Latent Space

The learned space of vectors and weights, hundreds or thousands of dimensions found in training rather than scripted, where a network holds what it extracts from inputs and generates outputs. Latent Spacecraft (Nina Beguš, Metahaven, Gašper Beguš) makes it philosophical in three moves. The geometry is physical: weights are matter, a motor-cortex neuron contributes a vector like a coordinate axis, so "the purely symbolic level does not exist." Invented twice, by nervous systems and in silicon, it makes brains and GANs isomorphs. And because it cannot be pictured or told, it is treated as architecture, an interior to walk, hence Non-Narratable Worlds; Somaini is cited: whoever controls it controls "the bounds of the visible." Its declared antagonist is de Monchaux's Black Box; its quiet one is Substrate Agency, since one geometry across tissue and silicon is a functionalist wager Tripaldi refuses. Agüera y Arcas calls embeddings a "geometry of meaning."