Learning and the mind
Chunking
Working memory holds only a few items, so experts get around the limit by grouping details into larger units they recognise at a glance.
George Miller, Psychological Review (1956); applied to chess by William Chase and Herbert Simon (1973); the limit revised to about four by Nelson Cowan (2001).
Miller's 1956 paper noticed that the number of items people can hold in immediate memory stays near seven whether the items are digits, letters or words, even though a word carries far more information than a digit. The unit is the chunk, a group the person already knows as one thing, and the way to hold more is to recode. He described a colleague named Smith who taught himself to read binary digits in groups and so stretched his span to about 40 digits. Chase and Simon took the idea to chess in 1973. A master shown a real position for five seconds reproduced about 16 pieces; a Class A player (a strong amateur) managed about eight and a beginner four. On random arrangements of the same pieces all three did worse than the beginner had on real positions.
The chess result is the mechanism. The master does not have a bigger memory; he has a larger vocabulary of familiar patterns (a castled king, a pawn chain, a pair of connected rooks) and sees a position as a handful of those rather than as 25 separate pieces. Chase and Simon reasoned that a master who holds 20 pieces with room for about five chunks must be using chunks of four or five pieces each. Random positions contain no familiar patterns, so the advantage disappears. The same story explains why an experienced programmer reads a function as three ideas where a beginner sees 40 lines, and why learning any field goes through a long phase of building vocabulary before anything feels easy. Expertise is largely a store of chunks.
Two corrections to the folklore. Miller said at the end of his paper that he suspected the recurring seven was only a "pernicious, Pythagorean coincidence", and Cowan argued in 2001 that the seven was a rough estimate and a rhetorical device; when people are stopped from recoding, the limit is closer to four. And chunks do not transfer: the master's advantage vanished on random boards, so skill in one domain says little about memory in another. The chess numbers also come from three subjects, one per level, and the master placed fewer pieces than de Groot's grandmasters had in the 1960s, which Chase and Simon put down to messier positions. The shape of the finding has held; the exact figures are from a small study.
Example
In Chase and Simon's 1973 experiments at Carnegie Mellon, reported in Cognitive Psychology, a master, a Class A player and a beginner each viewed positions from real games for five seconds and then rebuilt them on an empty board. On middle-game positions with 24 to 26 pieces, the master placed about 16 correctly on the first attempt, the Class A player about eight and the beginner about four. On random positions with the same number of pieces, every subject did worse than the beginner had on real ones. On a second set of nine positions from a book of master games, the beginner, the Class A player and the master scored 33, 49 and 81 percent correct.
Where it applies
When joining a codebase, the aim of the first weeks is to build chunks: learn the module names, the idioms and the three shapes a request takes, after which the code reads in paragraphs. A tokeniser is a recoding scheme in Miller's sense, turning characters into chunks so each slot of a model's context carries more. The memory game at /games/memory is a small exercise in the same thing.
Think about it
What in your work do you still read item by item that an expert would see as one pattern, and what would that pattern be called?
Applied it to something real? Log it.
Test yourselfWhy did the chess master lose his memory advantage on randomly arranged boards in Chase and Simon's 1973 study?
His advantage came from recognising familiar patterns as single chunks, not from a larger memory. Random boards contain no familiar patterns, so all three players fell to a few pieces each.
Related models
- Deliberate practiceExpert skill is built by structured, effortful practice aimed at specific weaknesses with feedback, not by hours spent doing the activity.
- The Feynman techniqueTest whether you understand something by trying to explain it in plain words to a beginner; where the explanation breaks, your understanding has a gap.
- Bottlenecks and the theory of constraintsA system's throughput is set by one constraint; improving anything else is wasted effort until that constraint is fixed.