Earlier quoted context omitted.
I've recently been enjoying thinking of GPT-3 utterances as creating short-lived consciousnesses for the length of an interaction. There's no consciousness during training (just filling in the blanks...) but when we interact with it in a generative fashion, there is - depending on your prompt - a somewhat coherent 'I' that is invoked, and maintained over the course of the conversation. Seeing as I have no strong clai…
I don't think it's maintained over the conversation. The recent history of the conversation is fed to the engine with each new interaction. So the potential flash of consciousness is even shorter than you think. Each reply is a new flash, without memory, only connected by the written content of recent text.
Quite reminiscent of 'Permutation City' in fact... (In which uploaded human consciousness is slowed down arbitrarily... And worse.)