Halliday & Matthiessen (1999: 555):
The significance of global probabilities in the grammar emerges in various ways, both historical and synchronic. Historical change in language is typically a quantitative process, in which probabilities in systems at every level are gradually nudged in one direction or another, now and again becoming categorical so that some systemic upheaval takes place. Each instantiation of a tense form, say, whenever someone is speaking or writing in English, minutely perturbs the probabilities of the system — because what we call "system" and "instance" are one and the same phenomenon, being observed from different depths in time.
There are, of course, more catastrophic types of change: languages become creolised, creolised systems in turn become decreolised, or a language ceases to be spoken altogether. At the "instance" end, a single highly-valued instance may exert a disproportionate effect: quotations from the Bible and from Shakespeare are familiar triggers of this "Hamlet factor" in English (in media discourse today almost every change is a sea change, which goes into our folk taxonomy of types of change).
But such qualitative effects take place against a background of microscopic quantitative pressures, the sort of nanosemiotic processes by which a language is ongoingly restructured as potential out of the innumerable instantial encounters of daily life — the "sheer weight of numbers", as we sometimes call it.
And in the ontogenetic dimension of history, the growth and development of language in a human child, an analogous dialectic can be observed: highly valued instances of text (rhymes, favourite stories and the like) interact with the quantitative pressures of the talk going on around (the child's access to the global probabilities — note that children do not begin learning the grammar by sorting out functional variation!) to yield a meaning potential that is a reasonable copy of the probabilistic system shared among the community.