There's a distinction between EDT and FDT. Evidential decision theory prefers that you take the action which maximize conditional expected utility, and is particularly susceptible to arguments on grounds like these, whereas FDT cares about, if you and similar entities ran a particular choice or way of thinking about choices, what decision function would yield the best outcome? Knowing if you have a specific differentiating factor, or whether you really did roll the extremely unlikely P(live) = (1/6)^n outcome, is the dominant factor in any kind of decision... (read more)
The main benefit, I think, of treating near-term incidents very seriously is because beginning to care about safety seriously makes caring about future AI level threats much easier. It's almost a necessary precondition, I think. So long as the stakeholders understand that, I think approaches like this can very much be the right track.