Scott linked an article from new scientist, Memories may be stored on your DNA, which relates with what has been on my mind lately.
I was reading about some hardware evolution experiments while at the bookstore the other day. They were using a FPGA to create (evolve) a sine wave generator, if I recall correctly. The end result of the first run worked perfectly, but had some bizzare attributes -- it programmed a section of the gate array but didn't wire that section to the rest. When the unconnected section was turned off, the circut no longer worked. It also didn't work when the same design was put onto an identical FPGA. As far as the experimenters could tell, the evolved solution depended upon some subtle property of that specific FPGA chip.
In other words: GP solutions cheat like hell.
They will use absolutely any available subtle or side-effect aspects of environmental conditions to push themselves to the next rung on the fitness ladder. An important lesson that I've taken from GP is that artificial / automatic program generation mostly results in programs that make almost no sense from a human programmer perspective, at least partly because of all this cheating. The solutions produced by GP are mostly twisted convoluted things which mostly don't look like they work at all, let alone solve anything.
When a human abstracts things, it ends up taking a much more layered structure, with the meat of the problem usually ending up on the highest layer. Automatic techniques do not succumb to such niceties.
Speaking of human abstraction, I recently stumbled upon the Discordia page of Mark-Jason Dominus which makes a keen observation with fascinating attributes - "In a well-designed computer program, No two components are alike, or even similar." Conceptually, that is, all similar aspects of a program are abstracted. Good programs are themselves not like the beautiful fractals of recursion that lay in the mind of their creator or the algorithms that they represent. You could perhaps think of them as the compressed version -- the most informationally dense form of the logic, information theory might say.
Now take it full circle... from the biological inspired computation back to biology. Biological evolution cheats just like artificial evolution does. At first glance we see all this lovely DNA-level abstraction... but as they dig in they find all sorts of trickery that operates on all levels low and high. Like throwing in some assembly (or even better, some binary executable blobs) into the middle of my blog software. The re-use of DNA (or perhaps the coating on DNA was it?) as memory storage in the brain is completely typical.
If the designers were indeed intelligent, they sure didn't plan on doing any maintenance.