Appendices›Appendix C

Reading Pathways

Eight curated routes through the book. Each one tells you what to read in what order, roughly how long it takes, and what you can do after.

A friendly robot hiker standing at a forest trailhead with eight wooden signpost
Figure C.0.1: Eight pathways branch from one trailhead. Pick the destination first, then the trail. s pointing in different directions, each decorated with a small symbol for a different reading pathway. The trails branch out across a colorful hillside and the robot consults a folded trail map.
Big Picture: How to use this page

Eight curated pathways, classified by goal: "ship a RAG bot in a weekend" is a different journey than "build a coding agent that passes SWE-bench." Each pathway names the skill you'll have at the end and the chapters that get you there, with a realistic time estimate.

Estimates assume comfortable Python, willingness to skim, and reading on a laptop rather than a phone. Add 30-50% for hands-on lab time. Halve them if you only need the concepts and not the code.

The Eight Pathways

Each pathway picks a complete subset of the book and orders it so the reader builds shippable artifacts as they go. Read the goal, scan the time budget, then follow the numbered steps in order; every step lands a measurable skill before introducing the next concept.

1. RAG Engineer (weekend)

Goal: Stand up a question-answering RAG system over your own documents that you can actually ship.

Time: 8–12 reading hours; 6–10 lab hours. Doable in one weekend.

What you can do at the end: Chunk a document corpus, embed it, store the vectors, retrieve relevant chunks, ground LLM answers in them, evaluate the result, and deploy behind a streaming endpoint.

  1. Chapter 11 (LLM APIs) · 1 hr · so you can call a model
  2. Chapter 12 (Prompt Engineering) · 2 hr · so the calls produce useful output
  3. Chapter 31 (Embeddings & Vector DBs) · 2 hr · so retrieval works
  4. Chapter 32 (RAG) · 3–4 hr · the main event
  5. Section 42.1 (Eval fundamentals) · 1 hr · so you know if it works
  6. Optional: Chapter 62 (Production Engineering) · 1 hr · so you can ship it
  7. Optional extension: Chapter 38 (LLM-Powered Recommender Systems) · 2 hr · applies your RAG pipeline to conversational recommendation; covers query enrichment, item retrieval, and conversational re-ranking over a catalog

Skip: Parts I (Foundations) and II (Understanding LLMs) entirely. You don't need to know how Transformers work to ship a RAG system; come back later when curious.

2. Agent Builder (2 weeks)

Goal: Build an autonomous agent that uses tools, recovers from errors, and could plausibly run in production.

Time: 20–28 reading hours; 15–25 lab hours. Two part-time weeks or one full-time week.

What you can do at the end: Design an agent loop, expose tools via function calling (or MCP), handle multi-step planning, recover from failed tool calls, sandbox dangerous operations, and instrument the whole system with traces and cost guards.

  1. Chapter 11 (LLM APIs) · 1 hr
  2. Section 12.2 (Reasoning + ReAct) · 2 hr
  3. Chapter 26 (Agent Foundations) · 4–5 hr
  4. Chapter 27 (Tool Use & MCP) · 3–4 hr
  5. Chapter 28 (Multi-Agent) · 2 hr (skim; most "multi-agent" is one good agent in a wig)
  6. Chapter 29 (Specialized Agents) · 3 hr (read the agent type closest to your use case)
  7. Chapter 49 (Agent Safety & Production) · 4 hr (don't skip this; it's the difference between a demo and a deployment)
  8. Chapter 42 (Eval) · 2–3 hr

Skip: Part I, Part IV (training), most of Part IX (strategy). You can ship an agent without training a model.

3. ML Practitioner Transitioning to LLMs (~10 days)

Goal: You know classical ML (scikit-learn, maybe PyTorch). You want to be productive with LLMs without unlearning your existing skills.

Time: 25–35 hours over 1–2 weeks.

What you can do at the end: Translate "I would have used XGBoost for this" into "should I use prompting, RAG, fine-tuning, or stay with XGBoost?" Defend your answer in a design review.

  1. Chapter 0 (ML + PyTorch foundations) · 2 hr (refresher; skim if you already know it)
  2. Chapter 2 (Attention) · 2 hr
  3. Chapter 3 (Transformer Architecture) · 4 hr (do the 300-line lab; this is where the lightbulb turns on)
  4. Chapter 6 (Pretraining + Scaling Laws) · 3 hr
  5. Chapter 7 (Modern LLM Landscape) · 2 hr
  6. Chapter 13 (Hybrid ML+LLM) · 3 hr (this chapter is for you specifically)
  7. Chapter 16 (Fine-tuning Fundamentals) · 4 hr
  8. Chapter 17 (PEFT) · 3 hr (LoRA is the rosetta stone for ML ↔ LLM transfer)
  9. Chapter 42 (Evaluation) · 2 hr
  10. Optional bridge: Chapter 38 (LLM-Powered Recommender Systems) · 2 hr (connects classical RecSys intuitions to LLM-based retrieval and generation; especially relevant if recommendation, ranking, or search is part of your ML background)

Skip: Parts VI–VII unless agents/multimodal are on your roadmap.

4. Researcher / Graduate Student (semester)

Goal: Build a strong enough foundation to read current LLM papers without effort and identify a feasible thesis topic.

Time: 60–80 reading hours plus paper reading, lab work, and discussion sections. Plan around a semester course.

What you can do at the end: Read an arxiv paper from this year, identify the contribution, judge whether the claim holds up, and reproduce a small version of the experiment.

