Model & vocabulary
Not built-
Embeds the word bank once and caches it locally. Required for Discover, Traverse, Arithmetic, and the Manifold view. On the WASM (CPU) backend, larger sizes take longer.
Tool
Neighbors20
Temperature0.00
Sampling statefrozen
Plots the nearest concepts from the bank. Raise the temperature to melt the deterministic lookup: frozen returns the same neighbors every time, liquid samples creatively nearby, vapor drifts toward the edge of relevance. Select any point to see its own neighbors; double-click to travel there.
Steps8
Interpolates a great-circle path between the two vectors and decodes each step to its nearest real word. The depth chart sounds the water along the way: where no real concept is nearby, the path crosses open water — a semantic gap.
Hops10
Looseness8
Free association as a walk: every hop jumps to the k-th nearest concept not yet visited. Looseness 1 circles the drain of synonymy; 8 wanders into adjacent territory; 25 free-associates like a poet. Divergent thinking rendered as a tunable parameter. The readout reports how far you drifted from the seed.
Voids to find6
Samples pairs of related-but-distinct concepts and sounds the water between them. The deepest voids are places where no word lives: candidate territory for a concept that does not exist yet. Click a void to chart it — its midpoint plots as (a ↔ b), the depth chart shows the crossing, and buoys mark every found void on the Manifold map.
Vector arithmetic decoded against the bank, excluding the inputs. Sentence models satisfy this loosely; treat results as suggestive, not exact.
Bring your own concepts. Works without the vocabulary bank.
View & projection
Clusters (k)4
Smoothing2.2
Sea level12%
Height is local concept density across the whole bank, drawn over its top two principal directions. Meaning pools into islands; most of the space is open water where no concept lives — the manifold hypothesis, rendered. Run any tool, then return here to see where its results sit on the map. Click an island to inspect it; double-click to explore it.
In Points view, axes are the top 3 principal components of the active points. Low explained variance means the 3D view is a shadow of a higher-dimensional arrangement.
Legend (cluster)
Run a tool to populate.
points 0 dim -
explained variance: -
PC1 right PC2 up PC3 depth
drag rotate · wheel zoom · click select · double-click explore
Latent Space Explorer
Build the vocabulary to begin exploring, then open the Manifold view to see the whole bank as terrain.
Selected concept
Click a point in the plot.
Cosine similarity (active set)
lowhigh
Notes
Meaning here behaves less like a filing cabinet and more like a fluid: concepts pool into dense basins, thin out into open water, and every point between two ideas is a valid coordinate. The Manifold view charts that ocean for the whole bank, Traverse sounds its depth, and temperature melts or freezes the search. The creativity instruments push further: Drift makes free association a tunable walk, the Gap hunter finds voids where no word lives yet, and the shoreline overlay marks the working hypothesis for where creativity happens — the band between cliché and noise. The argument behind all of it: The Shoreline Thesis → · background: Semantic Fluidity of Latent Space
Technically: this is contextual embedding space, not a generative model's activations. Vectors are mean-pooled and L2-normalized, so cosine similarity is a dot product. Neighbors are decoded by nearest-vector lookup against the bank, since embedding models cannot map a vector back to text directly. The terrain is a 2D shadow of the full arrangement: islands are real neighborhoods; exact shorelines are projection artifacts.
Created by Greg Robison for LatentSpaces Labs (2026)