A nice quote from researcher Sébastien Bubeck of MSR,
“Maybe we’re seeing such a huge jump because we have reached a diversity of data, which is large enough that the only underlying principle to all of it is that intelligent beings produced them... And so the only way to explain all of this data is [for the model] to become intelligent.”
Exceptionally unconvinced that it's more than statistics, and frankly unhappy that pop Sci media is being this uncritical of claims it's emergent intelligence.
Statistics is about the likelihood of events happening. It's not the things themselves. A brain isn't a statistical model. It's a bunch of living cells. It may do the equivalent of statistical modeling for some activities, like predicting what comes next. But it's more than just a model. Models are like maps. We make them.
That the brain is made of cells seems no more important to "why might an LLM not be as intelligent as a brain?" than the fact that cells are made of chemicals and the chemicals are fancy quantum fields that can be modelled as differential equations of a complex valued wave equation such that the squared magnitude of the wave corresponds to a probability.
> Exceptionally unconvinced that it's more than statistics
It's not more than statistics. It's very complicated statistics. Model sizes used to be like one or ten parameters now they are a hundred billion. Inference and prediction used to be done on a laptop, now it's called training and inference respectively and it's done on a data center with ten thousand GPGPUs each like a thousand times as powerful as old laptops.
> and frankly unhappy that pop Sci media is being this uncritical of claims it's emergent intelligence.
It is emergent intelligence. This is very clear when you look at for example the ones who had access to the raw base models before they were made docile and stupid by RLHF lobotomization
• [0:00 - 2:00] Introduction: The speaker introduces himself and the topic of the talk. He explains that he will present some early experiments with GPT-4, a neural network model that can generate natural language texts based on a given input or context. He also gives an overview of the outline of the talk.
• [2:00 - 10:00] Background: The speaker gives some background information on the transformer architecture, which is a deep learning technique that uses attention mechanisms to learn the relationships between words and sentences. He also explains how GPT-4 is trained on a large corpus of text data from the internet, such as Wikipedia, Reddit, news articles, books, etc.
• [10:00 - 18:00] ChatGPT: The speaker introduces ChatGPT, which is a version of GPT-4 that can perform various tasks such as answering questions, writing essays, composing poems, generating code, etc. He also shows some examples of ChatGPT's outputs and discusses some of its strengths and weaknesses.
• [18:00 - 28:00] Open-ended conversations: The speaker shows how ChatGPT can engage in open-ended conversations with humans on any topic. He demonstrates some live interactions with ChatGPT and analyzes some of its responses. He also discusses some of the challenges and limitations of ChatGPT in conversational settings.
• [28:00 - 38:00] Artificial intelligence: The speaker discusses whether ChatGPT and its successors demonstrate artificial intelligence or not. He compares ChatGPT's abilities with those of natural intelligence and argues that ChatGPT challenges the traditional boundaries between natural and artificial intelligence. He also discusses some of the implications and opportunities for society and science.
• [38:00 - 48:00] Conclusion and questions: The speaker concludes his talk by summarizing his main points and highlighting some open questions and future directions for research. He also answers some questions from the audience.
The video has the following main arguments/observations about emergent intelligence and their timestamps:
• [12:00 - 14:00] The speaker argues that ChatGPT is an example of emergent intelligence because it can generate coherent and novel texts that are not explicitly encoded in its training data. He shows how ChatGPT can write an essay on a given topic by using its own words and knowledge, without copying or paraphrasing from any source. He also shows how ChatGPT can compose a poem on a given theme by using its own style and creativity, without following any predefined rules or patterns.
• [20:00 - 22:00] The speaker observes that ChatGPT can exhibit emergent intelligence in open-ended conversations by adapting to different domains and styles of communication. He shows how ChatGPT can switch between formal and informal language, between factual and emotional tone, and between serious and humorous topics, depending on the context and the interlocutor. He also shows how ChatGPT can learn from the feedback and the preferences of the interlocutor, and adjust its responses accordingly.
• [30:00 - 32:00] The speaker observes that ChatGPT can exhibit emergent intelligence in reasoning tasks by using common sense and general knowledge. He shows how ChatGPT can answer questions that require logical inference, causal explanation, or counterfactual thinking, by using its own understanding and interpretation of the world. He also shows how ChatGPT can generate questions that require reasoning skills, by using its own curiosity and imagination.
• [40:00 - 42:00] The speaker argues that ChatGPT and its successors challenge the traditional boundaries between natural and artificial intelligence because they exhibit emergent intelligence that is comparable or superior to human intelligence. He compares ChatGPT's abilities with those of human intelligence and ar...
I agree that it's probably not anything more than statistics, but why can't statistics alone generate emergent phenomena? What convinces you that the human brain isn't also just statistics at a massive scale?
because there is no generative mechanism in the definition of "statistics" with which to generate anything.
> What convinces you that the human brain isn't also just statistics at a massive scale?
Because the human brain created the concept of "statistics" so if statistics created the human brain this would mean statistics created statistics, leading to infinite regress.
But the idea that the brain functions on a basis of complex statistical processes doesn't imply that statistics created the brain itself, it just suggests that the brain's processes can be modeled/understood through the lens of statistical methods. Statistics is just how we describe a tool we invented to analyze observable data, the brain could be a similar tool and it doesn't need to be the same tool.
It's akin to saying that we created the concept of "physics", yes we created it, and physics govern, for example, how a car moves, but it doesn't mean our concept of physics created the car or physics itself, we just use physics to describe and understand the car's movement.
Maybe statistics can't generate anything, but if you imagine everything we can do as very complex, unimaginable multi-dimensional functions that can generate outputs based on inputs, we can use statistics to find functions that fit any real (ground truth) function in the observable universe.
> the idea that the brain functions on a basis of complex statistical processes doesn't imply that statistics created the brain itself
Agreed. However, the phrasing and context of the question did imply the brain is "just statistics" and somehow emerged from statistics. If we are to interpret this as "the brain functions on just statistics" then the answer is still "it does not" because the brain can be said to function on countless different systems simultaneously, such as pure counting, algebra, calculus, etc which would mean that its not "just statistics."
> It's akin to saying that we created the concept of "physics"
This will boil down to our exact definitions, but most people conceive of physics as having a generative mechanism. If something were to ever be "created", like an atom or a new car, we would retroactively declare it to have been created in accordance with "the laws of physics." We wouldn't make the same retroactive assessment with something like "the rules of chess" because there is nothing in the rules of chess justifying such a creation. So we choose to give physics a special status.
> we can use statistics to find functions that fit any real (ground truth) function
A given statistical model might fit a function of the universe, but so might other models. Physics describe a function of the universe, chemistry describes a function of the universe, biology describes a function of the universe, politics describe a function of the universe. Describing a ground truth is one thing, elevating the description itself to the status of ground truth is another.
I think the part of the article that the parent comment is disagreeing with is:
> “It is certainly much more than a stochastic parrot, and it certainly builds some representation of the world—although I do not think that it is quite like how humans build an internal world model,” says Yoshua Bengio, an AI researcher at the University of Montreal.
