On Tuesday, Meta AI announced the development of Cicero, which is the first AI to achieve human-level performance in the strategic board game diplomacy. This is a notable achievement as the game requires deep interpersonal negotiation skills, implying that Cicero has gained some command of the language required to win the game.
Even before Deep Blue defeated Garry Kasparov at chess in 1997, board games were one useful measure of AI performance. Another barrier fell in 2015 with AlphaGo defeated Master Lee Sedol. Both games follow a relatively clear set of analytical rules (although Go’s computer AI rules are usually simplified).
But with Diplomacy, A large part of the gameplay involves social skills. Players must show empathy, use natural language, and build relationships to win – a difficult task for a computer gamer. With that in mind, Meta asked, “Can we build more effective and flexible agents that can use language to negotiate, persuade, and collaborate with humans to achieve strategic goals, much like humans do?”
According to Meta, the answer is yes. Cicero learned his skills by playing an online version of diplomacy on webDiplomacy.net. Over time, it became a master of the game, reportedly scoring “more than double the average score” of human players and ranking in the top 10 percent of players who played more than one game.
To create Cicero, Meta brought together AI models for strategic thinking (similar to AlphaGo) and natural language processing (similar to GPT-3) and combined them into one agent. During each game, Cicero looks at the state of the game board, the course of the conversation, and predicts how other players will behave. It creates a plan that it executes through a language model that can generate human-like dialogue that allows it to coordinate with other players.
Meta calls Cicero’s natural language skills a “controllable dialogue model” where the heart of Cicero’s personality lies. Like GPT-3, Cicero draws from a large corpus of Internet texts scraped from the Web. “In order to build a controllable dialogue model, we started with 2.7 billion parameters beard-like language model pre-trained on text from the web and tuned to over 40,000 human games on webDiplomacy.net”, writes Meta.
The resulting model mastered the intricacies of a complex game. “Cicero can, for example, deduce that later in the game it needs the support of a particular player,” says Meta, “and then develop a strategy to win that person’s favor — and even recognize the risks and opportunities that player sees.” from their particular point of view.”
Metas Cicero Research appeared in the journal Science under the title “Human-level play in the game of diplomacy by combining language models with strategic thinking”.
As for broader applications, Meta suggests his Cicero research could “break down communication barriers” between humans and AI, e.g. B. Having a long-term conversation to teach someone a new skill. Or it could power a video game where NPCs can speak like humans, understanding the player’s motivations and adapting as they go.
At the same time, this technology could be used to manipulate people by impersonating humans and, depending on the context, tricking them in potentially dangerous ways. With that in mind, Meta hopes other researchers can build on top of its code “in a responsible manner” and says it has taken steps to detect and remove “toxic messages in this new domain,” likely related to dialogue , which Cicero learned from the Internet texts, it included – always a risk for large language models.
This article was previously published on Source link