Researchers from the Meta-Fundamental AI Research (FAIR) Unit, a research arm of Meta, Facebook’s parent company in the field of artificial intelligence, from leading American universities (MIT, Harvard, Stanford, Columbia, California-Berkeley, Carnegie-Mellon) made a relevant publication in the journal “Science”.

Although AI has so far achieved considerable success and success against humans in two-player competitive games such as chess and Go that do not require communication with others, games such as diplomacy are more difficult for an “intelligent” engine because they require the ability to negotiate with colleagues in natural language, cooperation, multi-level competition with multiple players at the same time, Actions involving dishonesty and stupidity are necessary.

The new system developed by the FAIR team has two subsystems: a dialogic one and a strategic reasoning one.

Thanks to its new learning algorithms, It is not only capable of simulating natural languageBut – more importantly – The goals, beliefs, and intentions of the seven human teammates can be analyzed based on their movements on the game mapThis information is then used to prepare a strategic action plan that takes into account diverse and often conflicting interests to propose a coordination based on mutual benefit.

Playing anonymously against humans in 40 online games of a diplomacy “tournament” between August and October 2022, Cicero The average “score” (25.8%) was more than double that of humans (12.4%).. Cicero “passed as human” to all 82 human players it encountered, none of whom realized it was playing an AI program rather than a human.

They found out later, and were certainly surprised. Only one indicated (after the break…) that he had some doubts that one of his colleagues was not human.

Link to scientific publication:

www.science.org/doi/10.1126/science.ade9097