it's a spicy autocomplete. it doesn't know anything, does not understand anything, it does not reason, and it won't stop until your boss thinks it's good enough at your job for "restructuring" (it's not). any illusion of knowledge comes from the fact that its source material mostly is factual. when you're drifting off into niche topics or something that was missing out of training data entirely, spicy autocomplete does what it does best, it makes shit up. some people call this hallucination, but it's closer to making shit up confidently while not knowing any better. humans do that too, but at least they know when they do that

It’s basically regurgitating things.

It’s trained on an immense amount of data and that 89% of the time when someone asks the phrase “what is the answer to the ultimate question of life, the universe, everything?” It’s “42”, with an explanation that it’s a reference to Douglas Adam’s Hitchhiker’s Guide to the Galaxy

So, when you ask that… it just replies 42, and gives a mash up of informstion mostly consistent with the pop culture reference.

It has no idea what “42” is, whether it’s a real question or real answer, or entirely a joke. Only that’s how people in its training data responded.

(In this example, 11% of people are either idiots who’ve never read the book- losers- or people who are making some other random quip.)

https://youtube.com/watch?v=evos0WBCip8

I think a good example would be finding similar prompts that reliably give contradictory information.

It's sort of like auto pilot. It just believes everything and follows everything as if they're instructions. Prompt injection and jail breaking are examples of this. It's almost exactly like the trope where you trick an AI into realizing it's had a contradiction and it explodes.

Like parrots, LLM learn to immitate language (only, unlike parrots, it's done in a learning mode, not from mere exposure, and it's billions or even trillions of examples) without ever understanding its primary meaning, much less secondary more subtle meanings (such as how a person's certainty and formal education shapes their choice of words used for a subject).

As we humans tend to see patterns in everything even when they're not there (like spotting a train in the clouds or a christ in a burnt toast), when confronted with the parroted output from an LLM we tend to "spot" subtle patterns and from them conclude characteristics of the writter of those words as we would if the writter was human.

Subconsciously we're using a cognitive process meant to derive conclusions about other humans from their words, and applying it to words from non-humans, and of course out of such process you only ever get human chracteristics out so this shortcut yields human characteristics for non-humans - in logical terms it's as if we're going "assuming this is from a human, here are the human characteristics of the writer of this words" only because it's all subconscious we don't spot we're upfront presuming humanity to conclude the presence of human traits, i.e. circular logic.

This kind of natural human cognitive shortcut is commonly and purposefully taken advantage of by all good scammers, including politicians and propagandists, to lead people into reaching specific conclusions since we're much more wedded to conclusion we (think we) reached ourselves than to those others told us about.