By the most-accepted definition of AGI, if it's stupid sometimes, it's not AGI. At least, not if it's stupid as often as LLMs are.
For example, one of the stupid things we do is fail to realize that "intelligent" is a word with multiple senses. Which then makes a muddle of conversations like this because we all - myself included - tend to switch back and forth between senses of the term without even realizing we're doing it. So we have sense A of "intelligent", which means something like "capable of abstract reasoning". And we have sense B of "intelligent", which means something more like "the opposite of stupid". These two meanings don't actually have much to do with each other. But a lot of the hype around LLMs and the possibility of them being a form of AGI tends to hinge on tacitly assuming that an artificial system that can do abstract reasoning tasks won't do stupid things. Simply because sense A has one characteristic, sense B has the other, and we happen to use the same word to describe both.
But the definition I had in mind was the one in Wikipedia, "matches or surpasses human capabilities across virtually all cognitive tasks." Well, if it's going to do that, it can't be stupider than capable humans nearly as often as it is.
It’s things like that that leave me convinced it’s not really AGI. The symbolic reasoning element doesn’t really seem to be there. And I’m not prepared to call something general intelligence if it can’t do that.
Also, maybe even more to the point, it can’t even learn. No, pre- and post-training are not the same thing.