More precisly, it tells how the relative speed difference between 2 algorithms change when n goes to toward infinity. An O(n) will win over O(n^2), but the latter case might have a huge constant factor that makes it slower than the former for e.g. n <= million.
And very often, n is smaller than a million.
Big O notation, or Landau notation, is not just O(g(n)). It's also o(), Ω(), ω() and Θ(). It's understood, though, that n is considered to usually be rather large - another name for this notation is, after all, "asymptotic notation." In any case that you're really worried about speed, you should probably be calculating the speed of your algorithm directly rather than using mere asymptotic generalizations.
Aside: I can't think of any algorithms with huge constants like you described; in theory, you're correct, but in practice, the asymptotic generalizations apply for n <= 1000 or even often 100.
Err, no it doesn't. Big O is a tool for saying how quickly an algorithm runs based on the size of n, and it's usually used to measure real things, like what will happen when the filesystem has to handle 100,000 files, then 1,000,000, etc.