🧱 DiffusionBlocks: Training Neural Networks One Block at a Time https://pub.sakana.ai/diffusionblocks

DiffusionBlocks is a novel training method that enables neural networks to be trained sequentially, one architectural block at a time, rather than end-to-end. This approach, developed by Sakana AI, leverages diffusion-based techniques to allow each block to learn independently while maintaining compatibility with the overall network, potentially reducing computational costs and enabling more modular network design. The method represents an alternative to traditional backpropagation-based training that could have implications for distributed and efficient deep learning.

Source: David Ha (X) | 2026-05-28