We consider a population of agents that can choose between two risky technologies: an old one for which they know the expected outcome, and a new one for which they have only a prior. We confront different environments. In the benchmark case agents are isolated and can perform costly experiments to infer the quality of the new technology. In the other cases agents are settled in a network and can observe the outcomes of neighbors. We analyze long–run efficiency of the models. We observe that in expectations the quality of the new technology may be overestimated when there is a network spread of information. This is due to a herding behavior that is efficient only when the new technology is really better than the old one. We also observe that between different network structures there is not a clear dominance.