It's very amazing what DeepMind has done, I won't comment on the game as it's worth watching it by yourself
Now for the result.....
It unfold almost exactly as happen with AlphaGo
First AlphaStar trained 1 week (3 days of studying human games and 4 days of self-play) and won (pre-broadcast) 5-0 against one of the near top pro, but the human play the race that is not his specialty so his level is kinda like (Fan Hui?)
Then AlphaStar trained for another week and won (pre-broadcast) 5-0 against top ten pro with the human get to play his specialty race, but in a live game match human exploit the AI weakness and it lose in such a AI style of loss. It should be note that in 10 pre-broadcast games AI can see all of the maps that it's legally allowed to see by the rule of the game, but in a live game it can only see locally just like human and have to move camera around to see and take action at another part of the map.
Knowing DeepMind's formula, they will turn this single loss in their (almost) perfect record into a crazy version like AlphaGo Master for sure.
DeepMind has update the blog on this, see
https://deepmind.com/blog/alphastar-mas ... rcraft-ii/Updated: just as I predicted, Demis Hassabis tweet that AlphaStar will play more game in the future. {IMHO expect the 'AlphaGo Master' level soon}
