A tech CEO recently argued that video games could provide superior training data for artificial general intelligence (AGI) compared to traditional internet text. The claim centers on a critical weakness in current large language models (LLMs) like ChatGPT and Claude: their inability to fully grasp how objects move through space and time—a fundamental requirement for AGI.

The CEO explained that LLMs excel at processing static text but lack the embodied understanding of physical dynamics necessary for true intelligence. Video games, with their interactive 3D environments, offer rich data on spatial relationships, causality, and temporal sequences, potentially overcoming these limitations.

For content creators, this insight has practical implications. Many AI tools in use today rely on text-based models that struggle with tasks involving motion, layout, or physics—think video editing, animation, or virtual set design. Training on game data could lead to AI that better understands these elements, enabling more intuitive assistants for multimedia production.

Such advancements might empower creators to automate spatial work, like generating realistic scene transitions or simulating object interactions, streamlining workflows in gaming, film, and mixed reality. However, the proposal remains theoretical; the shift from internet to game training is still exploratory, with no immediate products or timelines specified.

As the industry debates the future of AI training, creators should monitor how this perspective shapes tool development. If adopted, it could redefine AI’s role in crafting immersive content, offering new efficiencies for those working in dynamic visual media. For now, the CEO’s comments underscore a growing focus on bridging the gap between text and physical intuition in AI research.