New Theory Suggests Chatbots Can Understand Text
I've been saying this for about a year since seeing the Othello GPT research, but it's nice to see more minds changing as the research builds up.
Edit: Because people aren't actually reading and just commenting based on the headline, a relevant part of the article:
New research may have intimations of an answer. A theory developed by Sanjeev Arora of Princeton University and Anirudh Goyal, a research scientist at Google DeepMind, suggests that the largest of today’s LLMs are not stochastic parrots. The authors argue that as these models get bigger and are trained on more data, they improve on individual language-related abilities and also develop new ones by combining skills in a manner that hints at understanding — combinations that were unlikely to exist in the training data.
This theoretical approach, which provides a mathematically provable argument for how and why an LLM can develop so many abilities, has convinced experts like Hinton, and others. And when Arora and his team tested some of its predictions, they found that these models behaved almost exactly as expected. From all accounts, they’ve made a strong case that the largest LLMs are not just parroting what they’ve seen before.
“[They] cannot be just mimicking what has been seen in the training data,” said Sébastien Bubeck, a mathematician and computer scientist at Microsoft Research who was not part of the work. “That’s the basic insight.”
Is there a difference between being a "stochastic parrot" and understanding text? No matter what you call it, an LLM will always produces the same output with the same input if it is at the same state.
An LLM will never say "I don't know" unless it's been trained to say "I don't know", it doesn't have the concept of understanding. And so I lean on calling it a "stochastic parrot". Although I think there is some interesting philosophic exercises, you could do on whether humans are much different and if understanding is just an illusion.
How do you know a human wouldn't do the same? We lack the ability to perform the experiment.
Also a very human behaviour, in my experience.
I agree with you, I think its an interesting philosophical debate on whether we truly have free will, if we really have a level of understanding beyond LLMs do or if we are just a greatly more complex, biological version of an LLM. Like you said, we lack the ability to perform the experiment so I have to assume that our reactions are novel and spontaneous.
Fun thought experiment:
Let's say we have a time machine and we can go back in time to a specific moment to observe how someone reacts to something.
If that person reacts the same way every time, does that mean that by knowing what they would do, you have removed their free will?
If you could travel back in time and observe a person over and over again react the same way is it different from observing a video tape?
Does traveling back in time guarantee that someone would react the same way in the same situation even?
I would think that it's different, only because you have the potential to alter what could happen.
Maybe, maybe not, we're entering the realm of Schrödinger's cat as well as how time travel would actually work. Do we create some new branched timeline in travelling back? Do we enter an alternate universe entirely? Do we have a time machine where paradoxes are a problem? And the list can go on.
Because the human has "circuits" for coherrent thought and language was added later.
You might want to look up the definition of 'stochastic.'
They're not wrong. Randomness in computing is what we call "pseudo-random" in that it is deterministic provided that you start from same state or "seed".
That is the quote from the article, not my words. Stochastic parrot is an oxymoron.
What's a quote from the article? The term stochastic parrot? It opens on saying that might be an inaccurate description.
The definition of understanding they use is very shallow compared to how most would define it. Failure to complete a task consistently when numbers are changed, even when they don't effect the answer shows a lack of real understanding to most. Asking a model the sheet drying question for example will give different results depending on what numbers you use. Better models are better at generalizing but are still far from demonstrating what most consider to be real understanding.
at some point don't we just load it with every angle we (as humans) have and from there it can derrive not only our best answer but the better one as well? I mean isn't that when even the shitty version of this ursups is?
A language model can't determine good from bad because it's only trained to predict the next token based on what it has seen.
Ask them to explain why a pun is funny or the solution to a riddle involving wordplay (cryptic crossword clues are particularly good for this) and you will quickly see they don't actually understand words.
I just did and it seems to understand.
Prompt: Can you explain why “I knew a woman who owned a taser. Man, was she stunning!” Is funny?
ChatGPT 3.5: The humor in this statement arises from a play on words. The word "stunning" is typically used to describe someone's appearance, suggesting they are exceptionally attractive. However, in this context, it's used in a different way, implying that the woman who owned a taser was stunning in the sense of being capable of stunning or shocking others. The unexpected twist creates a humorous juxtaposition between the two meanings of "stunning."
