 WordPiece, Unigram, and… 3.4 Vocabulary Size, Merge … 3.5 The Embedding Table and… 3.6 Special Tokens, Chat Te… 3.7 Packing, Masking, and T… 4.1 Self-Attention: Q, K, V… 4.2 Multi-Head Attention an… 4.3 Causal Masking and the … 4.4 Positional Encodings: S… 4.5 The Feed-Forward Block … 4.6 Layer Normalization: Pr… 4.7 Decoder-Only vs. Encode… 4.8 A Close Reading of 'Att… 4.9 Mixture-of-Experts Laye… 5.1 Training View vs. Infer… 5.2 Prefill vs. Decode: The… 5.3 Why One New Token Means… 5.4 The KV Cache from First… 5.5 Sampling Strategies: Te… 5.6 Speculative Decoding 5.7 Continuous Batching 5.8 Prefix Caching and Prom… 5.9 Structured and Constrai… 6.1 What an Inference Engin… 6.2 Kernel Fusion and GPU O… 6.3 PagedAttention and the … 6.4 Memory Management for L… 6.5 Quantization: INT8, INT… 6.6 Tensor, Pipeline, and E… 6.7 Throughput vs. Latency:… 6.8 Comparing Engines: vLLM… 6.9 Benchmarking Your Own I… 7.1 Pretraining Data: Mixtu… 7.2 Distributed Training: F… 7.3 Supervised Fine-Tuning 7.4 The RLHF Pipeline: Rewa… 7.5 DPO and Its Variants: S… 7.6 Constitutional AI and R… 7.7 Tool-Use Fine-Tuning 7.8 Long-Context Training T… 7.9 Synthetic Data and Dist… 7.10 LoRA and Parameter-Effi… 8.1 Loss vs. Benchmarks: Tw… 8.2 Eval Harnesses: lm-eval… 8.3 Contamination and Leaka… 8.4 Evaluating Reasoning 8.5 Tool and Agent Evaluati… 8.6 Human Preference Evalua… 8.7 Calibration and Abstent… 9.1 Embeddings from Scratch… 9.2 Vector Search and Appro… 9.3 Dense, Sparse, and Hybr… 9.4 Chunking Strategies 9.5 Rerankers and Cross-Enc… 9.6 RAG Architectures End t… 9.7 GraphRAG and Knowledge … 9.8 Context Windows: Advert… 9.9 Context Engineering: Co… 10.1 A Short History of Agen… 10.2 Function Calling as Str… 10.3 The Agent Loop: Model, … 10.4 Transcript Formats: How… 10.5 MCP: A Cross-System Sta… 10.6 Remote MCP, OAuth, and … 10.7 Planning and Reasoning … 10.8 Computer Use and Browse… 10.9 Multi-Agent Systems and… 11.1 OpenAI's API: From Chat… 11.2 Anthropic's API: Messag… 11.3 Google's API: Gemini, G… 11.4 Open-Weight Models: Lla… 11.5 The Hugging Face Ecosys… 11.6 Bedrock, Azure AI, and … 11.7 Lock-In, Portability, a… 12.1 Scaling Laws: What They… 12.2 Inference Economics: Co… 12.3 Open vs. Closed: Ecosys… 12.4 Safety and Alignment: T… 12.5 The EU AI Act and Globa… 12.6 Standards Wars: MCP, A2… 12.7 A Short History of the … 12.8 What Comes Next: Honest… Tap or hover any post to see what it builds on (indigo) and what builds on it (amber). A click pins it here; walk the chips backwards until you hit ground you already know. Dashed rings mark the featured posts.
Prefer a list? The series index has all twelve arcs with curated starting paths.