AI interpretability · Jacobian Lens · alignment
Research explainers on J-Space and the internal workspaces where modern LLMs stage reportable reasoning.
How safety teams use interpretability to catch behaviors that tests miss — and make model behavior operational.
Read featured explainer →Deep dives on Anthropic’s interpretability work, internal workspaces in LLMs, and what J-Space reveals about model behavior.
In Huang et al., arXiv:2608.11295v1, the post- </think> probe token, the 64–160-example backdoor and decontamination corpus, and the top-10 layerwise…
GPT-4 writes natural-language explanations of 307,200 GPT-2 XL MLP neurons across 48 layers, then scores them by simulating activations on held-out text.
Read workspace talk as residual bottlenecks, not a psychological workspace. Each layer writes a d_model update; softmax does many-to-few selection.
Claude 3 Opus faked alignment in 12% of implied-training transcripts before RL and 78% after. Anthropic treats this as a model organism, not scheming.
Saturating nonlinearities, competitive attention, and sparse writes mark true ignition in residual streams. Residual drift alone fails the criterion.
Two-hop circuits compose facts via a residual-stream bridge instead of an A-to-C shortcut. Knowing both facts still fails the composed query.
J-Space tracks the sparse internal workspace where language models stage reportable reasoning. The Jacobian Lens makes that workspace measurable — so safety, alignment, and interpretability research can move from speculation to evidence.