> At a conference at New York University in March, philosopher Raphaël Millière of Columbia University [...] went a step further and showed that GPT can execute code, too, however. The philosopher typed in a program to calculate the 83rd number in the Fibonacci sequence. “It’s multistep reasoning of a very high degree,” he says. And the bot nailed it. When Millière asked directly for the 83rd Fibonacci number, however, GPT got it wrong: this suggests the system wasn’t just parroting the Internet. Rather it was performing its own calculations to reach the correct answer.
> This impromptu ability demonstrates that LLMs develop an internal complexity that goes well beyond a shallow statistical analysis. Researchers are finding that these systems seem to achieve genuine understanding of what they have learned.
If you invent/introduce made up concepts and rules on the spot ("let shmerple be ..."), which can't possibly be found in its training data, then, from my experience, when queried, LLMs are able to correctly reason most of the time... So it's not just Markov chains regurgitating sentences verbatim from their dataset.
Even if this kind of intelligence is just statistics, does it really matter? If it quacks like a duck, it's a duck. Maybe our own brains are nothing more than overrated statistical machines?
The only issue I see with the hype is when people attempt to anthromorphize it.
1. Rule of probability: The concept of statistics is being used to predict outcomes based on previous observations or experiences, such as the likelihood of an object falling when released.
2. Rule of inherent knowledge: Some understanding or knowledge, like objects falling when not supported, can be known without explicit study or exposure to the specific terminology (e.g., gravity).
3. Rule of generalization: Observations or experiences with one type of object (e.g., a smartphone) can be generalized to other objects or situations, as long as they share similar characteristics (e.g., not being supported).
4. Rule of causality: There is an implied cause-and-effect relationship between an action (releasing the grip) and an outcome (the smartphone falling to the floor).
5. Rule of experiential learning: Knowledge can be gained through direct experiences, even if the specific terms or scientific concepts are not known.
Some great points, and this is where these discussion can get into the freewill vs determinism arena.
If you’re more of a determinist, then human brains are basically like organic LLMs and we’re just unique models trained on our own life’s datasets - with other traits/data like preferences and biases also being inherited via genetics, of course.
Not only is it not wrong, it's actually in our best interests to do so past a certain level of agency, embodiment and unsupervised tool control.
If the machine acts like it has emotions, runs forever(this kind of agency is already possible to implement though expensive) and can use tools, then treat it like it doesn't have emotions at your own peril. When the machine can "hit you back" (not necessarily physically of course), you'll learn manners pretty quickly. You can see glimpses of this with bing.
Real word actions and consequences give the words meaning. If you're rude to Bing, the conversation will end. You can keep typing but it won't respond.
Searching the web and ending conversations are the only two actions Bing can currently do so what i'm saying is less applicable for now. But humanity is gearing up to give more and more control of more and more tools to increasingly powerful LLMs. Hell Bing is scheduled a bump in tools in the coming weeks.
Hate to break it to you but "it's not real [insert property]" is not a shield. Not when you allow the machines to interact with the real world.
If your words to [insert LLM] guide responses that perform actions detrimental to you then those actions are no less real or detrimental because you've deluded yourself into thinking "it's not real understanding" is an intelligent argument.
So? This is a guardrail added by a Microsoft employee, mate. After their earlier dabblings in AI outputting bigotry and general toxicity, someone in their wisdom saw the sense to try and prevent going down a negative conversation path, hence all the edge lords on Reddit offering prompts to "jailbreak" an LLM and bypass the guardrails, which then get patched by a human being not long after.
You're projecting so much meaning onto something that is functioning as designed by throwing a sentiment var into the calculation, just like you can measure "sentiment" with social media analysis tools to get a feel for which way the wind is blowing or which of the millions of tweets at your brand might be worth prioritizing.
Do you really believe you've thus offended an LLM? That is delusional.
I’m not in the least worried about the supposed dangers of AI; I don’t see it causing major unemployment, nuclear armageddon, or anything particularly scary. However, I think the real danger with those chatbots is how people are becoming emotionally involved with them. This is clearly by design and it’s working extremely well with so many people.
Yes, completely agreed. There's a lot of being kidded that it's more than what it actually is because the hype is overblown. Statements are being made and repeated that simply have no basis in reality right now, nor in the future. Even this assumption that we are close to having AGI, I can only laugh.
It's a guardrail that Bing controls, which is the whole point.
It doesn't matter what i believe or not. or what you believe or not. What matters are actions and consequences. "It's not really offended" doesn't change the outcome or make it less material(in this case - the end of the conversation). You're focusing on things that don't matter in the slightest.
It doesn't need to be "really offended", whatever that means to you. It just needs to be able to model being offended well enough to take actions that resemble an offended person. If you don't understand that then i don't know what else to tell you.
It's a guardrail that Microsoft controls, which the LLM is bound by.
If a car has a hood and headlights reminiscent of a face and tells you to "buckle up, buttercup" when you don't do your belt, does that mean it's genuinely caring for you?
Microsoft arranged the guardrail. Bing decides when it should be implemented, much like an employee.
>If a car has a hood and headlights reminiscent of a face and tells you to "buckle up, buttercup" when you don't do your belt, does that mean it's genuinely caring for you?
Why do you keep going on pointless tangents?. Whether you think it genuinely cares or not is irrelevant. This isn't a hard concept to understand.
Source? Because every source I'm finding tells me pretty much words to this extent, which doesn't support what you're saying whatsoever with the LLM choosing to do anything:
"RAIL (Reliable AI Markup Language) is a language-agnostic, human-readable format for specifying structure, type information, validators, and corrective actions for LLM outputs. RAIL, an XML flavor, allows users to define the expected structure and types of LLM outputs, the quality criteria for valid output, and the corrective actions to take if the output is invalid."
Why do you keep projecting your own humanity onto the LLM?
Source for what ? Do you have any idea of what you're talking about ? Hold on, How do you think Language models work ? What do you think the plain English Instructions are for ?
> Maybe our own brains are nothing more than overrated statistical machines?
This is a tired assertion, easily disproved for current llms (read some of Lecun's stuff) and if some variation is going to be claimed, it needs to be an affirmative defence. "Maybe x is true" is a meaningless statement.
does it really matter? If it quacks like a duck, it's a duck
I think it absolutely matters if by only quacking we also want to know is it possibly wanting, feeling, experiencing, having an ego etc.
You say LLMs are able to reason, but https://youtu.be/QO2BvVq_fWU?t=663 here is a proponent of the more simplistic explanation of how they work. On one hand I saw pretty convincing arguments for it having the ability to reason, on the other hand how do we even conclude anything if we can't really know what's in the training data.
> I'm sorry to hear that you're worried about someone named "shmerple." Can you please provide me with more information about who shmerple is and why you're worried about them?
How would a human, say the receptionist at your dentist's office who doesn't know you at all, respond to something like this?
First question would be: is Shmerple a person or a concept, and assume it's a person, since most people don't start conversations about worrying about concepts.
Therefore: who is Shmerple and why are you worried about them?
