The article describes approaches by various AI organisations to harness the depth and breadth of games environments in order to train their algorithms. Games contain a wealth of rich data based on a huge range of simulated (for which read readily assimilated by AIs) real-world environments. From the mundane task of recognising a stop-sign (of fundamental importance for driverless cars) to the complexity of a multi-purpose algorithm, training by way of games-playing is giving AI a big jump forwards.

Link to article: