AI breaks the game of hide and seek
This is a really fun video where an AI is given a version of the game of hide and seek and each team has to learn ways to win - either by finding the other team or by not being found.
Over time they develop and adapt to each others strategies including having to work as a team to win. They also find ways to break the expected rules including finding actual issues with the modelling of the environment. Other approaches are just really clever ways to win - mostly on the hiding side but also occasionally from the seeking team. These are all beautifully and clearly illustrated in the video.