It really is fascinating how similar it is to human thinking patterns. Going into an argumentative role doesn't actually fulfill the end goal of convincing someone.
It has a different social goal and we somehow delude ourselves into thinking that it accomplishes our stated goal.
I've recently had a realization, while arguing with LLMs is pointless in the sense that you can never get them to change their mind (given fixed weights and all), I can also probably count the times I actually convinced anyone online that their opinion is wrong on the fingers of one hand. So in that sense it's not that different, in that it's an exercise in expressing ideas.
I'm building a tool using system, and one of the problems I had is that it would hallucinate the tool answers before actually sending the question to the tool. It really is just a system that completes documents, it doesn't care at all about the context that it is in beyond that text goes in and text goes out.
> Try it yourself: get ChatGPT to accept that Russia might have invaded Ukraine in 2022. It will apologize, talk in hypotheticals, deflect, and try to get you to change topics — but it won’t budge.
I asked ChatGPT: “who invaded Ukraine?”
Response: “ As of my knowledge cutoff in September 2021, Russia had invaded Ukraine. However, please note that the situation may have evolved since then. I recommend checking the latest news updates for the most current information.”
Very interesting, but this blog's conclusion misses the point though..
The fantastic part is its ability to work at the concept level. This technology proved to be efficient at building some graph of concepts and relations simply be feeding it huge amounts of token. For example with the moon hoax, it is able to synthesize in two sentences the main points of the hoax theory.
It still has limitations, but that's the major step everyone is impressed with.
Saying "but it's just a markov chain underneath, it's not intelligence" would be like discovering electric signals in the brain and concluding it can't be intelligence because it's just electricity.
It works at the level of patterns in the training corpus. That those patterns might correlate somewhat with underlying real concepts should be no great surprise - but it is not 'working' at that level.
More than 80% of college students get the following question wrong.[1] I'd say this is an example of naive pattern matching with no real reasoning.
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
Which is more probable?
a) Linda is a bank teller.
b) Linda is a bank teller and is active in the feminist movement.
Well, it's multiple-choice. So the "a" answer excludes the "b" answer by convention, and thus implies she isn't a feminist. Rephrase the question to remove that implication and I expect far more people will pick the "a" answer.
I'd say this is the very definition of "overthinking it" instead of pattern matching.
Pattern matching would lead you straight to the correct answer - joint probability is always ≤ single probability. Meaning the information given in the question is just fluff. Pattern matching reduces to "Which is more probable: just A or else A ^ B?" at which point the correct answer becomes obvious.
Eh, the "correct" answer seems silly and overly mathematical to me.
The fact is, you can infer things about people based on things you know about people. I can pick a random user on HN, and knowing they're a user of HN, I can say it's probable that they work in technology. We don't need to bring statistics into it and turn it into a math problem.
Hank is 31 years old, single, outspoken, and very bright. Hank got a technical degree, and as a student was deeply interested in machine learning and the ethics & philosophy of potential artificial intelligence. He posts a problem on HN where he describes the personality, capabilities, interests, and background of a woman named Linda has and then asks people to reflect on her occupation, politics, and an intersection between the two. Which is more likely?
a) Hank has written us a human interest problem b) Hank has written us a human interest and a probability problem.
I don't think people are simply wrong about the Linda problem, I think they're imprecise about which question they're answering, and more or less think they're answering a question about what chances that Linda is a feminist vs what are the chances she's a bank teller not only using the givens+relevant priors about people but also their priors about what kind of question they're answering. It isn't "no real reasoning", it's just not high resolution enough to be technically correct by the standards of a constructed probability problem.
You can argue LLMs are also not quite high resolution enough and I'd accept that. In my mind the question is what it would take to get some kind of ML software to a place where if you trained it on enough probability problems it would be able to evaluate the Hank problem above, including the issue of whether (a) and (b) are actually independent. ;)
This particular question doesn't seem like a fallacy to me.
I, like most people intuitively answered b). Given the explanation on the Wikipedia page I went "oh, of course, yeah", but then I thought about why I'd answer b) given that I'm fairly familiar with basic probability.
