Concept Index

Each concept has one introductory chapter; later chapters build on it. The application route previews model concepts in the Model Crash Course, so you can skip Part I. The table marks these previews. Chapter references use “(← S2 Transformers and Decoding)” for prior explanations and “(→ S16 Security and Governance)” for later detail.

Use this index to find a concept’s chapter. The chain rule, softmax, log plots, standard error, bootstrap and cosine similarity are collected separately in the Math Refresher, with section links in the relevant rows. See How to read the codes for the labels before chapter titles.

Models

Concept	First taught	Deepened in
Backprop, optimizers, initialization and regularization (Math Refresher: chain rule)	S1 Neural Network Basics	S3 Pretraining and Scaling
Experiment tracking (W&B)	S1 Neural Network Basics	—
Perplexity (Math Refresher: log scales)	S1 Neural Network Basics	S3 Pretraining and Scaling, S5 Inference and Serving
Word embedding (word2vec)	S1 Neural Network Basics (○)	—
Tokenization / BPE	S2 Transformers and Decoding (application route: previewed in the Model Crash Course)	S6 LLM APIs and Application Basics (token pricing)
Attention, the Transformer block	S2 Transformers and Decoding	S5 Inference and Serving
Causal mask (what happens to training and generation without it)	S2 Transformers and Decoding	—
Decoding / sampling (temperature, top-p; Math Refresher: softmax)	S2 Transformers and Decoding	S4 Post-training and Reasoning, S5 Inference and Serving, S6 LLM APIs and Application Basics, S7 Prompts, Structured Output and Workflows
Limits on sampling parameters (some APIs reject temperature, top_p, top_k)	S6 LLM APIs and Application Basics (S2 Transformers and Decoding, S4 Post-training and Reasoning and the Model Crash Course mention it first and link here)	S7 Prompts, Structured Output and Workflows, S11 Context Engineering
RoPE, GQA / MLA, MoE	S2 Transformers and Decoding	S5 Inference and Serving
MoE compute and memory (compute follows active parameters, memory follows total parameters)	S2 Transformers and Decoding	—
Resource accounting (parameters, FLOPs, memory)	S2 Transformers and Decoding	S3 Pretraining and Scaling, S5 Inference and Serving
Quantization concepts	S2 Transformers and Decoding	S5 Inference and Serving (implementation), S6 LLM APIs and Application Basics (GGUF / local models), S18 Deployment and Architecture
Multimodal architectures	S2 Transformers and Decoding (○)	S9 Retrieval and RAG (document parsing), S10 Tools and MCP (computer use)
Data engineering (filtering, deduplication, PII)	S3 Pretraining and Scaling	S4 Post-training and Reasoning (synthetic data)
Scaling laws, emergence (Math Refresher: log–log plots)	S3 Pretraining and Scaling	—
Parallel training	S3 Pretraining and Scaling	—
SFT, LoRA	S4 Post-training and Reasoning	S9 Retrieval and RAG (embedding fine-tuning)
Chat template	S4 Post-training and Reasoning (application route: previewed in the Model Crash Course)	S7 Prompts, Structured Output and Workflows (what tool calls look like underneath)
RL basics, RLHF, DPO (including DPO’s limits: offline learning, no exploration)	S4 Post-training and Reasoning	S14 Agent Evals, Data and Optimization (agentic RL)
GRPO / RLVR, reasoning models	S4 Post-training and Reasoning	S6 LLM APIs and Application Basics (using them in applications)
Test-time compute, verifiers	S4 Post-training and Reasoning	S14 Agent Evals, Data and Optimization
Reward hacking, sycophancy	S4 Post-training and Reasoning	S13 Memory and Personalization (personalization that panders), S14 Agent Evals, Data and Optimization (optimization that games the metric)
Synthetic data and distillation	S4 Post-training and Reasoning	S8 Evals Methodology, S14 Agent Evals, Data and Optimization (router distillation)
A mental model of the GPU	S5 Inference and Serving	S18 Deployment and Architecture
Prefill / decode, KV cache	S5 Inference and Serving (S0 AI Systems Overview and the Model Crash Course preview it)	S6 LLM APIs and Application Basics, S11 Context Engineering, S18 Deployment and Architecture
Continuous batching, PagedAttention	S5 Inference and Serving	S18 Deployment and Architecture
TTFT / TPOT / goodput	S5 Inference and Serving	S6 LLM APIs and Application Basics (measuring TTFT), S18 Deployment and Architecture
Prompt caching (KV prefix reuse)	S5 Inference and Serving (application route: previewed in the Model Crash Course, taught mainly in S11 Context Engineering)	S18 Deployment and Architecture
How constrained decoding works	S5 Inference and Serving	S7 Prompts, Structured Output and Workflows
A map of speedups (FlashAttention, speculative decoding)	S5 Inference and Serving	—

