Precisely, it's a tech that aims to write the most average, most likely human text.
That’s why it’s repeated. It’s kind of correct if you squint and it’s easy to understand
Even very simple ones may require you to twist the definition of "correctness". I open a REPL and type "1/3.0*3.0" and get "0.9999999999". Then you have to do mental gymnastics like "actually it is a correct answer because arithmetic in computers is implemented not like you'd expect".
Exactly. The fact that language is fuzzy is why LLMs work so well.
The issue is that most people expect computers to not make mistakes. When you write a formula in an excel sheet, the computer doesn’t mess up the math.
The average non tech person knows that humans make mistakes, but are not used to computers making mistakes.
Many people, maybe most, would see an answer generated by a computer program and assume that it’s the correct answer to their question.
In pointing out that LLMs are guessing at what text to write (by saying “average”) you convey that idea in a simplified way.
Trying to argue that “correct” doesn’t mean anything isn’t really useful. You can replace the word “correct” with “practically correct” and nothing about what I said changes.
I don’t think people are dumb, just that the vast majority of people dont have knowledge of statistics and AI.
“Random” isn’t good either. Obviously the text gpt and other LLMs generate isn’t random.
Statistical works too, sure.
Sure, computers are better at arithmetic humans, but let's be honest, nobody uses chatgpt as a calculator. Last 20 years AI is getting everywhere, we keep laughing that sometimes AI systems make very obvious stupid mistakes. Now we finally have a system that makes subtle mistakes very confidently and suddenly people are like "I thought computers are never wrong". I can't fathom how anyone would expect that.
I don’t really think there’s much left for us to talk about.
If a LLM returned something really unusual for Shakespeare when you didn't ask for it, you'd say it's not performing well.
Maybe that's tautological but I think it's what's usually meant by "average".
I'm sure LLMs with something different is on the near horizon but I don't think we're there quite yet.
The point was that no, you wont (necessarily) get some "average" shakespeare. A sampler may introduce bias and look for the "above average" shakespeare in the distribution.
Please write <specifics of the text you want LLM to write>, <style instruction>.
There's no way those "AI detectors" could determine whether the text was written by AI from text itself, as it's trivial to make LLM output have any style imaginable.
Other statistical anomalies probably exist; it is certainly possible to tell that an average is from a larger or smaller sample size (if I tell you X fair coin flips came up heads 75% of the time, you can likely guess X, and can tell that X is almost certainly less than 1000).
But in practice it doesn’t look possible, or at least the current offerings seem no better than snake oil.
It takes a lot of skill and a long time to develop a unique style of writing. The purpose of language is to be an extremely-lossy on-average way of communicating information between people. In the vast majority of cases, idiomatic style or jargon impairs communication.
That's only true in the aggregate. Within a single answer, LLMs will try to generate a word choice which is more likely _given the preceding word choices in that answer_, which should reduce the blending of idioms.
> where most of us are trained and reinforced in a particular region.
The life experience of most of us (at least here in HN) is wider than that. Someone who as a child visited every year their grandparents in two different regions of the country could have a blend of three sets of regional idioms, and that's before learning English (which adds another set of idioms from the teachers/textbooks) and getting on the Internet (which can add a lot of new idioms, from each community frequented online). And this is a simple example, many people know more than just two languages (each bringing their own peculiar idioms).
Someone who learned English mostly through books and the Internet could very well have such a mixture, since unlike native speakers of English, they don't have a strong bias towards one region or the other. You could even say that our "training data" (books and the Internet) for the English language was the same as these LLMs.