This looks... incredibly interesting. Having a multitude of neural networks as agents competing with each other to maximize a general objective for the entire "

This looks... incredibly interesting. Having a multitude of neural networks as agents competing with each other to maximize a general objective for the entire "society of agents" seems very appealing to me as a mechanism for scaling computation. Call it Minsky's pyrrhic victory, if you will.[a] One obvious next step is to partition both network and resources within the this "society of agents," so that no agent competes with all other agents and no agent competes for all resources, creating sparsity and small-world dynamics at runtime. The authors are probably already considering this. Looking forward to digging in :-) -- [a] Minsky, who developed the "Society of Mind" theory, is often credited as having caused a decline in neural net research and a shift to symbolic AI research in the 1970s and early 1980s, after the publication of his book "Perceptrons: an introduction to computational geometry," with Seymour Paper. I find it a bit ironic that the "society of mind" in this new paper consists of descendants of the perceptron. See https://en.wikipedia.org/wiki/Perceptrons_(book)#Reception_and_legacy https://en.wikipedia.org/wiki/Perceptrons_(book)#Reception_a... for details.