Follow a guided notebook to find the induction heads in a small transformer, the tiny circuit that lets models continue patterns they have seen earlier in the text.
Induction heads are the classic “hello world” of mechanistic interpretability. Finding them yourself teaches the tools (attention patterns, ablations) that interpretability researchers use.
Output: A notebook showing the attention pattern, an ablation, and your explanation.
Open the ARENA interpretability materials (or the TransformerLens docs) in Google Colab. Load a small model such as GPT-2 small.
Feed the model random tokens repeated twice. Check that loss drops on the second half. That is the induction behaviour you are hunting for.
Plot attention patterns and look for heads that attend from a token to the token after its previous occurrence. Score every head this way.
Ablate (switch off) your candidate heads and show that the in-context pattern-copying gets worse. Correlation is not enough. This is the causal test.
Write a short post: what induction heads do, how you found them, and one thing that confused you.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Next: Tackle a beginner-graded open problem in interpretability Next steps
Neel Nanda’s list of 200 open problems grades each one by difficulty, including problems marked as good for beginners. Pick one, scope it down and investigate with TransformerLens.
Use Neuronpedia, a free browser tool, to see which internal “features” of a language model light up for different ideas, then try to break your own explanation of one.
Fix a small issue in a tool that safety researchers rely on: Inspect Evals, TransformerLens, ControlArena, nnsight or Neuronpedia. No ML research experience needed, just software skills.
Bender and Koller argue a system trained only on text cannot learn meaning. Use Neuronpedia’s free interface to Gemma Scope to inspect features for concepts that seem to need grounding (colours, spatial relations) and check whether they respond across very different wordings.
