OpenAI published a technical report on 15 February 2024 describing its video generation model as a "world simulator" [source]. The report frames the model's ability to generate video as simulating physical environments, stating that "we explore large-scale training of generative models on video data" and that "we find that video models exhibit a number of interesting emergent capabilities when trained at scale."

The document describes the model's capacity to generate videos depicting "3D consistency, long-range coherence and object permanence" and claims these outputs suggest the model is learning "a simulator of the physical world." OpenAI states the model can generate videos conditioned on text prompts and can extend existing videos forward or backward in time.

The report does not provide quantitative benchmarks for physical accuracy or define what constitutes a "world simulator." No comparison is made between the model's outputs and ground-truth physics simulations. The document acknowledges current limitations, noting that "Sora sometimes creates physically implausible motion" and "may struggle with accurately simulating the physics of a complex scene."

The framing has drawn attention in the research community. The report positions video generation as a step toward "simulators of the physical and digital world, and the objects, animals and people that live within them," but does not specify how generated video outputs differ from statistical pattern matching in training data.

OpenAI has not released the model publicly. The report includes sample videos and technical architecture details but no independent verification of the claimed emergent capabilities. The document was published on OpenAI's blog and remains accessible as of this report.