Then how do you expect to build an AI system without user data? How do you expect Siri to know you when it doesnt know anything about the user? Some smart agent eh?
Going back the topic on hand, this site and many of the other tech blogs are blowing things out of proportion. John isnt a great hire for Apple when it comes to improving Siri. His big contribution to Apple may be more toward search technology then AI. He wasnt part of the original group in the AI department. He was a third party when Google bought his company.
Apple actually has a very good answer for this: machine learning.
Here are the talks (if you're not a developer, just ignore the code stuff - the examples are amazing enough!):
https://developer.apple.com/videos/play/wwdc2017/703/
https://developer.apple.com/videos/play/wwdc2017/710/
Here are some of the things machine learning is enabling today in iOS:
- Photo search: e.g. "roses", "ball" or "dog"
- Image captioning: e.g. photo -> "dog in a field playing with a ball"
- Language recognition: e.g. "...." -> Spanish
- Object tracking: (they have an amazing video of tracking a bunch of grapes on a windy day)
- Object detection: (they have a jaw-dropping video where they go around a house, and the camera recognises his integrated kitchen appliances as "fridge" or "microwave" and detects a "banana" in real-time)
- Apple Watch handwriting recognition (in my experience, incredibly accurate)
The most amazing thing is that every one of these features is 100% offline, on-device processing. They are not sending your text to a server to detect the language, or your image to generate a caption, like Google might do.
Instead, Apple create neural-networks which train machine-learning models which you download and run as part of an App which uses these features. They gather data from iOS devices to train the models and improve them - but, and this is the key difference - that data is anonymised and sampled together random junk, so Apple can't trace it back to you (and it's deleted after a short period of time, once the model doesn't need it any more).
You can also create and train your own models (there is already a big community doing this, before Apple), and they have tools which allow you to easily integrate pre-trained models from the research community.
They can also do some limited training on-device, so the model will evolve to suit you. That's what the keyboard does, for example.