Semantic Axes
Type a word. The tool reads it out of a word-embedding space and shows you the
named axes it leans on, the words nearest to it, a 2-D map of its neighborhood, and
the axes hiding in that neighborhood. Where the projector makes you guess an axis, this one
proposes them. Then keep clicking to wander.
try:
king ·
mother ·
ocean ·
scientist ·
paris ·
money ·
wolf ·
hug
loading embeddings…
substrate:
static embeddings
SAE features
compare ⇄
▶ companion: Why vector search beats keyword search — an intuition pump.
two decompositions of king — ours vs the model’s
our axes — curated, named, signed
the model’s features — learned, sparse, emergent
where king sits — its most salient axes
bar points toward the pole the word leans to; length = strength. Click an axis to fan its neighbors out along it.
map king ’s neighborhood on two axes
x
y
click any point to explore that word
build your own axis — two opposing words define a direction
axes discovered in the neighborhood
— unsupervised PCA, model-proposed, not from the curated bank
Static substrate: GloVe 6B (Wikipedia + Gigaword, 300-d), centered → L2-normalized → int8.
SAE substrate: live feature activations via
Neuronpedia .
Axis method: difference-of-anchors / semantic projection
(Grand 2022 ,
SemAxis 2018 ,
POLAR 2020 ).
Free & open source.