They really are. Distance functions used on the encoded embeddings of sentences work stunningly well. While a primitive use case compared to true AI, the sheer trajectory of the effectiveness of these models is breathtaking.

Except that, as soon as you step out of the manifolds of embeddings representing the data, the model is happy to provide you with a completely clueless answer (which is bad) but even worse, is not even aware that it is being clueless (which is the really bad news).

Yes but that’s not necessarily progress towards consciousness.

I argue that the effectiveness of the matrix factorizations (creations of abstractions) in the embeddings points to the fact that the math works. Meaning, abstractions that are meaningful are created. Which is a building block of consciousness.

It's kind of a question of progress. Maybe being able to extract a complex abstraction through statistical means is close to 'understanding' or 'consciousness', or maybe abstractions are to consciousness what an amino acid is to a living organism. We don't understand consciousness well enough to confidently say.

What's your definition of consciousness?

I don't know what the correct definition is, but I will take the author's definition for the sake of argument:

>In this view, consciousness isn’t a mysterious ghost in the machine, but merely the word we use to describe what it’s “like” to model ourselves and others.

>When we model others who are modelling us in turn, we must carry out the procedure to higher orders: what do they think we think? What might they imagine a mutual friend thinks about me?

While I certainly buy that this sort of modeling is one effective way to be good at the language modeling task, I am unconvinced that it is the only way, and thus the fact that the networks are undeniably very good at the language modeling task does not necessarily imply they are doing so through consciousness (as defined above), or something on the road to consciousness.