Through the history of deep learning, the frontier has been networks that can be trained to some feedback of whether they're working or not in ~6-8 days. Within a few years there will almost certainly be ASICs that allow us to train hundreds of thousands or millions of networks in several weeks, relative to several weeks for one of the same network in 2017. This will let us start exploring things like evolving networks seriously.
Fundamentally, we know neural networks can instantiate general intelligence, and we know genetic search is capable of finding the right neural networks. There are big differences between the CS and biological versions of each, but it's striking that the big breakthrough in "AI" was deep neural networks and not anything else.
When I think about the difference between AlphaZero and human intelligence, I don't think it's "more intelligence." AlphaZero seems perfectly intelligent to me: I think the difference is more about the selective pressures that produced us. AlphaZero is a reflection of its environment and the process by which it developed in that environment. I would be shocked the future of deep learning continues to be hand-design from human intuition.
Edit: looking at the link, I want to caveat the above by saying this paper may or may not be "it," but directionally I think the idea is underrated and erroneously out of favor now as neural networks were in 2009. Ken Stanley (coauthor on the linked paper) in particular has been hung up on one particular approach, NEAT, since like 2002 that is kind of interesting but definitely not the end-all.