Imagitation

Gašper Beguš's portmanteau for what a speech-generating GAN does when imitation overshoots into invention. In Latent Spacecraft (with Nina Beguš and Metahaven) the evidence is a fiwGAN trained on eight English words that uttered start, carrot and watery, nonce forms such as sart and greachy, and, at negative latent values, unprompted two- and three-word strings: a connectionist route to concatenation and so to proto-syntax, with no Chomskyan language organ. The mechanism is Informative Imitation: a generator that never sees the data must satisfy a discriminator while encoding a recoverable code, and its probing of the latent space yields words that were never there. Children's babbling and speech play are the biological counterpart, Joyce's Finnegans Wake the literary one; FinneGAN imagitates in Joyce's idiom. The term sits awkwardly against Predictive Intelligence: the paper insists GANs are not autoregressive predictors, the very capacity Agüera y Arcas makes definitional of intelligence.