On Tuesday, Meta AI announced the development of Cicero, which it considers to be the first AI to achieve human-level performance in a strategy board game. Diplomacy. It is a remarkable achievement as the game requires in-depth interpersonal negotiation skills, which means that Cicero acquires the specific language skills necessary to win the game.
Even before Deep Blue defeated Garry Kasparov in chess in 1997, board games were a useful measure of AI achievement. In 2015, another barrier fell when AlphaGo defeated Go master Lee Sedol. Both games follow clear analytical rules (although Go’s rules are simplified for computer AI).
But with Diplomacy, A large part of gaming involves social skills. Players must show empathy, use natural language and build relationships to win, which is a difficult task for computer players. With this in mind, Meta asked, “Can we create more efficient and flexible agencies that can use language to negotiate, persuade, and work with people to achieve strategic goals similar to the way people work?”
According to Meta, the answer is yes. Cicero learned his skills by playing the online version Diplomacy On the Diplomacy.net website. Over time, it has become the owner of the game, reportedly achieving “more than twice the average score” of human players and ranking in the top 10 percent of people who play more than one game.
To create Cicero, Meta pulled together AI models for strategic reasons (similar to AlphaGo) and natural language processing (similar to GPT-3) and rolled them into a single agent. During each game, Cicero looks at the status of the game board and the history of conversations and predicts how other players will react. It creates a plan that it executes through language models that can create human-like boxes that allow it to coordinate with other players.
Meta calls Cicero’s natural language skills an “controllable box model” that is at the heart of Cicero’s personality. Like the GPT-3, Cicero downloads a large amount of Internet content that has been downloaded from the Web site. Meta writes, “To create a manageable conversation template, we started with 2.7 billion BART-like language templates, pre-trained on text from the Internet, and edited over 40,000 games on webDiplomacy.net. “.
The result model masters the mastery of complex games. “For example, Cicero can assume that later in the game it will need the support of a certain player,” Meta said, “and then devise a strategy to win that person’s favor and even acknowledge the risks and opportunities.” “That players see. From their unique point of view.”
Meta’s Cicero research appeared in the journal Science under the heading “Playing the human level in a game of diplomacy by combining linguistic patterns with strategic reasoning.”
For a wider application, Meta suggests that its Cicero research can “break down barriers” between humans and AI, such as maintaining long-term conversations to teach someone new skills. Or it could energize video games where NPCs can speak like humans, understand players’ motivations, and adapt in ways.
At the same time, this technology can be used to manipulate people by deceiving people and deceiving them in ways that can be dangerous depending on the context. Along those lines, Meta hopes other researchers can generate their code “responsibly” and says it has taken steps toward detecting and removing “toxic substances in this new domain” that are likely Refers to the Cicero box learned from the Internet article. It’s swallowed – always a risk for large language models.
Meta provides a detailed website to explain how Cicero works and also has Cicero open source code on GitHub. Online Diplomacy Proponents – and others may be wary.