Runway co-founder and co-CEO Anastasis Germanidis says the company has shifted from taking on Hollywood to real-time physics simulation with its general world model.
All major film studios use Runway across pre-production, production and post-production.
Runway's GWM-1 world model is in private beta and will be generally available in the coming months, with talks underway with Physical Intelligence.
Runway moves from generating video to simulating the world
Runway began as three art school friends trying to make deep learning models usable inside a creative tool, years before generative AI had a name. Film studios now use it across production, but the company's larger bet is the world model, an interactive system that predicts the outcome of actions in an environment rather than returning a finished clip. Runway argues that is the route to simulating physics, training robots and running experiments faster than the real world allows.
Takeaways
Runway's next bet is the world model. GWM-1, the company's world model, is in private beta and will be generally available in the coming months. It builds representations of environments and predicts the outcome of actions, unlike a video model that returns a finished clip from a prompt.
World models demand real-time interaction. A video model takes a minute or two to return output, while a world model has to be controlled continuously, and Runway's recent models generate at around 100 milliseconds per frame. Low latency is the hardest research and engineering problem.
World models train robots and cars. Runway generates data for robotics labs including Physical Intelligence, simulating tasks like folding clothes that traditional simulators handle badly. Simulated data only needs to correlate with reality, and world models can produce rare driving scenarios such as crashes and bad weather on demand.
Runway competes through focus and know-how. The team is about 10 times smaller than the big labs and has roughly 100 times less compute, so Runway leans on its in-house research team and years of accumulated know-how, and on being more data efficient. It released the first publicly available text-to-video model.
Every major studio already uses Runway. Runway is used throughout film production, from pre-production through post-production, for VFX and editing work, and has spread into marketing, advertising, games and architecture. A growing share of final frames is generated in Runway.
AI multiplies artists rather than replacing them. Germanidis cites Photoshop and photography, which he says created modern art rather than ending it. He expects Runway-made feature films in cinemas, and says the constraint on the world is supply rather than demand, with thousands of films released a year against millions of books.
From art school to a research lab
Germanidis left San Francisco tech after finding a monoculture and a drive to commercialise everything, and went to New York to study art.
He and his two co-founders, all self-described failed filmmakers, began collaborating around 2016.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Their first goal was an interface that made early deep learning models easy to use as a creative tool.
They realised they needed an in-house research team, and it took until 2022 or 2023 for results usable in production.
What Germanidis expects next
An ecosystem of infrastructure companies and aggregators on top of world models, mirroring what happened with earlier image and language models.
Feature films made with tools like Runway reaching cinemas, plus new media forms sitting between film and games.
World models trained on all observations, from human-scale video to scientific experiment data, so experiments run faster than in reality.