The capability of neural networks to generate texts and images by learning from large amounts of data is often framed as both the most significant contribution and the most obvious flaw of contemporary artificial intelligence research. Much critical work thus starts from a reading of training datasets – but the mapping from training data to trained model is always messy and indirect. Bias is not just a question of what is represented but also of the logic of representation itself, of the peculiar ways of knowing that emerge from training neural networks on unprecedented amounts of multimodal data.

Vector Media considers the ideological charge of these peculiar ways of knowing and thus writes a new historical epistemology of generative artificial intelligence. It reframes artificial intelligence systems as determined by a central, overarching paradigm: that the production of new knowledge can be operationalized as a geometric interpolation between fragments of existing knowledge. Vector mathematics, of the kind employed in the famous analogy query of natural language processing, “men + king = woman + ?”, is imagined as a tool to achieve this – but its prerequisite is the elimination of all media specificity, going far beyond the kind of alignment required by mere digitization. We are used to technologies creating new (or emulating old) media. Artificial intelligence systems are instead designed to collapse all media into a universal space of commensurability, the vector space, enabling their circulation, translation and manipulation within artificial intelligence infrastructures.