Capstone projects

Follow Karpathy's 'Let's build GPT' end to end, then train on Shakespeare. If you do exactly one capstone, do this one — every later module assumes the intuition it builds.

Train the ReLU toy model from Toy Models of Superposition and reproduce the sparsity/importance phase transitions and feature geometry (digons → triangles → pentagons).

Train a sparse autoencoder on one layer of GPT-2 small with SAELens. Find five interpretable features, characterize their activation distributions, and validate one causally.

The field's rite of passage: replicate the Indirect Object Identification analysis in TransformerLens — name movers, S-inhibition, backup heads — with activation patching.

Pick a trait, build a contrastive activation steering vector on a small open model, and ship a demo page with a steering-strength slider. Watch coherence fall apart past the sweet spot.

Have a friend (or an agent) LoRA-finetune a hidden quirk into a small model. Find it using only interpretability tools, then write up your audit like the Anthropic auditing-games paper.