Daily AI CatchupKyutai's MIRA Renders a Playable Rocket League Match With No Game Engine
Kyutai and General Intuition released MIRA, an open-source world model built with Epic Games that generates live 2v2 Rocket League entirely inside a neural network — no game engine underneath. Trained on 10,000 hours of AI-vs-AI gameplay, it renders boost meters, crashes, and multi-screen sync on a single GPU at 20fps, with code, training data, and a playable demo all open-sourced.
Why it matters
💻 Developer · The code, training data, and playable demo are all open-sourced — worth a look if you're exploring world models for simulation or synthetic environment generation.
📦 Product · This isn't really about replacing video games — it's a proof point for simulating physical environments to generate robot training data, which is a much bigger and more durable market.
🎨 Design · A neural network rendering coherent, physically plausible game state (boost meters, collisions, multi-screen sync) at playable framerates is a striking demonstration of how far generative world models have come.
📈 Business · World models capable of simulating physical environments are a foundational technology for robotics training data — worth tracking as an adjacent bet if you're in the physical AI space.
🤔 Just Curious · Researchers built an AI that can generate a playable video game — a full soccer-with-cars match — entirely by imagining it frame by frame, with no actual game engine running underneath.
Sources: MIRA world model