.@originalmaderix just got Apple’s Neural Engine to do training, not just inference. He reverse-engineered the private ANE stack and ran forward + backprop directly on the ANE. (Previously, Apple mainly exposes ANE through Core ML, and doesn’t provide a public training API/docs for ANE.) If this scales, local fine-tuning and always-on experimentation become cheap, quiet, and private. It’s still an early PoC: single-layer demo, some CPU fallback (dW/Adam), and private APIs that could break. But the efficiency upside is real, super exciting!