John Hopfield's 1984 paper might be the starting point to everything we need to know to understand human organization. It is five pages long. 
The paper takes off of Hopfields's 1982 paper (two pages long), which folded a McCulloch-Pitts-type network unto itself, turning a feedforward network into a feedback network. The modification for the 1984 paper: graded response. Moving from digital to analog.
What does graded response mean? Sigmoid functions. We are moving from a binary panic response to a measured, gradual response. And yes, the connection to Tom Schelling's dying seminar paper on critical mass models is not a coincidence.
At a certain level of abstraction, the brain is a electrical circuit. Tbf, this is generally too much of an abstraction, but what we have here is a blueprint for an asynchronous peer-to-peer consensus mechanism. With all the features and bugs of such a mechanism.
Not going to post too much of the math, but this is an important one in that it introduces the "energy function": the starting point to a Monderer-Shapley potential that drives equilibrium convergence. This needs tweaking, but it already has the all-important "½" factor.
But how does equilibrium convergence look like? Something like this for a simple two-node scenario. 1. Convergence doesn't happen perpendicular to the contour lines: humans are not hillclimbers. 2. Yes, this is Nash equilibrium convergence in a pure coordination game.
And yes, this is an early application of replicator dynamics as an interaction model outside the evolutionary game theory context of Maynard Smith et al from the mid 1970s.
The paper also touches on, but doesn't carry thru, the notion of injecting quantal noise. Mutation.
The paper also touches on, but doesn't carry thru, the notion of injecting quantal noise. Mutation.
At this juncture we are at the jumping-off point to a variety of trajectories. Quantal noise leads us to (simulated) annealing, Boltzmann machines and eventually to quantum annealing. Mutation leads us to Kandori-Mailath-Robb and their "myopia, mutation, and learning" framework.
Coordination leads us to Elinor Ostrom's commons. Combinatorics leads us to Joan Robinson's complaint about the other Cambridge and their simplistic production function. It leads us to Edith Penrose, and to Ailsa Land and Alison Doig's branch and bound heuristic.
But if John Hopfield's network provides so many jumping-off points, why is it so obscure? Bc it's unpredictable, inefficient, and doesn't scale. And industrial use of machine learning favors efficient networks over architectural beauty. Backpropagation, deep learning: hierarchy.
But as deep learning is currently running into the same limit of hierarchical organization, we can take a step back and look at Hopfield's network to ask ourselves why, how and when organizational types: vertical vs horizontal, centralized vs decentralized, work. Or fail.
The answer lies in the difference between one-sided and two-sided Bayesian updating, in directed vs undirected graphical models, in directed acyclic graphs vs Markov random fields, in military discipline vs filter bubbles. In Shannon entropy.
But that's for another thread.
But that's for another thread.
I read Hopfield's paper in 1992, and it's become the jumping-off point for most of the things I did. That's where the 1992 in my profile comes from.
Five pages well worth reading, bc it might provide a jumping-off point to a theory of organization without a jarring disjuncture.
Five pages well worth reading, bc it might provide a jumping-off point to a theory of organization without a jarring disjuncture.
The "jarring disjuncture" is Milton Friedman's conjecture of the rationally optimizing plant; from a chat with @_julesh_ last year. It sent Chicago into a Kuhnian cycle where its attempts to square theory with the real world became ever more convoluted.
