By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Runway’s Bet Beyond Video: World Models, Robotics, and the Neural OS — Anastasis Germanidis

From making one of the earliest bets on generative video to building world models that simulate physics, power robots, and render software interfaces directly as pixels, Runway is pushing far beyond AI video generation. In this episode, Runway co-founder Anastasis Germanidis joins swyx and Vibhu to unpack how the company went from experimental creative tools to frontier video models — and why Runway now sees video generation as a path toward general-purpose world simulation. We go deep on the evolution from Gen-1 and Gen-2 to Gen-3 and Gen-4.5, including the thousand-A100 bet that helped kickstart Runway’s video research, the surprising weekend hack behind Gen-2, and the existential moment when OpenAI released Sora and people declared “Runway’s done.” Anastasis explains why scaling video models may be enough to learn physics, why real-time generation is inevitable, and how Runway is turning those models into simulators for robotics and interactive worlds. We also explore Interface World Models — software rendered directly by neural networks with no HTML, CSS, or React — and the possibility of a fully neural operating system. Anastasis lays out his “Lucid Dream Test” for world models, explains how third-person video could unlock robotics at scale, and discusses video agents, omni models, AI-native creative workflows, open-source world models, and what happens when models learn from more than just text, images, and video. We discuss: • How Runway went from creative tools to building frontier generative video models • Why Runway made a massive early bet on a cluster of 1,000 A100s • The Stable Diffusion story and Runway’s role in its development • Why text-to-video alone was never enough for professional creative work • How Gen-2 emerged from a weekend experiment combining text, depth, and video • Why camera control helped lead Runway toward world models • Why learning directly from the physical world may unlock capabilities language cannot capture • Whether simply scaling video prediction can teach models physics • What happened inside Runway when OpenAI released Sora • How the team scaled Gen-3 by roughly 10× in a three-month push • Why Anastasis believes real-time video generation is inevitable • Interface World Models and generating software pixels directly instead of generating code • Why the endgame could be a fully neural operating system • How Runway is using world models as simulators for robotics • Why third-person internet video could be the largest training source for robotic intelligence • World Action Models and turning video models into robot policies • The “Lucid Dream Test” for knowing when world models are good enough • Video agents, omni models, and why orchestration eventually moves inside the model • Open-source world models, Chinese video models, and the global race in generative video • Why artists are responding differently as AI tools become more controllable • The future of real-time creative tools, physical AI, multimodal models, and scientific simulation — Anastasis Germanidis • LinkedIn: Anastasis Germanidis • X: @agermanidis Timestamps 00:00:00 Hook 00:01:44 Introduction 00:07:01 Runway’s Origins and the Bet on Generative Video 00:14:07 The Stable Diffusion Story 00:20:28 Gen-2, Controllability, and the Weekend Hack 00:24:46 From Video Generation to World Models 00:29:47 Learning From the World, Not Just Language 00:36:48 Sora, Runway’s Existential Crisis, and Gen-3 00:41:23 Why Real-Time Video Is Inevitable 00:44:50 Interface World Models: Software Without Code 00:52:09 The Fully Neural Operating System 00:56:55 World Models for Robotics 01:04:16 Robot Policies and World Action Models 01:09:31 The Lucid Dream Test 01:13:25 Video Agents and Omni Models 01:24:56 Artists, AI, and Creative Workflows 01:28:58 Physical AI and the Future of World Models

Anastasis GermanidisguestswyxhostVibhuhost

Sep 25, 20261h 38mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

WHAT IT’S REALLY ABOUT

Runway’s path from generative video to world models and robotics

Anastasis Germanidis explains how Runway evolved from giving artists access to early open-source vision models into building a research organization that trains large-scale generative video models.
He describes key lessons from deploying video generation—especially that text-to-video alone is insufficient without controllability features like camera motion, motion brushing, references, and prompt-completion pipelines.
He frames Runway’s expansion from video generation into “world models” as a direct consequence of scaling: better video prediction requires richer simulation of physics, dynamics, and counterfactual outcomes.
He recounts how OpenAI’s Sora created an existential shock that accelerated Runway’s shift to diffusion transformers, distributed training infrastructure, and a larger-scale Gen-3 effort.
He outlines future directions including real-time video generation, interface world models that render software UIs directly as pixels, and robotics applications where video-pretrained models are adapted into simulators and action/policy models.

IDEAS WORTH REMEMBERING

5 ideas

Interfaces will become model-generated, not coded.

He argues today’s language models are general-purpose, but users still interact with them through rigid, hand-built interfaces; he expects the interface layer (rendering and pixels) to become learnable and delivered end-to-end with the model.

Runway’s trajectory follows a scaling-first thesis—from creative tools to frontier video training.

He describes moving from early creative tooling (open-source CV models for artists) to building a research organization and eventually training large-scale proprietary video models, driven by the belief that quality scales predictably with compute.

Control and conditioning mattered as much as raw video generation quality.

He says early text-to-video alone wasn’t enough for real creative workflows, pushing Runway to add conditioning methods like camera control and motion brush, and to use prompt-completion pipelines because “humans are terrible at prompting.”

Video prediction is treated as a path to general world simulation.

He frames “world models” as the next step: if models predict video well, they must increasingly simulate physics and dynamics; he claims Runway sees physics-benchmark improvements as model/compute scale increases.

Sora forced a fast organizational and technical scaling transition at Runway.

He describes Sora’s release as triggering a brief existential crisis, followed by a rapid push to scale Runway’s training (including diffusion transformers and model parallelism) and release Gen-3, emphasizing how much infrastructure Runway had to build itself.

WORDS WORTH SAVING

5 quotes

It was a indication that you could take a model that was trained on this very boring data set essentially of like not that many interesting things happen when you're on the road, and then you can repurpose it and go very out of distribution and make something that was artistically compelling.

— Anastasis Germanidis

If scaling laws apply on video just like they apply on language models, then as we scale the computer we put into those models, then they're gonna be able to simulate physics.

— Anastasis Germanidis

Understanding the dynamics of the world is very different than, uh, generating, uh, cute videos. Uh—Yeah, they're, they're the same thing.

— Anastasis Germanidis

If you ask anyone what is their favorite time at Runway that was there during that time, it was that, that push in like three months to get to a model better than, than Sora.

— Anastasis Germanidis

Effectively, the, the end game of something like interface world models is you have, uh, a fully neural operating system.

— Anastasis Germanidis

Runway origin thesis: generative media reshapes creative toolsGreen Screen/segmentation for VFX and rotoscopingGen-1 depth-conditioned video-to-videoGen-2 text-to-video via two-stage depth pipelineControllability: camera control, motion brush, referencesScaling infrastructure: diffusion transformers and model parallelismSora shock and Gen-3 scale-up push

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.