Multiverse: The First AI Multiplayer World Model (enigma.inc)

The Enigma team has unveiled Multiverse, the first AI multiplayer world model designed to enable agents to interact and adapt collaboratively in shared environments. Built on traditional single-player architectures, the model introduces innovative changes that allow simultaneous frame generation for multiple players, ensuring internal consistency in visual outputs. By merging the players' actions and frames into a unified input, Multiverse addresses the challenges of multiplayer state representation, a key step towards more immersive AI-driven gaming experiences. This development is significant for the AI and machine learning community as it enhances the capability of models to simulate real-world interactions among multiple agents, promoting advancements in cooperative AI systems. Technical innovations include an efficient training process that extends the model's prediction horizon from 0.25 seconds to 15 seconds, allowing it to model complex interactions like vehicle dynamics over longer durations. The training utilized a modified version of Gran Turismo 4, collecting data using a novel method that reconstructed player inputs from on-screen indicators, further showcasing the model's potential for real-time responsiveness and adaptability in dynamic environments.