Why this CEO thinks video games make better training data than the internet
TL;DR
General Intuition is betting that video game data can teach AI more about space, motion, and time than scraped web text. CEO Pim de Witte argues that LLMs like ChatGPT and Claude are strong with language but weak at modeling how physical scenes unfold. The New York startup spun out of gaming platform Medal TV. Its goal is to train world models for physical AI, robotics, and potentially defense-related use cases, which makes the ethical boundary question more than theoretical.
Nauti's Take
The thesis is sharp because it hits a real weakness in the current LLM wave: text describes the world, but it does not force a model to predict collisions, motion, or follow-up actions cleanly. Still, a lot of this sounds like a funding narrative.
Game data is not the real world by default; it is a designed world with rules, shortcuts, and bias. Worth watching, but the AGI framing should not be treated as proof.
Briefingshow
The point is bigger than gaming: interactive worlds contain actions, consequences, and physics, not just descriptions of reality. That could matter more for robotics, simulation, and AI agents than another jump in text-only model scale. The catch is control: whoever owns these world models may influence systems that eventually act in physical environments.
Sources