Optical receiver updates AI model memory directly from light

Cornell Tech researchers built an optical receiver that updates its own memory directly from photocurrents produced by a beamed light array, rather than just reading encoded data like a QR code. The design, presented at the IEEE/JSAP Symposium on VLSI Technology & Circuits, aims to let AI model parameters be loaded optically to cut memory-update energy costs.

Why it matters: Energy and power constraints are already a live bottleneck for AI infrastructure, as seen in the PJM grid strain and Google's first negative-cash-flow quarter tied to AI spending. A hardware approach that cuts the energy cost of updating model parameters in data centers, self-driving cars, or edge robots targets that bottleneck directly, though it's early-stage research rather than a deployable product.