Right, I think the Skyrim example another commenter linked is a good showcase of how that model had no real conception of what it was "doing" and no real cognitive ability to seek an understanding.

I do think that some kind of "embodiment" or at least the addition of some kind of structured semantic model on top of/in addition to the raw token/vector embeddings is probably going to be necessary.

I am interested though, in how much of that internal/semantic understanding can be "emulated" by a sufficiently rich text model over the course of a conversation. There was an interesting article a little while back about improving GPT-3's (IIRC) performance on math problems by structuring the prompts so that it would describe working out all the intermediate steps in solving the problem.

I do suspect that whatever additional pieces we need to reach real general AI, the first really convincing examples of it will come not from looking at the architecture of the raw model, but at how it can be prompted to generate a believable accounting of it's own internal states over the course of an interaction.