this post was submitted on 25 May 2024
819 points (97.7% liked)
Technology
59589 readers
3148 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Can we swap out the word "hallucinations" for the word "bullshit"?
I think all AI/LLM stuf should be prefaced as "someone down the pub said..."
So, "someone down the pub said you can eat rocks" or, "someone down the pub said you should put glue on your pizza".
Hallucinations are cool, shit like this is worthless.
No, hallucination is a really good term. It can be super confident and seemingly correct but still completely made up.
for it to "hallucinate" things, it would have to believe in what it's saying. ai is unable to think - so it cannot hallucinate
Hallucination is a technical term. Nothing to do with thinking. The scientific community could have chosen another term to describe the issue but hallucination explains really well what's happening.
huh, i kinda assumed it was a term made up/taken by journalists mostly, are there actual research papers on this using that term?
Yup. Loads of them! https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=hallucinations+llm&btnG=
It used to mean all generated output though. Calling only mistakes hallucinations is new, definitely because of hype.
So how do you prove it can't think? Or that you actually can?
because it's a text generation machine..? i mean, i wouldn't say i can prove it, but i don't think anyone can prove it's capable of thinking, much less of reasoning
like, it can string together a coherent sentence thanks to well crafted equations, sure, but i wouldn't qualify that as "thinking", though i guess the definition of "thinking" is debatable
It's an interesting question. I am inclined to believe that the faster it gets at running those equations, over and over and over, reanalysing is data and responses as it goes, that that ultimately leads to some kind of evolution. You know, Vger style.
It can tell you how to stack things on top of each other the best way to get a high tower. Etc.
Those are not random sentences. If you can not define thinking in a way this machine fails at, then stop saying it does not think.
A parrot can be trained to tell you how to stack things on top of each other the best way to get a high tower.
This is just an electronic parrot, millions of times faster to train than the biological parrot, specialized in repetition alone (can't really do anything else a parrot can) and which has been trained on billions of texts.
You're confusing one specific form in which humans externally express cogniscence with the actual cogniscence itself: just because intelligence can produce some forms of textual communication doesn't mean that the relationship holds in the opposite direction and such forms of textual communication require intelligence, or if you will, just because you can photograph a real pizza to get a picture of a pizza doesn't mean a picture of a pizza is actually of a real pizza and not something with glue to make it look like it has stringy melted cheese.
Again, it is absolutely capable to come up with it's own logical stuff, hence my example. Stop saying it just copies existing stuff, that is simply wrong.
interesting, in my experience, it's only been good at repeating things, and failing on unexpected inputs - it's able to answer pretty accurately if a small number is even or odd, but not if it's a large number, which indicates it's not reasoning but parroting answers to me
do you have example prompts where it showed clear logical reasoning?
Examples showing that it comes up with it's own solutions to an answer? Just ask it something that could not have been on the Internet before. Professor talking about AGI in GPT 4
Personal examples would be to code python to solve a 2D thermal heat flux problem given some context and constraints.
Here is an alternative Piped link(s):
Professor talking about AGI in GPT 4
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I'm open-source; check me out at GitHub.
well, i just tried it, and its answer is meh --
i asked it to transcribe "zenquistificationed" (made up word) in IPA, it gave me /ˌzɛŋˌkwɪstɪfɪˈkeɪʃənd/, which i agree with, that's likely how a native english speaker would read that word.
i then asked it to transcribe that into japaense katakana, it gave me "ゼンクィスティフィカションエッド" (zenkwisuthifikashon'eddo), which is not a great transcription at all - based on its earlier IPA transcription, カション (kashon') should be ケーシュン (kēshun'), and the エッド (eddo) part at the end should just, not be there imo, or be shortened to just ド (do)
this paper says it is capable of original thought. It also "speaks" of it in high regard in other things. That is also my experience using it for... over a year?! now.
Sure, whatever.
Amazing reply, given the context.
I'm actually a domain expert on AI whilst your "assertive denial without a single counter-argument" answer to my simplified explanation together with your "understanding" of the subject matter shown in the post before that one, shows you're at the peak of the Dunning-Krugger curve on this domain and also that you do not use analytical thinking or the scientific method in any way form or shape when analysing a subject.
There is literally no point in explaining anything to somebody who reasons like that and is at that point of that curve.
You keep your strongly held "common sense" beliefs and I'll keep from wasting any more of my time.
this paper clearly says it is capable of original thought. It also "speaks" of it in high regard in other things.
Re-read it: it says AI is capable of "originality" and does not mention "thought" at all.
You're the one presuming that "originality" requires cognition and hence understood "originality" as meaning "original thought" even though they're different concepts (specifically the latter is a subset of the former).
In your interpretation of that paper you did the exact same logical mistake as you seem to be doing in your interpretation of LLMs - you made assumptions backed only by gut feeling thus taking a leap to reach a conclusion ultimately supported only by your gut feeling.
"electronic parrot" and outperforming almost all humans in creativity and originality is an extreme contrast to me, regardless of my misuse of terms. So I fail to understand what you want to say, since this contrast must be apparent to you too.
The original context of my comment was even more basic and to me proven by what the paper says: Those are not things it copied somewhere. Also, I still think there is no test to prove it can/can't think.