AlphaZero was miraculously good, almost to the point of straining credibility. AlphaGo and AlphaStar are more like normal advances. They are mostly engineering,

AlphaZero was miraculously good, almost to the point of straining credibility. AlphaGo and AlphaStar are more like normal advances. They are mostly engineering, although theoretical contributions are not trivial. (Using reinforcement learning for value network in case of AlphaGo, and multi-agent self-play setup in case of AlphaStar, since straight self-play doesn't work.)