OK - so "It learns" - let us just consider for a moment what that means in the context of a computer based application. Is it learning in manufacture or learning in practice.
All neural network applications (software or hardware) learn, or in the parlance of the discipline are trained. This means a set of example data with known results are given to the application which then is trained to recognise correct results and reject wrong ones. Note that during this process an external "programmer" (a person) decides when an application is sufficiently train such that training stops. This requires a personal subjective decision and the application can be over trained such that it is too particularised to the training set and cannot generalise to new case, or it can be undertrained and not distinguish separate cases.
OK, so this is all done in the programming / preparation / configuration / manufacture phase.
Now consider two dentical brainchips manufactured with the same training being put into practice in the real world and subjected to random input. Question do we want both chips to provide the same answer to the same input.
If you want them to always provide the same answer they need to either stop learning, or they need to have exactly the same pattern of random input AND the identical learning mechanism. If they are subject to different inputs over time and they continue learning in practice they will diverge over time in the answers they give to the same question.
It is for this reason that you do not want learning to continue in practice or the brainchip becomes as fallible as a human.
Think of judges sitting on the High Court. They go through the same training, but are subject to different experience in practice, but keep learning. A case comes before the court and you get a 5-2 split. Same problem with the brainchip if it learns in practice, the answer you get depend on the history of the chip you have; another one can give a different answer.
So be careful what you mean when you say it "learns!"
I do think that the technology has great potential, particularly in pattern recognition areas like voice, writing, and facial recognition. I certainly think hardware implementation offers significant advantages over software implementation.
But it is still a generic tool that must be fitted to application and the device, and that is where the engineering problems will come, and time and testing be required.
I have read the announcements and having worked in the marketing division of a large international computer company for 15 years I understand how to project hype. The announcements read to me as marketing bumph, the devil is in the detail and a good marketer does not allow the customer to understand the full implications of the detail until after the order is signed. "Don't confuse selling with installing"
Be cautious, ask questions, and get informed.
That's me done, I just intended to sound a warning for the sake of balance. I'll keep my knowledge and experience to myself from here I, but I do watch progress with interest and will continue my own research.
(20min delay)