Models From First Principles — Working Ledger
Book intention
Understand modern AI models by opening the abstractions and building the mechanisms in layers, while keeping the mathematics, architecture, optimization, and working code connected.
The review question is not merely whether each chapter is well written. It is whether each chapter does the job it is supposed to do at this point in the sequence and whether the sequence leaves a useful model or mechanism missing.
Editorial state: ACTIVE — REVISION
Review ladder: UNREVIEWED → STRUCTURE OK → CONTENT OK → PROSE OK → DONE
Current intended arc
The Model Inside the Model → MR.Q → EBT → SICQL → HRM → Tiny → Inside Tiny → PACS → Preference Rankers → comparison / model choice
The final comparison currently lives at 15-chapter.md; its front matter still identifies it as chapter 08. Do not silently normalize this. Review its new role after Preference Rankers first, then decide whether it should become 09-chapter.md and be updated to include the new model.
Chapter ledger
| # | File | Canonical title | Status | Job / durable notes |
|---|---|---|---|---|
| 00 | 00-chapter.md |
The Model Inside the Model | UNREVIEWED | Establish the first-principles method: useful model abstractions are made from smaller mechanisms that can be opened, traced, and understood. Set the conceptual contract for the book. |
| 01 | 01-chapter.md |
MR.Q — Building a Neural Quality Model From Two Embeddings | UNREVIEWED | Introduce a compact quality model and make scoring/ranking from representations concrete before the book moves into more composite architectures. |
| 02 | 02-chapter.md |
EBT — From One Score to Q, V, Policy and Advantage | UNREVIEWED | Expand from a single learned score into related decision quantities and show how Q, V, policy, and advantage fit together. |
| 03 | 03-chapter.md |
SICQL — Building a Model From Q, V and Policy Networks | UNREVIEWED | Turn the quantities from EBT into a composed architecture with specialized networks; show how model behaviour emerges from the relationship between components. |
| 04 | 04-chapter.md |
HRM — Hierarchical Reasoning With Fast and Slow Recurrent State | UNREVIEWED | Introduce a different reasoning mechanism: recurrent state operating at different timescales. Broaden the reader’s notion of what a reasoning model can be. |
| 05 | 05-chapter.md |
Tiny — Recursive Reasoning With a Small Neural Network | UNREVIEWED | Show how repeated computation through a small model can produce reasoning behaviour; challenge the assumption that capability must come primarily from scale. |
| 06 | 06-chapter.md |
Inside Tiny — Residual Blocks, Attention and Sparse Autoencoders | UNREVIEWED | Open Tiny itself and connect its larger behaviour to recognizable internal mechanisms. This chapter must deepen 05 rather than merely repeat it. |
| 07 | 07-chapter.md |
PACS — Building an Optimizer From Gradient Statistics | UNREVIEWED | Shift from architecture to learning dynamics and show an optimizer as another model-like mechanism that can be derived and inspected from first principles. |
| 08 | 08-chapter.md |
Preference Rankers — Learning Which Answer Is Better | UNREVIEWED | New chapter. Introduce pairwise preference learning and connect embeddings/quality scoring to ranking, reward modelling, and systems that learn relative rather than absolute targets. |
| 09? | 15-chapter.md |
Which Model Should You Use? MR.Q, EBT, SICQL, HRM, Tiny and PACS Compared | UNREVIEWED | Finale/comparison. Re-evaluate after adding Preference Rankers. It currently omits the new chapter and internally calls itself chapter 08. Decide its final number, title, comparison set, and whether the conclusion still represents the revised book. |
Book-level questions to resolve during review
- Is the book really a coherent progression of model mechanisms, or currently a collection of interesting architectures?
- Does PACS belong where it is, given that it shifts from model architecture to optimization?
- Does Preference Rankers naturally follow PACS, or should the comparison/sequence change around it?
- Which important model family or learning mechanism is conspicuously absent once the existing chapters are understood as a set?
- Does the finale compare the things the preceding chapters actually taught, using criteria the reader now understands?
Current next action
Start with 00-chapter.md. Review structure first. Do not begin by polishing sentences.
