Skip to content

Jonny-English/learn-interpretability

Repository files navigation

Learn Interpretability: Distill Circuits

A bilingual, Colab-first repository for working through the entire Distill 2020 Circuits thread with strict live-only notebooks.

This repository now does one thing only:

  • map the Distill Circuits thread into a single D00-D09 sequence
  • keep mirrored English and Chinese notes
  • generate live notebooks for D01-D09
  • refuse pre-rendered figures/, precomputed artifacts/, and slideshow-style walkthroughs

Start Here

Why This Repo Exists

The Distill Circuits thread is still one of the clearest public introductions to neural circuits. The problem is that many readers meet it as a sequence of beautiful static pages. This repository rebuilds the thread as a runnable study path: each live notebook has to earn its claims with public weights, public data, and runtime-generated analysis.

Distill Sequence

ID Article Paper Mode Docs Colab
D00 Thread: Circuits Thread: Circuits reading-only D00 -
D01 Zoom In: An Introduction to Circuits Zoom In: An Introduction to Circuits live D01 Colab
D02 An Overview of Early Vision in InceptionV1 An Overview of Early Vision in InceptionV1 live D02 Colab
D03 Curve Detectors Curve Detectors live D03 Colab
D04 Naturally Occurring Equivariance in Neural Networks Naturally Occurring Equivariance in Neural Networks live D04 Colab
D05 High-Low Frequency Detectors High-Low Frequency Detectors live D05 Colab
D06 Curve Circuits Curve Circuits live D06 Colab
D07 Visualizing Weights Visualizing Weights live D07 Colab
D08 Branch Specialization Branch Specialization live D08 Colab
D09 Weight Banding Weight Banding live D09 Colab

Repo Shape

  • content/course.json: the single source of truth for D00-D09
  • docs/en and docs/zh: mirrored article notes and navigation pages
  • notebooks/en and notebooks/zh: mirrored live notebooks for D01-D09
  • utils/distill_circuits.py: shared runtime helpers for InceptionV1, stimuli, searches, and weight analysis
  • scripts/: notebook generation, README/docs rendering, validation, link checks, audit, and smoke tests

Regenerate And Verify

python3 scripts/render_readmes.py
python3 scripts/generate_course_notebooks.py
python3 scripts/validate_course.py
python3 scripts/audit_realtime_policy.py
python3 scripts/check_links.py
python3 scripts/smoke_notebooks.py

About

Bilingual Colab-first mechanistic interpretability course with paper reproductions, research scaffolding, and P6-ready evidence gates

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors