On a technical point, I was impressed by their CoreML model format. The specification is open, optimized for maximum transferability, it can convert models from Keras (with TensorFlow), scikit-learn and others, and their Python model converter is also open source: https://pypi.python.org/pypi/coremltools

Once you have your model as a CoreML file, it is stupidly simple to incorporate it into an app (the live demo with the flower model was very impressive), and Xcode will convert it to a machine-optimized version.

I was skeptical when I saw the announcement, but honestly it seems like a game changer -- to be able to drag-and-drop models that other people have trained into an app and use them with virtually zero boilerplate is just great.

Video here: https://developer.apple.com/videos/play/wwdc2017/703/