Current neural network hardware simulates synaptic weights as floating-point numbersstored in silicon memory, applying them through sequential binary arithmetic. This paperproposes a fundamentally different computational primitive: the photorefractive frequencyselective optical relay, in which weights are physically instantiated as refractive indexgratings in crystalline material. Computation is not performed on physical signals — it isperformed by them, through Bragg diffraction at optical timescales. We describe the operatingprinciple, a tabletop two-stage demonstrator buildable for under 2,000, and three engineeringconfigurations: a quantum chemistry simulation substrate, a volumetric AI inference core,and a fully optical input/output interface that eliminates the digital conversion boundary.The architecture offers three properties unavailable in silicon: weight application at the speedof light, energy expenditure concentrated at weight-writing rather than weight-application,and graceful degradation under damage. These are not performance improvements — theyrepresent a categorical change in the relationship between computation and physics.
Petřina Jaroslav (Tue,) studied this question.