Meta has launched a new AI capable of beating human opponents in the game of diplomacy.
The game was considered “an almost impossible grand challenge” for the AI, Meta wrote in an AI blog post, called CICERO.
Diplomacy is particularly difficult, even compared to complex games like chess and go, because it requires mastery not of hard and fast rules, but of soft skills. Players must know the art of understanding the perspectives and needs, desires and habits of others; make complex, living plans that can change with human whims; then persuade other players to work with them and against others. (The latter is trickier for an AI than you or me.)
In short, it is primarily a test of social skills, not strategic, logical or mathematical ones.
Because it relies on social skills – not strategic, logical or mathematical ones – diplomacy has long been considered an “almost impossible” challenge for an AI.
“Despite many advances in training AI systems to mimic human language, creating agents that use language to intentionally communicate with humans in interactive environments remains a major challenge,” the team wrote. Meta research in an article on CICERO, published in Science.
An AI in a game like Diplomacy must speak like a real person, the researchers wrote in their blog post, demonstrating empathy and knowledge of the game to build relationships. And the language imperative works both ways: if the AI can’t recognize a bluff, or the subtext of what others are saying, it will be quickly outmaneuvered.
To create a winning AI, the CICERO team combined a natural language processing model (think the famous GPT-3) and a strategic reasoning model (like the Stockfish engine used to play chess), as summarized by Ars Technica.
Meta calls this ability a “controllable dialog pattern”. Throughout the game, CICERO turns its analytical gaze on the game state and player chat histories, predicts how others will act, and then devises a plan around those scenarios.
“CICERO can deduce, for example, that later in the game he will need the support of a particular player, and then strategize how to win that person’s favor – and even recognize the risks and opportunities that that player sees from his particular point of view,” the team wrote.
An AI in diplomacy needs to speak like a real person to build relationships. And it goes both ways: if the AI can’t recognize a bluff, or the subtext of what others are saying, it will be quickly outmaneuvered.
With his plan in place, CICERO then uses natural language processing to craft messages that are human enough to set his plan in motion. The model was pre-trained on the Internet’s chaotic corpus, much like GPT-3, and then cut its teeth on 40,000 games archived on webDiplomacy.net.
The system seems to be working quite well. When CICERO played 40 games of webDiplomacy against human competitors, he averaged more than double the average score of human players, ranking him in the top 10% of online gamers who played more than one game.
It’s not as dominant as today’s chess engines – chess has precisely quantifiable correct moves, and modern computers will hardly ever lose (unless they’re programmed to make it easy for you to task).
But scoring among the top performers in a game like Diplomacy is a huge step up, given how fuzzy and subjective the situation is. In fact, the CICERO AI was so good that people prefer this as an ally.
CICERO combines a natural language processing model (think GPT-3) and a strategic reasoning model (like the Stockfish engine used to play chess), to develop and then execute a plan.
“CICERO is so good at using natural language to negotiate with people in diplomacy that they often preferred working with CICERO over other human participants,” the researchers wrote.
The Meta team thinks their model could eventually help streamline communication between people and AI — imagine an AI that can hold a convo long enough to teach you something, they suggest. Or, for a less ambitious but honestly more fun goal, it could power more realistic video game NPCs that will adapt to your player.
But, as Ars Technica has pointed out, the ability to cooperate is a stone’s throw from the ability to manipulate. To that end, they’ve taken steps to detect harmful speech that CICERO might encounter, and the team has “high hopes” that people will “build responsibly” with their work, which is available as open source on GitHub. .
We would love to hear from you! If you have a comment about this article or have a tip for a future Freethink story, please email us at [email protected]
#Diplomacy #game #negotiate #form #alliances #persuade #people