Here is one example of AI at work.
An experiment was done using evolutionary techniques to allow a computer to "design" a
simple circuit, using logic gates as building blocks. The design goals for
the circuit were simple, and the best design that engineers could come up
with used about 14 "gates", the smallest components that engineers use in
logic circuit design.
In the evolutionary technique, a computer manipulated a sequence that
served as the blueprint or DNA for the circuit, describing how the gates
were to be connected. The computer could then test the resulting circuit,
compare the results with the desired results, and rearrange the gates
using "genetic algorithms" on the blueprint sequence, and try again, over
and over. The computer was not programmed with any knowledge about how the
gates function, or given any "intelligence" other than the relativly
simple genetic algorithms (similar to gene crossover, etc).
The resulting circuit took only 12 gates. Engineers couldn't figure out
how it worked at first. What happened is that some of the gates ended up
being fed intermediary voltages, in a way they weren't designed to
operate, so they no longer operated like simple logic gates. So, the
resulting design was "simpler" in needing fewer components -- but vastly
more complex in understanding how it worked. No part of the design made
much sense by itself. The circuit had to be analyzed as a whole, using
much more complex techniques for circuit analysis than are needed for
traditional logic circuits.