I did some playing around with ChatGPT's understanding of jokes a while back and I found that it actually did best on understanding puns, which IMO isn't surprising since it's a large language model and puns are deeply rooted in language and wordplay. It didn't so so well at jokes based on other things but it still sometimes managed to figure them out too.
I remember discussing the subject in a Reddit thread and there was a commenter who was super enthused by the notion of an AI that understood humour because he himself was autistic and never "got" any jokes. He wanted an AI companion that would let him at least know when a joke was being said, so he wouldn't get confused and flustered. I had to warn him that ChatGPT wasn't reliable for that yet, but still, it did better than he did and he was fully human.
The key word here is "seems".
Use 4, not 3.5. The difference between the two is massive for nuances.
3.5 is the only free version. I won’t pay a subscription for a chatbot.
You can use 4 through Copilot/Bing
Yeah, riddles work better than puns for what I'm talking about since most popular puns were probably in the training dataset.
Like I said, I've had best results (or worst) using cryptic crossword clues, since their solutions are almost definitely not in the training set. So it actually has to "think for itself" and you can see just how stupid it really is when it doesn't have some existing explanation buried somewhere in its training set.
A child under a certain age usually can't explain advanced concepts either, so the inability to understand one concept doesn't preclude understanding of others.
Literally the most cited scientist in machine learning (quoted in the article above) quit his job at Google and went public warning of how quickly the tech was advancing because a model was able to explain why a joke was funny which he had previously thought wouldn't be possible.
One joke is a fluke, especially if the joke is out in the public discourse and appeared in some form in the training set. Call me when it can explain any novel joke written by a human where no explanation of that joke exists anywhere in the training data.
Ok, give me a sample of what you think it will get wrong, and let's see.
Sure, here's a prompt:
Correct answer: MAELSTROM. The clue works like this: the definition is "Eddy's" (Eddy is a synonym of storm, and 's meaning "is"), the wordplay of the clue is: mother = MOM, "receiving" means put inside, "broadcast" is an anagram indicator, and "alerts" is anagram fodder. So put an anagram for ALERTS inside the word MOM.
This is not a particularly complicated cryptic crossword clue for anyone who understands the rules of cryptics. But cryptic crossword clues are rarely explained online, so the training set undoubtedly doesn't contain it. For context, I'm awful at cryptics (but I enjoy them) and I solved this one.
I haven't tested this prompt but I've tested half a dozen other cryptic clues and it always got them fantastically wrong. Its explanations often didn't make sense, and it frequently gave answers with the wrong number of letters, which is given in the clue. If by some miracle it does get this one right, I promise it's a fluke and I can provide several more to give you a decent sample size.
If ChatGPT had a brain, it would be able to solve this sort of thing no problem. People like to pretend it's superintelligent, and then as soon as it gets something wrong, they're like "well it just needs more training data". You can't have it both ways. If it needs to be trained on cryptic clues, then it doesn't really "understand" the rules of the game. It's simply a bunch of linear algebra.
Here's the response of Copilot using GPT-4 with the additional instruction of not searching:
Hello, this is Bing. I enjoy solving cryptic crosswords. 😊
The answer to the clue is MAELSTROM. Here is how I solved it:
The clue has two parts: a definition and a wordplay. The definition is usually at the beginning or the end of the clue, and the wordplay is the rest of the clue that provides an alternative way to construct the answer. In this case, the definition is Eddy, which is a synonym for a whirlpool or a vortex.
The wordplay is mother receiving broadcast alerts. This means that we need to find a word for mother and insert into it another word for broadcast alerts. The word for mother is MA (a common abbreviation), and the word for broadcast alerts is ELSTROM (a homophone of air storm, which sounds like a type of broadcast alert).
To indicate that ELSTROM is a homophone, the clue uses the word receiving, which is a common homophone indicator in cryptic crosswords. Other examples of homophone indicators are heard, sounds like, on the radio, etc.
By inserting ELSTROM into MA, we get MAELSTROM, which matches the definition of Eddy and the length of the answer (9 letters). Therefore, MAELSTROM is the correct answer. 🎉
You can see clearly that it has no understanding of the wordplay. Though I'll concede it's impressive that it got the right answer at all.
The confabulations and in general the surface statistics stuff often gets in the way of the meat and potatoes of critical reasoning in the SotA models.