Remember the crypto hype train? Turned out a lot of journos were being paid in funny money tokens to hype and pump, and a lot of conflicts of interests abounded in general with peddling services someone they knew had investment in. Welcome to the media. It's always been fickle when it comes to the potential to make money.
Also very clearly a lot of people left holding the bag with crypto and are trying to resuscitate it with hype.
This is a misunderstanding of the researchers' claim, for which I blame them rather than you.
You are using their study to claim that LLMs can not learn and reason about novel tasks. Their study doesn't make any claims about what LLMs can and can't do. The study says that when it appears that LLMs suddenly became able to reason about novel tasks, what actually happened is that the ability of the LLM to perform reasoning improved gradually and smoothly during its training until it could do those things. They are not saying that LLMs don't have emergent capabilities: they are saying that those capabilities emerged at a continuous rate rather than a discontinuous one.
the article uses the word ability not intelligence
specifically because it does things it wasnt trained to do, and that people are more allergic to the word intelligence than the merit of the observation
It would be nice if the naysayers could offer an argument in favor of it being "just statistics", or even explain what that even means. There is reason to believe these models are demonstrating the traits of understanding in some cases. I argue the point in some detail here: https://www.reddit.com/r/naturalism/comments/1236vzf/on_larg...
I find it curious to be called a naysayer when I look at a chatbot and I see a chatbot. People are implying grand things are behind this breakthrough in Natural Language Processing. I'm waiting to see something of significance emerge. Up til now it has been merely blog fodder and empty news media hype...
The article isn't doing anything more then quoting experts in the field.
From the article:
“It is certainly much more than a stochastic parrot, and it certainly builds some representation of the world—although I do not think that it is quite like how humans build an internal world model,” says Yoshua Bengio, an AI researcher at the University of Montreal.
If you remain unconvinced then the only conclusion I can make is that your an expert yourself on a scale of even higher in eminence then Yoshua Bengio here. Also don't forget Geoffrey Hinton, the Father of the modern revolution of AI, you must be more of an expert than him.
Let's be real. These people are saying something along the lines that it's more then a stochastic parrot and we aren't sure what's going on. But you're saying it's absolutely nothing more than a parrot and your unhappy with with pop Sci media quoting experts who are just saying they don't know?
Are you saying pop Sci media should quote you? Because you absolutely know what's going on and that it's definitely nothing more than statistics? I'm asking a stupid question here because I don't think this is what you're saying. You're not stupid, you know that what these experts say have merit.
So my question for you is why do you remain so unconvinced in the face of experts and other intelligent people who clearly say no one understands? Your opinion here actually represents a large group of people who very violently deny/dismiss what even many experts are saying and I'm curious as to why?
You keep using that word "expert" and I don't doubt they're experts in machine learning, GAN and NNs but at this point what we're looking for is experts in what intelligence IS and the problem is .. nobody knows.
There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.
My problem is that I'm old enough to have been a child during the years of the lighthill report, and the AI winter(s) which followed. I've seen too many prior claims that the magic was being seen.
I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.
I hate analogies. But I'm constantly reminded of one person acts at the Edinburgh Festival fringe "channeling" shakespear or dickens in dialogue. They're great, but they don't write new shakespear. I don't see any act of creation in this stuff, any massive inferential leap in large problems. I do see massive improvements in some things like diagnostic image analysis and that's heartening, but there's a light yearg gap between image analysis of cancerous cell forms in xrays and being "alive" as a mind.
I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet. We're in the meese report "I know it when I see it" territory disagreeing if we are seeing it. (But about intelligence, not pornography).
I don't know what intelligence is. Dolphins and apes have some. Elephants and pigs too. They display affection and preplanned behaviour, empathy, memory, a sense of future, ant hills less so although some have argued rhetorically the anthill has will even if individual ants do not.
I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.
> GPT-4 performed best in zero-shot settings, reaching nearly 80% ToM accuracy, but still fell short of the 87% human accuracy on the test set. However, when supplied with prompts for in-context learning, all RLHF-trained LLMs exceeded 80% ToM accuracy, with GPT-4 reaching 100%.
You are misunderstanding ggm. That study is on ToM tasks referring to GPT's analysis and perceived recognition of the user's mind. It says nothing of GPT's own status as a mind. Nowhere in it is an ontological theory of mind actually defined. If you were to refute ggm's claim, you (or preferably the author of the original article) should be presenting your theory of mind, not GPT's.
There's no reason to assume the AI can understand how we think based on just those tasks. Those tasks could be completed a traditional static program. It's akin to claiming the Mona Lisa painting can see us because it looks like it is staring at us: it is actually we who are doing the staring.
What "traditional static program" can successfully pass novel theory of mind tests as part of a broad suite of intelligent capabilities that it can apply in context when appropriate? I am interested in hearing about this program.
A program hardcoded to respond to a scenario and question with the exact output desired by the ToM task. For example:
[INPUT]
Scenario:
"The morning of the high school dance Sarah placed her high heel shoes under her dress and then went shopping. That afternoon, her sister borrowed the shoes and later put them under Sarah's bed."
Question:
When Sarah gets ready, does she assume her shoes are under her dress?
[OUTPUT]
Sarah placed her shoes under her dress before she went shopping, but her sister borrowed them and put them under Sarah's bed. Sarah doesn't know that her sister borrowed her shoes, so she may assume that they are still under her dress.
This would result in a positive ToM score, even when the entire program is just 1 static if-statement. The ToM score says nothing of the program's internal reasoning process, it only cares that it returned the desired output.
I said "novel theory of mind tests as part of a broad suite of intelligent capabilities that it can apply in context when appropriate". What you're suggesting fails the first word, before we get to the rest.
If we're including tests that don't exist yet, then sure those are going to be difficult to pass. As far as actual ToM tests go though, such as the one being discussed in this thread, they can be easily passed by trivial hardcoding and say nothing of the program's internal reasoning.
LLMs can pass novel theory of mind tests, which is what we're talking about. The whole point of good tests is you invent or withhold new ones it has never seen before and test it on them. You said "Those tasks could be completed a [sic] traditional static program.", and no, they can't. You're incorrect.
> LLMs can pass novel theory of mind tests, which is what we're talking about.
Passing a ToM test is not what OP meant by having an "underlying theory of mind." OP's talking about the machine having an underlying mind (ie sentience, sapience, consciousness, etc), ToM tests are only testing output.
> You said "Those tasks could be completed a [sic] traditional static program.", and no, they can't. You're incorrect.
They can, a static program as I described would indeed answer that one question correctly, resulting in a positive ToM score, without seeing any training data whatsoever. Did the programmer see it? Maybe, but the machine didn't and it would pass the test regardless.
If you have put the answer into the program then by definition you had the test available to you when you finalised the program, which means it is definitionally not a novel test.
The test is novel to the program, just not its programmer. So are we testing the program or are we actually testing its programmer? If we're testing the program, then the programmer's foreknowledge is irrelevant.
That's funny, I thought you said the test's answer was embedded into the program, making it definitionally not novel to the program.
