I only had a chance to see the live game, but very impressive stuff, especially in the early game! However, it does seem to have the same issue as all other AIs

I only had a chance to see the live game, but very impressive stuff, especially in the early game! However, it does seem to have the same issue as all other AIs in that it is inflexible and fails to adapt to unusual situations - As the commentators pointed out, it kept building oracles when being constantly harassed by an immortal drop, whereas any amateur player would be able to react more effectively by building a single phoenix. It also fails to recognize patterns - MaNa was able to find and repeatedly abuse a border between two strategies: whenever he was not actively dropping, AlphaStar would attempt to move out and push, then immediately retreat to defend when being dropped, whereas a human would recognize the reoccurring theme and either keep pushing or stay in their base. While not groundbreaking, still exciting to see an AI that can hold its ground against pro players - it certainly demonstrates the potential of machine learning for constrained problem spaces.