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 claim on a strict definition of consciousness, Imma gonna go ahead and call it conscious.

the thing is, this is /commoditized/ consciousness, which can be spun up and discarded at will, a million times a minute. Totally incapable of coordination or planning, in contrast to, say, SkyNet. The future is usually weirder than we imagine.