What's this about?
Code as Worlds has agents discover *executable* world representations for physical reasoning — instead of predicting pixels, write a program that models the situation and run it. CLAP shows cross-embodiment video world models can act as zero-shot physical simulators, escaping the usual restriction of action-conditioned video models to a single robot body.
Why write a program instead of predicting pixels?
Because a program is checkable, composable and cheap to run a thousand times. If the model's understanding of a situation is expressed as code — these objects, these masses, this constraint — you can execute it, test it against observation, and edit it when it is wrong. A pixel prediction gives you an image you must judge by eye. It is the difference between a model that has a theory and a model that has a memory of similar-looking things.
Why does this matter to you?
- Executable representations are inspectable, which is the property every safety and auditing argument this month has been asking for.
- CLAP's cross-embodiment claim addresses the biggest practical limit in robot learning: data collected on one robot is useless on another.
- Both are the "structure beats scale" position that Anima Anandkumar argued for physics on Latent Space last week.