computations would then translate almost directly into building actual physical structures out of atoms.

In the past biology—with all its details of DNA, proteins, ribosomes and so on—has provided our only example of programmable construction on an atomic scale. But the discoveries in this book suggest that there are vastly simpler systems that could also be used.

And indeed my guess is that the essential features of all sorts of intricate structures that are seen in living systems can actually be reproduced with remarkably simple rules—making it for example possible to use technology to repair or replace a whole new range of functions of biological tissues and organs.

But given some form of perhaps complex behavior, how can one find rules that will manage to generate it? The traditional engineering approach—if it works at all—will almost inevitably give rules that are in effect at least as complicated as the behavior one is trying to get.

At first biology seems to do better by repeatedly making random modifications to genetic programs, and then applying natural selection. But while this process does quite often yield programs with complex behavior, I argued earlier in this book that it does not usually manage to mold anything but fairly simple aspects of this behavior.

So what then can one do? Occasionally some kind of iterative or directed search may work. But in my experience there are so many different and unexpected things that can happen with simple programs that ultimately the only way to find what one wants is essentially just to do an exhaustive search of all possibilities.

And with computers as they are today one can already often look at trillions of cases—as on page 833. But while this is enough to see a tremendous range of behavior, there is no guarantee that one will in fact run across whatever specific features one is looking for.

Yet in a sense this is a familiar problem. For especially early in their history many branches of technology have ended up searching the natural world for ingredients or systems that serve particular purposes—whether for making light bulb filaments or drugs. And in some sense the only difference here is that in the abstract world of simple programs doing a search becomes much more systematic.