Why this CEO thinks video games make better training data than the internet
TL;DR
General Intuition argues that LLMs like ChatGPT and Claude are strong at text but weak at space, time, and motion. Gaming data is pitched as a missing ingredient for world models and physical AI. The startup spun out of Medal TV and uses gameplay clips plus button and mouse actions, so models can learn behavior rather than only infer it from video. General Intuition is valued at $2.3 billion, raised $320 million, and lists Coatue, Eric Schmidt, MIT researchers, and Google DeepMind researchers among its backers.
Nauti's Take
The idea is strong because it hits a real weakness in the current AI wave: text models can talk about the world, but they do not have a body. Gameplay data at least contains traces of action, timing, and consequence.
Still, parts of the story smell like a Valley grand narrative: a $2.3 billion valuation, AGI language, and robotics demos. The practical proof starts when these models behave reliably outside carefully chosen demos.
Briefingshow
If AI can learn from gameplay, training is no longer just about text, web pages, or synthetic datasets. Millions of tiny human-machine decisions become valuable: jump, wait, dodge, aim, fail. For robotics, simulation, and agents, that is a different data source than the open web, but it also raises new power questions around proprietary platform data.
Sources