Google hype, their scientists believe their own smoke now. This is unscalable, non adaptive narrow AI where brute force was used together with machine learning. Realize the game here is fixed, it’s a closed system. If the game agents suddenly could do unexpected behaviors then the Google AI would crash and burn, just like any other system based upon catastrophic re-training, which is why self-driving cars fail too.
Self driving cars work great when they are working within their training paradigm but fail miserably when something outside the training paradigm happens. This is why Google saying they next want tackle weather prediction with this stuff is complete hype nonsense. First off weather is chaotic and not predictable with digital computers. Digital computers can only approximate chaotic and fractal systems because they are analog super Turing in nature.
You cannot predict a super Turing system with a Turing limited machine. The additional degrees of super Turing problem solving freedom make that impossible. Next they will play jeopardy, woops someone (IBM) already did that and it didn’t scale either. QAI SPECTRE™ BRAIN CHIP to blow their AI smoke away.