@EtherealResonance Who can tell? (Certainly not me)
But how would YOU spot and tell the difference between a reply and an uttering that looks like a reply?
@torstentorsten can't.
What feels scary is that chatgpt can do lots of math 80% of the time very accurate. Mostly with help of creating its own python code and letting it run.
I find that alone very crazy
Re: "but isnt that the same with (some) Humans ?"
Kind of, yes. There are probably humans in every field who can fake the ambience of knowledge well enough to fool other humans who _don't_ know the field.
But it isn't a brilliant idea to go to a _human_ bullshitter for advice either :-)
The difference isn't that LLMs can produce plausible-sounding bullshit and humans can't. Both can.
It's more like, most people already _know_ that some confident-bullshitter bloke in the pub may not be reliable in explaining their physics homework :-)
(or providing case law for their legal case, or telling them which mushrooms are safe to eat.)
The way LLMs have been sold as "intelligent", it might not be quite so obvious at first that they don't actually know what they're talking about - and that whether their answers are right or not is a roll of the dice. That's why it's worth explaining.
"Without enougth information, they misinform."
This sentence implies that there's an "enough information" which could stop LLMs from misinforming people. But that isn't the case. Correct or incorrect information isn't the basis on which they function.
Is your argument that limiting its task to "summarise this specific text" means it will have "enough" information and won't get anything wrong?
Hmm interesting. I don't think I would ever entirely trust the summary of an LLM, but then I would retain some scepticism about a summary from most humans too.
I don't think "They are an really good interface to interact with Humans" though. Not currently. For that to be the case, the average human would have to have a significantly better understanding of the limits of what an LLM can and can't do. Otherwise, the "learning" you refer to is going to produce a lot of damage along the way.
I tend to explain it as "What would a reply that a significant amount of the students in this room would by a good chance not reject because it sounds plausible look like".
@rpin42 Humans might sometimes resemble this process, but it is not at all accurate to say we apply “exactly the same approach” because we plainly do not. We can remember facts. We can detect inconsistencies. We can detect and ignore superfluous information. An LLM cannot do any of these things.
Sometimes the output looks like they do these things. But the fact that the output looks like the output of thinking doesn’t mean it was the result of thinking, or even the result of a process analogous to thinking. We think. It doesn’t.
@mcnees Nature today linked to an interesting article and good analogy — "Think about it like a multiple-choice test. If you do not know the answer but take a wild guess, you might get lucky and be right. Leaving it blank guarantees a zero."
Some time ago (I do not remember the source) I have read about an interesting teaching approach. The assignment was to use LLM for a given project and then to discuss where the LLM was wrong.
@scottfweintraub @mcnees >> what is not clear is the degree to which people also answer questions that way.
Yes, it is. They don’t.
>> LLMs are definitely not what we believe intelligence to be, but could it be that that belief is incorrect?
No.
@scottfweintraub @mcnees LLMs assign “tokens” to words and work on a kind of map of which tokens are associated with each other, in what sequence. But they don’t “know” what the words mean. They’re not even words, just tokens.
Humans don’t work like this.
@MisuseCase @scottfweintraub @mcnees Current research suggests our brains do actually work a *little* like that. Essentially, they have an internal representation of a concept or a relationship between concepts, which is then turned into an external representation (speech, writing, gestures, etc.) to express that concept or relationship to others. Paraphasia and aphasia are believed to be misfires or disconnections in this token-to-language mapping. This is also believed to be why aphasia affects only the ability to use language, but not intelligence.
Of course, our brains are far more complicated than just their language centers, and the language centers are definitely more efficient than LLMs (both in training and in use).
@MisuseCase
@scottfweintraub @mcnees I think it depends on the context and the human. I know some people that will BS an answer to not look bad. Sometimes they're right. I joke that someone I know who regularly does that is an LLM.
Generally, no I don't believe this is how humans think, even if it may mimic one mode we use sometimes. But I do believe there is a lot to learn about ourselves from how we are reflected in the machine.
I'm not sure that's entirely the case. I had a... chaotic childhood, and there was definitely a period where I was, especially under stress, inclined to give plausible answer-shaped replies for which actual truth was irrelevant. Around this time I had also read a lot of joke books and could confidently land dirty jokes that I had zero knowledge of.
So I suspect the LLM expectation-influenced, consistency-driven glibness is similar to part of how we answer, but it only dominates in pathological conditions (compulsive liars, fabulists, some kinds of illness or brain damage).
I am really getting salty about this kind of comment.
EVERY TIME a discussion about LLMs gets even slightly philosophical someone comes up with this "what if we're really like LLMs" with an implied naughty snigger.
No, LLMs do not build models of reality, the way basically every animal more complex than a sea-slug manages.
@mcnees
Yes, LLMs are trained to be convincing, not trained to be correct. When they are, it's by accident.
Can I borrow this?
@mcnees @tchambers Imagine an #LLM as an improv actor who’s read every single script ever written. They’re given a scene to act out, and they have to come up with the next line that fits, even though they don’t have a clue about what’s going on.
Is this from a larger document available online?
Kevin Cheek
Robert McNees
IainB
Der Ulli
Torsten Torsten
EtherealResonance
reiterator
Jennifer Moore 😷
Nik | Klampfradler 🎸🚲
Rich 🔶UK #RejoinEU
Paco Hope
xenophyte
Escarpment
Adam Millerchip
Frieke
Frank Davies
Dr Suzanne she/her
Scott Weintraub
Misuse Case
Zimmie
Ted
William Pietri
Resuna
Richard Hendricks
Marcus Richter
Bryan 
Blender Dumbass ( J.Y.Amihud )
Koantig
GuySoft
Guerric Haché
Emily🌱
Stu Watts
Karen Campe
Drey
okanogen VerminEnemyFromWithin
Chris Bussard
Mark Gardner
Casey Jennings
Shafik Yaghmour