05/11/2024
Generative artificial intelligence encompasses models designed to synthesize new text, images, sounds, or other kinds of content according to the datasets they have been trained on [...] For users of generative models, who might not have direct access to neither the model itself nor the training data, this entails an epistemological challenge, as all that is available for interpretation is the input and the output, with everything in between hidden away inside nested black boxes.
This essay proposes that these nested black boxes can be, if not opened and examined, at least shaken for clues about their functioning. My argument is that, while the high-dimensional nature of latent spaces makes them fundamentally impenetrable to human cognition, the correlation between inputs and outputs can be operationalized to obtain some insights into the data a model has been trained on, what the model has learned from it, and how the model draws upon it to synthesize new information. As a qualitative researcher, I approach these questions from the perspective of everyday use at the human scale.
Gabriele de Seta, i Bergen
https://sociologica.unibo.it/article/view/19512
Synthetic Probes: A Qualitative Experiment in Latent Space Exploration Authors Gabriele de Seta Department of Linguistic, Literary and Aesthetic Studies, University of Bergen https://orcid.org/0000-0003-0497-2811 Gabriele de Seta is, technically, a sociologist. He is a Researcher at the University o...