j-space

Watch a model think

Turn a jlens chat export into an animation of the concepts a model holds in its layers — flaring and fading a beat before each word is written. Bubble size tracks accumulated decode probability; colour is salience relative to the strongest concept.

Drop a jlens export (.json)
or browse files · or watch the demo
No export yet? Run a prompt on an open-weight model in Neuronpedia's Jacobian-lens explorer, click the download icon to export it as JSON, and drop the file here.
Runs entirely in your browser — nothing is uploaded anywhere.
j-space
00:00.0 / 00:00.0 · tok —

Export

Renders the full sequence from the start.

WebM video
Records one real-time playthrough at 60 fps. Fast and small.
Animated GIF
Rendered offline, frame by frame. Bigger file, plays anywhere.
Preparing…

What am I seeing?

Each bubble is a word the lens caught the model holding in mind, decoded from its layers as it reads the prompt and writes its answer.

size
Accumulated decode-probability mass. Every time a layer's state decodes toward the word, the bubble gains mass; it deflates when the layers go quiet about it. Big bubbles are thoughts the model kept returning to.
colour
Salience, relative to the strongest concept on screen at that moment: cool slate for background murmurs, incandescent amber for the model's current preoccupation.
ring flare
The concept was just reinforced by the token being processed right now.
peak decode
Hover a bubble. "Peak decode 88% at layer 57" means the model's layer-57 state, read out through the lens and the model's own vocabulary head, put 88% of its next-word probability on that word.
None of this is what the model said. It is what the model's own machinery was leaning toward saying, caught before the final layers got their veto. Decode percentages are lens readouts (a first-order approximation of what the remaining layers would do), not the model's actual output probabilities.
Make your own: run a prompt on an open-weight model in Neuronpedia's Jacobian-lens explorer, click the download icon to export the run as JSON, then load it here with Load file (or drop it anywhere on the page).