Why this CEO thinks video games make better training data than the internet

Artificial intelligence researchers are increasingly recognizing that video games may provide more effective training data than traditional internet sources for developing sophisticated AI systems. This shift in perspective challenges conventional approaches to machine learning and raises important questions about how artificial general intelligence (AGI) might be achieved beyond current large language models.

Large language models like ChatGPT and Claude excel at processing text but struggle with spatial reasoning and temporal understanding—critical capabilities for more advanced AI systems. Video games, with their complex 3D environments, physics simulations, and dynamic interactions, offer rich training grounds for teaching machines how objects move through space and time. Unlike static internet data, gaming environments provide continuous, interactive scenarios where AI systems can learn cause-and-effect relationships, physics principles, and environmental navigation in ways that mirror real-world complexity.

The structured nature of video game worlds also provides clear rules, measurable outcomes, and consistent feedback mechanisms that facilitate more efficient machine learning. This contrasts sharply with the messier, less structured data found across the internet, which often contains contradictions, bias, and contextual ambiguity.

• Video game developers may become key partners in AGI research, fundamentally altering the tech landscape
• AI training methodologies could shift from internet-scale data collection to curated, physics-based simulations
• Companies investing in game engines and virtual environments gain competitive advantages in AI development
• The resource requirements for training AI systems may decrease through more efficient, focused data sources
• Spatial reasoning capabilities could dramatically improve across robotics, autonomous vehicles, and embodied AI applications

This perspective represents a significant departure from the prevailing strategy of scaling models on internet data. If video games prove superior for developing AGI, it suggests that quality and relevance of training data may ultimately outweigh raw quantity. This could democratize AI development by shifting focus from massive data collection to sophisticated simulation design, potentially accelerating progress toward more capable AI systems while raising important questions about how these technologies are developed and deployed responsibly.