Developer Manjeet Singh, with Claude Opus, has successfully conducted neural network training on Apple's Neural Engine (ANE) on the M4 chip through reverse engineering. This marks the first time training has been achieved on the ANE, which is typically used for inference. By bypassing Apple's CoreML framework, the team directly mapped over 40 private classes to the IOKit kernel driver, enabling model compilation in memory—a crucial step for training. The project implemented training for a single transformer layer, achieving 9.3ms per step with 11.2% ANE utilization. The ANE's core computational primitive was found to be convolution, not matrix multiplication, leading to significant throughput improvements. Despite being in early stages, the project is open-sourced under the MIT license and has gained significant attention, with around 2,800 stars on GitHub in five days.