2 ms·
Upon reflection, I can see the argument that to be strictly superior than a human (i.e. not just superhuman on an arbitrarily large subset of tasks that humans
by aidenn0 4y ago
Upon reflection, I can see the argument that to be strictly superior than a human (i.e. not just superhuman on an arbitrarily large subset of tasks that humans do), you need to be able to, in some way, emulate SILTBness, because the SILTBness is necessary for human-level performance on at least some tasks.
If this is the argument TFA is making then:
1. I don't agree with it
2. It was both not very well stated, and assumed to be true.
3. Even if we take this argument as true, an AI that can outperform humans on the "right" subset of tasks can be dangerous enough for thinks like alignment to be a concern.