Existentially dangerous doesn't mean the benefits still don't outweigh the costs. If there's a 95% chance that uFAI kills us all, that's still a whopping 5% chance at unfathomably large amounts of utility. Technology still ends up having been a good idea after all.

Each level adds necessary nuance. Unfortunately, at each level is a new chance for unnecessary nuance. Strong epistemic rationality is the only thing that can shoulder the weight of the burdensome details.

Added: Your epistemic rationality is limited by your epistemology. There's a whole bunch of pretty and convincing mathematics that says Bayesian epistemology is the Way. We trust in Bayes because we trust in that math: the math shoulders the weight. A question, then. When is Bayesianism not the ideal epistemology? As humans the answer is 'limited resources'. But what if you had unlimited resources? At the limit, where doesn't Bayes hold?