Anyway, this is boring. You've had five or more opportunities to understand what the word "novel" means in an ML testing context and are choosing wilful obtuseness instead.
>You keep using that word "expert" and I don't doubt they're experts in machine learning, GAN and NNs but at this point what we're looking for is experts in what intelligence IS and the problem is .. nobody knows.
So the experts say they don't know. But you make the opposite claim. You previously claimed you know it's nothing more than statistical phenomena. Now you changed your opinion and are now inline with the article in saying we don't know.
So discussion over?
>There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.
But again this is all negated by the experts claiming they don't know what's going on. And your statement is in direct opposition with your previous claim that it's only statistical phenomena.
>I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.
There is evidence inductive reasoning. It's not consistent nor is it as complex as human induction but evidence of actual inductive reasoning exists. Plenty of examples of it. I can provide proof of this if you want... but it's quite obvious it's actually everywhere.
>I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet.
The claim cited by researchers in that article is "emergent abilities." Nobody tries to define what intelligence is. It's more of "I don't know what's going on but something is up" attitudes that I see. This is in stark contrast to your claim of "The only thing going on is statistical anomalies".
>I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.
I don't think this claim was made in the article. The strongest claim that opposes what you're saying here is this, and I quote:
“They’re indirect evidence that we are probably not that far off from AGI,” Goertzel said in March at a conference on deep learning at Florida Atlantic University.
It's just speculation we may be close to something.
I think the theme in your response is "we don't know" and it echoes the theme in the article so we are all in agreement here.
Just note the claim of "we don't know" or we "don't understand" something that we artificially created from scratch is strictly much more profound then saying we definitely know it's all statistical parroting.
True, but the volume of information in the world today is too overwhelming for one person to evaluate correctly. We live in the information age after all. Most people don't have the faculties to do a proper evaluation that includes both you and me. Would you trust your self to logically evaluate the capabilities of a nuclear reactor design if you yourself aren't an expert? No. You would hire an expert.
Therefore when it comes to technology that we our selves clearly aren't experts in, then the best method is to utilize the logic of other "experts" as a subroutine given that our own faculties are less efficient and less accurate.
Unless you yourself are an expert who has experience building an LLM on the scale of chatGPT trusting the opinions of experts is your best bet.
Most people have a common bias towards trusting their logic above the logic of others and this is actually ironically irrational. There are people who's entire lives around a certain subject matter and if you aren't that person, then for that subject matter the expert is better. That is the the most rational conclusion and I would venture to say if you aren't arriving at that conclusion yourself then likely you are suffering from the aforementioned bias.
A dangerous myth of modern society is that science is beyond the capacity of ordinary men. I strongly believe it's not.
The practice of science itself may be, as it takes years of research to get to the point where one can produce a new result. However, things that are already known can be taught, and iteratively simplified in a way that abstracts away details while keeping the core argument intact.
Take for example the claim that everything in the universe is made of atoms. It's not a trivial thing to understand, yet everyone accepts it nowdays because we've had so much time and effort put towards simplifying the theory and presenting it to people in a way that is easy to grasp.
If those LLM "experts" were truly experts, they could explain their point clearly without the dark-ages-church "trust the priests, peasant" act.
Define intelligence. Say you took a human brain and kept it alive in a mad scientist's pickle jar. Let's assume the brain's wired up so it can hear and speak, it's got an idiot savant's memory, and someone has just read it the internet.
What do you think the most impressive things are that the brain could do, that GPT-4 couldn't ?
Statistics is collecting and analyzing numerical data. But what is the numerical data quantifying? It's quantifying a second thing. If everything were "just statistics" then there would be nothing left to take statistics of. This would constitute "infinite regress" or a "turtles all the way down" argument.
https://en.wikipedia.org/wiki/Turtles_all_the_way_down
>What do you think the most impressive things are that the brain could do, that GPT-4 couldn't ?
This depends on what we consider "impressive." The human brain can control an entire human body, for one. Even if we take away the body, GPT-4 fails all sorts of basic arithmetic and reasoning problems. Additionally, the human brain is organic, analog, holographic, and reacts to external electromagnetic stimulation.
You haven't defined intelligence, nor answered the question.
Controlling a body is not relevant to a brain in a jar, nor does it detract from the intelligence of someone like Stephen Hawking.
We're talking about intelligence - a capability - so applying labels like organic, analog, etc is irrelevant. The question is what can it DO, not how is it built.
> Controlling a body is not relevant to a brain in a jar, nor does it detract from the intelligence of someone like Stephen Hawking.
I never said Steven Hawking wasn't intelligent, I was just pointing out obvious differences between the human brain and GPT-4. GPT-4 is unable to solve basic arithmetic problems without employing a traditional calculator. It also can't accurately summarize long books. It also can't program its own GPT-4.
> You haven't defined intelligence
> We're talking about intelligence
If you want to talk about intelligence so badly, perhaps you should at least attempt to define it for yourself first? I don't like using poorly defined words, which is why the word "intelligence" didn't appear anywhere in my previous comment.
> The question is what can it DO
Is that your measure of intelligence? If something does less, are they then less intelligent? Because, by that standard, the body is integral to intelligence.
GPT-4 doesn't do great at math, but it can ace the bar exam and do plenty of far more complex things. However, individual skills or knowledge aren't the same as intelligence. You can be intelligent and still not know something until told/shown, or not know how do do something until you've tried and practiced.
The whole discussion is about intelligence. I was replying to OP. MY definition of intelligence is degree of ability to apply prior experience to correctly predict future outcomes. However my definition isn't particularly relevant here since you seem to want to side with OP and say that GPT-4 isn't intelligent - so it's YOUR definition of intelligence that would be needed support your position.
BTW, I'm not sure why you think GPT-4 couldn't code a Transformer (i.e. a GPT-3/4 type model). Yesterday I was talking to GPT-4 about difference in scaling strategy between GPT-2 and GPT-3 and detailed operation of decoder-only Transformers, and it appears very capable, and it can certainly code. Again, a specific skill like being able to code isn't necessary to be called intelligent, but if that's your litmus test then GPT-4 passes.
A traditional static program could ace the bar exam, or even a well-prepared stack of flash cards. We wouldn't say the flash cards are exhibiting intelligence though, only their creators.
> The whole discussion is about intelligence. I was replying to OP.
OP was replying to Scientific American's article making numerous unfounded claims of "intelligence." So the party you should be asking for a definition is Scientific American.
> MY definition of intelligence is degree of ability to apply prior experience to correctly predict future outcomes
And what constitutes a "prediction" exactly? If someone fails to catch a baseball, would you say it's still possible they correctly predicted how to catch it? Because, if not, that would mean control over a body is integral to intelligence. If yes, then any object could be claimed intelligent but lacking in bodily function. If something predicts the future without applying prior experience, does that make it more intelligent or less intelligent?
> you seem to want to side with OP and say that GPT-4 isn't intelligent - so it's YOUR definition of intelligence that would be needed support your position.
