From my understanding AlphaStar used a LSTM to determine the actions of its agents, but it had a transformer to draw attention to a subset of the inputs

It would be nice to get an in-depth overview of the techniques used behind AlphaStar...