it does understand their relationships. the difference are the subjects of the domain it learns about exist purely as 2 dimensional images. once these models

it does understand their relationships. the difference are the subjects of the domain it learns about exist purely as 2 dimensional images. once these models get larger and include wider ranges of sensory data beyond just imagery (as can be seen with models like GATO), they are clearly better able to "glue together" concepts across multiple domains. i would argue we absolutely do nothing different with regard to 'gluing things together'. we just have a wider range of sensory inputs.