Show HN: Zero-power photonic language model–code (zenodo.org)

Entropica is a novel generative language model whose forward pass can be executed physically with a passive linear‑optical interferometer — meaning essentially zero electrical power during inference. The research demonstrates a 1024‑mode, 32‑layer unitary network built only from Reck‑scheme Mach‑Zehnder interferometer (MZI) meshes, using Born‑rule readout (optical intensity → probability) to produce TinyStories‑style text. The model was trained in under 1.8 hours on a single laptop GPU, and the authors provide all code, weights, and dataset scripts. They also show a clear optical implementation path, including printed phase masks and a basic ~$30 laser‑pointer demo, to illustrate real‑world, passive inference. This matters because it shows a physically realizable route to very low‑energy neural inference by encoding linear transforms as unitary optics and relying on measurement for probabilistic output — a blueprint for analog photonic language models. Key technical details: the network is strictly unitary during propagation (1024 modes × 32 layers), uses only standard MZI meshes (Reck decomposition) for programmable linear optics, and converts amplitudes to token probabilities via Born‑rule readout. While training was done digitally, the work opens a reproducible, hardware‑aware direction for energy‑efficient ML, enabling researchers to explore scaling, noise/attenuation effects, and co‑design of training algorithms for photonic inference.