> "People originally thought that we were doing a game that used some form of long-term machine learning in which, if you played the game for hours on end, it would eventually become you and copy exactly how you play. We didn't want to do that, though. We had a few early prototypes in which the game became a representation of how you interacted with it on a macro scale, but we found that became extremely punishing."
> These iterations of the game weren't only punishing, they were predictable and dull. If you performed a certain action at the very start then you could likely never perform it again, given that the machine would be able to see it coming. As such, your quota of potential moves reduced with each play session. By the end of the game, you barely have any strategies left to use at all.
At this point it sounds like they abandoned that idea, but (!):
> Emborg and his team opted for a system in which the AI learns and unlearns from you.
> These iterations of the game weren't only punishing, they were predictable and dull. If you performed a certain action at the very start then you could likely never perform it again, given that the machine would be able to see it coming. As such, your quota of potential moves reduced with each play session. By the end of the game, you barely have any strategies left to use at all.
At this point it sounds like they abandoned that idea, but (!):
> Emborg and his team opted for a system in which the AI learns and unlearns from you.
Fascinating!