Meta AI has announced the development of Cicero, which it claims is the first artificial intelligence to achieve human-level performance in the strategic board game Diplomacy. In fact, this is a remarkable achievement and should be emphasized because Diplomacy requires deep interpersonal communication skills, not superficial like other games.

Even before Deep Blue beat Garry Kasparov at chess in 1997, board games were a key measure of AI’s success. Another hurdle was overcome in 2015 when AlphaGo defeated Go master Lee Sedol. Both of these games have a relatively clear set of analytical rules, and probability-based approaches are at the core of these games.

Cicero performs at a human level at Diplomacy

But in Diplomacy, the main challenge of the game is that it demands advanced social skills abilities. To win, players must show empathy, use natural language, and build relationships. With this in mind, Meta asked, “Can we create more effective and flexible agents that can use language to communicate, persuade and work with people to achieve strategic goals similar to what humans do?” she asked. According to meta, the answer is yes.

The developed Cicero AI was trained by playing an online version of Diplomacy on webDiplomacy.net. Over time, Cicero mastered the game, achieving “more than twice the average score” of human players.

To create Cicero, Meta brought together AI models for strategic reasoning (similar to AlphaGo) and natural language processing (similar to GPT-3) and combined them into a single agent. During each game, Cicero progresses by analyzing the state of the game board, their chat history, and how other players will behave. Because Cicero can create human-like dialogues, he can work in coordination with other players and execute his plan.

Meta describes Cicero’s natural language skills as a “controllable model of dialogue.” Like GPT-3, Cicero uses libraries obtained from the web. Meta also states that they started Cicero’s training with a BART-like language model with 2.7 billion parameters.

Cicero develops strategy within strategy

As the meta researchers point out, Cicero’s abilities and strategies are highly detailed. According to the reports, Cicero can understand that he will need the support of a certain player in the later parts of the game, and he creates a separate strategy to win the favor of that person. He can even take a position by noticing the risks and opportunities that that player may face.

Cicero’s future may be good or bad

As for broader applications, Meta suggests Cicero’s research “could ease barriers to communication” between people. Or it is stated that it can be used for more realistic NPCs in training areas or even games. In this way, the motivation of the player can be adapted.

At the same time, this technology can be used to manipulate people. To that end, Meta hopes other researchers can “responsibly” create their own code, and says it’s taking steps to detect and remove “toxic messages in this new space.” Meta’s “steps” here probably mean dialogues learned from internet texts. Creating large language models with an internet database involves such negative risks. Finally, Meta launched a website to explain how Cicero works, and also open-sourced Cicero’s code on GitHub.