# Subtext

A real-time instrument that visualizes what a language model is internally 'thinking' word by word during a live conversation, using the Jacobian lens on residual-stream activations.

Subtext continuously applies the Jacobian lens, a technique that decodes a residual-stream activation at any layer through the model's own unembedding to reveal which vocabulary words that internal state is disposed to produce, to a live conversation with a local language model. On every token, both while the model reads the user's message and while it generates its reply, the lens is read at nine depths and rendered as it happens, so a viewer can watch verdicts form during reading, planned words held at strength before they are emitted, and unspoken intermediate reasoning steps surface.

Unlike existing interactive lens demos, Subtext is conversational and continuous: it covers the reading phase over the user's message as well as generation, streams at generation speed using a KV cache, and pairs the visualization canvas with a per-token ledger and a per-word inspector. Sessions can be exported and replayed later in any browser without needing a GPU.

Website: https://github.com/ninjahawk/Subtext