We study a dynamic learning model in which heterogeneously connected Bayesian players choose between two activities: learning from one's own experience (work) or learning from the experience of others (search). Players who work produce an inflow of information which is local and dispersed around the society. Players who search, instead, aggregate the information produced by others and facilitate its diffusion, thereby transforming what inherently is a private good into information that everyone can access more easily. The structure of social connections affects the interaction between equilibrium information production and its social diffusion in ways that are complex and subtle. We show that increasing the connectivity of the society can lead to a strict decrease in the quality of social information. We link these inefficiencies to frictions in peer-to-peer communications. Moreover, we find that the socially optimal allocation of learning activities can differ dramatically from the equilibrium one. Under certain conditions, the planner would flip the equilibrium allocation, forcing highly connected players to work, and moderately connected ones to search. We conclude with an application that studies how resilient a society is to external manipulation of public opinion through changes in the meeting technology.