On the Golom forum they said it is less buggy if you train the neural-net AI on your own computer. You might lose some intelligence but gain fewer bugs.
The problem is that the game had a 100 games of 100 turns training AI. If you try that on your PC it would take weeks, during which time you can't use your PC. I tried 20 turn games 30 times to see what that did. It ran overnight. Now it does seem less buggy but possibly dumber. My opponents are doing strange huge military builds. All concentrate on one unit and go all out on production that will take many weeks. The US is building a huge carrier fleet. I noticed in a write up of another game and looking at a few units in my games the AI hasn't learned the importance of carriers moving only with support vessels. I hope the AI is watching battles to learn tactics.
The AI plays a turn and monitors what problems each country had and what action it took and how that affected the problem. This gets stored into its advice list for that problem which countries use. It will learn from the game you play so it get's smarter as it goes along. The big need is for the training to be able to be broken down into stages - to take an AI you have trained overnight and run it the next night for more training etc. EDIT - Sorry, I looked further and it does start with the last trained AI values. So I can just set it to get a liitle more training each night.
SuperPower contains 4,407 military designs. 94 unit types are used for modeling the United States' forces while 153 are included for Russia.
389 Light Armored Vehicles
311 Battle Tanks
239 Mobile Launchers
160 Attack helicopters
193 Transport helicopters
317 Attack aircrafts
146 Patrol crafts
15 Aircraft Carriers
For game play they have redefined corvettes as the only ships that carry AA missiles. I am curious what brought that on.