Improbable Marks: the recursive latent

Datasets of text, images or drawings can create collections of variation. Through their patterns of similarity and difference, generative AI can be tasked with learning what is significant. Multimodal mixtures enable the correspondence between different types of expression to be measured. These can go beyond text and image, combining collections of different affordances to construct an architectural understanding of multimodality. These novel latent spaces can be designed, sampled and explored. When a creative act, such as speech, drawing or design, is tokenised, each mark is no longer read simply as representation but also as a sequence of action. These recursive latent spaces measure the probability not of outcomes but of process. Every mark is measured twice: against the cultural milieu the model learnt, and against the patterns the work establishes. From a recursive self-patterning we might establish the poetic. Rather than using tokens to simply predict what comes next, these models can measure expression against the probability field it disturbs. The latent space is not just static and generative but traces, mirrors and redraws creative agency.