Entropica becomes first zero-power AI language model using light-based computing; trained in under 2 hours on laptop GPU, runs on passive optical interferometer needing no electricity for inference

Entropica: 1024-mode unitary evolution with Born-rule readout, trained on TinyStories in under 2 hours on a laptop GPU

Published November 29, 2025 | Version v1 Technical note Open Description Entropica is the first generative language model whose forward pass is physically realizable as a passive linear-optical interferometer (zero electrical power during inference). A 1024-mode, 32-layer unitary network using only Reck-scheme MZI meshes and Born-rule readout learns coherent TinyStories-style generation in under 1.8 hours on a single laptop GPU. Model is ready for a full optical implementation path demonstrate...