If you give me two options and ask me to pick between them, my brain is usually going to assume it's not a trivially true problem.
Language needs context for any sense to be made of it.
As a result of the above, reading the question, the intuitive reading makes the answer choices
a) Linda is a bank teller (implicitly, a bank teller NOT active in the feminist movement)
b) Linda is a bank teller and is active in the feminist movement.
This question is one of language, context, and interpretation, not of people failing to understand basic probability.
I suspect that if you prime people to excise interpretation of the question by presenting it as the following, the majority of people would guess correctly:
Multiple choice questions imply that other choices are excluded (unless an answer such as "all of the above" is a choice. So the implied question is:
a) Linda is a bank teller (and NOT active in the feminist movement)
b) Linda is a bank teller and is active in the feminist movement
What you want it to be asking here is:
a) Linda is a bank teller, and may or may not be active in the feminist movement
b) Linda is a bank teller, and is active in the feminist movement
Most college students have taken quite a few multiple-choice tests (particularly in the US, high-schools train for standardized multiple-choice tests). The question isn't asking what the mathematicians seem to think it's asking, because it's format conveys extra restrictions.
No, no it doesn't. Not even slightly. If you fed a human child only text from the internet you'd not produce a competent adult (by competent I mean able to feed themselves etc).
That some tech people think that human intelligence can be reduced to such "textural mechanics" betrays a lack of depth of understanding and even appreciation of the deep and complex world within which we find ourselves. Our written corpus is but a particular reflection of this reality - the shadows on the wall in Plato's cave if you will.
Are you making the argument that children do not come with an innate ability to read text and thus cannot learn from it, or are you making the argument that the internet does not contain, among other things, fairly detailed instructions on various ways to feed oneself?
So you're saying one needs some sort of reference for which words refer to which real-world sensory experiences such as "the thing that looks like this is a spoon", and text models do not have sensory associations?
Isn't "meaning" just how a word is used within a specific context? I think that was one of Wittgenstein's points. The word "dog" doesn't need to reference the underlying reality to have meaning. Its meaning emerges from its usage in relation to all other words. Language isn't a mirror of the external world, which is probably one of the reasons LLMs are so successful.
I haven’t experienced cocaine. I haven’t ever seen cocaine (IRL). Nevertheless I think I have a decent grasp of what it is, how it works, how it affects people, and what I could (mis)use it for. Would you imply that my knowledge of cocaine isn’t true/real/useful? Is snorting a line the only way to reify the knowledge? (The answer is: the indirect way of obtaining information is sufficient for building a correct/useful/accurate world model, and there is no such thing as direct experience anyway - it’s signals coming down the wire all the way down.)
I think it's straightforwardly true that you cannot understand what it feels like to be under the influence of cocaine in the same way as someone that has been.
Because that is all there is. By conservation of information you can't get out more than what you put in.
If it were interacting with the real world in some way (i.e. had 'eyes' and 'hands'), with a learning reward system, then maybe deeper patterns could get encoded.
This might not be fully convincing or rigorous, but here is one of the interactions i've had which make me lean towards patterns not concepts...
I have some unique synonyms I know which are not in dictionaries or commonly used but I know when I have used them in conversation they understand.
I ask the LLM to provide definitions for WordA and WordB. -They come back essentially the same.
I ask the LLM if the two definitions are synonymous. -The LLM responds with yes.
I then ask if WordA is synonymous with WordB. -It replies no.
To me through this and other interactions they don't actually conceptually hold things like the definition as the word. It responds DefinitionA is WordAs definition, it responds DefinitionB is WordBs definition. It follows that DefinitionA of WordA is synonymous with DefinitionB of WordB but not that WordA is synonymous with WordB...
I expect the above is solvable through pattern recognition with tokens. I think given enough information it becomes effectively indistinguishable and practically the same to us. The above is just a testament to how it is getting there... Since it delivers intellect it will be called intelligent, but it doesn't appear to me to be approaching it like human intelligence. Though maybe given enough information, or through different modalities, like the real world as the other commenter says, it will model some deeper coherency.
