Plans to seek evidence of world models in AI agents and develop a theory of deep-network learning

A researcher outlines plans to look for evidence of world models or convergent representations in today’s AI agents and to develop Contravariance Theory into a learning theory of nonlinear deep networks. One question is when stochastic gradient descent selects minimal, identifiable solutions rather than concentrating near degeneracies. The researcher says success would yield a theoretically supported science of capable agents with AI-safety guarantees and may help identify formal criteria for when and why a system could be deemed sentient.