Background
Germanidis's early work (2015–2018) was in creative applications of machine learning — interactive film installations, GAN experiments, video-analysis tools — before co-founding Runway with Valenzuela and Alejandro Matamala Ortiz out of NYU ITP. Runway shipped Gen-1 through Gen-4 and landed on TIME's 100 Best Inventions of 2024 for Gen-3 Alpha.
By 2026 Runway is valued at around $5.3B and has repositioned around world models, robotics and real-time video, with a dedicated Runway Robotics effort and a research claim that simulated and real-world outcomes correlated at 0.95 across eight robot policies. His essays 'Towards Universal Simulation' (Oct 2024) and 'Real-World Superintelligence' (Jul 2026) set out the thesis.
What they run now
Runway's research direction: general world models (GWM Worlds 2, with what he calls a general action space merging navigation, manipulation and human interaction), real-time video, Runway Robotics, and a developer platform. He explicitly frames large-scale video pretraining as the advantage over domain-specific logs — a model trained on general video already knows how a pedestrian stumbles — and frames world models as evaluation substrate for policies, not just content generation.
What they buy
Inference. Runway's buying is shifting: less 'is this shot beautiful' and more 'does this simulation behave correctly'. That points toward physical-plausibility annotation, action-conditioned video data, robot-policy evaluation, and world-consistency rubrics. Creative taste data still matters for the film and enterprise business (Runway Studios, brand work) but it is no longer where the research money is going. No data vendor is named publicly.
Career
- 2018 – presentCo-founder & Co-CEO, Runway — Previously CTO; co-CEO with Cristóbal Valenzuela by 2026
- c. 2016 – 2018Researcher / creative technologist, NYU ITP
- c. 2016–2018New York University (ITP), Graduate work, interactive telecommunications (unconfirmed)
On the record
Says world models will become the RL environments for essentially everything. source
Argues broad video pretraining covers long-tail events that a driving fleet would log only a handful of times. source
Cites 0.95 correlation between simulated and real outcomes across eight robot policies as evidence world models can replace physical evaluation. source