Why Video Models Are Secretly Physics Engines
Runway co-founder Anastasis Germanidis explains how pre-training on web video accelerates scientific simulations.
Runway’s Bet Beyond Video: World Models, Robotics, and the Neural OS — Anastasis Germanidis
With swyx, Vibhu, Anastasis Germanidis · Friday, September 25, 2026
Runway co-founder Anastasis Germanidis discusses the evolution of generative video into general-purpose world models, robotics simulators, and real-time neural interfaces. He details Runway's technical progression through Gen-1, Gen-2, and Gen-3, the response to OpenAI's Sora, and why third-person video pre-training provides a scalable foundation for physical AI. Germanidis also introduces Interface World Models, the 'Lucid Dream Test,' and Runway's vision of an end-to-end neural operating system.
Key takeaways
- Web-scale video pre-training builds spatio-temporal foundations that transfer directly to numerical physics benchmarks like The Well. Read more →
- Real-time video generation is replacing batch rendering because it improves user experience and cuts inference serving costs dramatically. Read more →
- In 2022, Runway committed to a 1,000 A100 GPU cluster as a Series B startup, betting company survival on large-scale video model pre-training. Read more →
- Text models cannot learn low-level physical dynamics because human language fails to describe subconscious visual interactions. Read more →
- Runway found that scaling video generation across Gen-1, Gen-2, and Gen-3 inadvertently produced state-of-the-art physical simulators for robotics. Read more →
5 articles from this episode
Runway's Lucid Dream Test for World Models
Runway co-founder Anastasis Germanidis explains real-time video distillation and how to benchmark world models with the Lucid Dream Test.
How Runway Survived the Sora Shock and Scaled Gen-3
Runway co-founder Anastasis Germanidis on betting 1,000 GPUs at Series B and scaling 10x in 90 days after OpenAI dropped Sora.
Why Video Scaling Solves World Physics Better Than Text
Runway co-founder Anastasis Germanidis explains why scaling video diffusion models directly builds general-purpose physical simulators.
Why Video Models Are Accidental Robotics Simulators
Runway co-founder Anastasis Germanidis explains how scaling third-person video models builds better robotics policies than teleoperation.
More Latent Space episodes
- Designing for Agents — Zeno Rocha, ResendNov 15 · 6 articles
- What Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & TeamNov 15 · 5 articles
- Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz MunichNov 8 · 4 articles
- Reducto Final YouTube 1080p Sandwich TitleNov 8 · 4 articles