5 ms·

Why call it anything other than an error or bug? When my server spits out data I didn’t expect or want, I don’t say it hallucinated or confabulated. When the h

Why call it anything other than an error or bug? When my server spits out data I didn’t expect or want, I don’t say it hallucinated or confabulated. When the hurricane model gets the landfall site wrong, what do we call that? There was just a HN thread on police crime models. When those are wrong are they “hallucinating” crime? Why personify these statistical models at all?

bloqs 3y ago
Because you can't manipulate errors to gain from them in the same way I think. Hallucination indicates influence is possible, it's dynamic. Error is simply "this didnt work". I'm not certain on this but I feel like the concepts are somehow distinct.
HPsquared 3y ago
Because they are already mimicking human behaviour (speech / 'thinking'). It lends itself to making psychological analogies.
wang_li 3y ago
Because no one is going to give you five hundred million dollars unless you dress up all your work with words that imply way more than is actually there. The language used in AI fields confuses the non-experts, reporters, bloggers, investors, and the like. Who then proceed to infer and extrapolate and write their own stories about how AI is currently, actively, displacing and replacing expert white collar workers. Those stories then inspire the next iteration of writers/investors to produce yet more bullshit. The end result is an edifice of crap sustained by a feedback loop of fabulists and fantasists.
TeMPOraL 3y ago
Because your server is following a recipe, and if it does it wrong, it's either a problem with the recipe, or a hardware issue - which itself is also built from recipes - or something exceptional like cosmic radiation. LLMs, on the other hand, are big amorphous blobs of numbers that we throw example inputs and outputs at, hoping they'll eventually tune in on a pattern that we ourselves don't understand and can't formalize. Statistical models are somewhere in between on this spectrum - there is a recipe implementing inaccurate model of reality; some parts of it we understand, the others we hope the model will tell us. And for those models, there are different classes of errors - the usual software and hardware errors, and separately, situations where the computation is done correctly, but results just don't match experimental evidence. This second class is more similar to LLM hallucinations. As for personification angle, it's because LLMs are developed for text completion, and recently optimized for chat-based interface and fine-tuned to interpret instructions. I.e. the personification starts with how (and why) they're being developed.
empath-nirvana 3y ago
> Why call it anything other than an error or bug? Is the point to create an oracle or something that matches the capabilities of a human? Because humans make mistakes and confabulate all the time, and there might be something _inherent_ to general intelligence that requires some level of behavior that we would classify as "mistakes".
stonogo 3y ago
It's good marketing to use psychiatric terminology here, since it feeds into the handwaving about AGI. It's vital to the sales model of this field that people with purchasing authority be gaslit into believing that there is a nascent mind at the core of this software.
Karellen 3y ago
Because distinguishing between types of error is important and useful. A buffer overrun is different from a stack overflow, which is different from a division by zero, which is different from "else die()" on unexpected input, which is different from a power failure, which is different from a cosmic-ray bit-flip. The kinds of problems these errors cause, the kinds of defences you can use to protect against them, and the kinds of reactions you need to have when they occur are different. Using the right name for the type of error you're getting helps categorise it, talk about it, and come up with solutions that are relevant to that problem.
dwheeler 3y ago
There are many kinds of errors. It's helpful to have specific names for specific kinds of problems. Otherwise it's hard to discuss and fix.

bloqs 3y ago

Because you can't manipulate errors to gain from them in the same way I think. Hallucination indicates influence is possible, it's dynamic. Error is simply "this didnt work". I'm not certain on this but I feel like the concepts are somehow distinct.

HPsquared 3y ago

Because they are already mimicking human behaviour (speech / 'thinking'). It lends itself to making psychological analogies.

wang_li 3y ago

Because no one is going to give you five hundred million dollars unless you dress up all your work with words that imply way more than is actually there. The language used in AI fields confuses the non-experts, reporters, bloggers, investors, and the like. Who then proceed to infer and extrapolate and write their own stories about how AI is currently, actively, displacing and replacing expert white collar workers. Those stories then inspire the next iteration of writers/investors to produce yet more bullshit. The end result is an edifice of crap sustained by a feedback loop of fabulists and fantasists.

TeMPOraL 3y ago

Because your server is following a recipe, and if it does it wrong, it's either a problem with the recipe, or a hardware issue - which itself is also built from recipes - or something exceptional like cosmic radiation. LLMs, on the other hand, are big amorphous blobs of numbers that we throw example inputs and outputs at, hoping they'll eventually tune in on a pattern that we ourselves don't understand and can't formalize. Statistical models are somewhere in between on this spectrum - there is a recipe implementing inaccurate model of reality; some parts of it we understand, the others we hope the model will tell us. And for those models, there are different classes of errors - the usual software and hardware errors, and separately, situations where the computation is done correctly, but results just don't match experimental evidence. This second class is more similar to LLM hallucinations. As for personification angle, it's because LLMs are developed for text completion, and recently optimized for chat-based interface and fine-tuned to interpret instructions. I.e. the personification starts with how (and why) they're being developed.

empath-nirvana 3y ago

> Why call it anything other than an error or bug? Is the point to create an oracle or something that matches the capabilities of a human? Because humans make mistakes and confabulate all the time, and there might be something _inherent_ to general intelligence that requires some level of behavior that we would classify as "mistakes".

stonogo 3y ago

It's good marketing to use psychiatric terminology here, since it feeds into the handwaving about AGI. It's vital to the sales model of this field that people with purchasing authority be gaslit into believing that there is a nascent mind at the core of this software.

Karellen 3y ago

Because distinguishing between types of error is important and useful. A buffer overrun is different from a stack overflow, which is different from a division by zero, which is different from "else die()" on unexpected input, which is different from a power failure, which is different from a cosmic-ray bit-flip. The kinds of problems these errors cause, the kinds of defences you can use to protect against them, and the kinds of reactions you need to have when they occur are different. Using the right name for the type of error you're getting helps categorise it, talk about it, and come up with solutions that are relevant to that problem.

dwheeler 3y ago

There are many kinds of errors. It's helpful to have specific names for specific kinds of problems. Otherwise it's hard to discuss and fix.