Great talk! The point that the system uses all the qubits no matter the problem size clarifies why one expects the best performance for the largest problem sizes. It brings to mind Gustafson's law for parallel computing.

I built a discrete simulated annealer and watched it walk right out of a solution because the parameters were less than optimal. Catherine's pointed out parameter tuning is a big part of simlated annealing. However one should include a ridiculous amount of tuning time as part of the simulation time or create an AI turner for simulated annealing as the process is useless without tuning. The fact a quantum annealer does not really need tuning is a huge plus.