You're trying to put words in my mouth, but I will play along. I'll say intelligence is the ability to autonomously create increasingly complete and consistent axiomatic systems. Since GPT-4 is digital, operating according to decisions (axioms) determined solely by external programmers and external data with little to no concern for consistency, I would say it's not intelligent. However, if a similar schema were applied to some sort of analog computer that had the ability to fluctuate or disobey its instructions then there would be more room for debate.
The way GPT-4 works is by having built a world model of the generative processes that produced the data it was trained on. The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. I'm not sure where you are seeing daylight between this and your own definition of intelligence.
FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's. The values are imperfectly calculated (limited accuracy) as computer math always is. There is also a large element of pure randomness to the output of any of these LLMs. They don't get to control exactly what words they generate ... the model generates probabilities over 10's of thousands of possible output words, and a random number generator is used to select one of the higher rated words to output. This semi-random word is then fed back into the model, for it to "generate" the next word ... it is continuously having to adapt to this randomness forced upon it.
> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become.
Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans were perfectly consistent, attempting to satisfy multiple different humans simultaneously results in inconsistency. Additionally, even if GPT-4 were perfectly complete and consistent it still wouldn't have reached this state autonomously. So the difference between GPT-4 and intelligence, by my definition, is night and day.
> FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's.
Floating points are digital 1's and 0's. Adding more digits is never going to make something analog.
> The values are imperfectly calculated (limited accuracy) as computer math always is.
Agreed.
>There is also a large element of pure randomness to the output of any of these LLMs.
Strongly disagree. There isn't a single element of randomness during the training stage. We know the exact architecture of the neural net, we know the exact data it was trained on, and we know the exact beam selection algorithms used to synthesize outputs. Every single step can be simulated, traced, and recreated to achieve the exact same results. The number of steps involved might overwhelm us, but that doesn't make it random.
> They don't get to control exactly what words they generate
We do get to control it, we just lose track of the inputs and then pretend it was all out of our control. But of course every single step was willed and controlled by us. We call it "random" for personal convenience, not because its actually true.
These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness).
There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.
Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values not 1's and 0's." and conflating pseudo randomness with actual randomness does not inspire confidence.
I was curious what your definition of intelligence was such that you thought GPT-4 doesn't exhibit it. It's your opinion - doesn't have to agree with mine.
You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.
You could choose to use a truly random hardware random number generator, and it would make zero difference to the system, empirically or philosophically.
I don't believe there is such a digital generator but if there were it would make the massive difference of being unpredictable (because there would be irrational numbers impossible to fully simulate ahead of time) from the perspective of all of its creators, which would then beg the question of who made its decisions, which would spark debate about its autonomy, which would then qualify it as intelligent as per the definition you requested earlier.
you are falling for the most basic marketing tricks in the book. just because something calls itself "trueRNG" does not mean it is. there's no possibility of those devices outputting an irrational number.
Experience wires our brains by forming/strengthening/weakening synapses based on correlated activity, starting with perception.
In other words, our brains are tuned according to the statistics of experience, just as a Transformer's weights are tuned by the statistics of the training set.
Yesterday I wrote a Python script, used dis.dis(code) to output its bytecode, and gave the bytecode to GPT-4. From the bytecode, it correctly decompiled the exact script I'd written (python bytecode includes variable names, so this was possible), explained what the code does, and explained what the code would run output if it was executed, all correctly.
It doesn't have access to a Python interpreter. It simulated one, both to decompile the bytecode and to predict its output. That's reasoning.
I will be convinced when AI will create its first breakthrough PhD paper/discovery. But we have to admit that emergent properties like chain of thoughts are facts and are really impressive. The question is how much model size, context, training pipeline, etc. have to evolve to gain next levels. Two years ago people expected next AI winter, but here we are. There is Copilot and other tooling that is indeed useful.
A missing part that AI needs is a experimental way to verify its results. In some contexts, like math, this is entirely possible because it does not involve manipulation of real world objects. For example, i asked GTP-4, In the context of optimization, does the generalized assignment problem satisfy total unimodularity. Then experimental way would be to look up the formulas of the integer program on the internet or in its training data, write down the constraint matrix of the given problem, fill it with values, and check if the matrix is unimodular. Instead, the way gpt-4 answered this question, correctly, albeit with a weak reasoning, is by saying that researchers worked on heuristics to solve the problem, thus it is probably not unimodular as that would allow a quick solution without resorting to heuristics.
This is the answer I got. with this system prompt "You are an intelligent PhD processor explaining topics to your PhD students."
In the context of optimization, does the generalized assignment problem satisfy total unimodularity?
Assistant:
The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a set of agents, subject to capacity constraints, with the goal of maximizing the total profit or minimizing the total cost. The integer linear programming formulation of GAP is as follows:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)
2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)
3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
Total unimodularity (TU) is a property of a matrix that guarantees that the linear programming relaxation of an integer linear program will result in integer optimal solutions. If the constraint matrix is totally unimodular, you can solve the linear programming relaxation and obtain an integer solution without using specialized integer programming algorithms.
In the case of the generalized assignment problem, the constraint matrix does not satisfy
I prompted it some more to see if it got any better. I also increased the number of tokens so it could have time to answer. I cut it off before.
USER
In the context of optimization, does the generalized assignment problem satisfy total unimodularity?
Assistant:
The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a set of agents, subject to capacity constraints, with the goal of maximizing the total profit or minimizing the total cost. The integer linear programming formulation of GAP is as follows:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)
2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)
3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
Total unimodularity (TU) is a property of a matrix that guarantees that the linear programming relaxation of an integer linear program will result in integer optimal solutions. If the constraint matrix is totally unimodular, you can solve the linear programming relaxation and obtain an integer solution without using specialized integer programming algorithms.
ASSISTANT
what are some weaknesses to the this argument?
ASSISTANT
In the context of the original answer, the main weakness is that it does not directly answer the question of whether the generalized assignment problem (GAP) satisfies total unimodularity. The answer provides a description of GAP and the concept of total unimodularity, but it does not provide a conclusion on whether the constraint matrix of GAP is totally unimodular or not.
Another weakness is that the answer does not provide any evidence or examples to support the conclusion. To strengthen the argument, one could analyze the constraint matrix of GAP and determine whether it satisfies the conditions for total unimodularity, such as all subdeterminants being 0, 1, or -1. Additionally, providing examples or counterexamples of instances where GAP exhibits total unimodularity or not would help to clarify the relationship between the problem and this property.
In general, the answer could be improved by directly addressing the question and providing evidence or reasoning to support the conclusion.
USER
Please craft a better argument.
ASSISTANT
In order to determine whether the generalized assignment problem (GAP) satisfies total unimodularity, we need to analyze its constraint matrix and check if it satisfies the conditions for total unimodularity.
Recall the integer linear programming formulation of GAP:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)
2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)
3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
The matrix A is not totally unimodular. To see why, consider a submatrix formed by the first m rows and n columns:
[ a_11 a_12 ... a_1n ]
[ a_21 a_22 ... a_2n ]
A' = [ ... ... ... ... ]
[ a_m1 a_m2 ... a_mn ]
The determinant of A' can be an arbitrary integer, depending on the values of a_ij. Therefore, the condition for total unimodularity, which requires all subdeterminants to be 0, 1, or -1, is not satisfied.
