As you know, the Meta AI team has been working on a translation for a long time. Cicero, the product of the company’s latest artificial intelligence work, negotiates and collaborates with people, and more importantly, he can persuade them. All these skills of Cicero were tested with the strategy game Diplomacy, and Cicero received a passing grade from its creators. Let’s take a look at the technical details of Cicero and what it offers to users.
Meta AI’s speech and translation-oriented artificial intelligence studies
Before examining the technical details of Cicero, it is useful to briefly remind the speech and translation-oriented artificial intelligence studies of Meta AI. Announcing that it is working on an artificial intelligence-assisted “universal speech interpreter” (UST) in February 2022, Meta AI will launch in July 2022 as part of the “No Language Left Behind” (NLLB) project, including low-resource languages such as Asturian, Luganda, and Urdu. announced that he would make translations for a total of 200 languages. In October, Meta AI showed us that it simultaneously translated unwritten spoken languages.
Cicero, perhaps, took all this work a step further, simultaneously translating 6 languages while playing Diplomacy. Let’s say that hundreds of messages were sent within the scope of the game. Using natural language processing, Cicero developed relationships with other players throughout the game.
How does Cicero work?
Meanwhile, Cicero consists of two components, the planning engine, and the dialog component. These two components work together to create meaningful dialogues. As the name suggests, the planning engine generates strong ideas for both itself and the partners it works with, thus feeding the dialogue component with strong ideas.
We can say that Meta gathers new techniques in two completely different fields of artificial intelligence under the umbrella of Cicero. The company used strategic reasoning as used in AlphaGo and Pluribus, and natural language processing as used in GPT-3, BlenderBot 3, LaMD, A, and OPT-175B.
At every point in the game, Cicero looks at the game board and conversation history to model how other players will behave. He then uses the models he uncovers to control a language model that generates free-form dialogs. The language model in question informs other players while suggesting logical actions to them, and more importantly, it can coordinate with other players.
To reveal the controllable language model, the team trained a BART -like language model with text from more than 40,000 games people play online and on webDiplomacy.net.
Those who want to work with Cicero can take a look at the article by Meta tea, which has released the code as open source on GitHub. Likewise, it should be noted that researchers can access data via CICERO RF.
What can Cicero offer?
Artificial intelligence assistants, who have answered short queries so far, can have more comprehensive and long-form conversations with Cicero’s model in the coming period. In addition, non-player characters in the game can gain the feature of being a player with Cicero. These characters may develop the capacity to understand your speech and act accordingly.
In addition to these, let’s say that Cicero can also be used for educational purposes. Cicero can guide users who want to develop skills through long conversations and support the training process.