Both drop to roughly the same low level. On random boards, masters reconstruct barely more than novices. The masters’ advantage on real positions wasn’t a bigger buffer or trained visual memory — it was chunking. A real midgame decomposes, for a master, into a handful of familiar patterns (“fianchettoed bishop,” “castled kingside with pawn storm coming”): ~25 pieces, maybe 5 chunks, comfortably inside the budget. Scramble the pieces and the patterns vanish; the master is back to memorizing individual pieces with the same ~4-chunk buffer as everyone else.
Chase & Simon’s estimate: mastery ≈ tens of thousands of stored patterns, built over years. This reframes what it means to be good at something: expertise is a vocabulary, not a bigger CPU. The same result replicates in bridge, electronics, programming — show developers real code for seconds and seniors recall far more than juniors; shuffle the lines and the gap nearly closes.
Three consequences worth keeping:
- The bottleneck shapes the architecture. A tiny working memory forced human cognition into strategies — externalize (write things down), chunk (compress via long-term knowledge), and serialize (one hard thing at a time). Cognitive architectures like ACT-R and SOAR are largely theories of how a system with exactly this bottleneck still gets things done. When those systems arrive two lessons from now, working memory will be a named component in them.
- The agent-harness parallel is almost embarrassingly direct. A context window is a working memory: fixed capacity in tokens, unbounded in what a token can point to. Summarization, retrieval handles, and structured notes are chunking — swapping raw items for compact pointers into a larger store. And the chess result carries a warning: an agent’s “expertise” lives in whether its stored patterns match the situation. Off-distribution, your master is a novice.
- The pitfall: “7±2” as a design law (“menus must have seven items!”) misreads Miller — his paper was partly a joke about the number’s ubiquity. The load-bearing findings are: the buffer is small, its unit is the chunk, and chunks are made of knowledge.
Where this goes: next lesson builds the first machine with an explicit working memory — the production system, Newell & Simon’s rule engine, which is the atom every later architecture in this track is built from.