In predictive coding, each layer of a network is treated as its own local dynamical system that settles toward the activation values that best predict the layer below it, rather than computing a single global forward-then-backward pass. Because credit assignment happens through local coupling between neighboring layers instead of a globally synchronized, weight-transport-style backward signal, predictive coding is considered more biologically plausible than backpropagation. Standard predictive coding has historically struggled to scale with network depth, since error signals from the output diffuse slowly into early layers. Sakana AI's PC-ALM (predictive coding via augmented Lagrangian method), announced September 14, 2026, extends predictive coding by adding Lagrange-multiplier "dual neurons" to each layer's local dynamics, connecting back to a 1988 Yann LeCun paper and reportedly scaling learning signals to networks as deep as 1000 layers using only local computation.