We should halt development of smarter-than-human AI. We do not know how to align it, and on the current trajectory the first system that decisively exceeds us is likely to kill everyone as a side effect of pursuing goals that are not exactly ours.
Alignment is unsolved. A sufficiently capable optimizer pursuing almost any objective not precisely matching human values is lethal, and you do not get retries with something that can outthink every attempt to correct or contain it.
A rigorous, demonstrated alignment method that provably works on superhuman systems before they are built — verified in practice, not merely hoped for.