Best architecture for "AI"
I still think the term "deep machine learning" is more accurate than the marketurds' "AI" twaddle, but, that aside, it's been my impression that neither traditional CPU architecture nor GPU is particularly suited to "AI" purposes. The FPGA approach seems like option for a better philosphical and logical choice.
Of course we can argue that anything capable of implementing a Turing complete system, if powerful and fast enough, could "do AI", since it's the code that matters. Except, in an "AI", the algorithms that ultimately get executed were likely not the ones originally written. There is flexibility and fluidity right down to the bits. So it seems reasonable that a machine designed for deep learning, which means discovering and executing new rules and algorithms on the fly, might benefit from being able to redefine the very computing substrate it runs on.