Roadmap

I am writing this book in the open, one chapter at a time. That is partly discipline: a public draft is one you have to keep honest. And partly an invitation: the field is moving fast, and I would rather think it through where you can push back. This page is the map — the plan of the book, what is ready to read, and what comes next.

Preface — why this book exists, and how it is being written. · Published

Introduction — two ways to relate to a language model; who this book is for; how to use it. · Later

I. Foundations

1. Old Problems, New Object — the argument of the book: interpretability’s open questions are measurement questions. Constructs, validity, reliability — and the difference between representing and intervening. · Next

2. How Language Models Work — the object itself: neural networks, embeddings, attention, the residual stream, and superposition, only as far as the methods require. · Later

II. Methods

3. Observation — descriptive inference about a model’s internals: probes, the logit lens, sparse autoencoders, and comparing representations. · Later

4. Intervention — causal inference about a model’s internals: activation patching, ablation, circuits, attribution graphs, and steering. · Later

5. Validation — how to know a result is true: benchmarks, metrics, and reliability. · Later

III. Application

6. The Ideology of Language Models — a full study, from research question to design to what the result licenses. · Later

Conclusion — what social science should build next. · Later

---

Every methods chapter closes the same way: what does this technique license you to claim, and what does it not?

If you want to follow along, the source lives on GitHub; every page has a “Report an issue” link, or write to me at tapan.y@northeastern.edu.