Concept note · the lab vocabulary
Latent space
The geometry inside a model where its actual reasoning happens.
What it is
The high-dimensional activation space a language model computes in. Every answer is a path through it — token trajectories climbing the layer axis, bending as attention and feed-forward blocks act on them. The weights are the terrain; the activations are the weather. Interpretability is the problem of saying, for a specific prompt, which route through that space actually carried the thought.
In the lab
- IMAXING Latent Space →
- The instrument built on the reframing: a prompt rendered as a volume, contrast prompts as fog, a dye emitter at a chosen feature — and the lit route as the answer to “which path carried the thought?”
Our position
Currents, not circuits. “Circuit” implies fixed topology and on/off switches; activation vectors behave like fluid, not wire — so we borrow the mathematics of flow: viscosity shaping the path, laminar versus turbulent as the difference between clean inference and hallucination, solitons as concepts that hold their shape. And we observe the way fluid dynamicists do: not by measuring the whole velocity field at once — that yields statistics, not comprehension — but by injecting a tracer and watching where the medium carries it. One dye pulse, one legible thread: branching where currents split, pooling where they stagnate.
Honest limitations
The honest accounting: reducing the space without losing structure is research-grade work — two principal components carry roughly two-thirds of the variance, which means a third of the story is not on screen. Aligning per-bottle maps into one coherent atlas is unsolved, real-time refresh at meaningful rates is hard, and a lit path is strong evidence of a route, not a causal proof. The framework is the rare part and it exists; the rest is a team-scale engineering project, and we say so.
Related in the vocabulary
- Instrumentation →
- Meets it in IMAXING Latent Space — building the instrument that makes a hidden state legible.
The experiments are the proof
Every claim in this note is made concrete somewhere in the lab — the idea exists because the machines needed it.
The map of the lab →