Late but, regarding reasons for not being conscious, a repeated reason in that list is the inability to learn from experience, or learn even labeled data after training is complete... theres general problems thats specific to how training is done in an LLM now, but there are alternative architectures are around but are not used by tech giants as theyre unproven or not viable to train at scale yet. Particularly Id recommend reading about alternative learning architectures from the current meta, being back propagation, into something like forward forward; theres many variations of it but the idea is mostly the same. Its an architecture that does not require a backwards pass, and as a result do not require extra infrastructure at training time vs inference time. Additionally it can use constrastive methods to learn at inference time based on unlabeled data, so this can be used to "learn on the fly"
Statefulness is another thing that can be addressed in RNNs, which exist in LLM formats like RWKV, but need some adjustments to have a true state that can expand and collapse in scope in a structured way, so statefulness is definitely on the horizon...
The real issue is externalized knowledge lookup, the model cannot untagle syntactical knowledge from trivia or contextual knowledge, and so its not simple to develope a sort of lookup db that can be expanded on the fly... so I believe thats still the missing piece.
My 2 cents. conscious as you said is not well defined, but from my readings on it I can say that I believe in levels of consciousness, and that these components would bring it to a higher level of being than it is now.
