I think basically it requires machines that work in fundamentally different ways that can't be reduced to statistics, and that we'll get better at approximating intelligence for a while until we hit a wall where no amount of scale makes any big difference.