Mechanistic Interpretability
Michael Igorevich Ivanitskiy
2026-06-22
What you’ll learn
- Some motivation for why you should care about interpretability
- overview of current techniques, mostly on interp for language
models
- open problems that might make for good projects
Talk Outline
- Introduction
- “Classical” Interpretability, where it fails
- Transformer notation & review
- mechanistic interpretability
The Problem:
- We don’t “make” LLMs, we find them
- These models are getting smarter, very fast
Gradient descent finds a set of parameters which minimize the
training objective, but it’s just a point in a billions-dimensional
space of possible models. We don’t build models the way we build cars or
planes.