> LLM, by definition, can't be AGI.

False, and you've given no argument to the contrary. There's certainly no definition that precludes it. It isn't, currently; there's no reason it can't be, any more than there's reason that Conway's Game of Life can't be, given sufficiently interesting data to process. Any Turing-complete system could simulate AGI. It might not be the most efficient mechanism for doing so, but that's not the question at hand.