Game Footage as Control Supervision

The important signal is that General Intuition is proving a cheaper way to teach causality to action models. MIRA shows that a model can watch gameplay paired with exact button presses, then learn not just what Rocket League looks like, but how steering, boosting, collisions, and scoring unfold step by step. That matters because robots and agents also need input to outcome data, not just pretty video, if they are going to act reliably in the world.

• Gameplay clips are unusually useful training data because they already contain the missing supervision layer. Medal records the video and the player actions together, which turns each clip into a small cause and effect lesson instead of a passive movie.
  1 medal 2 mirawm 4 sacra
• Rocket League is a hard testbed, not a business line. It has fast motion, multiple players, partial observability, and constant object interaction, so a model that stays coherent there is showing it can learn dynamics that look more like real control problems than static video generation.
  2 mirawm 3 arxiv 5 kyutai
• The monetizable output is the action model trained inside the learned simulator. The same workflow appears in the company's robotics stack, where agents are pretrained and evaluated in simulation first, then adapted to real hardware with much less live data collection.
  4 sacra

This points toward a broader data strategy where game footage becomes a factory for embodied AI training. If General Intuition can keep pairing action labeled interaction data with world models, it can expand from game environments into robotics and other control systems where the winning product is not the simulator itself, but the policy trained inside it.