Defeating 90% of humans, Meta's first 'AI diplomacy model” is on Science! Netizen: Please take Xiao Zha to court

For decades, diplomacy has been considered "a job that AI can never take over."

Because it requires players to master and understand other people’s perspectives and motivations behind them, formulate complex plans and make timely adjustments, then use language to achieve cooperation with others, and finally Convince them to form partnerships, alliances, etc.

To Communication, Trust and Betrayal's attention makes diplomacy completely different from more rule-oriented "games" such as Go and chess.

However, the latest research from Meta shows that AI may also be able to do the work of diplomats!

In the online diplomacy game competition from August to October 2022, CICERORanked among the top 10% of all "players". Its average score of 25.8% is more than twice the average score of its 82 opponents (12.4%). What’s even more worth mentioning is that during the actual game, not a single player discovered that artificial intelligence was playing the game!

Now, this latest result has also been published in the form of a paper in

Science

.

The launch of CICERO will surely become a major achievement in the field of natural language processing.

Because this indicates that artificial intelligence has the potential to "cooperate with humans better and more naturally" and represents a big step towards AGI for humans.

Defeated 90% of humans, no one found out that AI was involved

"Diplomacy" is a seven-player classic strategy game, which can be said to be the board game Risk and Solitaire. A combination of the game poker and the TV show Survivor, it was developed in the 1950s by Hasbro, a famous American toy company.

Through the "role-playing" of the seven European powers in the early 20th century, players need to build trust, negotiate and cooperate with other players and occupy as much territory as possible.

To avoid being blocked by counterattacks from their opponents, players will communicate privately, discuss potential coordinated actions, and then put their actions on paper, complying or violating the rules of the other players Promise of.

Such a game full of deception and power tactics is also regarded by some players as an ideal way to lose friends. It can be called a "Friends' Game"!

As mentioned above, unlike games such as chess and Go, diplomacy is more about

"people"

It's not a

"rules" game. If the model cannot recognize that someone may be bluffing, or accurately identify the other player's aggressiveness in a certain move, it will obviously lose the game very quickly.

Likewise, if it doesn't speak like a real person, show empathy, build relationships, and talk about the game, it won't find other players willing to work with it.

Over the past few decades, researchers have been building an “AI diplomat” with natural language communication capabilities. However, because this major challenge is far beyond the capabilities of existing AI, no researcher has ever succeeded.

#It wasn’t until the recent emergence of CICERO that this fact was completely overturned.

CICERO is essentially a "chat robot" that can communicate with other diplomatic players to take effective actions in the game.

Cicero was a famous politician, philosopher, and orator in ancient Rome. He was born on January 3, 106 BC. He was famous in Roman political circles for his eloquence. .

Meta will be named after this AI model, and the meaning is self-evident.

From August to October 2022, CICERO participated in a total of 40 competitions in the online "Diplomacy" competition organized by webDiplomacy, ranking among the top 10% of all participants ; Of the 19 people who played five or more games, Cicero ranked second.

In 40 games, CICERO's average score was 25.8%, more than double the average score of its other 82 opponents (12.4%), and combined its strategic dialogue and gameplay abilities Shown vividly.

Who can not love AI that can think and express?

CICERO is based on a 2.7 billion parameter BART-like language model that is pre-trained on text from the Internet and uses a dataset of more than 40,000 diplomacy games played online at webDiplomacy.net Expanded.

The data also includes more than 12 million messages generated when players communicate with each other.

CICERO’s model mainly consists of two parts, namely "Strategic Reasoning" and "Natural Language Processing" ".

The integration of the two technologies enables CICERO to reason and strategize around player motivations, then use natural language to communicate, agree to achieve common goals, form alliances and coordinate plans, Mainly reflected in "Cooperation", "Negotiation" and "Coordination" three sides.

For example, CICERO can infer that later in the game it will need the support of a particular player, and then develop a strategy to win that person's favor—even identifying that player's risks and Chance.