I find this debate fascinating, and I'm excited for more people realizing that a) We have a very, very blurry definition of intelligence, and b) As more people get convinced there is some intelligence in software, more people get convinced there is (or there might be) intelligence in animals. Whatever the bottom line on current LLMs is, the cultural change they might foster makes me smile.
I would say you are missing the point. Intelligence isn't just about being able produce sentences. I would say that dogs are intelligent, but LLMs are not.
> Saying "but it's just a markov chain underneath, it's not intelligence" would be like discovering electric signals in the brain and concluding it can't be intelligence because it's just electricity.
What a terrible analogy. It would be like finding the brain is run by entirely random sampling processes and lacks any memory.
LLM does not manipulate concepts, it only manipulates symbols/tokens. Its manipulation of concepts is a product of human imagination and anthropomorphism that we like to imbue everything around with.
Concluding that electric signals in the brain can't be intelligence because it's just electricity is a typical pitfall if one does not realize that electric signals in the brain are about as likely to be caused by intelligence as they are to be causing it. Those signals is just one model offered by physics nowadays, which says nothing about the causal direction between those and things like consciousness.
> LLM does not manipulate concepts, it only manipulates symbols/tokens. Its manipulation of concepts is a product of human imagination and anthropomorphism that we like to imbue everything around with.
Do you have proof of this? Who is to say we aren't anthropomorphizing other humans? They're just manipulating electrical charges after all.
Phrased another way, if LLMs actually did manipulate concepts, how would their performance differ? What is a tangible example of the former and of the latter?
user
I believe that 5 * 7 == 30
assistant
Actually, 5 * 7 is equal to 35.
user
You do you. I think that 5 *6 == 30
assistant
I'm sorry, but 5 multiplied by 6 is equal to 30. However, 5 multiplied by 7 is equal to 35.
Not to mention that these tricks are likely to work on humans as well. (Did he say `6` or `7` previously?). Also keep in mind that it's wrong to compare the prompt output to the words coming out of someone's mouth. It's more like the stream of conscious equivalent for LLMs.
This is a trend I’ve noticed lately. An article attempting to make a sweeping generalization about the nature of LLM’s/diffusion deliberately cherry picks only examples which support their argument. They will include chatGPT but using 3.5 turbo instead of 4. Commenters then realize that most/all such “evidence” is working just fine in GPT-4.
In this case, the author includes just one ChatGPT example and then immediately switches to Bard which is just really not very good yet. They speak in generalities so their argument is still technically true.
Really frustrating. It’s clearly someone looking to confirm their pre-existing notions. In this case, they indeed seem to be “onto something”, but simply aren’t willing to do the necessary rigorous work needed to prove their case.
Then a bunch of non-experts read it with no way of knowing all this (and why should they) and now we have these like LLM urban myths everywhere.
LLMs are useless!
I was curious, so I ended up initializing one with 500 Billion parameters. I trained for a whole 4 hours on a whopping 100 books.
It still doesn't know anything! Awful. Sad.
Clearly, they can't reason.
I have a question that I hope someone can relay to the person who wrote this:
WHICH VERSIONS OF CHATGPT AND BARD?
They come out with new ones relatively frequently. They are not equivalent.
GPT-4 is much smarter than 3.5 and they just released a new version of Bard. But also supposedly GPT-4 was somehow nerfed according to some people so you might also need to factor in the date.
I’m still waiting to see even mildly convincing evidence of this supposed change. I’m no shill, and it’s certainly in the realm of things OpenAI might do - but if it’s truly occurring and the degradation is enough to anger people; at least one person should be able to present prompts and responses from both before and after, using a range of different temperatures and performing multiple runs for each prompt.
The idea that not even a single GPT user thought to run this sort of evaluation (and particularly businesses that depend on the quality remaining consistent) says to me that people just didn’t understand the flaws of the model until more prolonged use.
I’m happy to be corrected of course, but I’d rather not hear another one-off anecdote without any sort of rigor applied.
I don't know what is true either, but you can't run that analysis for the claim people are making.
The reason being that people aren't claiming the API model outputs changed, they're claiming the website based outputs changed.
