Goal. Pair the seven reference books to the chapters they serve, and suggest
an order to read them in relative to the book's five authoring waves.
I lean on seven reference books. None is quoted at length anywhere in the text;
instead each chapter carries a read-along pointer by short key, and this
appendix is the master index behind those pointers. Local copies live in
references/ at the repo root (see that directory's README.md); they are not
distributed with the book.
Sutton & Barto, Reinforcement Learning: An Introduction (2e)
Part V
[RLHF]
Lambert, RLHF: LLM Alignment and Post-Training
Parts IV–VI (esp. ch. 5–8)
[BRM]
Raschka, Build a Reasoning Model (From Scratch)
Parts III, V, VI (ch. 3–8; App. C Qwen3 source)
[BLLM]
Raschka, Build a Large Language Model (From Scratch)
Part I
[MADL]
Chaudhuri, Math and Architectures of Deep Learning
Part I math spine (ch. 2–9)
[GAIA]
Hurbans, Grokking AI Algorithms (2e)
Pre-reading only; RL intuition (ch. 10)
[CAI]
Ness, Causal AI
Part IV
Two of these earn special notes. [GAIA] is deliberately light-duty: it
overlaps the other texts at lower depth, so it serves as optional pre-reading
for RL intuition, never as a chapter dependency. [CAI] earns a full part of
its own (Part IV) because judge-model and benchmark-comparison pitfalls are
confounding problems, and framing evaluation claims causally is a committee-grade
differentiator.
The books are not read cover-to-cover front-to-back; they are read alongside the
parts they serve, in roughly the authoring-wave order (spec §8).
Before anything (optional): skim [GAIA] ch. 10 for gentle RL
intuition. Nothing downstream depends on it.
With Part I (the theory spine): read [MADL] ch. 2–9 as the math
backbone and [BLLM] ch. 2–5 as the from-scratch build companion. These
two run in parallel with Part I chapter for chapter.
With Parts II–III (inference and evals):[BRM] ch. 1–3 and its
App. C (Qwen3 source) back the model-anatomy and eval-framing chapters.
With Part IV (causal):[CAI] Parts 1–3, in order, are the spine.
With Part V (RL foundations):[S&B] is the backbone (ch. 2–4 and
ch. 13 carry the most weight), with [BRM] ch. 6–7 and [RLHF] ch. 6
arriving as the material turns to LLM-specific policy optimization.
With Part VI (post-training):[RLHF] ch. 1–8 is the throughline;
[BRM] ch. 4–5 and ch. 8 cover reasoning-time scaling and distillation.
Every entry below mirrors the read-along admonition in the named chapter.
Chapters with no reference pairing are hands-on or self-contained (Part 0, most
of Part II, and Part IX lean on tooling and the book's own artifacts rather than
outside reading). Part VII does carry read-along pointers, because the loop
chapters lean on [RLHF], [S&B], and [MADL] for the algorithm and kernel math
even as the labs are hands-on; its per-chapter rows are below.
When a chapter's read-along pointer and this table disagree, the chapter is
the source of truth: update this appendix, not the chapter. This index accretes
as chapters are drafted, so pairings for chapters still in stub form may be
refined when their read-along blocks are written.