Young children learn the meanings of thousands of words by the time they can run down the street. Many efforts to explain this rapid development begin by assuming that the computational-level problem being solved is acquisition. Consequently, work in this line has sought to understand how children infer the meanings of words from cues in the communicative signals of the speakers around them. I will argue, however, that this formulation of the problem is backwards: the computational problem is communication, and language acquisition provides cues about how to communicate successfully. Under this framing, the natural unit of analysis is not the child, but the parent-child dyad. A necessary consequence of this shift is the realization that the statistical structure of the input to the child is itself dependent on the child. This dependency radically simplifies the computational problem of learning and using language.