Abstract: Classical machine learning typically focuses on extracting insights from static datasets, where the underlying distribution is independent of the learner's actions. This assumption fails in modern economic systems such as recommender systems and prediction/betting markets, where data is dynamic and being consistently updated and learned by multiple strategic agents. This necessitates a convergence of learning, game and market theories, connected via mathematical frameworks of optimization and dynamical systems. In this talk, we demonstrate how Adam Smith's "invisible hand" from 2.5 centuries ago can be concretized as a global optimization process, where classical market mechanisms such as tâtonnement correspond to gradient descent, while proportional response and multi-logit choice correspond to mirror descent. In the absence of a global optimization objective, the coupled behaviors of learners can be proven to be Lyapunov chaotic even in the simplest game settings via a volume expansion argument.