A memorable line from a talk by Ewa Dabrowska that I went to recently:
"It is generally agreed that the representations assumed by generative theories cannot be learned from the input. For generative linguists, this fact is a fundamental premise of arguments for the innateness of at least some aspects of these representations: since they cannot have been learned from the input, they must be available a priori. An alternative conclusion, of course, is that we need a better theory - one that does not assume representations that are unlearnable."
Her answer is construction grammar: kids first learn individual low-level constructions like "What's ___ doing?" as unanalysed units, and only later come up with higher-level schemas of which these constructions are special cases (the next stage in this case would be "What's ___ ___ing?") Judging from the evidence she presented, showing that the vast majority of a 3 year old's utterances could be accounted for solely on the basis of simple substitutions within sentences they are known to have already heard, "children's [linguistic] creativity seems to involve superimposing and juxtaposing memorised chunks." This view of language more or less inverts the usual grammarian's perspective: the most general rules are developed only after specific cases have been learned, and the specific cases presumably continue to be stored independently. It strikes me as a rather promising way of thinking about historical syntax.