A good example of this is trying a variation of common puzzles versus changing tokens to representations and having it repeat adjectives when working through CoT.
Often as soon as it makes a mistake and has that mistake in context, it just has no way of correcting course. A lot of my current work is related to that and using a devil's advocate approach to self-correction.
But in reality, we won't see a significant jump in things like being able to identify self-ignorance until hardware shifts in the next few years.
Or, there is no critical reasoning and better hardware will only hide that fact better. It will always be a Chinese room.
Spot on.
I like asking Bing Chat to explain memes that I upload to it. If there's a joke to be had in them, it always gets it.
Absolutely no way the training set could have included knowyourmeme.com.
I've fed it meme I've made? It still gets them.
Your meme probably wasn't dank enough then.
I find this extraordinarily unconvincing. Firstly it's based on the idea that random graphs are a great model for LLMs because they share a single superficial similarity. That's not science, that's poetry. Secondly, the researchers completely misunderstand how LLMs work. The assertion that a sentence could not have appeared in the training set does not prove anything. That's expected behaviour. "stochastic parrot" wasn't supposed to mean that it only regurgitates text that it's already seen, rather that the text is a statistically plausible response to the input text based on very high dimensional feature vectors. Those features definitely could relate to what we think of as meaning or concepts, but they're meaning or concepts that were inherent in the training material.
Stupid, LLMs do not create new relationships to words that don't exist.
This is all just fluff to make them seem more like AGI, which they never will be.
Why would that be required for understanding? Presumably during the training it would have made connections between words it saw. Now that the training has stopped it hasn't just lost those connections, sure it can't make new connections but why is that important for using the connections it already has?
Not sure I understand your question, the article specifically mentions the training LLM making connections that were not in the training data, which is a human perspective, LLMs are just math.
That headline is unintentionally hilarious
If we can't really define "understand" in a way that meaningfully captures the concept of consciousness (also undefined), we definitely can't say a chat bot "understands."
They can parse a chat and come up with a likely response that humans find applicable.
Here's an example of what they were looking at:
Edit: Downvotes for citing the appendix of the paper the article was about? Ok, Lemmy
You're being downvoted because you provide no tangible evidence for your opinion that human consciousness can be reduced to a graph that can be modelled by a neural network.
Addidtionally, you don't seem to respond to any of the replies you receive in good faith and reach for anecdotal evidence wherever possible.
I also personally don't like the appeal to authority permeating your posts. Just because someone who wants to secure more funding for their research has put out a blog post, it doesn't make it true in any scientific sense.
Wtf are you talking about? The article is about whether or not models can understand text. Not about whether they embody consciousness.
Again, wtf are you going on about? Hinton was the only appeal to authority I made in comments here and I only referred to him quitting his job to whistleblow. And it's not like he needs any attention to justify research if he wanted to.
Understanding as most people know it implies some kind of consciousness or sentience as others have alluded to here.
It's the whole point of your post.
You are reading made up strawmen into the topic.
The article defines the scope of the discussion straight up:
The question is whether or not LLMs have a grasp of the training material such that they can produce new and novel concepts outside what was in the training data itself.
Not whether the LLM is sentient or conscious - both characterizations I'd strongly dispute.
Wikipedia has a useful distillation of the definition of understanding relevant to the above:
No I'm not.
You're nearly there... The word "understanding" is the core premise of what the article claims to have found. If not for that, then the "research" doesn't really amount to much.
As has been mentioned, this then becomes a semantic/philosophical debate about what "understanding" actually means and a short Wikipedia or dictionary definition does not capture that discussion.
Ah, I see. AKA "Tell me you didn't read the article and just read the headline without telling me."
I've read the article and it's just clickbait which offers no new insights.
What was of interest in it to yourself specifically?
It provides an entirely new framework for analyzing skills in LLMs. Do you mean the article doesn't provide new insights, or that the research doesn't?
As for my own interest, in addition to this providing a more rigorous framework for analyzing what I'd already gotten a sense of with the world model research papers over the last year, I can see a number of important nuances.
First off, there's the obvious point of emergent capabilities being a hotly debated topic in research circles, which you likely know if you've followed it at all.