  1. Full Part I (Chapters 0–5) · 12–15 hr (no shortcuts; you'll need this)
  2. Full Part II (Chapters 6–9 + 31) · 14–18 hr (this is the meat)
  3. Chapter 10 (Interpretability) · 4 hr (now part of Part II)
  4. Full Part IV (Chapters 13–16) · 10–12 hr
  5. Chapter 75 (Emerging Architectures & Frontiers) · 6–8 hr (thesis-topic gold mine)
  6. Section 42.10 (Research Methodology for LLM Papers) · 2 hr (seriously, read this before you publish)

Skip: Parts VIII–IX (Production / Strategy) unless your research has industrial co-authors.

5. Interpretability & Safety Specialist (3–4 weeks)

Goal: Develop the toolkit to explain why a model produced a specific output, and to red-team agent systems for failure modes.

Time: 35–45 reading hours.

What you can do at the end: Run a probing classifier, visualize attention patterns, identify induction heads in a small model, sketch a sparse autoencoder for feature discovery, design a red-team test suite for a production agent.

  1. Chapter 3 (Transformer Architecture) · 4 hr
  2. Chapter 6 (Pretraining) · 3 hr
  3. Chapter 10 (Interpretability) · 8–10 hr (the heart of this pathway)
  4. Chapter 49 (Agent Safety) · 5 hr (prompt-injection, sandboxing, blast-radius)
  5. Chapter 47 (Safety, Ethics & Regulation) · 6 hr (red-teaming frameworks, privacy attacks)
  6. Section 10.2 (Mechanistic Interpretability) · 3 hr

Skip: Parts III, V, VII unless they directly inform your safety surface.

6. Founder / Product Manager (1 week)

Goal: Make defensible decisions about what to build with LLMs, what to buy vs. build, and how to ship it without setting fire to the unit economics.

Time: 12–18 reading hours. One focused week.

What you can do at the end: Read a model card and judge whether it fits your use case. Estimate token cost per user-task. Decide build vs. buy. Defend "we'll use Claude, not GPT, for this workflow" with reasons.

  1. Chapter 7 (Modern LLM Landscape) · 2 hr
  2. Section 12.1 (When prompting is enough) · 1 hr
  3. Section 16.1 (Fine-tune-or-not decision tree) · 1 hr
  4. Chapter 13 (Hybrid ML+LLM Architectures & Decision Frameworks) · 3 hr (build-vs-buy framing; the chapter for PM decision-making)
  5. Part XIV (LLM Applications by Industry) · 2 hr (skim; read the chapter closest to your space)
  6. Chapter 38 (LLM-Powered Recommender Systems) · 1–2 hr (high business relevance: if your product has search, discovery, or recommendation, this chapter shows what is now possible and what it costs to build)
  7. Chapter 62 (Production Engineering for LLM Systems) · 2–3 hr (scaling, cost guards, what shipping actually costs)

Skip: Parts I, II, IV, VI (technical depth). Loop back if a specific decision needs it.

7. Course Instructor (1 month)

Goal: Build a 12–14 week graduate or upper-undergraduate course around this book.

Time: 30–40 reading hours to design the syllabus; ongoing for teaching.

What you can do at the end: Hand out a syllabus on day one that draws from the right chapters, with labs that match each week's content.

  1. FM.7 Copyright & Legal · 1 hr (this exists; start here)
  2. Appendix B (Course Syllabi) · 2 hr (syllabus templates for five course tracks)
  3. Full Part I (Chapters 0–5) · 12–15 hr (the first half of the course)
  4. Selected chapters per topic · 12–18 hr (pick depth of coverage based on course goal: research-leaning → Part II + IV + X; engineering-leaning → Parts III + V + VI + VIII)
  5. Every chapter's end-of-chapter exercises, mine for homework problems

Skip: Nothing on first pass. Cross-reference what your students will actually use after graduating.

8. The Curious Generalist (a long weekend)

Goal: Get the conceptual sweep of the field without committing to building anything yet.

Time: 6–10 hours.

What you can do at the end: Hold an informed conversation about LLMs at a dinner party. Distinguish hype from substance. Know which deeper rabbit hole to fall down next.

  1. Foreword · 15 min
  2. Chapter 7 (Modern LLM Landscape) · 2 hr (read for the lay of the land)
  3. Chapter 8 (Reasoning Models) · 2 hr (the most "what's new" topic)
  4. Chapter 26 (Agent Foundations) · 2 hr (skim)
  5. Chapter 75 (Frontiers) · 2 hr (the "where this is going" chapter)

Skip: All code blocks. Skim, don't write.

How to Choose a Pathway

Key Insight: The 30-Second Decision Tree

Do you have a deadline? Pick pathway 1 (RAG) or 2 (Agent) for the fastest route to a shipped artifact.

Do you have classical ML chops? Pathway 3 minimizes overlap with what you know.

Are you in school? Pathway 4 (Researcher) or 7 (Instructor) depending on which side of the lectern.

Are you a founder/PM? Pathway 6.

Just curious? Pathway 8.

Working in trust & safety? Pathway 5.

Combine Pathways

The pathways are not exclusive. Pathway 1 (RAG) plus pathway 2 (Agent) is the natural sequence for someone building agentic RAG products. Pathway 3 (ML transition) plus pathway 6 (Founder) is the right sequence for an experienced data leader who is now founding an AI company.

The chapters appear once in the book regardless of how many pathways visit them. Reading the same chapter for two pathways is fine; cross-references within each chapter assume you may arrive in any order.