Hierarchy of Compression
In 1960, Miller co-authored Plans and the Structure of Behavior with Eugene Galanter and Karl Pribram, proposing that all human behavior is organized by hierarchical structures of test-operate-test-exit units. Before acting, the mind tests the current state against a desired state. If mismatched, it operates, then tests again. A plan is a nested structure of these units, arranged from abstract goal to concrete action. The connection to chunking was immediate: a chunk is a compressed plan, a sequence of test-operate cycles executed so often that the entire sequence collapses into a single unit. The experienced driver does not consciously step through a turn; the entire maneuver occupies one slot. The hierarchy of compression is the central organizing principle of human cognition. At the bottom sit raw operations — individual lines of code, brush strokes, words. At the top sit abstract plans — strategic intentions, architectural visions, meanings organizing entire systems. Between them lie layers of chunking, each level compressing the complexity below so the level above can manipulate it.
In the AI Story
Programming, understood through the hierarchy, is distributed across multiple compression levels simultaneously. The developer holds a plan at the top: build a system that does X. The plan decomposes into architectural components: database, business logic, presentation. Each component decomposes into modules, functions, and finally individual operations. The developer's cognitive challenge is maintaining coherence across levels — ensuring low-level operations serve mid-level modules, which serve high-level architecture, which serves the original plan. The hierarchy may contain thousands of elements across a dozen levels. Seven slots must manage all of it.
The management strategy is selective decompression. The experienced developer holds a compressed representation of the entire hierarchy — a small number of high-level chunks summarizing essential structure while omitting details. When she needs to work at a lower level, she decompresses the relevant chunk, expanding it into sub-chunks that temporarily occupy working memory slots. When she finishes, she recompresses, collapsing details back into summary and freeing slots for navigation to a different part of the hierarchy. This process of selective decompression and recompression is the cognitive essence of not only software development but every complex creative activity.
The gradient from mechanical to judgmental work runs through the hierarchy. At the lowest levels, work is primarily mechanical — apply a known technique to a specific situation. At the highest levels, work is primarily judgmental — decide what should exist, why, and what trade-offs are acceptable. AI tools do not compress the hierarchy uniformly. They compress from the bottom up, eliminating mechanical levels while leaving judgmental levels untouched. The developer's cognitive activity shifts upward, concentrating on levels where judgment dominates.
This shift creates both liberation and destabilization. The developer freed from managing lower levels has more cognitive resources for upper levels — more attention to architecture, ethics, user experience. But she may lose the ability to selectively decompress into lower levels when necessary. Decompression requires familiarity with those levels, a chunking vocabulary built through recoding. If the AI has always handled the lower levels, the compressed representation she holds is not a chunk she built. It is a label she was given. And labels, unlike chunks, cannot be decompressed.
Origin
The hierarchical structure of action was a central concern of the first generation of cognitive scientists, who were looking for alternatives to the stimulus-response sequences of behaviorism. Plans and the Structure of Behavior drew on contemporary work in control theory, artificial intelligence (particularly Newell and Simon's Logic Theorist), and neurophysiology (Pribram's research on frontal lobe function).
The TOTE framework was meant to replace the stimulus-response reflex as the basic unit of behavioral analysis. A TOTE unit embedded feedback, goal-directedness, and hierarchical organization into a single mechanism. The proposal was bold for 1960 and has since been refined but not displaced; contemporary theories of motor planning and cognitive control continue to operate within broadly similar architectures.
Key Ideas
Nested TOTE units. Plans are hierarchies of test-operate-test-exit units, each testable and operable at its own level, each embeddable in higher-level units.
Compression from the bottom up. Chunks form at lower levels first and provide raw material for higher-level chunks. The hierarchy is built cumulatively.
Selective decompression. Expertise involves moving fluidly between levels — holding a compressed high-level chunk, decompressing it when necessary, then recompressing to navigate elsewhere.
The gradient from mechanical to judgmental. Lower levels are more constrained and mechanical; higher levels involve more open-ended judgment. AI compresses the mechanical levels first.
The risk of labels replacing chunks. When compression eliminates the experience of operating at lower levels, the user may hold labels that cannot be decompressed when the tool fails.
Appears in the Orange Pill Cycle
Further reading
- George A. Miller, Eugene Galanter, and Karl H. Pribram, Plans and the Structure of Behavior, Holt, Rinehart and Winston, 1960
- Allen Newell and Herbert A. Simon, Human Problem Solving, Prentice-Hall, 1972
- David Marr, Vision, W. H. Freeman, 1982
- Andy Clark, Being There: Putting Brain, Body, and World Together Again, MIT Press, 1997