Cosine grid
What does “close in the geometry” actually mean?Build your own
Selected words
·
Word math
Is this geometry actually computing something, or is “closeness” just suggestive?A
−
B
+
C
=
?
A 2D slice of the 300-dimensional math, oriented along the difference. Subtracting builds the arrow; adding carries that same arrow to the third word. ✕ is where the math lands; the dotted line hops to the nearest real word. The picture is a flattening — on-screen distance can lie, so trust the numbers: they are the true 300-dimensional similarity.
Nearest matches in the lexicon
Fill the three slots, or pick a scenario below.
Try a scenario
Neighborhood explorer
If you give me a word, what’s the model’s sense of what it’s like?Any word
Click any orbiting word to walk there.
Real nearest neighbors
from the full 400K-word space. Boxed words are in this tool’s 500 — click to walk there; gray ones aren’t shipped.
One word, many meanings
When a word has several senses, which one does the geometry remember?Add context — shift the meaning
Averaging another word into the center vector is a hand-made version of what contextual models do automatically. Try bank + river, court + wedding, measure + amendment.
Each word here has two, three, even four unrelated senses — but only one vector. The sense cluster that sits closest is the one the training text fed most; a sense parked far away barely exists to the model.
One vector serves every sense. Each sense's words sit at their true cosine distance from the center — closer means that usage dominated the training text. Hover any word for the exact number.