Most machine-learning models rely on mountains of data to replicate human text, but new research suggests the recipe for learning language might be simpler