It might ultimately be useful for it to be in the neural architecture, though, so the machine can imagine those things.

You could be right, though. Maybe it's not needed, and the inputs combined with generative language/audio/imagery is enough 🤷‍♀️