I think the main problem is that deep learning is not AI in general meaning. Andrej Karpathy's "software 2.0" is a much better name for it. I think we cannot reach the real human level AI by directly training deep neural networks, but deep learning is a very promising tool for extending the human brain by computers, and in long term mimic parts of the brain or the whole human brain. So, I think software 2.0 is not a tool that learns like a human, but a tool that can learn how to act like a human brain, and I think it is a possible way to real AGI. (One of my articles about it: https://thebojda.medium.com/will-artificial-intelligence-exceeds-human-intelligence-aka-can-we-reach-the-technological-49f86f0dca33)