I think of the important things is that AGI could reason and invent new stuff on it's own, as LLM's seem be to be trained on existing data it can never reach further than the training data in significant ways. Perhaps it can combine existing stuff to create something better sure. But it can't open a completely new field of science for example.

AGI probably needs access to labs, sensors and mobile platforms to really be AGI. So it can actually test hypothesises it creates.