Which is why I keep saying that anthropomorphizing LLMs gives you good high-order intuitions about them, and should not be discouraged.
Consider: GP would've been much more correct if they said "It's just a person on a chip." Still wrong, but much less, in qualitative fashion, than they are now.
> It's just a database. There is no difference in a technical sense between "hallucination" and whatever else you imagine.
It's like a JPEG. Except instead of lossy compression on images that give you a pixel soup that only vaguely resembles the original if you're resource bound (and even modern SOTA models are when it comes to LLMs), instead you get stuff that looks more or less correct but just isn't.
I get what you're saying but I think it's wrong (I also think it's wrong when people say "well, people used to complain about calculators...").
An LLM chatbot is not like querying a database. Postgres doesn't have a human-like interface. Querying SQL is highly technical, when you get nonsensical results out of it (which is most often than not) you immediately suspect the JOIN you wrote or whatever. There's no "confident vibe" in results spat out by the DB engine.
Interacting with a chat bot is highly non-technical. The chat bot seems to many people like a highly competent person-like robot that knows everything, and it knows it with a high degree of confidence too.
So it makes sense to talk about "hallucinations", even though it's a flawed analogy.
I think the mistake people make when interacting with LLMs is similar to what they do when they read/watch the news: "well, they said so on the news, so it must be true."
To quote Luke Skywalker: Amazing. Every word of what you just said is wrong.
Which is why I keep saying that anthropomorphizing LLMs gives you good high-order intuitions about them, and should not be discouraged.
Consider: GP would've been much more correct if they said "It's just a person on a chip." Still wrong, but much less, in qualitative fashion, than they are now.
Just a weird little guy.
1 reply →
> It's just a database. There is no difference in a technical sense between "hallucination" and whatever else you imagine.
It's like a JPEG. Except instead of lossy compression on images that give you a pixel soup that only vaguely resembles the original if you're resource bound (and even modern SOTA models are when it comes to LLMs), instead you get stuff that looks more or less correct but just isn't.
This comes from not having a specific area or understanding, if you ask it about an area you know well, you'll see.
I get what you're saying but I think it's wrong (I also think it's wrong when people say "well, people used to complain about calculators...").
An LLM chatbot is not like querying a database. Postgres doesn't have a human-like interface. Querying SQL is highly technical, when you get nonsensical results out of it (which is most often than not) you immediately suspect the JOIN you wrote or whatever. There's no "confident vibe" in results spat out by the DB engine.
Interacting with a chat bot is highly non-technical. The chat bot seems to many people like a highly competent person-like robot that knows everything, and it knows it with a high degree of confidence too.
So it makes sense to talk about "hallucinations", even though it's a flawed analogy.
I think the mistake people make when interacting with LLMs is similar to what they do when they read/watch the news: "well, they said so on the news, so it must be true."