I think there is much talk about sophisticated parallelization and using GPUs when that isn't necessary in most cases.
I think there is a good deal of task and data level parallelism (high-level/granularity) that can be used without being a software-guru.
I agree, that the low hanging fruits will not scale much >4 cores in most cases, but that is good for now.
Take that Game for example, why all the effort to parallelizes one aspect of it, when it is easy enough to utilize 4 cores by doing AI, physics, world-setup, miscellaneous. The stuff doesn't happen continuous, but in time-steps. Is the player going to notice, if the world-setup uses the physics data of the last iteration instead of the current, if the NPC reacts with one iteration/frame latency?
Likely this is being done today, in games, but that was just an example.
I just think that there is much talk about how difficult it is when there are simple ways for many problems. Sure there are some problems that are hard or not to be parallelized, and some here are working on such, but are they really that common?