The actually useful question is not whether there are humans who, in their domains of expertise, can out-perform a general-purpose compiler.
The actually useful question is how many humans actually possess or can reasonably be expected to gain the knowledge to write screaming-fast bare-metal hand-tooled compiler-beating code safely on modern processor architectures, and figure out the trade-off if we decide to abandon optimizing compilers -- even though there will always be some nonzero number of humans who can write faster code -- in favor of either forcing every programmer to hand-roll performance-critical code or re-tooling everything and everyone everywhere to a hypothetical language (and probably a hypothetical hardware architecture) you don't even know how to describe in general terms yet.
(for the record, my money is on the "useless" compilers winning that cost-benefit analysis)
The actually useful question is how many humans actually possess or can reasonably be expected to gain the knowledge to write screaming-fast bare-metal hand-tooled compiler-beating code safely on modern processor architectures, and figure out the trade-off if we decide to abandon optimizing compilers -- even though there will always be some nonzero number of humans who can write faster code -- in favor of either forcing every programmer to hand-roll performance-critical code or re-tooling everything and everyone everywhere to a hypothetical language (and probably a hypothetical hardware architecture) you don't even know how to describe in general terms yet.
(for the record, my money is on the "useless" compilers winning that cost-benefit analysis)