How AI founders on adversarial landscapes of fog and friction

Abstract

Formal analysis, based on the asymptotic limit theorems of control and information theories, uncovers sufficient conditions for punctuated failure across the full spectrum of real-time cognitive process – essentially a generalization of the Yerkes-Dodson law – challenging recent assertions that instabilities in AI deep learning paradigms can be easily remedied, permitting their use in real-world critical systems. A temperature analog for cognition that is itself an order parameter is determined by rates of internal information transmission, sensory or intelligence input, and material resource availability. Phase transitions driven by the synergisms of such parameters express symmetry-breaking changes in groupoids characteristic of cognition at and across scales and levels of organization, significantly extending models abducted from physical theory. No modifications of current – or future – AI or other cognitive systems can or will be immune to failure when facing sophisticated adversarial challenge under conditions of friction and the fog-of-war. We indicate how to reconfigure these results for study of long-term conflict on ‘Sun Zu Landscapes’ of deception, deceit, and subtle influence under Lamarckian selection.