Runway’s Bet Beyond Video: World Models, Robotics, and the Neural OS — Anastasis Germanidis
Source
youtube.com
Author
Latent Space
Date
Context
Runway is exploring whether generative models can render software interfaces directly as pixels. In the interview, its cofounder discusses interface world models that produce the visible software surface instead of generating conventional webpage code. That expands the company's video-generation research toward interfaces whose appearance is itself a model output. A fully neural operating system is presented as a possible destination for this work, rather than an established replacement for existing operating systems.
Key takeaways · AI-distilled
Runway made an early bet on a cluster of 1,000 A100s to kickstart its video research.
Gen-2 grew out of a weekend experiment combining text, depth and video, and after OpenAI released Sora the team scaled Gen-3 by roughly 10x in a three-month push.
Germanidis argues third-person internet video could be the largest training source for robotic intelligence, and describes World Action Models that turn video models into robot policies.
He proposes a "Lucid Dream Test" for judging when world models are good enough, and says real-time video generation is inevitable.
