AGWM-2 was trained on 2 times the data of the original AGWM, with a focus on enhancing the model's ability to understand and generate complex text-based environments. This version aims to provide a more robust foundation for future developments in interactive AI systems.

AGWM (Artificial Generative World Models) is a novel AI framework that transcends traditional text-based reasoning by situating large language models within fully interactive, simulated 3D environments generated through advanced world modeling systems such as Google DeepMind’s Genie 3. Unlike standard chain-of-thought approaches, AGWM introduces experiential reasoning: the model does not merely think, it lives, acts, and learns within a persistent, physics-consistent virtual universe. Upon receiving a user query, e.g., “How do you build a Dyson Sphere?”AGWM deploys the reasoning agent into a rich, Earth-like simulation where it may spend simulated centuries constructing infrastructure, solving engineering bottlenecks, and generating novel theories. While this process spans thousands of in-simulation years, it unfolds in mere seconds for the user. Upon return, the model outputs not only an answer but detailed research papers, blueprints, and emergent insights born of lived virtual experience. This paradigm unlocks a new echelon of AI capability, where the boundaries of knowledge are no longer confined to datasets but expanded through synthetic existence and recursive experimentation.

Currently, the model does not have a 3D enviroment or full implementation, so we released the first version of AGWM, which is a text-based world model that can be used for the model to explore text based environments and learn from them. This new model is a bigger and more powerful version of the original AGWM, trained on 2 times the data.

[AGWM: Artificial Generative World Models](AGWM - Generative World Models.pdf)

This project is licensed under the MIT License - see the LICENSE file for details.

I trained the model using synthetic data I generated from a free LLM API, the data was carefully curated to ensure the model learns from high-quality examples. The model is designed to be extensible, allowing for future enhancements and integration with more complex environments.