“Once trained, world models are expected to be more lightweight and have fewer parameters than LLMs, leading to less expensive inference.”