Applications

Concept	First taught	Deepened in
Messages / roles, token pricing	S6 LLM APIs and Application Basics	S8 Evals Methodology (cost accounting)
Streaming (SSE, partial JSON)	S6 LLM APIs and Application Basics	S17 Product and AI UX
Retries, rate limits, idempotency	S6 LLM APIs and Application Basics	S12 The Agent Loop and Harness, S17 Product and AI UX
Cancellation and interruption	S6 LLM APIs and Application Basics	S17 Product and AI UX
Using reasoning models in applications (thinking budget, interleaved thinking)	S6 LLM APIs and Application Basics	S7 Prompts, Structured Output and Workflows, S12 The Agent Loop and Harness
Collecting feedback (minimal version: thumbs up / down)	S6 LLM APIs and Application Basics	S14 Agent Evals, Data and Optimization, S17 Product and AI UX
Local and small models	S6 LLM APIs and Application Basics	S18 Deployment and Architecture
Ollama’s context length (the default may be only 4K; beyond it, earlier content is dropped)	S6 LLM APIs and Application Basics (S5 Inference and Serving mentions it first and links here)	Later chapters link here whenever they give tips on local models
Tracing (OpenTelemetry GenAI)	S6 LLM APIs and Application Basics	S8 Evals Methodology, S12 The Agent Loop and Harness, S14 Agent Evals, Data and Optimization
Testing nondeterministic code	S6 LLM APIs and Application Basics	S8 Evals Methodology
Prompt structure, few-shot, CoT	S7 Prompts, Structured Output and Workflows	—
Prompt sensitivity	S7 Prompts, Structured Output and Workflows	—
Structured output, function calling	S7 Prompts, Structured Output and Workflows	S10 Tools and MCP, S13 Memory and Personalization (memory extraction)
Prompts as code, migrating across models	S7 Prompts, Structured Output and Workflows	S18 Deployment and Architecture
The five workflow patterns	S7 Prompts, Structured Output and Workflows	S12 The Agent Loop and Harness, S15 Multi-agent Systems (orchestrator-worker), S19 Real Agent Systems and the Capstone
Multimodal input (feeding images and PDFs to the model directly)	S7 Prompts, Structured Output and Workflows (○)	S9 Retrieval and RAG
Error analysis (open / axial coding)	S8 Evals Methodology	S14 Agent Evals, Data and Optimization, S15 Multi-agent Systems (the MAST failure taxonomy)
LLM-as-judge and calibration	S8 Evals Methodology	S13 Memory and Personalization (the pairwise judge in step 0 of A13 Layered Memory, for evaluating personalization), S14 Agent Evals, Data and Optimization (agent-as-judge)
Capability vs regression evals, pass@k / pass^k	S8 Evals Methodology	S10 Tools and MCP, S12 The Agent Loop and Harness, S14 Agent Evals, Data and Optimization, S18 Deployment and Architecture
Benchmark contamination and construct validity	S8 Evals Methodology	S14 Agent Evals, Data and Optimization
Error bars, bootstrap confidence intervals (Math Refresher: standard error, bootstrap; used once earlier in S4 Post-training and Reasoning)	S8 Evals Methodology	S9 Retrieval and RAG, S12 The Agent Loop and Harness, S14 Agent Evals, Data and Optimization, S18 Deployment and Architecture
Sample size planning (statistical power)	S8 Evals Methodology (○)	—
Multiple comparisons (Bonferroni)	S8 Evals Methodology (○)	—
Inter-annotator agreement (Cohen’s κ)	S8 Evals Methodology (○)	—
Cost accounting (cost per task, budgets)	S8 Evals Methodology	S13 Memory and Personalization, S15 Multi-agent Systems
Uncertainty (logprobs, calibration)	S8 Evals Methodology	S17 Product and AI UX (per-item uncertainty signals and how to show them, step 0 of A17 Complete Frontend), S18 Deployment and Architecture (cascades)
Calibration metrics (ECE, Brier score)	S8 Evals Methodology (○)	—
Document parsing, OCR	S9 Retrieval and RAG	—
Embeddings, vector search, hybrid search, reranking (Math Refresher: cosine similarity)	S9 Retrieval and RAG	S12 The Agent Loop and Harness (compared with agentic search)
Retrieval eval vs generation eval	S9 Retrieval and RAG	—
Long context vs RAG	S9 Retrieval and RAG	S11 Context Engineering, S13 Memory and Personalization (the memory baseline)
Choosing between prompting, RAG and fine-tuning	S9 Retrieval and RAG (the Model Crash Course first previews what fine-tuning can do)	S11 Context Engineering
Text2SQL	S9 Retrieval and RAG (○)	—
Indirect prompt injection 🔒	S9 Retrieval and RAG	S13 Memory and Personalization, S16 Security and Governance
Tool design and risk tiers	S10 Tools and MCP	S12 The Agent Loop and Harness, S16 Security and Governance, S17 Product and AI UX (confirmation UI)
MCP (transports, auth)	S10 Tools and MCP	S16 Security and Governance (supply chain), S17 Product and AI UX (MCP Apps)
stdio MCP server (core; the HTTP version with OAuth is in the advanced exercise)	S10 Tools and MCP	—
Tool search / deferred loading	S10 Tools and MCP	S11 Context Engineering
Code execution as a tool interface	S10 Tools and MCP	—
Agent-computer interface (ACI)	S10 Tools and MCP	S19 Real Agent Systems and the Capstone
Sandbox, egress 🔒	S10 Tools and MCP	S12 The Agent Loop and Harness, S16 Security and Governance
Context budgets and failure modes	S11 Context Engineering	S12 The Agent Loop and Harness, S15 Multi-agent Systems
Progressive disclosure	S11 Context Engineering	S12 The Agent Loop and Harness (skills)

