Level 2 · Small-model research on one GPU
Can you locate the components that implement a specific behaviour inside a small transformer, and prove it by intervening on them?
Work through ARENA’s transformer interpretability chapter: build a transformer, find induction heads in a 2-layer model, then reproduce the indirect-object-identification circuit in GPT-2 small.
ARENA is the curriculum many alignment researchers trained on. Doing it properly gives you the vocabulary and tools (TransformerLens, activation patching) that every interpretability project below assumes.
01Set up the ARENA exercises
Clone the repo and run chapter 1 notebooks in Colab or locally. Do the exercises yourself before opening solutions.
02Build a transformer from scratch
The first section has you implement GPT-2’s components and load real weights. Compare with your Level 1 GPT-2.
03Find induction heads
Use TransformerLens to inspect attention patterns in a 2-layer attention-only model and identify the heads that copy repeated sequences. Confirm with ablation.
04Reproduce the IOI circuit
Work through indirect object identification in GPT-2 small: activation patching, path patching, the name-mover heads.
05Pick a new small behaviour
Choose a behaviour not covered in the course (for example, a simple date or plural pattern) and find which heads and MLPs matter.
06Prove it
Ablate or patch the components you claim are responsible and show the behaviour breaks; patch them back and show it returns. Report effect sizes.
07Write a short LessWrong-style post
Claim, method, evidence, what you couldn’t explain. Keep it under 2,000 words with clear figures.