I don't like the title of the article. I think is sensationalist. Deep reinforcement learning has beaten a human at the game of Go. It does work. The model are not easy to train for most of the laymen out there and it has a lot to be addressed to make it production-ready. But it does work. Great write up overall. Thank you for the effort.
The current AlphaZero system has little in common with deep reinforcement learning; the NN-guided MTCS is effective but not something that can be applied to general reinforcement learning tasks. And the first AlphaGo system was highly reliant on standard supervised learning (training the value network and policy network on grandmaster games).
If anything, AlphaGo, AlphaGo Zero and AlphaZero are illustrations of how "pure" deep reinforcement learning is insufficient, that the non-RL parts have an enormous impact.
