CICERO from the research team’s POV | Meta AI
There’s been a long history of AI playing games as a benchmark, a way of evaluating progress in the field. If you look at games like chess, go poker, these were all truly adversarial games, where it’s all about moving pieces on more. At meta AI we built Cicero, the first AI agent to achieve human level performance in the game diplomacy. Our goal is really to push the limits of what AI can do to drive these incredible breakthroughs in these areas which can then drive impact. I became interested in this project because it sets out a really interesting intersection between different fields of AI. What really makes this different is that once you get into a cooperative setting, you really have to model other people and understand how they’ll react to the proposals that you make and how you can coordinate with them and you have to understand really how humans work. We wanted to, on one hand, have a model of human behavior, but at the same time be better than humans. And so we worked on the techniques of this goal to inform the learning. And that is to keep in mind that a lot of the agent improves its performance but just playing with itself. The way that our team is able to integrate the idea of actually having intentions in what it’s trying to accomplish, both for itself and what it thinks other people should do into the agent so that you can have conversations about it is really remarkable. I think Cicero is a step forward for human AI interaction and that no AI can actually cooperate with you using natural language, which is how people communicate and build their own plans. Cicero is an AI agent that works with you, collaborates with you. It actually brings people together. Metavers is kind of an incredible way for people to work together to collaborate and be. And you can imagine technology like Cicero playing a big part in them.