Cicero, an AI program developed by a team lead by Mark Zuckerberg, has proven to be a formidable player in the game of Diplomacy. In Diplomacy, players take on the roles of the Great Powers of Europe just before World War I and must make agreements to collaborate with each other in order to succeed. The game requires both strategic planning and communication skills.
Cicero was trained using reinforcement learning, a method where the program learns by repeatedly playing against itself and improving its performance over time. However, the Meta Fundamental AI Research Diplomacy Team that developed Cicero also incorporated natural language processing to give the program the ability to communicate in a way that mimics human language.
To play the game, Cicero looks at the board, remembers past moves, and predicts what other players might do next. It then simulates different moves and their possible outcomes before deciding on the best move to make. Additionally, the language model incorporated into Cicero allows it to generate possible messages to send to other players and choose the most persuasive ones based on its algorithms and past experiences.
Although Cicero did not exhibit the Machiavellian ruthlessness of some human Diplomacy players, it was able to withhold information from its opponents strategically. The program played 40 games in an online Diplomacy league and was able to rank in the top 10% of players without being detected as an AI program.
Cicero's success in Diplomacy highlights the potential for AI to excel in complex strategy games that require both planning and communication skills. The incorporation of natural language processing into the program allows it to interact with humans in a way that feels natural and could have applications beyond just gaming.
I'm looking forward to a trash-talking, Monopoly-playing version of Cicero.
