One of the most remarkable discoveries in biology recently is evolvability, or the theory that evolution itself evolves. Selection for evolvability yields the so-called "structured" (as opposed to purely random) variation seen in nature, modular dissociation, standardized component architectures and many other biological features of interest to engineering. The most striking feature of organisms with high evolvability is how environmental variation is reflected in their genotypes. For example the bacteria E. coli are said to contain a "swiss-army knife" of latent genomic functionality that can be switched on to deal with extreme environmental conditions. In computational terms, organisms with high evolvability are capable of detecting and exploiting underlying patterns in the function being optimized. The evolution of evolvability is similar to learning procedural bias in Machine Learning, and in reinforcement learning it parallels value-function decomposition. Natural evolvability, however, includes feedback between the hypothesis representation and search operators, which allows the representation to restructure itself to yield better variation patterns in response to mutation. Direct representations, which are the most commonly employed in search, cannot provide such feedback; since all parameters are committed to representing an aspect of the candidate hypothesis, no parameter can be co-opted into storing meta-information about the search space. Thus, employing more complex representations such as developmental encodings is a promising research direction for solving difficult search problems.