Adiabatic Quantum Computing: why not just set the system in its problem Hamiltonian $H_{P}$ immediately?

Most NP-complete problems can be formulated as finding the ground state of some Hamiltonian. If you create a physical system that has such a Hamiltonian, it will be a "frustrated system". It will settle into a state that is a local minimum of the energy, and while quantum mechanics says that it will eventually decay into the ground state (assuming that it's not isolated; that is, there is some mechanism for it to lose energy), the time this takes may easily be many orders of magnitude larger than the lifetime of the universe.

I once asked the exact same question during a course on quantum computation. Systems only "fall into" their ground states when they are in thermal equilibrium at zero temperature. Both of these pieces are problematic: (a) many systems that have been proposed for quantum computation have energy scales low enough that getting them down to sufficiently low temperatures is extremely challenging, and (b) as Peter Shor pointed out, you have no idea how long it will take for the system to actually reach thermal equilibrium - you could have a physical equivalent of a Monte-Carlo sign problem, where it takes exponentially long in system size for local perturbations to get you into thermal equilibrium.

But if you can control the initial Hamiltonian $H_0$, you can "force" the system into its ground state much more quickly - in principle by measurement filtration, but more realistically by making $H_0$ unfrustrated and with a very large characteristic energy scale. For example, if you have a system of quantum spins and you apply a huge uniform field to the whole system ("huge" meaning much larger than the temperature and the relevant spin interaction scale), then the whole system will align with the field very quickly and you can be confident that you're in the ground state.