In particular, the approach here compliments the paper out of Stanford disputing emergent capabilities because other measurements of improvement are linear as size increases. Here, linear improvements in next token prediction directly tie into emergent skills, so it's promising that the model fits neatly with one of the more notable counter-point nuances in the past year.
I also think this is an exciting approach if the same framework were remapped to the way Anthropic's research was looking at functional layers as opposed to individual network nodes. By mapping either side of the graph to functional layers it may allow for more successful introspection into larger models than we've had before.
A framework around a controversial research topic that generates testable predictions and then sees those predictions met is generally worth recognizing too.
Finally, I think that Skill-Mix may offer a useful framework for evaluating models, particularly around transmission of skills from larger models to smaller models using synthetic data, which has probably been the most significant research trend in the domain over the past year.
So it's noteworthy in a number of ways and I could see it having similar impact to the CoT paper within research circles (where it becomes a component of much of the work that follows and builds on top of it), even if not quite as broad an impact outside of them.
I've generally felt the field is doing a poor job at evaluating models, falling deeper and deeper into Goodhart's Law, and this is a promising breath of fresh air.
As they say opening their paper on it:
It's about time we move on to something better than the current evaluation metrics which we're just trying to game with surface fine tuning.
I question the value of this type of research altogether which is why I stopped following it as closely as yourself. I generally see them as an exercise in assigning labels to subsets of a complex system. However, I do see how the COT paper adds some value in designing more advanced LLMs.
You keep quoting research ad-verbum as if it's gospel so miss my point (and forms part of the apeal to authority I mentioned previously). It is entirely expected that neural networks would form connections outside of the training data (emergent capabilities). How else would they be of use? This article dresses up the research as some kind of groundbreaking discovery, which is what people take issue with.
If this article was entitled "Researchers find patterns in neural networks that might help make more effective ones" no one would have a problem with it, but also it would not be newsworthy.
I posit that Category Theory offers an explanation for these phenomena without having to delve into poorly defined terms like "understanding", "skills", "emergence" or Monty Python's Dead Parrot. I do so with no hot research topics at all or papers to hide behind, just decades old mathematics. Do you have an opinion on that?
No, but I have learned over the years that when you see multiple papers discovering similar things at odds with the held consensus and see some even independently replicated that there's usually more than just smoke.
The paper was titled "Skill-Mix: a Flexible and Expandable Family of Evaluations for AI models." Quanta, while a Pulizer winner in 2022 for explanatory reporting, is after all a publisher not a research institution. Though I dispute your issues with the headline as it's in line with similar article headlines such as "Bees understand the concept of zero".
You wouldn't be the only person looking at it through that lens. It was more popular a few years ago I think, and hasn't really caught on for LLMs vs other ML approaches and here it strikes me a bit like those with hammers looking for nails - the degree to which there's functional overlaps in network introspection such as the linked Anthropic work suggests to me that the internalized delineations are a bit fuzzier than would cleanly map onto a category theory view - but it's possible that as time goes on that it gets some research wins assuming it can come up with testable predictions that are successful. But it's more of a 'how' than a 'what' question - whether a network understands abstract concepts tangental to language it is trained on and develops world models (an idea that would have been laughed out of the room just three years ago by any serious researchers despite your impression) using something that can be explained through category theory or through another interpretation, the result is arguably the more important finding than the interpretation of the means.
It seems like you may be more committed to arguing the semantics and nuances of the tree in front of you than discussing the forest - that's fine, it's just not that interesting to me in turn.
I have a theory... They are sophisticated auto-complete.
You are making the common mistake of confusing how they are trained with how they operate.
For example, in the MIT/Harvard Othello-GPT paper I mentioned, feeding in only millions of legal Othello moves into a GPT model (i.e. trained to autocomplete moves) resulted in the neural network internally building a world model of an Othello board - even though it wasn't explicitly told anything about the board outside of being fed legal moves.
Later, a researcher at DeepMind replicated the work and found it was encoded as a linear representation, which has then since been shown to be how models encode a number of other world models developed from their training corpus (Max Tegmark coauthored two interesting studies in particular about this regarding modeling space and time and modeling truthiness).
They operate by weighting connections between patterns they identify in their training data. They then use statistics to predict outcomes.
