Jim always says something that changes the way you think, despite the fact that he insists he knows nothing at all. In this case, he pointed out that ARM can deal with non-stationarity induced by climate change in at least two ways:
1) if we're smart enough to develop models of the trajectory (e.g. of rainfall or something else) then add a state variable to make it stationary in the expanded state space
2) if we're not smart enough, optimize over short time horizons with the expectation that you will revisit it frequently, and be careful about the state that you leave the system in.
Option 2 is the new one, that is both computationally manageable and perhaps palatable.