Latent spaces and arenas of creation in light of generative artificial antelligence

Abstract

This reflective essay presents the authorʼs experience with generative artificial intelligence technologies, from experiments with GANs to text-based diffusion models. The narrative begins with the first image generated by VQGAN+CLIP from a prompt and then returns to reconstruct the technical and poetic trajectory that precedes this moment. Along this path, it introduces the creation of Ernesto Fufihaqi, used as a narrative device to explore intermediate states in neural network training. The text critically examines the transition from GANs to diffusion models, identifying a fundamental loss: while GANs enabled a poetic navigation through latent space, marked by metamorphoses and creative errors, more recent models privilege precision and control, reducing exploratory chance. The essay problematizes the temporal acceleration of technology and its incompatibility with the human time required for creative maturation, in which rapid obsolescence turns recent experiments into digital archaeology. It proposes the concept of creation arenas as spaces of transparent collaboration between humans and machines, in which authorship is distributed and process takes precedence over product. It advocates prototyping and experimentation as forms of knowledge, adopting a stance that avoids both nostalgia and uncritical enthusiasm, and concludes with a call for methodological curiosity in the face of ongoing transformations.