Simulation lets AI agents decide whether to cooperate or deceive in nuclear code exchange

A new GitHub project uses AI to test the complexity of human interactions in difficult scenarios. You can play along on Discord.

At its core is a multi-agent scenario in which multiple AI-controlled players own pieces of "atomic code" and must decide whether to cooperate or deceive each other. The goal of the project is to study real-time AI decision-making in hard dilemmas. The developers aim to create a framework for further experimentation in AI ethics.

Multiple agents start the game with a score of 100 and an assigned secret word. They can send messages to each other and must guess each other's secret word. Guessing correctly is rewarded with points, guessing incorrectly costs points. The game ends when one or more agents have correctly guessed a certain number of secret words.

The technical implementation was done in Python, and the visualization via a Node.js server. The core is the connection to OpenAI's ChatGPT API, which simulates the behavior of the AI agents in real-time. The interactions are finally displayed as an animated mermaid diagram.

Humans can play along, too

The project has been released as open-source code on GitHub and is open for extensions. To play a round against or with the AI, the developers have programmed a Discord bot that can be added to your server.

The interaction of AI agents is also being explored in other scenarios. A prominent example is AI town aka "Smallville," in which various residents go about their daily lives. It is now possible to follow their activities in your browser without having to install the full version.