+On the other hand, consider a *top-level* definition
+ f = (\x -> x) :: T a -> T a
+If we don't abstract over 'a' it'll get fixed to GHC.Prim.Any, and then
+if we have an application like (f "x") we get a confusing error message
+involving Any. So the conclusion is this: when generalising
+ - at top level use tyVarsOfType
+ - in nested bindings use exactTyVarsOfType
+See Trac #1813 for example.
+