LATENT // LAMINAE: Sonification of Hidden Geometric Reasoning in High-Dimensional AI Systems

LATENT // LAMINAE is a hybrid scientific–artistic project that converts latent-space operations from modern machine-learning systems into an industrial-metal audio composition. The work is grounded in the idea that large-scale AI models compute through hidden geometries—high-dimensional manifolds whose internal transformations cannot be expressed cleanly through human-readable language or chain-of-thought reasoning. Instead of relying on textual explanation, this composition sonifies the underlying mathematical behaviour of latent spaces, including: • vector interpolation across embedding manifolds• latent-space curvature and discontinuity basins• residual shifts and semantic drift• compression dynamics of hidden representations• layer-wise transformations (“laminae”) in deep neural networks Through rhythmic structure, tonal design, and aggressive spectral shaping, the track translates these geometric operations into sound, offering a multimodal perspective on how AI systems internally manipulate meaning. The project provides a novel interpretability pathway: using audio as a medium to reveal hidden transformations in machine-learning models that are typically inaccessible to human audiences.