Volume II
The Inner Mechanics of LLMs
Understanding the mechanisms, building tools, checking their results
PaperbackAmazon
KindleComing soon
Volume 02 · The investigation
The Internal Mechanics of LLMs
Understand the mechanisms, build tools, check their results
Choose your edition Amazon · Paperback / Kindle
Choose your book’s language, then its format. Prices and delivery details are shown on Amazon.
English
Français
Volume II
La mécanique interne des LLM
Comprendre les mécanismes, construire des outils, vérifier leurs résultats
Español
Volume II
Coming soonLa mecánica interna de los LLM
Comprender los mecanismos, construir herramientas, verificar sus resultados
PaperbackComing soon
KindleComing soon
Português (Brasil)
Volume II
A mecânica interna dos LLM
Compreender os mecanismos, construir ferramentas, verificar seus resultados
Other Amazon stores
Kindle · amazon.com ↗The reading journey
Your tool answers. How do you check it?
Build, measure, inspect and test what you observe. A method for putting results to the test.
- Define before measuringThe task, the computation and the reference
- Build and put to the testClassification, retrieval, extraction and abstention
- Inspect and interveneRepresentations, patching and adaptation
- Transfer carefullyLanguages, formulations and counting
Who is this for? What do you need?
For readers who want to build and evaluate their tools. Python and familiarity with tables help when reproducing experiments. The explanations can be read without running the code.
A convincing answer is not proof. This volume starts with an explicit task, examines a tool’s results and asks which controls support a conclusion.
Its four parts connect building, measurement and investigation. The French edition analyzed describes twenty-four experiments. Small-model examples and controlled datasets teach methods; they do not rank current models.
Inspection and intervention are distinguished: what can be observed or decoded does not, by itself, establish what causes a result. The final chapters examine transfer across languages and estimation on a corpus.
Companion code on GitHub
Companion code on GitHubThe repository accompanies the book. Check its contents and instructions for the resources currently available.
Reference to the French edition; pagination in other editions may differ. PDF pages 6–14 · T2-FR-20260927.