Not that I agree they’re conscious, but this is an incorrect and overly simplistic definition of a LLM. They are probabilistic in nature, yea, and they work on tokens, or fragments, of words. But it’s about as much of an oversimplification to say humans are just markov chains that make plausible sentences that can come after [the context] as it is to say modern GPTs are.
Grok is closed source, I believe, so it’s hard to say. But, ignoring unknown architecture or latent space details, this could be a lot of things. The way you seem to be using the term hallucination effectively applies to EVERY output of a GPT. They effectively reason probabilistically across a billion dimensioned space mapping language components, with various dimensions taking on various semantic values due to a sort of mathematical differentiation during training. This could be the result of influence from any number of things tbh.
Yea, this is what gets me and what news media needs to actually talk about. The texts are a security leak and sign of unfathomable idiocy and incompetence, sure, but there are missiles literally designed to target and blender people in an area the size of a car and nowhere else. They deliberately did this in the most irresponsible, collaterally devastating way possible. These people carefree murdered a bunch of children and families just trying to survive. Every single participant of that chat should be swaying in the breeze under a highway overpass right now.
Even better, gorditas literally translate to little fat girls i believe (but maybe I’m wrong). Tbf, we have a processed sausage-alike that we call a hot dog so language and our perception of it is weird