About

Learn Mechanistic Interpretability is an open-source textbook on how language models compute what they compute: what is inside a transformer, the techniques researchers use to reverse-engineer it, and what the field currently does and does not know.

Mechanistic interpretability lives in papers, blog posts, and code repositories, and the order it was published in is not the order that makes it easiest to learn. The material here is arranged as a path instead. There are 70 articles across 15 blocks, and each one assumes only what came before it.

Structure and Content

The articles run from transformer fundamentals through probing, steering, and circuit analysis to safety applications, then finish with the tooling the field uses. You can read straight through, or start wherever you need and follow the "Read first" links backwards to fill in the gaps. The topics page lays out the whole map, the glossary collects the vocabulary, and search covers the full text of every article.

Each article is about a technique, a concept, or an idea, not about a paper. One influential paper might become a single article, or several, or none, depending on how much of it generalizes beyond its own experiments.

Contributing

Contributions are welcome, from a typo to a whole new article. Every article has an "Edit this page on GitHub" link at the bottom and credits the people who wrote it. The source lives in the GitHub repository, and the writing is licensed under CC BY-SA 4.0.

The Logo