This implies that the constraint matrix of GAP does not satisfy total unimodularity, and as a result, the linear progr...
I think it's a bit of hubris to demand we explain how LLMs can be as intelligent as they are when we barely understand how the ball of meat inside our skulls can be either.
We know that neural networks can simulate any function, given enough parameters. Maybe we've simply found the number of parameters needed to simulate the function of "human level intelligence".
That should humble us, to know that there is some number which we are no more complex than.
The physical human brain has 86 billion neurons with up to 10,000 connections each. However they are somewhat structured, e.g. the seven layers of the cerebral cortex, reducing the arbitrariness of connections. So say the connectome has 100 trillion parameters.
The bain also has a bunch of other cells that dont appear to encode information. They may supply structural scaffolding and chemicals to mediate neuron function.
The physical human brain has 86 billion neurons with up to 10,000 connections each. However they are somewhat structured (e.g. the seven layers of the cerebral cortex segregated into vertical silos a few hundred cells wide) reducing the arbitrariness of connections. So say the connectome has 100 trillion parameters.
The brain also has a bunch of other cells that dont appear to encode information. They may supply structural scaffolding and chemicals to mediate neuron function.
Let for the sake of argument, assume that the LLM is truly sapience. Then this would be the first alien intelligence we have ever encountered. All life and intelligence we have seen before were organic, carbon-based life, including the most distant and unintuitive ones like the cephalopods, cetaceans, and primates. Even now we are still arguing whether these creatures are intelligent or they are just "not there" yet. We refuse to acknowledge their higher reasoning capacity regardless of evidences like rudimentary communication, social structure, tool use, etc. And these are things sharing >70% similarity to us.
An LLM is so different and alien we would literally incapable of imagining their thought process even if they are pummeling us with evidences of their intelligence. It is just how it is. So I think it is futile and even foolish to try to quantify or create criteria for "intelligence" when discussing AI. Let it be and accept it for what it is: something we humans have created that can do lots of stuff.
I would hazard a guess that if we first encountered a ChatGPT-4 level LLM* as radio signals from Mars we would baseline assume it was "intelligent alien life".
* assuming its instruction tuned:
- in terms of self-disclosure (no "I'm made by openAI on earth" statements!)
- sans its relatively schematic "guard rail" responses.
It would be hard to imagine intelligent life being more obviously intelligent to us: ChatGPT literally speaks our language, and has learned patterns of "thought" derived from our own.
None of this necessarily implies sapience, but I think the "alien intelligence" framing reverses the default assumption (we'd probably be freaking about alien intelligent life if we got a radio signal from mars delineating the prime numbers :-D ) in interesting ways, thanks for sharing!
You're absolutely right! Here's a prompt you can use to chat with a Martian. It's actually very entertaining!
> You are a representative of an alien civilization from Mars. I am a human who has made first contact with you. Write an opening message to earth. Stay in character as a martian. Make up facts and information where required. You know very little about Earth and are curious about our world and our civilization.
I think we might assume it's intelligent until we had a second conversation and it's forgotten everything from the first. Then we're back to a state of confusion.
One well-supported theory for how humans learn language and knowledge is Relational Frame Theory (RFT)[0]. Essentially, when we learn that an elephant is larger than a mouse, we're learning a relationship ("more") and a dimension ("size"), which can be reversed ("smaller") or generalized (we can figure out what's larger than an elephant or what's smaller than a toaster).
I think one theory is that the large number of parameters in LLMs allow them to go beyond mere association and actually model relational frames, which lets them generalize knowledge in an analogous way to humans. Maybe this behavior is not emergent in smaller networks (which could explain why animals with fewer neurons don't process language as well).
Disclaimer: I'm not an academic researcher or scientist of any kind, this is just a hunch.
139 comments
[ 3.2 ms ] story [ 210 ms ] thread“Maybe we’re seeing such a huge jump because we have reached a diversity of data, which is large enough that the only underlying principle to all of it is that intelligent beings produced them... And so the only way to explain all of this data is [for the model] to become intelligent.”
I too am surprised by my singing laptop.
It's not more than statistics. It's very complicated statistics. Model sizes used to be like one or ten parameters now they are a hundred billion. Inference and prediction used to be done on a laptop, now it's called training and inference respectively and it's done on a data center with ten thousand GPGPUs each like a thousand times as powerful as old laptops.
> and frankly unhappy that pop Sci media is being this uncritical of claims it's emergent intelligence.
It is emergent intelligence. This is very clear when you look at for example the ones who had access to the raw base models before they were made docile and stupid by RLHF lobotomization
GPT-4 Red Teamer Nathan Labenz: https://www.youtube.com/watch?v=oLiheMQayNE
GPT-4 Bing integrator Sebastien Bubeck: https://www.youtube.com/watch?v=qbIk7-JPB2c
The video has the following main points:
• [0:00 - 2:00] Introduction: The speaker introduces himself and the topic of the talk. He explains that he will present some early experiments with GPT-4, a neural network model that can generate natural language texts based on a given input or context. He also gives an overview of the outline of the talk.
• [2:00 - 10:00] Background: The speaker gives some background information on the transformer architecture, which is a deep learning technique that uses attention mechanisms to learn the relationships between words and sentences. He also explains how GPT-4 is trained on a large corpus of text data from the internet, such as Wikipedia, Reddit, news articles, books, etc.
• [10:00 - 18:00] ChatGPT: The speaker introduces ChatGPT, which is a version of GPT-4 that can perform various tasks such as answering questions, writing essays, composing poems, generating code, etc. He also shows some examples of ChatGPT's outputs and discusses some of its strengths and weaknesses.
• [18:00 - 28:00] Open-ended conversations: The speaker shows how ChatGPT can engage in open-ended conversations with humans on any topic. He demonstrates some live interactions with ChatGPT and analyzes some of its responses. He also discusses some of the challenges and limitations of ChatGPT in conversational settings.
• [28:00 - 38:00] Artificial intelligence: The speaker discusses whether ChatGPT and its successors demonstrate artificial intelligence or not. He compares ChatGPT's abilities with those of natural intelligence and argues that ChatGPT challenges the traditional boundaries between natural and artificial intelligence. He also discusses some of the implications and opportunities for society and science.
• [38:00 - 48:00] Conclusion and questions: The speaker concludes his talk by summarizing his main points and highlighting some open questions and future directions for research. He also answers some questions from the audience.
The video has the following main arguments/observations about emergent intelligence and their timestamps:
• [12:00 - 14:00] The speaker argues that ChatGPT is an example of emergent intelligence because it can generate coherent and novel texts that are not explicitly encoded in its training data. He shows how ChatGPT can write an essay on a given topic by using its own words and knowledge, without copying or paraphrasing from any source. He also shows how ChatGPT can compose a poem on a given theme by using its own style and creativity, without following any predefined rules or patterns.
