Current large language models are trained on human-generated data and thus distill a mixture of human intelligences, constraining them to near-typical human intelligence until they are embodied and can interact with and query the environment independently.