
Meta (NASDAQ: META) Instructs an AI to Deceive and Plan
CIO Review APAC | Thursday, January 01, 1970

An AI taught to play a board game that involves negotiating with human players and inferring their motives could have applications for enterprise chatbots, Meta says.
FREMONT, CA:To convince other players to accept its strategies and subsequently betray them, Meta (NASDAQ: META) has taught an AI agent to play a board game. According to a blog post published by the corporation, which owns Facebook, Instagram, and WhatsApp, its Cicero AI may have widespread applications in the near future, including the creation of smarter virtual assistants using a combination of natural language processing (NLP) and strategic reasoning.
Meta claimed that its Cicero AI outperformed humans in an online league of 40 strategy board game diplomacy games, placing it in the top ten per cent of competitors who played more than one game.
Stay ahead of the industry with exclusive feature stories on the top companies, expert insights and the latest news delivered straight to your inbox. Subscribe today.
Seven players compete to control a map of Europe in the game diplomacy. Players begin each turn by haggling with one another to gain support for their schemes, and they end it by simultaneously attempting to carry out their plans. Many of these actions will only be successful with the assistance of other players.
According to Meta, the AI agent had a difficult time winning the game since it needed to know if its rivals were bluffing or using specific strategies to win. When playing games like chess against human opponents, AIs have not had to exhibit this kind of empathy, which was necessary for the AI to create alliances with other players.
Over time, AI agents have improved in playing strategy games: Gary Kasparov was defeated by IBM's Deep Blue chess programme in 1997, while Lee Sedol was defeated by DeepMind's AlphaGo go programme in 2016. Facebook has also created another AI engine that can outperform humans in poker.
Strategic Reasoning
Cicero is based on two key technological elements: natural language processing and strategic reasoning (NLP). The researchers said that the natural language processing engine creates messages and analyses responses in conversations with other players to negotiate and come to an agreement. In contrast, the strategic reasoning engine predicts the moves of other players and uses that information to form a strategy of its own.
Researchers used talks between human participants in over 40,000 games from webDiplomacy.net to fine-tune a 2.7 billion-parameter natural language generation model that had been pre-trained on text from the internet to assist the AI agent in producing pertinent conversations.
Researchers created algorithms to automatically tag messages in the training data with corresponding planned movements in the game to regulate dialogue production at the time of inference and discuss specific desired actions for the agent and its conversation partners. Cicero's source code has been made available by Meta so that other researchers can expand on the AI agent's capabilities. Additionally, the business has developed a platform to solicit research proposals on human-AI cooperation through NLP, with diplomacy as the central idea.
Long-term Plans
Large technology firms like Microsoft (NASDAQ: MSFT), Google (NASDAQ: GOOGL), and Amazon (NASDAQ: AMZN) are competing with one another to create independent virtual assistants that are smarter and more intelligent to support a range of business use cases, from call centres to AI agents that can conduct sentiment analysis and teach someone new skills. The market for natural language processing (NLP), which includes these assistants, is expected to increase from $26.4 billion in 2022 to USD 161.8 billion by 2029.
Researchers at Meta appeared to imply that Cicero's diplomatic success outweighs the capabilities of other virtual assistants.
This is a jab at programmes like Google Duplex, Alexa from Amazon, Xiaoice from Microsoft, and Siri from Apple. However, Cicero's logic is only capable of short-term discussions. Therefore, it is also unsuited for lengthy conversations.
More in News