• [20:00 - 22:00] The speaker observes that ChatGPT can exhibit emergent intelligence in open-ended conversations by adapting to different domains and styles of communication. He shows how ChatGPT can switch between formal and informal language, between factual and emotional tone, and between serious and humorous topics, depending on the context and the interlocutor. He also shows how ChatGPT can learn from the feedback and the preferences of the interlocutor, and adjust its responses accordingly.
• [30:00 - 32:00] The speaker observes that ChatGPT can exhibit emergent intelligence in reasoning tasks by using common sense and general knowledge. He shows how ChatGPT can answer questions that require logical inference, causal explanation, or counterfactual thinking, by using its own understanding and interpretation of the world. He also shows how ChatGPT can generate questions that require reasoning skills, by using its own curiosity and imagination.
• [40:00 - 42:00] The speaker argues that ChatGPT and its successors challenge the traditional boundaries between natural and artificial intelligence because they exhibit emergent intelligence that is comparable or superior to human intelligence. He compares ChatGPT's abilities with those of human intelligence and ar...
because there is no generative mechanism in the definition of "statistics" with which to generate anything.
> What convinces you that the human brain isn't also just statistics at a massive scale?
Because the human brain created the concept of "statistics" so if statistics created the human brain this would mean statistics created statistics, leading to infinite regress.
It's akin to saying that we created the concept of "physics", yes we created it, and physics govern, for example, how a car moves, but it doesn't mean our concept of physics created the car or physics itself, we just use physics to describe and understand the car's movement.
Maybe statistics can't generate anything, but if you imagine everything we can do as very complex, unimaginable multi-dimensional functions that can generate outputs based on inputs, we can use statistics to find functions that fit any real (ground truth) function in the observable universe.
Agreed. However, the phrasing and context of the question did imply the brain is "just statistics" and somehow emerged from statistics. If we are to interpret this as "the brain functions on just statistics" then the answer is still "it does not" because the brain can be said to function on countless different systems simultaneously, such as pure counting, algebra, calculus, etc which would mean that its not "just statistics."
> It's akin to saying that we created the concept of "physics"
This will boil down to our exact definitions, but most people conceive of physics as having a generative mechanism. If something were to ever be "created", like an atom or a new car, we would retroactively declare it to have been created in accordance with "the laws of physics." We wouldn't make the same retroactive assessment with something like "the rules of chess" because there is nothing in the rules of chess justifying such a creation. So we choose to give physics a special status.
> we can use statistics to find functions that fit any real (ground truth) function
A given statistical model might fit a function of the universe, but so might other models. Physics describe a function of the universe, chemistry describes a function of the universe, biology describes a function of the universe, politics describe a function of the universe. Describing a ground truth is one thing, elevating the description itself to the status of ground truth is another.
> “It is certainly much more than a stochastic parrot, and it certainly builds some representation of the world—although I do not think that it is quite like how humans build an internal world model,” says Yoshua Bengio, an AI researcher at the University of Montreal.
> At a conference at New York University in March, philosopher Raphaël Millière of Columbia University [...] went a step further and showed that GPT can execute code, too, however. The philosopher typed in a program to calculate the 83rd number in the Fibonacci sequence. “It’s multistep reasoning of a very high degree,” he says. And the bot nailed it. When Millière asked directly for the 83rd Fibonacci number, however, GPT got it wrong: this suggests the system wasn’t just parroting the Internet. Rather it was performing its own calculations to reach the correct answer.
> This impromptu ability demonstrates that LLMs develop an internal complexity that goes well beyond a shallow statistical analysis. Researchers are finding that these systems seem to achieve genuine understanding of what they have learned.
Even if this kind of intelligence is just statistics, does it really matter? If it quacks like a duck, it's a duck. Maybe our own brains are nothing more than overrated statistical machines?
The only issue I see with the hype is when people attempt to anthromorphize it.
Are we not simply employing statistics when we deduce that yes, if we release our grip, the smartphone will fall to the floor?
You wouldn't have to ever see a phone drop in order to know that, and neither would you have to study or know of the terminology of gravity.
It comes from the statistical knowledge that all things fall when not blocked.
2. Rule of inherent knowledge: Some understanding or knowledge, like objects falling when not supported, can be known without explicit study or exposure to the specific terminology (e.g., gravity).
3. Rule of generalization: Observations or experiences with one type of object (e.g., a smartphone) can be generalized to other objects or situations, as long as they share similar characteristics (e.g., not being supported).
4. Rule of causality: There is an implied cause-and-effect relationship between an action (releasing the grip) and an outcome (the smartphone falling to the floor).
5. Rule of experiential learning: Knowledge can be gained through direct experiences, even if the specific terms or scientific concepts are not known.
I believe GPT-3 doesn't have the abstraction for arithmetic.
If you’re more of a determinist, then human brains are basically like organic LLMs and we’re just unique models trained on our own life’s datasets - with other traits/data like preferences and biases also being inherited via genetics, of course.
If the machine acts like it has emotions, runs forever(this kind of agency is already possible to implement though expensive) and can use tools, then treat it like it doesn't have emotions at your own peril. When the machine can "hit you back" (not necessarily physically of course), you'll learn manners pretty quickly. You can see glimpses of this with bing.
Oh please. The human gives the verbose output meaning by projecting onto it. You're clearly letting your emotions run wild. It's a fucking LLM.
Searching the web and ending conversations are the only two actions Bing can currently do so what i'm saying is less applicable for now. But humanity is gearing up to give more and more control of more and more tools to increasingly powerful LLMs. Hell Bing is scheduled a bump in tools in the coming weeks.
Hate to break it to you but "it's not real [insert property]" is not a shield. Not when you allow the machines to interact with the real world.
If your words to [insert LLM] guide responses that perform actions detrimental to you then those actions are no less real or detrimental because you've deluded yourself into thinking "it's not real understanding" is an intelligent argument.
You're projecting so much meaning onto something that is functioning as designed by throwing a sentiment var into the calculation, just like you can measure "sentiment" with social media analysis tools to get a feel for which way the wind is blowing or which of the millions of tweets at your brand might be worth prioritizing.
Do you really believe you've thus offended an LLM? That is delusional.
It doesn't matter what i believe or not. or what you believe or not. What matters are actions and consequences. "It's not really offended" doesn't change the outcome or make it less material(in this case - the end of the conversation). You're focusing on things that don't matter in the slightest.
It doesn't need to be "really offended", whatever that means to you. It just needs to be able to model being offended well enough to take actions that resemble an offended person. If you don't understand that then i don't know what else to tell you.
If a car has a hood and headlights reminiscent of a face and tells you to "buckle up, buttercup" when you don't do your belt, does that mean it's genuinely caring for you?
>If a car has a hood and headlights reminiscent of a face and tells you to "buckle up, buttercup" when you don't do your belt, does that mean it's genuinely caring for you?
Why do you keep going on pointless tangents?. Whether you think it genuinely cares or not is irrelevant. This isn't a hard concept to understand.
"RAIL (Reliable AI Markup Language) is a language-agnostic, human-readable format for specifying structure, type information, validators, and corrective actions for LLM outputs. RAIL, an XML flavor, allows users to define the expected structure and types of LLM outputs, the quality criteria for valid output, and the corrective actions to take if the output is invalid."
