As detailed in a 2022 IEEE Spectrum article, “spin glass” nanomagnet arrays have long promised a low-power, self-learning alternative to digital AI, though training them has remained a major physical hurdle. A new study in Nature Communications by Jérémie Laydevant, Danijela Marković, and Julie Grollier finally breaks this barrier by successfully training an Ising machine (a hardware spin glass) using Equilibrium Propagation. This physics-aware algorithm allows the system to learn by relaxing into energy minima—effectively treating the hardware’s intrinsic physics as the computational engine—rather than forcing it to run digital backpropagation. Validated on a D-Wave system, the approach achieved supervised learning results comparable to digital networks, bringing the vision of fully autonomous, physics-based neuromorphic chips closer to reality.