Meta has announced what it calls a “breakthrough” in a particular area of game AI: software called Cicero, which is the first AI to achieve “human-level performance in the popular strategy game Diplomacy.” Diplomacy is originally a board game that has many official and unofficial digital successors, and the reason it’s such an interesting choice is because the core of the game is negotiation: that is, it’s a multiplayer game in which the players constantly have to negotiate each other.
The post announcing Cicero acknowledges various AI “victories” over humans (fact check: Deep Blue lost to Garry Kasparov before defeating him a few years later, prompting IBM to decline a rematch), but says, “Really useful, versatile ones.” Agents need to go beyond just moving pieces on a board”. As such, Cicero is said to be able to negotiate, persuade, and work with human players to achieve strategic goals in the same way a human would.
Diplomacy has long been considered one of the great AI challenges for precisely these reasons. You need to understand other players’ motivations, adapt strategies on the fly, and ultimately win them over to your side. Well…Cicero played on webDiplomacy.net, an online version of the game, and “achieved more than double the average score of human players and was in the top 10 percent of participants who played more than one game.”
Indeed: “Cicero is so effective at using natural language to negotiate with people in diplomacy that they often prefer working with Cicero over other human participants.”
Treason! Rank, foul treason!
Part of the achievement is that Cicero was not built on top of the traditional self-play reinforcement method by which AIs learn games (playing millions of games against themselves or humans and processing the data). Meta says it contains two main elements: “strategic thinking, as used in agents like AlphaGo and Pluribus, and natural language processing, as used in models like GPT-3, BlenderBot 3, LaMDA, and OPT-175B.” .
A particularly important part is that Cicero can identify which players it needs to win and strategize to get them on the side. The software “executes an iterative planning algorithm that balances dialog consistency with rationality”, and predicts the players’ future moves based on the dialog before creating a plan incorporating these predictions.
It won’t conquer the world yet: Cicero can only play diplomacy, although of course Meta’s ambitions for this software go far beyond an old board game. The company thinks this could have a big impact on AI chat assistants, allowing them, for example, to engage in learning conversations and dialogues that teach people new skills.
“Alternatively, imagine a video game in which the non-player characters (NPCs) could plan and converse like humans – understanding your motivations and tailoring the conversation accordingly – to help you in your quest to storm the castle help.”
Well, that’s kind of interesting: maybe Edge magazine was right about Doom. what if you could talk to the monsters? You can read more about the technical side of Cicero and find the research paper hereor watch it play against some human experts (opens in new tab).