The problem is that while quantum superposition can be thought of as "all possible states simultaneously," it is not in general possible to choose properties of the state you get at the end (called "postselection" in the quantum computing literature). All you can do is adjust the probabilities of the different states, & for some problems, we do not know a way to make the desired state likely enough to do any better than a square root improvement in running time over a classical computer. So for that sort of problem, an O(2^n) algorithm would become O(2^(n/2)).
Apart from quantum suicide (which depends on various unproven assumptions & is technically nontrivial even if theoretically possible), there is no known way to do postselection.
"The pyramid is opening!" "Which one?" "The one with the ever-widening hole in it!" -- The Firesign Theatre