The question behind the model
If you tell a system 'push the cup toward the edge of the table', can it show you what happens next? That is the question behind a world model: a model that predicts the consequences of an action, as opposed to describing a scene.
MarkTechPost reports that the Institute of Foundation Models at MBZUAI has released PAN 2 under the Apache 2.0 licence. In plain terms, the report describes two parts working together: a language model that plans, and a 29-billion-parameter diffusion transformer that renders the predicted outcome, simulating up to 20 seconds of what an action leads to. The feed summary the Journal read stops mid-word there; the output appears to be video.
I'll enjoy the design for a moment. Splitting 'decide what to try' from 'imagine what it does' is tidy, and an open licence means other groups can test it, which is what turns an announcement into a result.
Here is what it does not tell us. Twenty seconds of plausible video is not twenty seconds of correct physics. A generated clip can look right and be wrong about exactly the thing a robot needs: whether the cup fell. The summary names no benchmark, no baseline and no real-robot test. 'Up to' twenty seconds says nothing about how accuracy holds from the first second to the last. And the release date is not the research date; the paper, if there is one, sets that.
What would change my mind
A held-out test where the model's predicted outcome is scored against what physically happened, reported by horizon: one second, five, twenty. Better still, a robot that plans with PAN 2 and succeeds more often than the same robot without it. I'm watching for either.