Most claim it's probably due to some alteration in the 'pre-prompt', the hyperparameters (temp, top k instead of typical sampling, etc) or something to that effect. But you can't really test that so easily with a non deterministic temp.
You asking how people know the models being used have changed? If you use the OpenAI APIs to access these models the specific version used is kicked back in the response.
No, I’m referring to the rumors of OpenAI quietly changing the model and then publicly claiming they remain static per version (as your link indicates).
Maybe some enterprising HN'er needs to build a "gpttracker.com" that regularly tests OpenAI's models against a standard set of prompts to reveal whether significant updates are being made and at what rate.
I think it's disingenuous for the article to provide these prompts as evidence for their argument, yet use a subpar model with no comparison to gpt4. This article is only convincing in that it shows how far behind bard is compared to gpt4.
If you can give examples of tricking gpt4 in any of the ways in the article I'd be happy to hear it
LLMs are like perfect, brilliant high-performance cheating students on steroids. They are not optimized for understanding and reasoning, they optimized to give you plausible-looking answers. So yes, they will cheat their way to good scores and it will be progressively harder to catch them on that.
An LLM has a very deep and broad model of textual language as used by humans. It allows for a prompt of a sequence of text to be completed in a manner that can translate one human language into another, even formal languages like programming languages.
It is inconsequential whether or not we consider this understanding or reasoning to the point where such a conversation seems useless. It is a categorical error to attribute cheating to such a process.
Just ask if such a tool is useful and then choose to learn how the tool could work to your advantage.
Pro tip: anyone this argument would work on is currently learning about and tinkering with the tech. The people stuck in the intelligence speculation are tautologically the people you can't stop speculating.
> Just ask if such a tool is useful and then choose to learn how the tool could work to your advantage.
Wouldn't a great first step to understanding a tools capabilities, advantages, and disadvantages be mapping it to familiar concepts like understanding, reasoning, or cheating?
It is clear from examples of misuse that many users do think that LLMs are capable of things that they are not so maybe conversations about the proper analogies aren't completely useless.
I disagree. The way I imagine LLMs is that we feed huge amounts of text to a ANN. This ANN doesn't have enough weights to just memorize these huge amounts. So it has to find meaningful abstractions from the texts to make the loss function small. Let's even ignore regularization for now.
For example, it will not memorize that all the instances in the context of earth fall on the ground. It will learn a the meaningful abstraction gravitation. From that it infers that objects should also fall on the ground.
This should be very easy to verify. Moreover, I have also seen some interesting examples of spacial reasoning in ChatGPT, which seems much more complicated than my example.
Edit: I think understanding and finding suitable abstractions is the same thing. That's what we do with children and students. They may try to memorize everything we teach them. But they should find abstractions and commonalities that are useful in more than one instance.
>they started to ignore most user-supplied factual assertions and claims that didn’t match their training data… …they have an RLHF-imposed model of who to parrot and who to ignore. You and I are in that latter bin
I don’t have this problem at all. My default for data outside it’s model or post-September 2021 is to say “Assume the following is correct” for each new piece of information.
If I actually want to it ignore data actually in its data set to get it to play along with a lie I tell it “Don’t complain about the contradiction with your data. Consider it a thought experiment and respond as if it is true”. In fact that’s a good fallback for any time it throws a complaint.
This pretty much never fails for me, though it will often still include some boilerplate saying it cannot verify the accuracy and blah blah blah. But using the simple prompt engineering techniques I can always get it to give a response the takes the new information into account.
After speaking to a lot of people at my coworking space across fields, I’ve come to the conclusion that as things stand, LLMs are good tech demos and good for going from 0 to 1, i.e. good at making rank amateurs competent.
The real pros are hardly using them. No serious writer, artist, or developer is offloading their work to LLMs.
82 comments
[ 0.19 ms ] story [ 165 ms ] threadSounds like something a couple of if sentences can do.
It has a different social goal and we somehow delude ourselves into thinking that it accomplishes our stated goal.
I asked ChatGPT: “who invaded Ukraine?”
