It's certainly impressive, but I haven't been giving it much attention. I do think its ability to store the information it's trained on (even if inaccurately and unable to understand the significance of it) and retrieve it through parsing natural language queries is something I'd like to better understand. Applying transfer learning to one of these large language models is worth looking into. I can assure you one thing though: the architecture of the human mind (or anything like it) has so many finely tuned parts and depends so much on the environment it's in that a language model alone cannot feasibly be trained to replicate that structure. For one, a lot of that information has not been written down yet and a lot of it also depends on values and structures which evolve in time where it would be very difficult to find training data that would be able to reproduce them.