Meta taught artificial intelligence to lie and strategize

Meta has trained an AI agent to play a board game that involves chatting with other players to convince them to support her strategies – and then betraying them.

The company, which owns Facebook, Instagram and WhatsApp, says its Cicero AI may have wide-ranging applications in the near future including developing smarter virtual assistants with the combined use of technologies such as natural language processing (NLP) and strategic reasoning, according to the company. blog post issued by the company.

in Research article in the academic journal ScienceMeta said Cicero AI achieved human-level performance on strategic board strategy games in an online league where it played 40 matches against 82 people, ranking in the top 10% of participants who played more than one game.

Diplomacy pits seven players against each other for control of the map of Europe. Each turn begins with the players negotiating with each other to support their plans and ends with them simultaneously trying to execute their moves. Without the support of other players, many of these moves will fail.

Meta said the game posed a challenge for the AI ​​agent, because winning requires him to understand if his opponents are bluffing or scheming in a certain way to win the game. The AI ​​needed to increase a certain level of empathy while playing the game to form cooperation with other players, something the AI ​​did not need to do when playing games like chess against human opponents.

AI agents in strategy games have been getting better and better over the years: in 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov, and in 2016, DeepMind’s AlphaGo beat top Go player Lee Sedol. Facebook has also developed another AI engine that can outperform humans at poker.

Strategic thinking

Cicero is built on two main technical components: Strategic Reasoning and Natural Language Processing (NLP). The researchers explained that while the strategic inference engine predicts the moves of other players and uses that information to form its own strategy, the natural language processing engine generates messages and analyzes responses in conversations with other players to negotiate and reach an agreement.

In order to help the AI ​​agent form relevant conversations, the researchers started with a natural language generation model made of 2.7 billion variables that was pre-trained on text from the Internet and tuned by conversations between human players in more than 40,000 games from webDiplomacy.net.

“We developed techniques to automatically annotate messages in the training data with corresponding planned movements in the game, so that at inference time we can control dialogue creation to discuss the specific actions required of the agent and its conversation partners,” the researchers said in a More report. Detailed blog post.

Mita has Open source Cicero code For other researchers to build on the capabilities of the AI ​​agent.

In addition, the company has established Portal to invite proposals on research In the field of human-artificial intelligence cooperation through NLP using diplomacy as a basic concept.

Long term planning

Big tech companies like Microsoft, Google and Amazon are competing with each other to develop smarter standalone virtual assistants to support a variety of business use cases, from call centers to AI agents that can perform sentiment analysis and teach new skills to an individual. The global natural language processing (NLP) market, which includes these assistants, is expected to grow from $26.4 billion in 2022 to $161.8 billion by 2029, according to a report by Fortune Business Insights.

The Meta researchers seem to suggest that Cicero’s success in diplomacy supersedes the capabilities of other virtual assistants available today, saying in a blog post, “For example, existing AI assistants can complete simple question-and-answer tasks, such as telling you the weather — but what if they could From having a long drawn out conversation with the goal of teaching you a new skill?”

This is a search for tools like Google Duplex, Amazon Alexa, Microsoft Xiaoice, and Apple’s Siri. But Cicero does not lend itself to long-term conversations either, because his reasoning is completely short-term. As the Meta researchers put it in a paper in Science, “From a strategic perspective, Cicero thought of dialogue in terms of players’ actions at the current turn. He did not model how his dialogue might affect the relationship with other players over the long term of the game.”

Copyright © 2022 IDG Communications, Inc. All Rights Reserved.

Leave a Comment