I've come to the same conclusion. AGI(and each separately) is better understood as a epistemological problem in the domain of social ontology rather than a category bestowable by AI/ML practitioners.

The reality is that our labeling of something as artificial, general, or intelligent is better understood as a social fact than a scientific fact - even if purely the role of operationalization of each of these is a free parameter in their respective groundings which makes it near useless when taking them as "scientifically" measurably qualities. Any scientist who assumes an operationalization without admitting such isn't doing science - they may as well be astrology at that point.

Whether LLMs will be a base technology to AI, we should remember one thing: logically it's easier to convince a human that a program is sapient than to actually make a program sapient, and further, it's easier still to make a program do spookily-smart things than it is to make a program that can convince a human it is sapient. We're just getting to the slightly-spooky level.