It is commonly acknowledged that "what is said" is not the same as "what is meant." In conversation, listeners must infer what speakers intend to communicate, drawing on not only the information in the linguistic signal, but also their own knowledge about the immediate context, the speaker, the world at large, and the conversational conventions of their community. These inferences lead to "pragmatic enrichment" of sentence meanings at a number of levels, ranging from the reference of expressions to presupposed or implicated propositions. How do children acquiring their first language manage to extract compositional semantic meanings from conversations built around the exchange of pragmatically-enriched meanings? What basic linguistic or social knowledge do they need to "break in" to more complex layers of communication? I will focus in particular on how children learn the verb 'think'. Even if 'think' were always used to describe salient, concurrent instances of believing, it would be difficult for children to learn, since you can't see/hear/feel someone believing something. However, pragmatic enrichment makes the problem much worse: the vast majority of adults' uses of 'think' occur in contexts where the belief itself is irrelevant to the conversation, as in, "I think I'm gonna go to the store," or, "Is it going to rain, do you think?". I will report several experiments investigating 3- and 4-year-olds' interpretation of 'think', and discuss the implications for children's semantic and pragmatic development.