mech

Mechanistic interpretability

mech

Prompt a small language model and read its J-space — the silent words it holds in mind across layers, via Anthropic’s Jacobian lens. Prefer a quick tour? See the precomputed showcase.

What you’re looking at

The Jacobian lens maps residual activations at each transformer layer into the model’s output vocabulary using the average input–output Jacobian Jl = E[∂hfinal / ∂hl]. Unlike the logit lens, it corrects for how representations rotate across layers — so mid-depth cells show concepts the model is poised to talk about, not noise.

Based on Verbalizable Representations Form a Global Workspace in Language Models (Gurnee, Lindsey, et al., Anthropic). This demo runs Qwen/Qwen3-1.7B on CPU with a Neuronpedia-fitted lens. Anthropic’s published figures use Claude-scale models — expect richer silent concepts there. For interesting readouts, try the curated chips (especially silent modulation / multi-hop), not free-form “think deeply.”