Chunking (Miller)

In the AI Story

The chess experiments of Adriaan de Groot in the 1940s and William Chase and Herbert Simon in the 1970s provided the empirical foundation for treating chunking as the core mechanism of expertise. Masters presented with meaningful chess positions could recall them with near-perfect accuracy after five seconds of exposure. Presented with random arrangements of the same pieces, masters performed no better than novices. The difference was not general memory capacity but the availability of chunks appropriate to the material. Meaningful positions could be compressed into structural patterns. Random positions could not.

Every cognitive technology in human history can be understood as a chunking infrastructure. Writing chunked memory, allowing information to be externalized and later retrieved. Mathematical notation chunked quantitative reasoning, allowing Newton to hold in a single equation relationships that would have consumed paragraphs of verbal description. Programming languages chunked computational instruction. Large language models chunk entire software systems into conversational exchanges. Each revolution felt like an expansion of capacity but was actually a compression of demand on a capacity that never changed.

The distinction between a genuine chunk and a mere label is one of Miller's most underappreciated contributions. A genuine chunk preserves the structure of the compressed information — the chess master's chunk for a Sicilian Defense contains pawn structures, tactical motifs, strategic objectives, all decomposable on demand. A label, by contrast, discards structure. The novice who has been told 'that's a Sicilian Defense' holds the name without the contents. Both occupy one slot. Only one carries its own explanation within it. The difference is invisible under routine conditions and catastrophic under novel ones.

Chunking is built through the slow, error-driven process Miller called recoding — the effortful transformation of unfamiliar information into familiar patterns. Recoding cannot be skipped or downloaded. It requires exposure, repetition, failure, and correction. The chess master's chunks were built across thousands of games. The experienced programmer's mental library of design patterns was assembled across thousands of debugging sessions. When AI tools generate pre-chunked solutions, the question Miller's framework forces is whether the chunks arrive with or without the structural knowledge that recoding would have built.

Origin

Miller introduced the term 'chunk' in his 1956 paper as a neutral term for whatever unit of information occupied a slot in working memory. He deliberately chose a casual word to signal that the unit was defined functionally rather than substantively: a chunk was whatever the mind treated as a single retrievable item, regardless of how much raw information that item compressed.

The concept gained theoretical weight through Miller's subsequent collaborations, particularly his work with Eugene Galanter and Karl Pribram on Plans and the Structure of Behavior (1960), which located chunking within a broader hierarchical architecture of human action. Herbert Simon's later work on expertise extended chunking into a general theory of skilled performance across domains.

Key Ideas

Compression without capacity expansion. Chunking allows the mind to hold more information without holding more items. Seven slots can contain seven digits or seven entire theories, depending on the quality of compression.

Built through recoding, not received. Genuine chunks are constructed through effortful engagement with the material. They cannot be handed to a learner in finished form without losing the structural knowledge that makes them robust.

Earned versus borrowed. A chunk built through recoding preserves its internal structure and can be decomposed when necessary. A chunk received passively — a label — occupies the same slot but cannot be decompressed. The difference becomes visible only when conditions depart from the routine.

The substrate of expertise. Expertise is not more cognitive capacity but better chunking. The expert and the novice operate with the same seven slots. The expert's slots contain more compressed meaning.

Contextual quality. The value of a chunk depends on the match between its structure and the demands of the situation. A chunk that is high-quality for one task may be low-quality for another. Quality is relational, not inherent.

Debates & Critiques

The question of whether AI-generated outputs can produce genuine chunks or only labels is contested and consequential. Proponents argue that working with AI tools involves its own effortful engagement — specifying requirements, evaluating outputs, iterating on designs — and that this engagement builds chunks appropriate to the new practice. Critics, drawing on Miller's framework, argue that the chunks built through AI-mediated work are structurally thinner than those built through manual implementation, because the error signals that drive recoding are largely absent. The empirical resolution will require longitudinal studies of practitioners whose careers began in the AI era, and those studies are only now beginning.