The dialogue-aware strategy module helps CICERO predict what actions other players may take, and what other players think CICERO may do, given their past conversations and the state of the game board.

Thus, CICERO will develop mutually beneficial plans for itself and other participants based on these predictions. These plans not only allow CICERO to find opportunities for mutually beneficial cooperation, but also help it find effective measures when cooperation is impossible.

There is a controllable dialogue model in CICERO, which is combined with strategic reasoning algorithms that control dialogue generation.

The Controlled Dialogue Model allows CICERO to engage in dialogue within a carefully chosen set of plans, usually ones that benefit both CICERO and the other players.

CICERO's dialogue is deeply rooted in free-form conversations generated within the ongoing game.

For example, CICERO might negotiate a tactical plan with another player, reassure allies of its intentions, discuss broader strategic dynamics in the game, or even just engage in casual chit-chat - including Pretty much anything a human player might discuss.

"Cicero was so effective at using natural language to negotiate with diplomats that they often preferred working with Cicero rather than Not other human participants," Meta said on its Twitter.

Meta AI Vice President and Chief Artificial Intelligence Scientist Yan Lecun believes that "being able to perform human-level performance in a strategically extremely complex game like diplomacy indicates the great potential of human-artificial intelligence cooperation." .

Although CICERO can only play diplomacy, the technology behind this achievement is closely related to many real-world applications. For example, controlling natural language generation through planning and RL can ease human interaction with Communication barriers between AI models.

For example, today’s artificial intelligence assistants can only perform simple questions and answers, such as telling you today’s weather, etc., but what if they teach you a new skill through long-term conversations?

Or imagine a video game in which non-player characters (NPCs) can plan and converse freely just like people—understanding your motivations and adjusting dialogue accordingly to help you complete The mission of siege the city.

Of course, even Meta itself admits that "CICERO is not perfect yet" - at certain important moments in the game, CICERO often makes outrageous mistakes.

Therefore, Meta chose to release CICERO’s code as open source, hoping to further improve it with the help of the AI ​​developer community.

Netizen: Please take Xiao Zha to court!

The release of the world’s first “AI diplomat” that is at the same level as humans has also triggered heated discussions among netizens.

Many netizens expressed:

"I am really looking forward to the next development of this research."

"Beating humans can be said to be the most humane game. It is simply so fascinating..."

##Although CICERO is just starting out, some people are looking forward to the application prospects of this "AI black technology" in real life:

"It can build a version to Help address collective action challenges, such as #COP28?"

The "COP28" mentioned by this netizen should refer to the 28th United Nations Climate Change Conference General Assembly.

At the just-concluded 27th Climate Conference, after several days of intense negotiations, representatives from various countries finally agreed to establish a fund mechanism to compensate for losses and damages caused by climate change.

In addition, the launch of CICERO has also caused concerns among many netizens. "This will directly encourage researchers to build models that are good at deception."

"Cheating and winning the game of diplomacy in a way that mimics human behavior is cute and fun."

"I wonder what else it can be used for? We need to be alert to the development of such tools."

"Artificial intelligence is very good at creating Art, etc. But now, its power of persuasion is "activated"."

"If you can convince a person, you can control their choices and thus their life. "

"So the final outcome will be - AI enslaves humans through persuasion!"

Finally, many netizens joked:

"Is this reliable? Cicero In the end, he was beheaded!"

"Please send Xiaozhao to The Hague (International Court of Justice)!"

Just two days ago, Galactica, a large-scale language model launched by Meta AI, was hastily removed from the shelves only 3 days after it was launched because it stated lies as facts. Nowadays, the launch of CICERO can be said to have once again caused waves in the AI ​​technology circle.

The above is the detailed content of Defeating 90% of humans, Meta's first 'AI diplomacy model” is on Science! Netizen: Please take Xiao Zha to court. For more information, please follow other related articles on the PHP Chinese website!