Part of the problem is that AI research likes to use terminology that sounds like what people do, when that’s not what the AI actually does.
Large language models are not intelligent in any sense. They are autocomplete on steroids. This is a computer program that was fed a book someone wrote, then mathematically tweaked to be able to guess the next word in a sentence in a way that resembles that book. That’s all it does. It does not think or learn in any sense we’d apply to a human.
To me, LLMs sound like a massive plagiarism engine, and I think they should need to get a license from the authors whose works they used to make the LLM under whatever terms that author wants to give, just like a publisher needs to get permission to print a copy of the work. But copyright law has no easy “bright line” for what counts and what doesn’t. So the courts will have to decide whether what the AI “creates” is similar enough to the original works to count as a violation, or if the AI and its results are transformative enough to count as something new.
I am sick of this trope of trying to argue that system X is or isn’t intelligent because it was built to do something that can be done non intelligently. LLMs are autocomplete, that’s just literally what they do. The autocomplete on your phone isn’t very intelligent if at all. Humans are DNA replicators but so are bacteria, which aren’t very intelligent if at all. You can’t argue from the type and/or character of the task whether something that was built to do that task is intelligent or not. LLMs at least appear to be intelligent because they do just about everything the AI skeptics were demanding machines must do in order to prove intelligence just 5 years ago, if you want to argue they’re not intelligent you need to do much more work than just calling them names like fuzzy jpeg, stochastic parrot, and autocomplete on steroids.
I use the term “autocomplete on steroids” because it gets across a vaguely accurate idea of what an LLM is and how it works to people who are thinking of it like sci-fi movie AI. Sorry if it came across that was my whole reason for considering them not intelligent.
LLMs do seem to pass a lot of intelligence tests we’ve come up with. Talking with one for the first time is a really uncanny experience, it’s a totally different thing than the old voice assistants. But they also consistently fail at tasks that would indicate an understanding of a topic. They produce good looking equations, but the math underneath doesn’t make sense. They hallucinate facts that don’t fit with the rest of what they themselves are saying, but look similar to the way right answers are written and defended. They produce really convincing responses, but when they fail they betray some really basic failures to understand what they’re saying.
I feel that LLMs are brute-forcing the tests people designed to measure intelligence. They can pass the bar exam, but they also contain thousands of successful bar exams to consult and millions of bits of text to glue those answers together with. But if you ask the LLM to actually do the job of a lawyer, they start producing all kinds of garbage that sounds good but doesn’t stand up to scrutiny when someone looks up the hallucinated case references.
