Hopfield's 1982 associative memory is a system of Ising spins whose retrieval dynamics minimize an energy function borrowed from spin-glass physics

John Hopfield's 1982 paper announces its cross-domain move in its title: "Neural networks and physical systems with emergent collective computational abilities" (PNAS 79:2554 — Tier 1). The network is built from binary threshold units with symmetric pairwise connections — formally a system of Ising spins — and its recurrent dynamics descend an energy (Lyapunov) function to a local minimum. Stored memories are the minima; recall is relaxation to the nearest one. The computation is "emergent" and "collective" in the physicist's sense: no unit computes the answer, the settled configuration of the whole system does.

The lineage the physics framing points to is spin-glass statistical mechanics — the study of disordered magnets with competing interactions, whose energy landscapes have many metastable minima. Hopfield mapped that landscape onto memory: many stable configurations, each a basin an initial state falls into. The specific claim that the energy function was borrowed from the Sherrington–Kirkpatrick (1975) spin-glass model is held here as [unverified-mechanism] — it rests on a Tier-4 pointer, not a primary reading (see question-verify-hopfield-1982-energy-from-sherrington-kirkpatrick-spin-glass); the Ising-spin and energy-minimization framing above is what the PNAS paper itself supports.

This is the physics anchor of several vault threads. It is the object the 2024 Nobel Prize in Physics crowns; the architecture Amari proposed in substance a decade earlier and Hopfield did not cite; the energy surface systems biology later reused to formalize Waddington's epigenetic landscape; and — generalized to continuous states — the mechanism that turns out to equal transformer attention. A 1982 model of a magnet is the bedrock under all four.