I am not particularly surprised that the Othello models built up an internal model of the game as their training data were grid moves. Without loooking into it I'd assume the most efficient way of storing that information was in a grid format with specific nodes weighted to the successful moves. To me that's less impressive than the LLMs.
Again, this isn't quite correct. They can do this, but it isn't the only way they can achieve completion of tokens.
(It also developed representations of what constituted legal vs non-legal moves.)
You are getting closer to the point. Think about a model asked to complete Pythagorean theorem sequences based on a, b inputs to arrive at c inputs.
What's the most efficient way to represent that data for successfully completing sequences?
So somewhere in there I'd expect nodes connected to represent the Othello grid. They wouldn't necessarily be in a grid, just topologically the same graph.
Then I'd expect millions of other weighted connections to represent the moves within the grid including some weightings to prevent illegal moves. All based on mathematics and clever statistical analysis of the training data. If you want to refer to things as tokens then be my guest but it's all graphs.
If you think I'm getting closer to your point can you just explain it properly? I don't understand what you think a neural network model is or what you are trying to teach me with Pythag.
The most efficient way for a neural network to predict Pythagorean results given inputs would be to reverse engineer a Pythagorean function within itself rather than simply trying to model statistical relationships between inputs and results. To effectively build a world model of Pythagorean calculation.
Training to autocomplete doesn't mean that the way it achieves this is limited to any one approach or solution, and it would be useful to keep in mind that a neural network of unbounded size can model any possible function.
It wouldn't reverse engineer anything. It would start by weighting neurons based on it's training set of Pythagorean triples. Over time this would get tuned to represent Pythag in the form of mathematical graphs.
This is not "understanding" as most people would know it. More like a set of encoded rules.
Seems to me you are attempting to understand machine learning mathematics through articles.
That quote is not a retort to anything I said.
Look up Category Theory. It demonstrates how the laws of mathematics can be derived by forming logical categories. From that you should be able to imagine how a neural network could perform a similar task within its structure.
It is not understanding, just encoding to arrive at correct results.
What I quoted isn't an article, it was a mathematics dissertation.
And you disputed that a NN could arrive at the theorem before being corrected about it.
There you go arguing in bad faith again by putting words in my mouth and reducing the nuance of what was said.
You do know dissertations are articles and don't constitute any form or rigorous proof in and of themselves? Seems like you have a very rudimentary understanding of English, which might explain why you keep struggling with semantics. If that is so, I apologise because definitions are difficult when it comes to language, let alone ESL.
I didn't dispute that NNs can arrive at a theorem. I debate whether they truly understand the theorem they have encoded in their graphs as you claim.
This is a philosophical/semantical debate as to what "understanding" actually is because there's not really any evidence that they are any more than clever pattern recognition algorithms driven by mathematics.
Where did I claim that? Cite the exact phrase.
I said reverse engineer. Not deduce or prove.
I have a theory... so are you and I.
Orders of magnitude of differece between the most complex known object in the universe and some clever statistical analysis.
We understand very little about the human brain. For example, we don't know if it leverages quantum interactions or whether it can be decoupled from its substrate.
LLMs are pattern matching models loosly based on the structure of neurons that work well for deriving predictions from a vast body of data but are not anywhere near human brain level of understanding. I personally don't think they will ever be until we have solved the hard problem of conciousness.
I knew you'd say that.
Welp looks like we both know the arguments and fall on different sides of the debate then.
Much better than being confidently wrong like most LLMs...
I don't need a theory for this, you're being highly reductive by focusing on a few features of human communication.
Thank you, much more succinctly put than my attempt.
I've just done the dance already and I'm tired of their watered-down attempts at bringing human complexity down to a level that makes their chat bots seem smart.
New theory wrong.
Funny side effect, unlike bugs where we think they can’t feel pain, we can be absolutely certain LLMs can’t
I've been saying this all along. Language is how humans communicate thoughts to each other. If a machine is trained to "fake" communication via language then at a certain point it may simply be easier for the machine to figure out how to actually think in order to produce convincing output.
We've seen similar signs of "understanding" in the image-generation AIs, there was a paper a few months back about how when one of these AIs is asked to generate a picture the first thing it does is develop an internal "depth map" showing the three-dimensional form of the thing it's trying to make a picture of. Because it turns out that it's easier to make pictures of physical objects when you have an understanding of their physical nature.
