Expert Mental Representations

In The You On AI Field Guide

The concept originates in the Chase-Simon 1973 chess studies, which demonstrated that chess masters' superior memory for positions disappeared when the positions were randomized — revealing that the masters were not recalling visual data but meaning-laden patterns. Ericsson generalized the finding across domains and deepened it into a theory of how such patterns are constructed: through the sustained, boundary-testing engagement that deliberate practice provides.

Representations have a property that makes them invisible in routine output and visible only in specific diagnostic moments. A senior software architect can feel a codebase the way a physician feels a pulse — not through analysis but through an embodied intuition deposited layer by layer through thousands of hours of patient work. This geological understanding operates below the threshold of conscious articulation, producing evaluations experienced as feeling but actually the product of pattern-matching processes too complex for explicit reasoning.

The mechanism of construction is the friction requirement: representations grow only through encounters with problems the existing model cannot handle, which force the model to adapt. Each debugging session that revealed a gap, each surgical complication that defied the resident's understanding, each passage the musician could not yet phrase — these moments of productive failure are the developmental currency. When AI handles the difficulty, the currency stops being minted. The output continues. The architecture does not grow.

For practitioners who built deep representations through pre-AI deliberate practice, AI tools function as amplifiers: the architecture exists, the tool extends its reach. For practitioners entering the field in the AI era without undergoing the pre-AI developmental process, the situation is fundamentally different: they possess tool-assisted production capacity without the evaluative substrate that deep representations provide. The two classes are indistinguishable by output metrics and radically different by understanding metrics — a distinction invisible until the moment the tool fails.

Origin

The construct was formalized in Ericsson's work on long-term working memory with Walter Kintsch in the 1990s, which demonstrated that experts overcome normal working memory limits by storing information in domain-specific retrieval structures. This explained how chess masters can play simultaneous exhibitions blindfolded, how waiters can remember twenty orders without writing them down, and how medical experts can integrate complex patient histories during diagnosis — all by encoding information in representations structured by domain-specific meaning.

Key Ideas

Meaning-structured, not data-structured. Representations encode relationships, implications, and responses — not surface features.

Constructed through struggle. The representations form only through the effortful engagement with problems that exceed current capability.

Invisible in routine output. Representations manifest only in diagnostic moments — when tools fail, situations are novel, or independent judgment is required.

Substrate for iudicium. The capacity to evaluate AI output depends on representations that AI-assisted production does not build.

Decay without maintenance. Like physical capacity, representations require continued engagement at the level that built them, or they gradually atrophy.