anthropic published the J-space paper today. tl;dr: models have a small emergent set of internal “silent words” (~a few dozen concepts at a time, <10% of activations) that they can report on, control, and use for reasoning. the measurement tool is the jacobian lens and they open sourced it, and neuronpedia posted pre-fitted lenses for qwen. so the obvious next step was to wire it into a chat UI and just… look at it. subtext runs qwen3.5-4B in bf16 on a single 12GB GPU and reads the lens at ...