Toward Physical AI: When the Hardware Becomes the Neural Network
On 18 September 2026 Nature Reviews Physics published a 165-reference Perspective arguing that self-organising memristive networks can learn physically, with no separation between the model and the hardware running it. Six researchers from Los Alamos, INRiM Turin, UCLA, Politecnico di Torino, the University of Canterbury and the University of Sydney set out what has been measured: 93.4% on MNIST learned online on a nanowire device, reservoir computing that trains only a readout layer, and criticality as a tuning target. This piece separates the measured results from the ambition, sets the energy argument against the IEA’s data-centre projections, and lists the four open problems — variability, drift, scaling and readout — that stand between a review paper and a component you can buy.
