Why video games might make better training data for AGI than the internet

The article exposes a core tension in current AI development: large language models excel at text but fundamentally lack an understanding of how objects and agents move through physical space and time. This gap limits their ability to generalize to real-world tasks that require spatial reasoning and physical intuition, making internet text data insufficient for achieving AGI. The key insight is that video game data, which encodes dense interactions in simulated 3D environments, might be the missing ingredient.

General Intuition, a New York-based startup backed by Jeff Bezos and valued at $2.3 billion, just closed a $320 million round led by Coatue, with participation from Eric Schmidt, MIT researchers, and Google DeepMind. CEO Pim de Witte argues that gaming data is far richer for training world models than internet text, because games contain complex physics, causality, and multi-agent interactions. The company spun out of Medal TV, a gaming clip platform, and is now focused on building models that can understand the causal fabric of the physical world.

The takeaway for serious builders is that the next leap in physical AI may come from alternative data sources like games, not from scaling text pretraining alone. However, the article also highlights ethical red lines, particularly around defense applications, reminding the community that dual-use risks must be addressed early. For engineers and product leaders, this signals a shift toward synthetic, structured environments as a primary training ground for embodied intelligence.