Agent systems

Concept	First taught	Deepened in
Compound AI, the augmented LLM, workflow vs agent	S0 AI Systems Overview	S7 Prompts, Structured Output and Workflows, S12 The Agent Loop and Harness, S19 Real Agent Systems and the Capstone
Agent levels 0–4	S0 AI Systems Overview	—
The agent loop, ReAct (S0 AI Systems Overview previews it with mini-SWE-agent)	S12 The Agent Loop and Harness	S19 Real Agent Systems and the Capstone
Harness components	S12 The Agent Loop and Harness	S14 Agent Evals, Data and Optimization (ablation), S19 Real Agent Systems and the Capstone
Permission modes, hooks 🔒	S12 The Agent Loop and Harness	S16 Security and Governance
Stop conditions, escalating to a human, HITL	S12 The Agent Loop and Harness	S19 Real Agent Systems and the Capstone
Checkpoint / resume	S12 The Agent Loop and Harness	S18 Deployment and Architecture (durable execution), S19 Real Agent Systems and the Capstone
Compaction, context reset, sub-agent isolation	S12 The Agent Loop and Harness	S13 Memory and Personalization, S15 Multi-agent Systems (sub-agent isolation), S19 Real Agent Systems and the Capstone
Skills	S12 The Agent Loop and Harness	S13 Memory and Personalization (procedural memory), S16 Security and Governance (supply chain)
Agentic search	S12 The Agent Loop and Harness	S15 Multi-agent Systems
Long-running patterns (initializer / worker, generator–evaluator)	S12 The Agent Loop and Harness	—
Memory types and hierarchy	S13 Memory and Personalization	—
Personalization	S13 Memory and Personalization	S17 Product and AI UX (interface)
Memory poisoning 🔒	S13 Memory and Personalization	S16 Security and Governance
Trajectory evals, simulated users	S14 Agent Evals, Data and Optimization	S15 Multi-agent Systems, S19 Real Agent Systems and the Capstone
Infra noise, benchmark hygiene	S14 Agent Evals, Data and Optimization	S19 Real Agent Systems and the Capstone
Prompt optimization (DSPy, GEPA)	S14 Agent Evals, Data and Optimization	—
Logging routing decisions and router distillation (← the routing workflow in S7 Prompts, Structured Output and Workflows)	S14 Agent Evals, Data and Optimization	S18 Deployment and Architecture (model routing)
Data flywheel	S14 Agent Evals, Data and Optimization	S17 Product and AI UX (feedback UI)
Multi-agent architectures and token economics	S15 Multi-agent Systems	S19 Real Agent Systems and the Capstone
Deep research architecture	S15 Multi-agent Systems	S19 Real Agent Systems and the Capstone
Workspace / worktree isolation	S15 Multi-agent Systems	—

Trust, product, shipping

Concept	First taught	Deepened in
Lethal trifecta, threat modeling	S16 Security and Governance	S19 Real Agent Systems and the Capstone (audit)
Structural defense design patterns	S16 Security and Governance	—
Guardrails, approval fatigue	S16 Security and Governance	S17 Product and AI UX
Data governance and regulation	S16 Security and Governance	—
Agentic UX, generative UI	S17 Product and AI UX	—
Feedback design	S17 Product and AI UX	—
Proactive agents	S17 Product and AI UX	S19 Real Agent Systems and the Capstone
Model routing and cascades (← the routing workflow in S7 Prompts, Structured Output and Workflows; router distillation in S14 Agent Evals, Data and Optimization)	S18 Deployment and Architecture	—
Eval-gated CI/CD, canaries	S18 Deployment and Architecture	—
Multi-tenant isolation	S18 Deployment and Architecture	—
Gateways, caches (exact / semantic; prompt caching ← S5 Inference and Serving)	S18 Deployment and Architecture	—
Durable execution	S18 Deployment and Architecture	S19 Real Agent Systems and the Capstone