We can see the first (exclusive-or) result as a limit on learnability. The linear associator's architecture is restricted in
such a way that there are some things it just can never learn. Many such
results have been derived for individual learning systems, whether
connectionist or not: we're now beginning to see general theories of
learnability. The idea is that we can see learning as function-building: our system has to learn a certain mapping 
from
its inputs to its outputs. But it only has a certain set of primitive
functions available, those provided by its hardware and software. If the
function to be learnt can't be built up from these primitive
functions, it's logically impossible for the system to learn it.