Science is built on the scholarly consensus that shifts with time. This raises the question of how new and revolutionary ideas are evaluated and become accepted into the canon of science. Using two recently proposed metrics, we identify papers with high atypicality, which models how research draws upon novel combinations of prior research, and evaluate disruption, which captures the degree to which a study creates a new direction by eclipsing its intellectual forebears. Atypical papers are nearly two times more likely to disrupt science than conventional papers, but this is a slow process taking ten years or longer for disruption scores to converge. We provide the first computational model reformulating atypicality as the distance across latent knowledge spaces learned by neural networks. The evolution of this knowledge space characterizes how yesterday's novelty forms today's scientific conventions, which condition the novelty--and surprise--of tomorrow's breakthroughs.