General Intuition leverages 3.8 billion action-labeled game clips from Medal to train vision-based agents that perceive frames and output actions in real time, achieving human-level and sometimes superhuman performance in dynamic environments. These world models go beyond video generation by incorporating interactivity, memory, and partial observability—learning from smoke, occlusion, and camera shake as meaningful signals. The team has successfully distilled large policies into compact, real-time models deployable on consumer hardware, enabling transfer from arcade games to realistic simulations and real-world video. Rejecting a $500M offer from OpenAI, the founders chose independence to build proprietary capabilities, using their data moat to pioneer 'frames in, actions out' APIs for gaming, robotics, and simulation. They argue world models and LLMs are complementary: while LLMs excel at orchestration, world models provide spatial reasoning essential for real-world interactions. By treating game highlights as 'episodic memory of simulation,' they enable reinforcement learning at scale. With a $134M seed from Khosla, the vision is for spatial-temporal foundation models to power 80% of atoms-to-atoms interactions by 2030—primarily through scalable simulation rather than physical deployment.