Circuit Traces
The attribution-graph method behind Anthropic's Biology of a Large Language Model, on an open model you can trace yourself. Each case study below is a real causal graph through Gemma 2 2B: which features caused which, from the prompt to the predicted token.
Method: Ameisen et al. and Lindsey et al. (2025), Transformer Circuits; open-source implementation by the circuit-tracer project; graphs computed on Gemma Scope transcoders and served live by Neuronpedia. A caveat that matters: these graphs describe a replacement model built from transcoder features, pruned for legibility. They are strong hypotheses about the real computation, to be confirmed with interventions, not proofs.
Any graph on Neuronpedia loads here: paste a slug or a full graph URL. To trace a new prompt, generate a graph with a free Neuronpedia account (Gemma 2 2B, one click) or run circuit-tracer locally on any supported open model, then load the resulting slug. This is the workflow the institute will use for its open-model biology work; what you trace, you can publish.