Latent Space
The compressed coordinate system a model learns, in which each example is a short vector rather than raw data.
"Latent" means inferred rather than observed. An encoder maps data into the space, a decoder maps a point back out, and the narrow bottleneck between them is what forces the model to keep only what matters.
Its useful property is that distance means something. Shift every pixel of a photo one column right and you've moved an enormous distance in pixel space and barely moved in latent space — which is why interpolating two latent codes produces a plausible third example while averaging two images produces a double exposure.
Two cautions. Axes aren't named attributes unless you trained for that, so reading dimension 7 as "sentiment" is usually wrong. And the space is only defined up to rotation, so coordinates from two training runs aren't comparable even on identical data.