In my view, for a machine to learn something to an Ura level the knowledge would have to be learned to a depth that it reflects upon the machine itself. So thousands of hours of Karate leads to permanent musculature changes, thousands of hours of experience in a mental field leads to a re-wiring and growth of the brain to accommodate that task.
Most machines use static logical rules and operations. That is IF, OR, AND, THEN etc. A simple machine like a winding watch uses the following rule. IF there is sufficient compression in the spring THEN turn the dials on the clock face in accordance with the timing set by the gears. The watch does not understand that IF, THEN operation to a deeper, Ura, level after thousands, or hundreds of thousands of hours, of turning.
All of our mechanical and digital machines utilize some combination of the fundamental logical rules to operate. The rather clever AI that we now take for granted to suggest a route in Google Maps, which uses complex learning algorithms that constantly improve their output with greater amounts of data (including from your usage), are still made up of basic logic (and really fancy maths and probability, which I am in no way downplaying). But fundamentally it is an optimisation.
That is, even AI which utilizes complex techniques such as deep learning which will develop the output algorithm based on large amounts of data has a set bound within which the algorithm can train. Google Maps might output optimised routes based on its training, however its type of output is not moulded by the task that it is giving.