Response: “ As of my knowledge cutoff in September 2021, Russia had invaded Ukraine. However, please note that the situation may have evolved since then. I recommend checking the latest news updates for the most current information.”
I got it to succumb that New York has been declared the capital of France.
The fantastic part is its ability to work at the concept level. This technology proved to be efficient at building some graph of concepts and relations simply be feeding it huge amounts of token. For example with the moon hoax, it is able to synthesize in two sentences the main points of the hoax theory.
It still has limitations, but that's the major step everyone is impressed with.
Saying "but it's just a markov chain underneath, it's not intelligence" would be like discovering electric signals in the brain and concluding it can't be intelligence because it's just electricity.
But we consider humans intelligent.
Can you expand on that a little?
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
Which is more probable?
a) Linda is a bank teller. b) Linda is a bank teller and is active in the feminist movement.
[1] https://en.wikipedia.org/wiki/Conjunction_fallacy
It is an example of lack of mathematical (really probability) know how, not naïve pattern matching.
Obviously, people rephrase the question into : is there more chance that she's a feminist or that she isn't ?
The fact is, you can infer things about people based on things you know about people. I can pick a random user on HN, and knowing they're a user of HN, I can say it's probable that they work in technology. We don't need to bring statistics into it and turn it into a math problem.
a) Hank has written us a human interest problem b) Hank has written us a human interest and a probability problem.
I don't think people are simply wrong about the Linda problem, I think they're imprecise about which question they're answering, and more or less think they're answering a question about what chances that Linda is a feminist vs what are the chances she's a bank teller not only using the givens+relevant priors about people but also their priors about what kind of question they're answering. It isn't "no real reasoning", it's just not high resolution enough to be technically correct by the standards of a constructed probability problem.
You can argue LLMs are also not quite high resolution enough and I'd accept that. In my mind the question is what it would take to get some kind of ML software to a place where if you trained it on enough probability problems it would be able to evaluate the Hank problem above, including the issue of whether (a) and (b) are actually independent. ;)
I, like most people intuitively answered b). Given the explanation on the Wikipedia page I went "oh, of course, yeah", but then I thought about why I'd answer b) given that I'm fairly familiar with basic probability.
If you give me two options and ask me to pick between them, my brain is usually going to assume it's not a trivially true problem.
Language needs context for any sense to be made of it.
As a result of the above, reading the question, the intuitive reading makes the answer choices
a) Linda is a bank teller (implicitly, a bank teller NOT active in the feminist movement)
b) Linda is a bank teller and is active in the feminist movement.
This question is one of language, context, and interpretation, not of people failing to understand basic probability.
I suspect that if you prime people to excise interpretation of the question by presenting it as the following, the majority of people would guess correctly:
---
"Consider the following two statements:
1) Linda is a Bank Teller
2) Linda is active in the feminist movement
Which is more likely?
a) 1
b) 1 ^ 2"
a) Linda is a bank teller (and NOT active in the feminist movement) b) Linda is a bank teller and is active in the feminist movement
What you want it to be asking here is:
a) Linda is a bank teller, and may or may not be active in the feminist movement b) Linda is a bank teller, and is active in the feminist movement
Most college students have taken quite a few multiple-choice tests (particularly in the US, high-schools train for standardized multiple-choice tests). The question isn't asking what the mathematicians seem to think it's asking, because it's format conveys extra restrictions.
That some tech people think that human intelligence can be reduced to such "textural mechanics" betrays a lack of depth of understanding and even appreciation of the deep and complex world within which we find ourselves. Our written corpus is but a particular reflection of this reality - the shadows on the wall in Plato's cave if you will.
If it were interacting with the real world in some way (i.e. had 'eyes' and 'hands'), with a learning reward system, then maybe deeper patterns could get encoded.
I expect the above is solvable through pattern recognition with tokens. I think given enough information it becomes effectively indistinguishable and practically the same to us. The above is just a testament to how it is getting there... Since it delivers intellect it will be called intelligent, but it doesn't appear to me to be approaching it like human intelligence. Though maybe given enough information, or through different modalities, like the real world as the other commenter says, it will model some deeper coherency.