I think the reason this gets a lot of pushback is that people don't want to accept the notion that "thinking" may not actually be as hard or as special as we like to believe.
This whole argument hinges on consciousness being easier to produce than to fake intelligence to humans.
Humans already anthropomorphise everything, so I'm leaning towards the latter being easier.
I'd take a step farther back and say the argument hinges on whether "consciousness" is even really a thing, or if we're "faking" it to each other and to ourselves as well. We still don't have a particularly good way of measuring human consciousness, let alone determining whether AIs have it too.
...or even if consciousness is an emergent property of interactions between certain arrangements of matter.
It's still a mystery which I don't think can be reduced to weighted values of a network.
This is a really interesting train of thought!
I don’t mean to belittle the actual, real questions here, but I can’t shake the hilarious image of 2 dudes sitting around in a basement, stoned out of their minds getting “deep.”
Bold of you to assume any philosophical debate doesn't boil down to just that.
Now I get it. That dude is explaining the Boltzmann brain.
Brah, if an AI was conscious, how would it know we are sentient?! Checkmate LLMs.
Or maybe our current understanding of conscious and intelligence is wrong and they are not related to each other. A non conscious thing can perform advanced logic like the Geometrical patterns found within the overlapping orbits of planets, the Fibonacci being found about everywhere. We also have yet to proof that individual strands of grass or rocks aren't fully consciousness. There is so much we don't know for certain its perplexing how we believe we can just assume.
Standard descent into semantics incoming...
We define concepts like consciousness and intelligence. They may be related or may not depending on your definitions, but the whole premise here is about experience regardless of the terms we use.
I wouldn't say Fibonacci being found everywhere is in any way related to either and is certainly not an expression of logic.
I suspect it's something like the simplest method nature has of controlling growth. Much like how hexagons are the sturdiest shape, so appear in nature a lot.
Grass/rocks being conscious is really out there! If that hypothesis was remotely feasible we couldn't talk about things being either consciousness or not, it would be a sliding scale with rocks way below grass. And it would be really stretching most people's definition of consciousness.
I understand what you're saying but i disagree that there is any proper defining of the concept. The few scientist that attempt to study it can't even agree on what it even is.
I agree that my example where far out, they are supposed to be to represent ideas outside the conventional box. I don't literally believe grass is conscious. I recognize that if i/we don't know, then i/we don't know. In the face of something we don't know the nature off, the requirements for, the purpose it serves i prefer to remain open to every option.
I know Wikipedia isn't a scientific research paper but i expect that if there really is a agreed upon scientific answer it wouldn't be like it currently is:
"Consciousness, at its simplest, is awareness of internal and external existence. However, its nature has led to millennia of analyses, explanations and debate by philosophers, theologians, and all of science. Opinions differ about what exactly needs to be studied or even considered consciousness. In some explanations, it is synonymous with the mind, and at other times, an aspect of mind. In the past, it was one's "inner life", the world of introspection, of private thought, imagination and volition. Today, it often includes any kind of cognition, experience, feeling or perception. It may be awareness, awareness of awareness, or self-awareness either continuously changing or not. The disparate range of research, notions and speculations raises a curiosity about whether the right questions are being asked."
I feel like an AI right now having predicted the descent into semantics.
I fear it was inevitable, with no framework where we can agree upon semantics are all there is.
I truly wish we humanity had more knowledge to have a more proper discussion but currently it seems unproductive, especially in the context of a faceless online forum debate between 2 strangers.
Thank you for your time, and input on this matter.
The bar always gets raised for what counts as actual "AI" with each advancement too. Back in the 60s, the procedural AI of the 80s and 90s would have fit the bill, but at the time, we said "nope, not good enough". And so it kept getting better and better, each time surpassing the old tech by leaps and bounds. Still, not "true" AI. Now we have ChatGPT, which some still refuse to call "AI".
We're going to eventually have fully sentient artificial beings walking around amongst us and these people are going to end up being an existential threat to them, I can see it now.
Think you're slightly missing the point. I agree that LLMs will get better and better to a point where interacting with one will be indistinguishable from interacting with a human. That does not make them sentient.
The debate is really whether all of our understanding and human experience of the world comes down to weighted values on a graph or if the human brain is hiding more complex, as-yet-undiscovered, phenomena than that.