Meta’s Cicero can beat humans in the game of Diplomacy.
AI’s abilities, particularly ChatGPT, are a hot topic. Professionals fear automation, but not in the field of foreign policy, but maybe they should be afraid as well. Meta’s AI team recently made progress suggesting AGI can automate foreign policy design, analysis, and implementation.
“CICERO: An AI agent that negotiates, persuades, and cooperates with people”
The META AI team wrote a blog post, with the above title, about their new AI model that achieved a top 10% ranking in the game of Diplomacy. Previously all games that AI achieved superhuman results were two-player zero-sum games. Diplomacy is unique because it’s not, which makes CICERO’s achievement impressive. In Diplomacy, players need to collaborate and use natural language to develop winning strategies against other players. This makes Diplomacy more similar to real-life decision-making. Diplomacy shows us how AI could be useful in complex decision-making situations like real-world foreign policy games and strategies.
History
During my Ph.D., I had the pleasure of studying Robert Axelrod’s use of game theory to explain how collaboration emerges in competitive human games. Axelrod, a mathematician and political scientist, used game theory models to study human strategies in political conflicts. Although his early conclusions were later scrutinized by prominent scientists like Elinor Ostrom, Adam Przeworski, and Martin Nowak, Axelrod’s approach still significantly impacts political science research today.
Robert Axelrod’s approach was pretty simple. He used a two-player game to show cooperation results in better outcomes and higher rewards for players. Over time, even without prior arrangements or communication, players adopt a cooperative strategy. This approach, using game theory to study political problems as a repetitive game, had a significant impact on research about foreign policy and resource-sharing (Garrett Hardin’s “Tragedy of the Commons”).
Outsmarting Humans
So, how does the CICERO model achieve such high performance? Well, first, let’s look at the Diplomacy game. Diplomacy is a strategic cooperative game played among seven players. The objective is to control as many supply centers (SCs) on a map as possible. In each round, players converse privately to coordinate their next attack, then make their moves simultaneously. The game continues until players agree to draw or one player controls the majority of SCs.
The secret sauce of the CICERO’s triumph is its unique combination of a language model and a strategic reasoning model. The language model, a Transformer-based NLP generator, incorporates information about the players’ intents in addition to the game’s history and current state. This conditioning leads to improved quality of dialogue, reducing noise and significantly enhancing performance.
The strategic reasoning model is based on Reinforcement Learning (RL). Divorced from the language and conversations, the optimal strategy is estimated at the game’s current conditions, past intents of the user and other users, and the expected actions of other users. Stabilizing the strategic reasoning model using regularization was another novel method used in this research. Finally, a set of six classifiers filter the messages to ensure the message’s high quality and relevance to the model’s intent and limit the leaking of valuable information.
How about ChatGPT?
Can ChatGPT be used for foreign policy generation? While it’s a proficient language model, ChatGPT’s rational reasoning and strategy design capabilities are limited. It can highlight critical points in a strategy or summarize web conversations on a particular topic, but it doesn’t possess strategic thinking capabilities.
I had a conversation with ChatGPT about the United States’ optimal foreign policy regarding the Iranian Uranium enrichment program. The conversation was impressive and showcased the potential for the future. You can read the full conversation below.
What to learn, in case you are interested?
CICERO represents groundbreaking AI research. For a long time, different branches of AI, like Natural Language Processing (NLP) and Reinforcement Learning (RL), were developed in separate silos by different communities. But with the emergence of deep learning, these fields have substantial overlap. Modern AI pipelines incorporate multi-modals and hybrid applications of large language models and RL. This means to develop such models, interested practitioners must learn a wide range of topics, which is quite daunting and gruesome.
The specific knowledge needed to learn about Cicero and ChatGPT depends on your starting point and learning goals. However, the success of these models highlights the significance of Large Language Models (LLMs) and Reinforcement Learning in shaping our future. Going forward, Reinforcement Learning with Human Feedback (RLHF) will likely be the focus of the next decade.
The Future is Here
While many people will object to using Artificial Intelligence in sensitive and high-impact fields such as foreign policy, my thinking is, why not? AI is great at analyzing enormous amounts of data, something that’s not humanly possible, and can provide intelligent summaries of crucial facts. With its ability to analyze data and provide strategies without emotional bias, AI is a perfect solution for crafting detailed policy drafts and scenario variations for nuanced situations.
Although Large Language Models (LLMs) and other forms of AI are useful, blindly accepting their output can be dangerous. As any human expert knows, these systems can produce hallucinations and mistake speculations for facts. When a model’s high confidence level combines with human ignorance of its probabilistic nature, the consequences can be catastrophic. While AI can assist policy researchers in digging into facts and understanding relevant considerations, we should not solely rely on AGI to make decisions for us in any significant human endeavor.
My Chat with ChatGPT
Imagine that you are an American foreign policy expert. Write a brief outline for such a policy to deal with the Iranian development of nuclear arms.
What is your policy recommendation in this scenario?
As an AI language model, I do not have personal opinions or preferences, but I can provide an analysis of potential policy options based on available information.
In the case of the Iranian development of nuclear arms, a comprehensive policy that combines diplomatic engagement, economic pressure, and military deterrence may be the most effective option for the United States to pursue.
Diplomatic engagement with Iran may provide an opportunity to negotiate and compromise, and could potentially lead to a peaceful resolution to the issue. Economic pressure, in the form of sanctions, may also be effective in pressuring the Iranian government to reconsider its nuclear program. Additionally, a credible military threat may deter Iran from pursuing nuclear arms and could strengthen the position of the United States in negotiations.
It is important to note that any policy decision must take into consideration the potential risks and benefits of each option. Military action, for example, could have significant negative consequences and should only be considered as a last resort. Additionally, it is important to engage with international organizations such as the International Atomic Energy Agency (IAEA) to monitor the Iranian nuclear program and ensure compliance with international agreements.
Ultimately, any policy decision regarding the Iranian nuclear program should be carefully considered and tailored to the specific circumstances of the situation.