"In theory" doing a lot of work there. You don't know that would be analogous to AGI and how far we are from that being feasible in real time, computationally is anybody's guess. There are already multiple models running concurrently in ML-based applications.
See, the problem I have with this type of discourse is that subtle but critical leap you make halfway through your post between realistic, practical concerns and sci-fi. A LLM can absolutely cause harm if it's widely used, implicitly trusted and it responds to deliberate or accidental biases. Absolutely.
Granted, that is also true of every search engine and social media algorithm that's already in place. But it's true.
But the way you present it, sandwiched between the incorrect impression that AGI is just a matter of hyperlinking a bunch of neural networks makes it seem like the LLM would be doing this consciously, instead of stochastically in the same way other automated data processing does it. Or that this is a new concern that we aren't dealing with right now. Or that the major asterisks that this would require a much better implementation and a much broader adoption than we currently have are removed from play.
And that's the caveats for the problems that are genuine, real and practical. The sci-fi part is what people are actually scared about and we're seriously not there yet. And you haven't outlined a problem here that can't be fixed by power cycling a computer, which is an entirely different conversation as well.
Look, it's fine. Speculating about science and its impact in society is healthy. I'm just annoyed when things go memetic in unreasonable ways at the expense of similar, much more pressing issues that aren't as flashy. I lived through Y2K and the cloning panic, which both made daily headlines. And then I lived through the whole of humanity getting brainwormed by social media and you can barely get the EU to sometimes wag a finger at Facebook.