This paper reports the empirical discovery and formal characterization of the Epileptiform Synchrony Limit (ESL): a universal failure mode of recurrent Spiking Neural Networks (SNNs) in which scaling contextual memory (membrane time constant τₘ) without maintaining strict Excitatory/Inhibitory (E/I) balance causes irreversible collapse of sparse semantic representations into a ~500 Hz synchronous firing state. Using the PyNN/Brian2 framework on the DDIN phonosemantic architecture, I executed three systematic experiments: a τₘ sweep, an 80/20 E/I lateral inhibition architecture, and a global Winner-Take-All (WTA) inhibitory circuit. In all high-τₘ configurations, the network entered a seizure state destroying all semantic clustering, regardless of inhibitory weight magnitude. I prove formally that this behavior is not a tuning failure but a fundamental physical constraint: when continuous Poisson input drive exceeds the time-averaged inhibitory conductance, the membrane voltage integrates without bound until the threshold is saturated at every timestep. No static weight-based inhibitory circuit can prevent this collapse under continuous input. The condition for stable asynchronous firing requires dynamic E/I equilibrium, requiring either (i) Hodgkin-Huxley adaptation currents (M-current, AHP) for single-neuron homeostasis, or (ii) thermodynamic analog noise from BrainScaleS hardware as stochastic regularization. I further prove that the ESL is not pathological but constitutive: the postulate "Intelligence requires The Void (Sparsity) " holds at the biophysical level. Continuous, unspaced spiking activity mathematically precludes the formation of sparse distributed representations that carry semantic structure. These findings establish the ESL as a formal boundary condition for neuromorphic AGI design, and identify its resolution as the primary remaining prerequisite for physical EBRAINS hardware deployment.
Amit Kumar (Thu,) studied this question.