Why this CEO thinks video games make better training data than the internet
TL;DR
General Intuition is betting that video games provide better training data for AI agents than the open internet because games capture movement, space, time, and decisions in controllable worlds. CEO Pim de Witte argues that LLMs like ChatGPT or Claude are strong at text but weak at understanding physical dynamics and causality in a way that generalizes. The key is not just gameplay footage, but action labels: which button was pressed, when it was pressed, and what changed in the environment afterward.
Nauti's Take
The first thing to verify is transfer, not the demo: which out-of-game tasks improve measurably, and at what failure rate? For teams building robotics, simulation, or agent workflows, the action-labeled data is worth watching.
Rights to gameplay data, label quality, and the gap between game physics and real environments remain unresolved.
Briefingshow
If AI agents are meant to work in robots, factories, or vehicles, text training alone is not enough. Games offer a cheap, scalable middle ground: not real physics, but rich in actions, consequences, and spatial rules. The open question is whether that knowledge transfers reliably into the real world.
Sources