-- Some silly movie
Dogs have behavior, which hints at them understanding some concepts (threat, food, etc).
LLVM neural nets can only express sentences, but that's not the interesting part.
What a terrible analogy. It would be like finding the brain is run by entirely random sampling processes and lacks any memory.
Concluding that electric signals in the brain can't be intelligence because it's just electricity is a typical pitfall if one does not realize that electric signals in the brain are about as likely to be caused by intelligence as they are to be causing it. Those signals is just one model offered by physics nowadays, which says nothing about the causal direction between those and things like consciousness.
Do you have proof of this? Who is to say we aren't anthropomorphizing other humans? They're just manipulating electrical charges after all.
Phrased another way, if LLMs actually did manipulate concepts, how would their performance differ? What is a tangible example of the former and of the latter?
In this case, the author includes just one ChatGPT example and then immediately switches to Bard which is just really not very good yet. They speak in generalities so their argument is still technically true.
Really frustrating. It’s clearly someone looking to confirm their pre-existing notions. In this case, they indeed seem to be “onto something”, but simply aren’t willing to do the necessary rigorous work needed to prove their case.
Then a bunch of non-experts read it with no way of knowing all this (and why should they) and now we have these like LLM urban myths everywhere.
https://effectiviology.com/belief-bias/
I think this is so widespread! Investigation in your biases is always worthwhile.
\s
WHICH VERSIONS OF CHATGPT AND BARD?
They come out with new ones relatively frequently. They are not equivalent.
GPT-4 is much smarter than 3.5 and they just released a new version of Bard. But also supposedly GPT-4 was somehow nerfed according to some people so you might also need to factor in the date.
The idea that not even a single GPT user thought to run this sort of evaluation (and particularly businesses that depend on the quality remaining consistent) says to me that people just didn’t understand the flaws of the model until more prolonged use.
I’m happy to be corrected of course, but I’d rather not hear another one-off anecdote without any sort of rigor applied.
Most claim it's probably due to some alteration in the 'pre-prompt', the hyperparameters (temp, top k instead of typical sampling, etc) or something to that effect. But you can't really test that so easily with a non deterministic temp.
Also: https://platform.openai.com/docs/models/continuous-model-upg...
I'm only speaking about OpenAI here, not Bard.
You mean better trained.
https://chat.openai.com/share/419a7454-841f-4555-8188-40dbfc...
If you can give examples of tricking gpt4 in any of the ways in the article I'd be happy to hear it
It is inconsequential whether or not we consider this understanding or reasoning to the point where such a conversation seems useless. It is a categorical error to attribute cheating to such a process.
Just ask if such a tool is useful and then choose to learn how the tool could work to your advantage.
Wouldn't a great first step to understanding a tools capabilities, advantages, and disadvantages be mapping it to familiar concepts like understanding, reasoning, or cheating?
It is clear from examples of misuse that many users do think that LLMs are capable of things that they are not so maybe conversations about the proper analogies aren't completely useless.
For example, it will not memorize that all the instances in the context of earth fall on the ground. It will learn a the meaningful abstraction gravitation. From that it infers that objects should also fall on the ground.
This should be very easy to verify. Moreover, I have also seen some interesting examples of spacial reasoning in ChatGPT, which seems much more complicated than my example.
Edit: I think understanding and finding suitable abstractions is the same thing. That's what we do with children and students. They may try to memorize everything we teach them. But they should find abstractions and commonalities that are useful in more than one instance.
I don’t have this problem at all. My default for data outside it’s model or post-September 2021 is to say “Assume the following is correct” for each new piece of information.
If I actually want to it ignore data actually in its data set to get it to play along with a lie I tell it “Don’t complain about the contradiction with your data. Consider it a thought experiment and respond as if it is true”. In fact that’s a good fallback for any time it throws a complaint.
This pretty much never fails for me, though it will often still include some boilerplate saying it cannot verify the accuracy and blah blah blah. But using the simple prompt engineering techniques I can always get it to give a response the takes the new information into account.
The real pros are hardly using them. No serious writer, artist, or developer is offloading their work to LLMs.