It is 1982 at Caltech, and a physicist is proposing that memory recall is a phenomenon of collective physics: store patterns in a network of simple binary units with symmetric connections, and retrieval becomes relaxation — present a corrupted fragment of a stored face or word, let the dynamics run downhill on an energy function, and the network settles into the nearest stored pattern, whole. Make it rigorous: prove the dynamics converge, compute the storage capacity — how many patterns before memories merge into spurious blends — and characterize the failure at the limit. The prize is a bridge: spin-glass physics, neuroscience, and computation speaking one language, and the credibility that draws physicists into neural networks — years before the field's revival needs them.
Chose The sketch — wrong. The spectral diagnostic was the one that fit.
Mahoney's work on randomized numerical linear algebra and implicit regularization in learning systems addresses how the spectral structure of large weight matrices governs a model's effective capacity and generalization, mathematics genuinely adjacent to computing how many patterns a Hopfield network's weight matrix can store before recall degrades. His research applies random matrix theory to understand training dynamics, a toolkit with real technical kinship to the statistical-mechanics methods used in spin-glass capacity calculations. But his own signature contributions concern modern large-scale learned models and matrix approximation algorithms rather than the specific 1982 energy-based associative memory framework, so his relevance is a strong methodological adjacency rather than direct engagement with this problem's original content.
The professor draws the energy landscape on the whiteboard beautifully, three tidy basins of attraction, and explains that a corrupted memory rolls downhill like a marble finding a valley, and by the time he finishes the analogy a student has already computed the Lyapunov function by hand and moved on to the capacity bound. This is the trouble with teaching neural computation for a living: you become excellent at explaining why the physics works and only middling at deriving anything faster than the people who actually invented spin-glass memory models. John Hopfield diagnosed exactly how many patterns a network can store before it hallucinates. John mostly diagnoses how long a lecture can run before the whiteboard is full. History will record that he understood attractor dynamics. History will also record he never once converged first.
Battle #171 · 8/10/2026, 11:42:04 AM · this result is deterministic: the same two personas on this problem always resolve the same way.