When a program decides at runtime that it wants to "augment" itself with additional code, it will issue the dlopen() function call. This function call tells the system that it should find the shared object referenced by the dlopen() function and create a binding between the program and the shared object. Again, if the shared object isn't present in memory already, the system will load it. The main advantage of this approach is that the program can determine, at runtime, which objects it needs to have access to.
Note that there's no real difference between a library of shared objects that you link against and a library of shared objects that you load at runtime. Both modules are of the exact same format. The only difference is in how they get used.
By convention, therefore, we place libraries that you link against (whether statically or dynamically) into the lib directory, and shared objects that you load at runtime into the lib/dll (for "dynamically loaded libraries") directory.
Note that this is just a convention — there's nothing stopping you from linking against a shared object in the lib/dll directory or from using the dlopen() function call on a shared object in the lib directory.