General Intuition's game-trained world models aim to bridge AGI's physical gap

Large language models like ChatGPT and Claude excel at text but struggle with understanding how objects move through space and time — a critical gap for achieving artificial general intelligence (AGI). General Intuition, a New York-based startup valued at $2.3 billion, believes video game data can bridge that gap. The company just closed a $320 million funding round led by Coatue, with participation from Jeff Bezos, Eric Schmidt, researchers at MIT, and Google DeepMind. CEO Pim de Witte explains that gaming environments provide rich, physics-grounded interactions that train 'world models' — neural networks that grasp spatial and temporal dynamics. General Intuition spun out of Medal TV, a gaming platform, allowing it to leverage proprietary gaming datasets.

The company's approach targets physical AI applications, from robotics to simulation. World models trained on gaming data could enable machines to predict outcomes and plan actions in the real world, a key step beyond static text-based reasoning. However, de Witte acknowledges ethical red lines, especially as models might be used for defense. The $320M raise — one of the largest in the AI world-modeling space — signals strong investor belief that gaming data is the missing ingredient for AGI. General Intuition plans to scale its models and partner with enterprises needing embodied AI.