With Metal and Mantle, it was clear that graphics programmers were wanting to use GPU's at lower levels of abstraction to more efficiently utilize the design of GPU's, which have a much different architecture than a decade ago. Without a corresponding option that is standards compliant, low-level API's are threatening to fragment GPU programming.
One operation I didn't see any mention of that seems to have some future role, albeit completely uncertain, is the Heterogeneous System Architecture Foundation spear-headed by AMD and having seemingly every chip-maker except Nvidia and Intel on-board.
As GPU stream processors get more CPU-like but are natural at parallel processing, it is a matter of time before we get the right abstractions so that map gets scheduled across many stream processors and is reduced by a higher-single-thread performance CPU and suddenly computer vision and many other naturally parallel data synthesis workloads are programmed in a less heterogeneous software environment and executed on chips where the stream processors and CPU's share a large amount of commonalities, possibly down to micro-op compatibility through something Nvidia could be pushing towards in their Denver architecture (as yet highly speculative).
To somewhat less than enthusiastic coverage, Nvidia has been building up their partnerships with automakers like crazy. Computer vision is obviously one of the applications that will be required in self-driving cars. Tango and other projects are also quiet beneficiaries of Nvidia tech maturing into the Tegra platform.
We have far from conquered programming and CPU design just because JIT's are good, 8GB of RAM is expected or GPU's can mine MHashes/s etc. We're in some future's bad-old-days. The idea that Khronos is involved in the unification of graphics and computing API's as well only makes the exciting question of who will drive our cars more intriguing.