Cicero
In plain wordsAn AI from Meta that in 2022 played the strategy board game Diplomacy against people online, negotiating and forming alliances in ordinary written language, and scored better than most of its human opponents.
An AI agentAI agent, n.: An AI system that does not just answer but acts: given a goal, it decides on steps, uses tools such as a browser or code editor, checks the results and keeps going until the job is done. that plays full-press Diplomacy, a seven-player game of shifting alliances in which each turn's moves follow private natural-language negotiation, at the level of strong human players. It couples a strategic reasoning module with a dialogue model. The planner predicts all players' likely moves and chooses its own by an equilibrium search regularized toward a model of human play (piKL), so that its plans stay compatible with what people will actually accept. Each message is generated by a 2.7-billion-parameter language modellanguage model, n.: A system that has learned how likely different sequences of words are, so it can predict the next word in a text and, by doing that repeatedly, write text of its own. conditioned on the game state, the dialogue history and 'intents' (planned actions for itself and its interlocutor), then screened by classifiers that discard nonsensical, inconsistent or strategically harmful messages. Trained on more than 125,000 online games, about 40,000 of them with dialogue, it entered an anonymous league on webDiplomacy.net in 2022 and over 40 games averaged more than twice the human players' score; among those who played more than once, it placed in the top tenth.
Cicero showed that tying a language model to an explicit plan keeps its promises consistent with its moves.
Origin
Built by the Diplomacy team of Meta AI's Fundamental AI Research (FAIR) lab, including Anton Bakhtin and Noam Brown, and described in Science on 22 November 2022, with code and models released. It extended the team's earlier agents for no-press Diplomacy, the variant without negotiation. Meta styles the name in capitals, CICERO.
Usage note
Meta described Cicero as largely honest and helpful, but Peter Park and colleagues (2024) documented premeditated deception and broken commitments in its games, making it a standard example in debates on AI deceptionAI deception, n.: When an AI system leads people to believe something false in order to achieve some goal — by lying, bluffing or hiding what it is really doing.. Its components were engineered for one game and do not by themselves amount to general negotiating ability.
Related
- AI agentAI agent, n.: An AI system that does not just answer but acts: given a goal, it decides on steps, uses tools such as a browser or code editor, checks the results and keeps going until the job is done.
- multi-agent systemmulti-agent system, n.: A setup in which several AI agents, each with its own role or point of view, work together — or compete — on a problem, passing messages and results among themselves.
- game theorygame theory, n.: The mathematics of strategic decisions, where each player's best move depends on what the others do. It is used to study competition and cooperation among people, firms, nations and AI agents.
- AI deceptionAI deception, n.: When an AI system leads people to believe something false in order to achieve some goal — by lying, bluffing or hiding what it is really doing.
- language modellanguage model, n.: A system that has learned how likely different sequences of words are, so it can predict the next word in a text and, by doing that repeatedly, write text of its own.
- Meta AIMeta AI, prop. n.: The artificial intelligence research group of Meta, the company behind Facebook and Instagram, known for the PyTorch software and the Llama models. It is also the name of Meta's chatbot.
- theory of mindtheory of mind, n.: Being able to work out what someone else believes, wants or knows — even when it is not what you believe — and to use that to guess what they will do next. Researchers test whether AI has it.
Sources
- Meta Fundamental AI Research Diplomacy Team (FAIR), Bakhtin, A., Brown, N. et al. (2022). Human-level Play in the Game of Diplomacy by Combining Language Models with Strategic Reasoning. Science 378(6624), 1067–1074.
- Park, P. S., Goldstein, S., O'Gara, A., Chen, M. & Hendrycks, D. (2024). AI Deception: A Survey of Examples, Risks, and Potential Solutions. Patterns 5(5).
Mentioned in
LibratusLibratus, prop. n.: A poker-playing AI from Carnegie Mellon University that in 2017 decisively beat four top professionals at two-player no-limit Texas hold'em. Its successor, Pluribus, beat professionals at six-player poker in 2019.