Build1 publisher2 min readPublished
Runway extends its video models to robots with open-weight Praxis-1
Runway debuted Praxis-1, an open-weight world action model it says adapts to new robot bodies with light fine-tuning, alongside an ad agent called Runway Ads. Because the weights are open, outside robotics teams can measure what that fine-tuning costs on their own machines.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Runway is betting that the scaling laws behind large language models will carry video models to world models and robotic generalization, The Deep View reported.
- Runway showed the products Wednesday at its Runway AI Summit, along with other uses for its video technology outside creative fields.
- Co-CEO Cristobal Valenzuela said the same technology that makes compelling videos for Hollywood and films now powers robots.
- Valenzuela said Runway Ads goes past generating ads: the agent also edits them, publishes them and tracks their performance.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure An agent that publishes ads and tracks their results needs write access to ad accounts and read access to campaign data. The account owner has to grant both before it can run.
- constraint Runway's scaling bet needs more training compute, and Valenzuela says compute is already slowing Runway and others in ways the company cannot directly control.
- precedent Valenzuela expects agents in every part of Runway, so customers should expect the access questions Ads raises to come up again in Runway's other products.
"I'm particularly excited about realizing that video models and world models are generalization machines," Valenzuela told The Deep View [5]. Stated as engineering, the claim is narrow and testable. For it to hold, a model pretrained to predict video must have learned enough about motion and contact to carry over from pixels to motor commands. Runway calls Praxis-1 a world action model, and it is the company's test of that claim [2].
Runway says Praxis-1 adapts to a new robotic embodiment with light fine-tuning [2]. A robotics team would want "light" turned into numbers. That means demonstrations or teleoperation hours per new body, training steps, and a success rate on tasks held out from tuning. The launch coverage does not include those figures, a parameter count, or license terms. Shipping the weights is still the right call for a generalization claim, because any benchmark Runway ran itself would only describe Runway's robots and Runway's tasks [2].
The scaling bet also has a seam. Interviewer Nat Rubio-Licht described Runway's stance on language models as "we've reached a saturation point" [7]. The same newsletter describes Runway as betting on the scaling laws behind those models [4]. Both positions can be true only if returns to scale have flattened for text but are still rising for video.
Valenzuela's agent plans start with ads. He said "ads are our first stepping stone towards that," with "that" meaning agents that work in pixels the way coding agents work in code [9]. He points to chat as the model. Coding models "are now just agent models, and you can just text them in like iMessage, and they will do the work for you," he said [11]. For fiction, he also wants to build "an agent that's a creative partner" [12].
What to watch
- Runway publishing Praxis-1's license, parameter count and the fine-tuning data it needs per new embodiment.
- Independent fine-tunes of Praxis-1 on robots Runway never trained on, reported with success rates on held-out tasks.
- Which ad platforms Runway Ads publishes to, and what account permissions it requests to do so.