Human intelligence relies heavily on symbols and the manipulation of symbols. However, most neural network-based approaches to AI do not heavily leverage symbolic systems. Santoro et al. proposes a reinterpretation of symbols as entities whose meaning is established by convention for those who participate in the convention. They consider how AI can develop symbolic behavior without exploiting the syntactic manipulation of discrete tokens (common in Good Old-Fashioned AI methods), including through social and cultural engagement (achieved through human-agent interaction).