CEO argues video games provide better training data for AGI than internet text

The pursuit of artificial general intelligence (AGI) faces a critical challenge: while large language models (LLMs) excel at text-based tasks, they struggle to grasp dynamic spatial and temporal relationships—a foundational capability for adaptable intelligence. Models like ChatGPT and Claude, trained primarily on static text data, lack the ability to predict or reason about how objects interact in physical environments over time. This limitation has prompted researchers to explore alternative data sources to bridge the gap, with gaming data emerging as a promising solution.

General Intuition, a startup focused on advancing AGI, is leveraging gaming environments to train models in understanding motion, causality, and spatial reasoning. By simulating dynamic scenarios where objects move and evolve in real-time, gaming data offers a rich context for models to learn physics-based patterns and adapt to changing conditions. The company’s approach combines synthetic game environments with real-world tasks, aiming to develop systems that can generalize knowledge across diverse domains. This method addresses a key shortfall in traditional LLMs, potentially paving the way for more versatile AI capable of reasoning beyond text.

The shift toward gaming data highlights a growing recognition that AGI requires multimodal learning, integrating visual, spatial, and temporal information. While still in its early stages, this strategy underscores the importance of dynamic training environments in building intelligent systems that mimic human-like adaptability. As research progresses, the success of such approaches could reshape the trajectory of AI development, emphasizing the need for data that captures the complexity of real-world interactions.