Interpretability
Open the black box: circuits, superposition, sparse autoencoders, induction heads, grokking, and tracing what LLMs compute.
Before this: LLMs from Scratch.
-
4 min
Step 1What is interpretability?
The goal and the motivation, in under four minutes.
What is interpretability?Anthropic
-
24 min
Step 2The dark matter of AI
A visual tour of how researchers read model internals.
The Dark Matter of AI [Mechanistic Interpretability]Welch Labs
-
25 min
Step 3Mechanistic interpretability: an introduction
Neel Nanda on features, circuits, and open questions.
An Introduction to Mechanistic Interpretability – Neel Nanda | IASEAI 2025International Association for Safe & Ethical AI
-
22 min
Step 4A whirlwind tour
The core ideas and techniques in one fast talk.
Neel Nanda – Mechanistic Interpretability: A Whirlwind TourFAR․AI
TransformerLensOrg/TransformerLens★ 4.1k -
13 min
Step 5Superposition
Why models pack more features than neurons.
Superposition: More Features Than NeuronsDataMListic
-
1h 00m
Step 6Transformer circuits and induction heads
The circuit behind in-context learning.
Stanford CS25: V1 I Transformer Circuits, Induction Heads, In-Context LearningStanford Online
-
35 min
Step 7Grokking, reverse-engineered
How a small network learns modular addition with Fourier features.
The most complex model we actually understandWelch Labs
-
29 min
Step 8Sparse autoencoders
Pull interpretable features out of model activations.
Hoagy Cunningham — Finding distributed features in LLMs with sparse autoencoders [TAIS 2024]AI Safety 東京
-
38 min
Step 9Intro to mechanistic interpretability (ARENA)
A hands-on lecture to start coding experiments.
ARENA Lecture, Week 1 Day 2: Introduction to Mechanistic InterpretabilityARENA – AI Safety Education
ARENA-education/ARENA_materials★ 1.4k -
28 min
Step 10Steering vectors
Change model behavior by editing activations.
Steering vectors: tailor LLMs without training. Part I: Theory (Interpretability Series)Anastasia Borovykh
-
3 min
Step 11Tracing the thoughts of an LLM
What attribution graphs revealed inside Claude.
Tracing the thoughts of a large language modelAnthropic
-
56 min
Step 12Circuit tracing in practice
How the circuit-tracing methods work and what they found.
Anthropic: Circuit Tracing + On the Biology of a Large Language ModelLatent Space TV (see @LatentSpacePod for Pod)
Build it
Watching isn't learning. Pick one and put it on GitHub.
- Load GPT-2 in TransformerLens and find which heads attend to previous tokens on repeated text.
- Train a small sparse autoencoder on one layer of a small model and list five features you can name.
- Use activation patching to locate which layer decides the answer on a simple factual prompt.