自从1997年IBM的“深蓝”人工智能系统在国际象棋世界冠军(Garry Kasparov的比赛中击败他之后，不幸的是，人类年复一年地看着我们基于代码的下属在越来越复杂的比赛中战胜我们。但这里有一个陷阱。虽然人工智能机器人在面对面对弈和视频游戏比赛中越来越擅长战胜人类，但当被指示与其他人合作完成一项共同任务时，这些系统的表现通常要差得多。Meta公司的研究人员认为，他们的新人工智能“Cicero”可能会对此有所解释。

根据Meta的基础人工智能研究外交团队与Gizmodo分享的一项新研究，研究人员第一次在战争战略桌游《外交》中训练人工智能达到“人类水平的表现”。新的人工智能以目睹罗马共和国灭亡的古典政治家和学者命名，能够有效地与其他人类玩家沟通和制定战略，计划最佳的胜利方法，在某些情况下，甚至可以通过人类。现在，研究人员表示，西塞罗通过结合对话和战略推理模型完成了任务，成为了多人工智能体学习的“基准”。

在过去的二十多年里，从国际象棋和围棋到星际争霸等更现代的游戏，来自各种公司的人工智能系统越来越令人印象深刻。虽然内容不同，但这些游戏都有一个基本的相似之处:它们都是零和游戏，赢家通吃。

外交则不同。在《Diplomacy》中，7名玩家竞争控制大部分供应中心。玩家之间会不断地进行互动，每一轮游戏都以一系列的前轮谈判开始。重要的是，《外交》玩家可能会试图欺骗他人，也可能认为AI在撒谎。研究人员表示，《外交》尤其具有挑战性，因为它需要在“鼓励玩家不信任任何人的环境中”与他人建立信任。换句话说，人工智能要想在《外交》中“获胜”，它不仅需要有效地理解游戏规则，还需要从根本上理解人类的互动、欺骗和合作，并且知道如何在不听起来像故障洗碗机的情况下把句子串起来。

Cicero或多或少做到了。Meta表示，在40款匿名在线《外交》游戏中，Cicero的平均得分是人类玩家的两倍多，在玩多款游戏的玩家中排名前10%。Cicero甚至在21人参加的8场比赛中获得了第一名。在游戏的每个阶段，Cicero都会根据其他竞争对手的游戏表现和文本对话来模拟他们可能会如何行动。

研究人员于2022年8月19日和10月13日对40个匿名在线weDiplomacy.net游戏进行了研究实验，总计玩了72小时。研究人员表示，他们没有看到游戏中的信息表明人类玩家认为他们是在和人工智能比赛。Cicero显然“以人类玩家的身份过关”，在《外交》的40场比赛中，共有82名独特的玩家。在研究中强调的一个案例中，Cicero能够通过提出一个互利的行动来成功地改变人类玩家的想法。

Cicero使用了大量之前的《Diplomacy》数据进行训练，以便让它能够与其他玩家进行适当的交流。研究人员表示，人工智能是在125,261个匿名外交游戏的数据集上训练的，其中约有40,000个游戏包含对话。总的来说，这个数据集包含了超过1200万条人类玩家之间交换的信息。

Cicero并不完美。AI的对话主要局限于当前回合的行动，Cicero并不擅长模拟它与一名玩家的对话如何影响长期与其他玩家的关系。它偶尔也会发送包含“接地错误”的信息，或者其他与自己的计划相矛盾的信息。(值得注意的是，人类也经常犯同样的错误)。尽管存在这些警告，但研究人员表示，Cicero应该在人工智能棋盘游戏名人堂中占有一席之地，因为它具有与人类合作的独特能力。

虽然这只是对一款棋盘游戏的一项研究，但Meta的新发现标志着一个潜在的新颖的、不那么末日般的视角，可以看待人工智能的增量成功。Cicero并没有因为人工智能系统在我们曾经珍视的游戏中逐渐击败人类而感到毛骨悚然，而是暗示未来人类和人工智能可以作为合作伙伴并肩工作，或者至少是共同的熟人，来解决问题。

原文：

Ever since IBM’s Deep Blue artificial intelligence system defeated world chess champion Garry Kasparov at his own game in 1997, humanity has watched, haplessly, year after year as our code based underlings vanquish us in ever more complicated games. There’s a catch though. While AI bots increasingly excel at trumping humans in head-to-head adversarial board and video games matches, the systems typically fare far worse when instructed to cooperate with other humans to accomplish a shared task. Researchers at Meta believe their new “Cicero” AI may have something to say about that.

