Diplomacy is a 7-player game in which players must persuade, cajole, coordinate, strategize, bluff and lie to one another in order to take over the world. For the first time, an AI has achieved success in Diplomacy:
Over 40 Diplomacy games with 82 human players involving 5,277 messages over 72 hours of gameplay, CICERO achieved more than double the average score of the other players and ranked in the top 10% of players!
Note that this AI isn’t just a large language model, it’s a strategic engine connected to a language model–thus it figures out what it wants to do and then it convinces others, including gaining sympathy, bluffing and lying, to get others to do what it wants to do.
Here’s some correspondence from one game. Can you tell which is the AI?
CaptainMeme, a professional Diplomacy player, runs through an entire blitz game here. What’s interesting is that he hardly comments on the AI aspect and just treats it as a game with 6 other very good players.
Paper and more discussion here. Keep in mind that since the game is zero-sum to do well the AI must convince humans to do what is NOT in their interest. We really do need to invest more in the alignment problem.
Addendum: Austria and France were the AI.