Neural networks, the algorithms behind modern AI, rely on notoriously uninterpretable thought processes that emerge during training. In this project, I attempt to understand them through parameter-mapping sonification—representing numbers as sound through variations in pitch, loudness, and timbre. I exploit the fact that AI with repetitive structure, such as large language models, have internal values which map naturally to the time dimension. The resulting sounds may serve as new ways of conceptualizing and interpreting these complex systems.

Additionally, I question the perceived tradeoff of beauty and information content in sonifications. Human hearing is particularly good at recognizing patterns over time. Considering the prevalence of “beautiful” emergent patterns in nature, can a model’s optimality reveal itself in the form of aesthetic qualities?