For the first time, according to a new study shared with Gizmodo by Meta’s Fundamental AI Research Diplomacy Team, researchers trained an AI to attain “human level performance” in the war strategy board game Diplomacy. The new AI agent, named after the classical statesman and scholar who witnessed the fall of the Roman Republic, was able to effectively communicate and strategize with other human players, plan best methods for victory, and in some cases, even pass as a human. Now, the researchers say Cicero, which accomplished its tasks by combining dialogue and strategic reasoning models, stands as a “benchmark” for multi-AI agent learning.

Over the past twenty or so years, an increasingly impressive assortment of AI systems from a variety of companies have trounced human players in everything from Chess and Go to more modern games like Starcraft. While different in content, those games all share one fundamental similarity: they are all a zero-sum, winner takes all competition.

Diplomacy is different. In Diplomacy, seven players compete to control the majority of supply centers. Players are constantly interacting with each other and each round begins with a series of pre-round negotiations. Crucially, Diplomacy players may attempt to deceive others and may also think the AI is lying. Researchers said Diplomacy is particularly challenging because it requires building trust with others, “in an environment that encourages players to not trust anyone.” In other words, for an AI to “win” at Diplomacy it needs to both understand the rules of the game efficiently but also fundamentally understand human interactions, deceptions, and cooperation, and know how to string together sentences without sounding like a malfunctioning dishwasher.

Cicero more or less did that. Meta says Cicero more than doubled the average score of human players across 40 anonymous online Diplomacy games and ranked in the top 10% of players who played more than one game. Cicero even placed 1st in an eight game tournament involving 21 participants. At every stage of the game, Cicero would model how other competing players were likely to act based on their game performance and text conversations.

The researchers conducted their study experiment August 19 and October 13 2022 over 40 anonymized online weDiplomacy.net games totaling 72 hours of play. Researchers say they didn’t see in-game messages to suggest human players believed they were playing against an AI. Cicero apparently “passed as a human player,” in 40 games of Diplomacy with 82 unique players. In one case highlighted in the study, Cicero was able to successfully change a human player’s mind by proposing a mutually beneficial move.

Cicero was trained on a hefty chunk of previous Diplomacy data in order to get it to prepare it to properly communicate with other players. The researchers say the AI was trained on a dataset of 125,261 anonymized Diplomacy games, around 40,000 of which contained dialogue. In total, that dataset contained over 12 million messages exchanged between human players.

Cicero wasn’t perfect though. The AI’s dialogue was mostly limited to actions in its current turn and Cicero wasn’t great at modeling how its dialogue with one player could affect the relationships with others over the long run. It also occasionally sent messages that contained “grounding errors,” or others that contradicted its own plans. (It’s worth noting that humans often make those same mistakes). Still, despite those caveats the researchers said Cicero should take its place among the AI board gaming hall of fame due to its unique ability to cooperate with humans.

While this is just one study from one board game, Meta’s new findings signal a potentially novel, and less apocalyptic lens with which to view incremental AI successes. Rather than feel creeped out by AI systems gradually beating out humans at games we once held dear, Cicero hints towards a future where humans and AI’s can potentially work side by side as partners, or at the very least, mutual acquaintances, to solve problems. 

本文由数字化转型网（www.szhzxw.cn）翻译而成，作者：Mack DeGeurin；翻译：数字化转型网郑亚茹；翻译审核：数字化转型网默然。