To learn Morse code, you begin by hearing letters as combinations of dits and dahs, short and long tones. A is di-dah, Z is dah-dah-di-dit, and so forth.
But to go faster than a few words per minute you've gotta stop consciously listening to individual units of sound and start hearing each letter as a single entity. And then, to move on to real fluency, you must give up separate letters and start perceiving larger units of meaning: THE, YOU, AND, the -TION ending, etc.
That progression to bigger and bigger chunks is obvious in a learned skill like the radiotelegraph code, or in constructed artifacts such as hotels (repeated rooms, made of repeated bricks, ...), electronic circuits (boards, chips, devices, ...), and computer programs (subroutines, functions, blocks, ...). Hierarchical chunking is equally vital but far less visible in a host of other contexts: living biological systems, language, art, mathematics, thought, ...
It's about patterns and pattern recognition. The human brain (most any wetware processing unit, in fact) is good at that; current computers not so much.
The reason evolution favors pattern recognition is likely that it's efficient. It in any case leverages the way other natural processes self-aggregate structure from simpler building blocks. -- Bo Leuf