The Noise project released code and speed experiments for training small language models without backpropagation. The widely cited reduction from 1.76 hours to 1.6 seconds concerns optimizing a small model to a specified loss target on eight H100 GPUs. It does not show a 4,000-fold advantage over conventional backpropagation for every workload; the project also reports better quality from backpropagation under comparable optimization conditions. The repository has since updated experiments with a larger model and a different target. Model size, loss threshold and the comparison baseline all matter.