The problem with Carmack's argument is the days of super-scaling up the performance in single-threaded CPUs is over. Unless we fundamentally shift to something like RSFQ or rod-logic, I doubt we're going to see 10Ghz mega fast single threads. (Quantum computing also requires a paradigm shift in the programming technique)
Throughput oriented computing is our future now, massive numbers of threads, perhaps some assymetry for specialized blocks. GPUs are a classic example of massive threading.
That means if you want maximum performance in the future, you face a paradigm shift in the way you program, and must work to parallelize your algorithms as much as possible. (And if another person tries to bring up Amdahl's law again, they need to be slapped, because a) most of the time the so-called 'inherently serial' algorithms being claimed do in fact have parallel versions and b) most of the game engines aren't being limited by the inherently serial sections, but by the fact that the programmers have not made enough effort to parallelize.
Carmack himself suggests the program isn't parallelization, it's just the extra work needed to do it. So sure, for a few more years, you may get away with avoidance of concurrency or just use very coarse grained parallelism, but 10 years from now, you're going to have to do the work anyway.
So you better start learning now. This is an area where Sweeney is conceptually ahead. Sweeney has seen the writing on the wall and is seeking to build new languages and tools to make the work easier, whereas Carmack is still resisting change.
