Runway Leadership on GWM Worlds 2, WorldPrompt, and Real-Time World Models
In this deep dive featuring interviews with Runway CTO Kamil Sindi, Principal Research Scientist Robin Kahlow, and co-founder and co-CEO Anastasis Germanidis, Latent Space examines Runway's GWM Worlds 2 research preview. The system streams real-time interactive worlds in continuous 720p video at 24 fps alongside 48 kHz audio. A central innovation is WorldPrompt, an input format functioning as a control layer for characters, cameras, and environments using timestamped events. Architecturally, Runway adapts foundational bidirectional diffusion models into causal autoregressive models that generate frame by frame, followed by distillation methods such as reducing denoising steps from roughly 50 down to four. The leaders highlight primary engineering hurdles, including error accumulation as generated frames feed back into the model and managing context history to prevent GPU memory exhaustion.
Original sources
Podcast Latent Space
Episode:Runway’s WorldPrompt and the Engineering of Real-Time Worlds ↗
Newsletter Latent Space
Issue:Runway’s WorldPrompt and the Engineering of Real-Time Worlds ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.