Wirewalk Instituteof AI
mod-380 Expert Waitlist

Interpretability and Model Internals

For ML engineers and researchers who want to look inside transformer language models, find the features and circuits behind a behavior, and use that understanding to debug, steer and audit models.

The curriculum is being prepared and reviewed. Enrollment is not open yet.

About this course

Behavioral evaluation tells you what a model does; interpretability tries to tell you why. Over the last few years the field has moved from toy models to tooling that works on real open-weight LLMs: libraries that expose every activation, sparse dictionaries that decompose those activations into human-readable features, and attribution graphs that trace how a specific answer was computed. Used carefully, these tools can find the cause of a failure, suppress or amplify a behavior, and support safety claims with evidence. Used carelessly, they produce convincing stories that do not survive a control experiment.

This course teaches both the techniques and the skepticism. You start with the residual-stream view of the transformer and the classic findings (induction heads, superposition). You then work through the experimental toolkit: linear probes, the logit and tuned lens, activation patching and attribution patching, and causal scrubbing-style hypothesis tests. The middle of the course covers sparse autoencoders and transcoders: training them with SAELens, using the open Gemma Scope 2 suite for Gemma 3, browsing and labeling features on Neuronpedia, and building attribution graphs with the open circuit-tracer library. The final modules apply interpretability to steering, refusal and deception audits, and model debugging, and close with what interpretability evidence can and cannot support. Tool and release facts are current as of October 2026.

What you will be able to do

  • Explain the residual-stream view of a transformer and how attention heads and MLPs read from and write to it
  • Extract and edit activations in an open-weight model with TransformerLens or nnsight
  • Use linear probes and the logit and tuned lens, and design controls that rule out spurious probe results
  • Localize a behavior with activation patching and attribution patching and test the resulting circuit hypothesis
  • Train or load sparse autoencoders and transcoders and judge feature quality with reconstruction and interpretability metrics
  • Build and read an attribution graph for a specific prompt and validate it with interventions
  • Steer model behavior with activation additions or feature clamping and measure side effects
  • Assess what an interpretability finding does and does not justify in a safety or audit claim

Syllabus

6 modules · 21 lessons · about 20 hours

Module 1: The transformer as a readable machine

The mental model and vocabulary interpretability work depends on.
4 lessons
  1. Goals and levels of interpretability: attribution, representation and mechanism
  2. The residual stream, attention heads and MLPs as read-write components
  3. Superposition and polysemanticity: why neurons are not the right unit
  4. Classic circuits: induction heads and indirect object identification

Module 2: Instrumenting models

Getting at activations in real models and running experiments reproducibly.
3 lessons
  1. TransformerLens: HookedTransformer, hook points and activation caching
  2. nnsight and remote execution on NDIF for models you cannot host
  3. Gradient attributions with Captum and their limits for LLMs

Module 3: Reading representations

Probing what information a model encodes and where.
3 lessons
  1. Linear probes, control tasks and the linear representation hypothesis
  2. The logit lens and tuned lens: watching a prediction form layer by layer
  3. Finding directions: truth, sentiment and refusal as linear features

Module 4: Causal methods and circuits

Moving from correlation to mechanism with interventions.
3 lessons
  1. Activation patching: clean and corrupted runs, metrics and pitfalls
  2. Attribution patching and path patching at scale
  3. Testing a circuit hypothesis: ablations, faithfulness and completeness

Module 5: Sparse dictionaries and attribution graphs

Decomposing activations into features and tracing computations through them.
4 lessons
  1. Sparse autoencoders: objectives, TopK and JumpReLU, and training with SAELens
  2. Transcoders, skip-transcoders and cross-layer transcoders
  3. Working with Gemma Scope 2 and Neuronpedia: finding, labeling and evaluating features
  4. Attribution graphs with circuit-tracer: tracing a multi-step answer and validating it

Module 6: Using interpretability in practice

Steering, auditing and debugging, plus the limits of the evidence.
4 lessons
  1. Steering with activation additions and feature clamping, and measuring collateral effects
  2. Auditing for refusal, sycophancy and deception-related features
  3. Debugging a model failure from symptom to mechanism
  4. What interpretability evidence can support: illusions, dark matter and honest reporting

Exam and certificate

The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 60 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by TransformerLens, nnsight / NDIF, SAELens, Neuronpedia, Google DeepMind (Gemma Scope), Anthropic (circuit-tracer), Hugging Face Transformers, or Captum. Product and company names are trademarks of their respective owners.

More in Model Development

All 16 courses in this track
mod-201 Intermediate Available

Deep Learning with PyTorch

A hands-on course in modern PyTorch: tensors and autograd, nn.Module design, data loading, a training loop you understand line by line, GPU and mixed-precision training, torch.compile, and exporting a model for deployment.

14 h · 19 lessons See price at checkout
mod-202 Intermediate Available

Transformers from the Ground Up

Build a decoder-only transformer piece by piece: tokenization, embeddings, scaled dot-product attention, multi-head and grouped-query attention, positional schemes such as RoPE, and the training and sampling loop, then map it onto real open-weight model architectures.

16 h · 19 lessons See price at checkout
mod-203 Intermediate Waitlist

Classical Machine Learning in Production

Gradient-boosted trees and linear models still win most tabular problems. This course covers leakage-proof pipelines, honest validation, tuning, calibration and explainability, then packaging, serving and monitoring for drift.

14 h · 19 lessons Join the waitlist