Google’s DeepMind created AIs that excel at complex games like chess and Go by playing against themselves. However, these AIs struggle with simpler games like Nim, where players take turns removing matchsticks from a pyramid-shaped board. Nim is part of a group of “impartial games,” where both players share the same pieces and rules. The challenge for AIs is that, unlike chess, it’s easy to predict the winner at any point in these games. This discovery is important because it highlights the limitations of current AI training methods and suggests that improvements are needed, especially as AI becomes more involved in solving real-world problems.
QUESTION: How might the limitations of AI in simple games like Nim influence the way we trust AI in more complex, real-world situations?
Figuring out why AIs get flummoxed by some games
