Suppose that, after some hard work, EY or someone else proves that a provably-friendly AGI is impossible (in principle, or due to it being many orders of magnitude harder than what can reasonably be achieved, or because a spurious UFAI is created along the way with near certainty, or for some other reason).
What would be a reasonable backup plan?
Try really hard to get reasonably safe oracle AI? Focus on human uploading first?
If it's worth saying, but not worth its own post, even in Discussion, it goes here.