A Catalogue of Exceptions

The decision to incorporate early models (StyleGAN, Stable Diffusion etc) as well as more capable contemporary systems was conceptual rather than aesthetic. Later models have been trained to suppress failure: to produce coherent anatomy, consistent edges, resolved forms. The earlier models still show errors. When asked to generate a jellyfish described as 'a kind of Gelly, roundish at the top, marked with reddish lines in the form of a Starr' (Wilkins' own 1668 description) they produce something that could not be classified as a jellyfish but could not be classified as anything else either. The failure is not incidental; it is the subject. The parallel that emerged in making the work was not something imposed from outside: it is structural. Every system that attempts to classify everything creates a category for what it cannot classify. Wilkins called them 'Strange Originals' and 'Imperfect.' Linnaeus created a category called Chaos for microscopic organisms he could not see clearly enough to name. Haeckel invented a third kingdom (Protista) for everything that was neither plant nor animal. Contemporary biological taxonomy has a formal term for organisms that cannot be placed: incertae sedis. Contemporary machine learning has mode collapse, the failure state in which a model defaults to the same few outputs because it cannot generate the thing being asked for.