Its not just for the AI field, if you read their About Us page you will see that a very interesting application has to do with UAV and algorithms for motion det

Its not just for the AI field, if you read their About Us page you will see that a very interesting application has to do with UAV and algorithms for motion detection. The chip set design space for this is very low energy but efficient/accurate at a certain trained task, which is what makes bio-inspired applications so compelling. But going back to AI, I would say this: Short term: a wide variety of demonstrators of various 'easy' problems, eg image and simple logic function pattern matching, in chip as opposed to in theory or software like you say. This can be done with dozens or a few hundred memristive devices. Medium term: more interesting (my maybe not trying NP hard problems yet) application in image, sound processing and other non-linear signal processing that begin to rival best approaches coming from dozens of GPU cores. Such applications might require tens of thousands or even hundreds of thousands of devices. Pattern matching and classifications at this stage might rival that of a very simple brain or a small component of the neo-cortex. Longer term: Unknown, but some possibilities when you use tens of millions of such devices may include full (neo)-cortex implementation in hardware, remember that an entirely new generation of non-linear algorithms that implement due to topological and temporal effects at the nano-scale may allow for better results than what we get given an equivalent number of transistors (yes, transistors are also nano-devices, but despite a small community working on sub-threshold switching, they are being used in an almost uniformly binary approach).