Meta named the AI after the Roman statesman Cicero.
It’s a beautiful dream: strategy games in which we have real negotiations with NPCs. Who rip us off, wring Pyrrhic victories from us at the round table and remember our behavior.
In reality, diplomacy in strategy titles often means: ready-made buttons for dialogue options, weighing a handful of values in the background and always the same voicelines when we open the menu for negotiation.
But the dream of immersive battles over alliances, victory conditions, and non-aggression pacts may be closer than many think. The solution, how could it be otherwise nowadays, is called AI.
The reason for the hopeful look ahead is called Cicero. Named after the Roman statesman Marcus Tullius Cicero, the AI is a project of the Facebook group Meta.
The AI is primarily intended to work cooperatively with humans, but also to argue and negotiate on its own behalf. To do this, Meta splits the artificial intelligence into two parts:
The first part of artificial intelligence is a model that develops an advantageous strategy. It works in a similar way to existing game AIs and has similarities with the Go AI AlphaGo, for example.
The second part is a natural language processing system. It should communicate with people in the most natural and coherent way possible. Such models, above all the GPT-3, have been very popular in recent months. GPT-3 has already written a Stellaris article for us as KI-Micha.
An architectural model shows the two subsystems of Cicero. (Image source: Science.org)
The special feature of Cicero is the combination of the two systems: Cicero only uses the first model to come up with a clever strategy and plan human reactions in advance. With the second model, the AI then tries to implement the strategy in dialogue with six human players at the same time, i.e. to negotiate favorably with everyone.
Cicero in Action: Top results in Diplomacy
To put Cicero to the test, Meta lets his AI play a web version of the classic board game Diplomacy. This is a strategy game with relatively simple rules. The special feature of Diplomacy is the social component.
Between each round, players have time to discuss. They conclude secret treaties, open alliances or agreements on how land gained is to be divided up. Then all games execute their moves at the same time – and stand by their agreements or break them.
Cicero is already showing the potential of AI in such open negotiations. Artificial intelligence now beats most human players in Diplomacy, ranking in the top 10 percent of users on webdiplomacy.net.
Cicero in action: The AI makes arrangements with human players via chat. (Image source: Science.org)
Is Cicero paving the way for better strategy games?
mentioned meta in a press release on Cicero explicitly that Cicero is not an AI intended to defeat people in games. Instead, she should learn cooperation and communication. This fits the statement of three-time Diplomacy World Champion Andrew Goff:
Cicero learned Diplomacy in a similar way to many other players. In the beginning, you often lie, simply because the game lets you do it. But one of the great lessons and one of the great things about Diplomacy is that it teaches you to be honest. Because that is the best way to success.
Andrew Goff, three-time World Diplomacy Champion
Nevertheless, Cicero’s example could have an even greater influence on strategy games in the near future. Because although Diplomacy is about cooperation, it is a so-called zero-sum game: in the end, one player wins and the others lose.
Cicero isn’t the only AI that could change strategy games forever. Another project comes from Europey and is intended to make civilization a whole new challenge.
more on the subject
Learning AI for Civilization – AI revolution from Europey
In order for a successor to Cicero to revolutionize the way we approach diplomacy in strategy games, a few stumbling blocks still have to be cleared. Thus, in Diplomacy, Cicero communicates honestly almost throughout. Although this is the best strategy for victory, it only provides limited exciting moments in games.
In addition, it still occasionally happens that Cicero makes serious tactical mistakes, throws his own plans into disarray, or sends messages with erroneous content.
However, the dream of an immersive game where our computer opponents engage in real negotiations with us, implement their own plans and actually respond to what we say seems to have come true just a tiny bit.
The following article tells you how fans of Stronghold and Age of Empires have found a way to drive AI in strategy games to peak performance.
On YouTube, old strategy games become arenas for modern gladiatorial combat
Do you think we will see AIs like Cicero in strategy games any time soon? Will the genre be re-launched from the ground up in the near future? And what do you think of Meta’s approach to teaching an AI strategic thinking? Write it to us in the comments!
