Modern artificial neural networks are fundamental to technologies that recognize images, interpret language, and support automated decision-making. But how do these systems actually learn? This project investigates whether different training strategies, including traditional gradient-based optimization and algorithms inspired by evolutionary processes, enable networks to develop distinct forms of “reasoning” and “behavior.” Through carefully controlled experiments and interpretability techniques, we analyze the internal workings of these models to uncover what occurs inside the “black box.” Our primary objective is to determine whether and how the choice of training method impacts the performance and behavioral characteristics of neural networks. Insights from this work could inform the design of more dependable, transparent, and human values-aligned artificial intelligence systems.