About
About this site
mechinterp.io is a personal site for research and writing on mechanistic interpretability.
What mechanistic interpretability is
Neural networks learn to solve problems, but nobody writes down how. Mechanistic interpretability is the effort to reverse-engineer them: to identify the features a model represents internally, work out the circuits that combine those features into behaviour, and check each explanation by intervening on the model and seeing whether it responds as predicted.
The goal is understanding that is specific and testable: not only that a model can do arithmetic, but which components move the numbers around, which ones compute the answer, and what happens when you remove them.
What you'll find here
This site will host my personal research and blog posts. The research section is for fuller write-ups of experiments; the blog is for shorter pieces, explainers, and work in progress. The work is independent and focuses on open-weight models that anyone can download and inspect.
Approach
- Prefer measurements to intuitions, and say how large each effect is.
- Report negative results and corrections, not just successes.
- Separate what was shown from what is only suggested.
- Publish code and data where possible.