An explorable map of what a language model was "thinking" while it answered.
Every glowing dot is a word the model internally considered at some moment of a conversation — read out of its layers with a Jacobian lens and rendered as a flyable 3D field. Left to right is the conversation unfolding. Height is how close a thought got to being spoken (a faint plane marks where thoughts become words). Brightness is strength.
The colors are the story:
- Blue — background thoughts. Far more going on than what gets said.
- Gold — thoughts that made it into the spoken answer.
- Pink — feeling-words (emotions, feelings, feel) that the model held the whole time and never said once.
- White — the word it actually chose at that moment.
Click any dot for its dossier — including, for the main conversation, the phrase each depth of the model was about to say at that moment (decoded by letting each layer act as the output head).
There's also an anatomy view (the never-spoken vocabulary self-clustered into ten families — cognition, monitoring, policy, metaphor…) and a language filter: this model (Qwen3.6-27B) holds a substantial Chinese-language channel in its middle layers during English conversations — compliance checks, denial constructions — which never reaches the output. You can view it in isolation, with English glosses on hover.
Captures of Qwen3.6-27B, via Neuronpedia's public
Jacobian-lens API and locally-run full-vocabulary captures with Neuronpedia's published
lens (neuronpedia/jacobian-lens, fitted on wikitext-103). The app opens on a catalog
of every capture — each card shows the exact prompt, the intervention applied (read from
the recorded API request, e.g. +[hum] L40–54, strength 1, reading-only plant), and what
to look for. Eight groups:
- introspection — the three flagship full-vocab runs (canonical hum, baseline, verbalize), where the pink never-said feeling-words live
- day-1 trajectory — the first consciousness-mapping passes
- steering — boost / ablate arms for the 'consciousness' direction (causal control)
- jokes — punchline anticipation ("side of the dataset" is held ~8 tokens early), plus an ablate-'joking' arm where the memorized joke survives and the fresh one dies
- counting — "exactly four" vs "stop when you feel like it": a stop-urge (DONE) gated by a quota, and a phantom queue one run holds that the other actually speaks
- arithmetic — easy vs hard: (10+20)/2 is answered while reading the question; (17+29)/2 is computed at the equals signs — and arrives in Chinese first (四十, 二十三)
- plants · steered / reading-only — the planted-state battery: hum, mirror, emotions, om, and multi-token bundles, across strengths, seeds and temperatures
All data ships in data/ (~62 MB; web/public/data is a symlink to it — on Windows,
enable git config core.symlinks true or copy the folder). The capture/layout pipeline
is in tools/ (export_v4.py + data/catalog.json as the manifest).
🎬 2-minute guided walkthrough — landing catalog, the hum field, the Chinese channel, anatomy view, and both math runs.
Web (any platform):
cd web && npm install && npm run devNative macOS (Metal, smoother with the full field):
cd MindVolume-macOS && swift runRequires a recent Xcode toolchain. Drag to orbit, scroll to zoom, slider to scrub time.
Lens readouts are vocabulary projections of internal states — evidence about internal verbal content, not about experience, intent, or "thoughts" in any strong sense. The lens is fitted on generic web text and applied off-distribution here. These are single greedy generations from one model. The pink-particles-never-cross-the-plane pattern is a robust observation in this data; what it means is an open question, and we've tried to keep the labels honest about that line.
Built with the Jacobian lens technique from Anthropic's "Verbalizable Representations Form a Global Workspace in Language Models" (reference implementation) and lenses fitted by Neuronpedia.
MIT — see LICENSE.