Sakana AI reports training 1000-layer networks without backpropagation using layer-local predictive coding

Sources: primary, paper, code, discussion
Summary: Sakana AI describes PC-ALM, a layer-local predictive coding method it reports trains networks of 1000 layers without a backpropagation pass, which the authors claim is the first layer-local method to reach that depth. The 1000-layer result is reported on MNIST only, at width 32, in residual MLPs, while CIFAR-10 and Tiny ImageNet appear as ResNet-18 comparisons against standard predictive coding at ordinary depth. Both the 1000-layer result and the first-layer-local claim appear only in the write-up, which carries a September 2026 date: the preprint it cites, arXiv 2605.31022 by Seely and Gould submitted 2026-05-29, states it analyzes PC-ALM in nonlinear predictive coding networks up to depth 128. No third party has reproduced the numbers.
Why it matters: Every layer in PC-ALM communicates only with its neighbours, so the result bears on distributed and neuromorphic training hardware where backpropagation's global forward-then-backward ordering is the expensive constraint.
Follow-up: Track an independent reproduction of the 1000-layer result outside the authors' benchmark set.