In this episode of The Silicon Satoshi, we dive into a fascinating open-source research project that bypasses Apple's restrictions to train neural networks directly on the Apple Neural Engine (ANE). Discover how developers reverse-engineered private APIs (like ANEClient and ANECompiler) to perform backpropagation and run custom compute graphs for transformer training entirely on the ANE hardware without relying on CoreML, Metal, or the GPU. We discuss the engineering challenges, the impressive proof-of-concept benchmarks, and what this means for the future of edge AI optimization on Apple Silicon. Note: This is an experimental research hack, not a production framework!