Subtext Visualizes an LLM's Internal Reasoning in Real Time

Subtext is an open-source tool built on Anthropic's discovery of "J-space" — a small subset of internal representations in language models that functions like a global workspace, verbally reportable, deliberately modulatable, and causally used in multi-step reasoning. Using the Jacobian lens method, Subtext continuously decodes a model's residual-stream activations into vocabulary-space predictions across nine layers, rendering them live during both the reading and generation phases of a conversation.

Running on an open 4B-parameter Qwen model on consumer hardware, the tool reproduces phenomena Anthropic originally observed in Claude-scale models: verdicts forming before any output tokens appear, planned words held at high activation while unrelated tokens are being emitted, and two-hop reasoning chains revealing their unspoken intermediate terms (such as "Italy" before "euros" in a geography riddle). The interface encodes layer depth vertically, reading versus generation phase by color, and readout strength via size and opacity.

The implementation is validated against Anthropic's reference code, matching top-5 readouts exactly with cosine similarity above 0.99998. It requires roughly 10GB of VRAM, supports exporting sessions for GPU-free replay in-browser, and is released under Apache 2.0. Subtext turns abstract mechanistic-interpretability research into a concrete, interactive instrument — though its authors caution it demonstrates functional information availability, not subjective experience.