Skip to content
English

Switch theme

Discover▾

Not sure where to start? This week’s hottest software, models and stacks in three minutes.

AI Apps▾

AI software people use: from assistants like ChatGPT to tools for coding, video and workflows, picked by purpose.

For Agents▾

Software for AI: tools that let your agent (Claude Code, Codex, Cursor…) browse, read docs, run code and remember.

Models▾

Choose models, estimate cost, self-host: prices, rankings, benchmarks, free tiers and compute.

Solutions▾

A goal or an error message? Ready-made stacks, steps and fixes by scenario.

Learn▾

A systematic course from zero to the frontier, with an algorithm library, must-read papers and tutorials.

Frontier▾

Daily AI news, weekly highlights, the latest papers and how AI is changing every discipline.

// Stage 9 · Safety, Evaluation & Frontier

Interpretability: from Attribution to Circuits

Attribution, probes, the logit lens, superposition, SAEs, circuit tracing, knowledge editing and activation steering.

Read the lesson → Code & exercises ↓

about 10 hours9 algorithms4 paperslesson text ready

01Goals

Do these first: Transformers & Attention

02Read on this site

03Algorithms to master

Each card has the original paper, prerequisites and resources.

Feature Attribution (Integrated Gradients / SHAP) Measures how much each input feature contributed to a prediction. 2017core Grad-CAM Gradient-weighted heatmaps showing where a CNN looks. 2016ext Logit Lens & Probing Decodes intermediate layers into tokens, or probes them with small classifiers. 2020core Circuits & Induction Heads Finds circuits implementing algorithms, like induction heads that copy patterns. 2022core Superposition Models pack more features than dimensions by superimposing them. 2022core Sparse Autoencoders (SAE) Decomposes activations into many sparse, interpretable features. 2023core Circuit Tracing & Attribution Graphs Maps causal feature pathways behind a specific response. 2025core Locating & Editing Facts (ROME) Locates where facts are stored and edits those weights directly. 2022ext Activation Steering Adds a direction vector to activations at inference to change behavior. 2023core

04Must-read papers

05Courses & videos

06Hands-on: project, code & exercises

Free, official or author-published pages. Most run in the browser or on a free Colab GPU.

PROJECT

Reproduce induction-head detection on GPT-2 small, then run a steering experiment with SAE features from Neuronpedia.

  1. 1
    arena.education ↗Courses

    Do ARENA’s transformer-interpretability exercises.

  2. 2
    TransformerLens ↗Reference code

    Reproduce GPT-2 induction heads with TransformerLens.

  3. 3
    Neuronpedia ↗Tools

    Browse SAE features on Neuronpedia and try activation steering.

07Self-test

Answer in your head, then open to check.

What does the superposition hypothesis say?

There are more features than neurons, so single neurons are often polysemantic.

The two parts of the SAE training objective?

Reconstruction error + a sparsity penalty.

What behaviour do induction heads implement?

Seeing A B … A, they predict B.

On these routes: Algorithm Researchers

Data updated: AI algorithms & curriculum(UTC+8)