Runway co-founder and CEO Cristóbal Valenzuela presents a 17-minute AI Engineer talk on video models, interactive interfaces and early Continuum research.

From a walking orange to world models

  • He opens with Runway’s 2019 film, generated at 128 × 128 pixels and enlarged for display.
  • He argues that predicting successive video frames at scale teaches relationships involving motion, gravity, occlusion and object permanence.
  • Runway evaluates physical and geometric behavior alongside creative appeal, he says.
  • His claim that viewers can no longer distinguish generated footage comes without a study protocol or accuracy figures in the talk.

Separate generated clips from interactive demos

  • Valenzuela explicitly says the earlier clips took seconds or minutes to generate.
  • His next examples introduce actions during generation, making the output react to user input.
  • A car interface changes lights, orientation and the hood; another example lets users move objects in a room.
  • He describes these as world interface models, without an explicitly modeled 3D scene underneath.
  • Runway’s robotics work is another proposed application; the talk provides no robot-performance evaluation.

Continuum turns reading into visual exploration

  • Continuum is an early research prototype for generating interactive visual experiences from screen content.
  • A selected passage about San Francisco fog becomes a video “portal.”
  • Previous portals, including one about batteries, remain available in a memory-like history.
  • A photosynthesis example changes its presentation when the window is resized and supports navigation through the explanation.

Give agents and people a shared visual workspace

  • Runway connects Continuum to a coding agent so it can visualize a problem, illustrated with a proposed bike-model task.
  • The person can inspect and guide that visual representation alongside the agent.
  • A Mars-search example generates an explorable outpost based on the user’s apparent intent.
  • These are creator-presented research demos; convincing imagery alone does not establish scientific accuracy or a correct simulation.

Treat the operating-system pitch as a research direction

  • Valenzuela imagines future screens streamed from video models and calls the direction an operating system.
  • The talk does not establish a shipped general-purpose operating system or replace conventional application backends.
  • He says Runway plans to release most of the demonstrated work over the following eight to twelve months.

“This is still very early research.” — Cristóbal Valenzuela, 11:14