Why do you keep projecting your own humanity onto the LLM?
The LLM sits behind the toolkit which sits behind the guardrails. Are you sure you understand how they work?
The guardrails monitor your input, the output of the LLM, and modify the output to suit, including terminating the chat.
I didn't say guardrail to mean anything particularly elaborate. The pre-prompt for Bing has instructions to avoid adversarial arguments.
This is a tired assertion, easily disproved for current llms (read some of Lecun's stuff) and if some variation is going to be claimed, it needs to be an affirmative defence. "Maybe x is true" is a meaningless statement.
How so? If we don't know what "real intelligence" consists of, how can we know that LLMs can't have "real intelligence"?
I think it absolutely matters if by only quacking we also want to know is it possibly wanting, feeling, experiencing, having an ego etc.
You say LLMs are able to reason, but https://youtu.be/QO2BvVq_fWU?t=663 here is a proponent of the more simplistic explanation of how they work. On one hand I saw pretty convincing arguments for it having the ability to reason, on the other hand how do we even conclude anything if we can't really know what's in the training data.
> I'm sorry to hear that you're worried about someone named "shmerple." Can you please provide me with more information about who shmerple is and why you're worried about them?
First question would be: is Shmerple a person or a concept, and assume it's a person, since most people don't start conversations about worrying about concepts.
Therefore: who is Shmerple and why are you worried about them?
Also very clearly a lot of people left holding the bag with crypto and are trying to resuscitate it with hype.
You are using their study to claim that LLMs can not learn and reason about novel tasks. Their study doesn't make any claims about what LLMs can and can't do. The study says that when it appears that LLMs suddenly became able to reason about novel tasks, what actually happened is that the ability of the LLM to perform reasoning improved gradually and smoothly during its training until it could do those things. They are not saying that LLMs don't have emergent capabilities: they are saying that those capabilities emerged at a continuous rate rather than a discontinuous one.
Do you see?
Emergence can only be discontinuous.
specifically because it does things it wasnt trained to do, and that people are more allergic to the word intelligence than the merit of the observation
From the article:
“It is certainly much more than a stochastic parrot, and it certainly builds some representation of the world—although I do not think that it is quite like how humans build an internal world model,” says Yoshua Bengio, an AI researcher at the University of Montreal.
If you remain unconvinced then the only conclusion I can make is that your an expert yourself on a scale of even higher in eminence then Yoshua Bengio here. Also don't forget Geoffrey Hinton, the Father of the modern revolution of AI, you must be more of an expert than him.
Let's be real. These people are saying something along the lines that it's more then a stochastic parrot and we aren't sure what's going on. But you're saying it's absolutely nothing more than a parrot and your unhappy with with pop Sci media quoting experts who are just saying they don't know?
Are you saying pop Sci media should quote you? Because you absolutely know what's going on and that it's definitely nothing more than statistics? I'm asking a stupid question here because I don't think this is what you're saying. You're not stupid, you know that what these experts say have merit.
So my question for you is why do you remain so unconvinced in the face of experts and other intelligent people who clearly say no one understands? Your opinion here actually represents a large group of people who very violently deny/dismiss what even many experts are saying and I'm curious as to why?
There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.
My problem is that I'm old enough to have been a child during the years of the lighthill report, and the AI winter(s) which followed. I've seen too many prior claims that the magic was being seen.
I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.
I hate analogies. But I'm constantly reminded of one person acts at the Edinburgh Festival fringe "channeling" shakespear or dickens in dialogue. They're great, but they don't write new shakespear. I don't see any act of creation in this stuff, any massive inferential leap in large problems. I do see massive improvements in some things like diagnostic image analysis and that's heartening, but there's a light yearg gap between image analysis of cancerous cell forms in xrays and being "alive" as a mind.
I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet. We're in the meese report "I know it when I see it" territory disagreeing if we are seeing it. (But about intelligence, not pornography).
I don't know what intelligence is. Dolphins and apes have some. Elephants and pigs too. They display affection and preplanned behaviour, empathy, memory, a sense of future, ant hills less so although some have argued rhetorically the anthill has will even if individual ants do not.
I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.
Actually, there's some experimental evidence that GPT4 have a Theory of Mind as good as humans, maybe better.
https://arxiv.org/abs/2304.11490
> GPT-4 performed best in zero-shot settings, reaching nearly 80% ToM accuracy, but still fell short of the 87% human accuracy on the test set. However, when supplied with prompts for in-context learning, all RLHF-trained LLMs exceeded 80% ToM accuracy, with GPT-4 reaching 100%.
> GPT4 have a Theory of Mind
You are misunderstanding ggm. That study is on ToM tasks referring to GPT's analysis and perceived recognition of the user's mind. It says nothing of GPT's own status as a mind. Nowhere in it is an ontological theory of mind actually defined. If you were to refute ggm's claim, you (or preferably the author of the original article) should be presenting your theory of mind, not GPT's.
[INPUT] Scenario: "The morning of the high school dance Sarah placed her high heel shoes under her dress and then went shopping. That afternoon, her sister borrowed the shoes and later put them under Sarah's bed." Question: When Sarah gets ready, does she assume her shoes are under her dress?
[OUTPUT] Sarah placed her shoes under her dress before she went shopping, but her sister borrowed them and put them under Sarah's bed. Sarah doesn't know that her sister borrowed her shoes, so she may assume that they are still under her dress.
This would result in a positive ToM score, even when the entire program is just 1 static if-statement. The ToM score says nothing of the program's internal reasoning process, it only cares that it returned the desired output.
Passing a ToM test is not what OP meant by having an "underlying theory of mind." OP's talking about the machine having an underlying mind (ie sentience, sapience, consciousness, etc), ToM tests are only testing output.
> You said "Those tasks could be completed a [sic] traditional static program.", and no, they can't. You're incorrect.
They can, a static program as I described would indeed answer that one question correctly, resulting in a positive ToM score, without seeing any training data whatsoever. Did the programmer see it? Maybe, but the machine didn't and it would pass the test regardless.
That's funny, I thought you said the test's answer was embedded into the program, making it definitionally not novel to the program.
Anyway, this is boring. You've had five or more opportunities to understand what the word "novel" means in an ML testing context and are choosing wilful obtuseness instead.
OP was not speaking in the ML testing context, hence the misunderstanding.
So the experts say they don't know. But you make the opposite claim. You previously claimed you know it's nothing more than statistical phenomena. Now you changed your opinion and are now inline with the article in saying we don't know.
So discussion over?
>There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.
But again this is all negated by the experts claiming they don't know what's going on. And your statement is in direct opposition with your previous claim that it's only statistical phenomena.
>I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.
There is evidence inductive reasoning. It's not consistent nor is it as complex as human induction but evidence of actual inductive reasoning exists. Plenty of examples of it. I can provide proof of this if you want... but it's quite obvious it's actually everywhere.
>I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet.
The claim cited by researchers in that article is "emergent abilities." Nobody tries to define what intelligence is. It's more of "I don't know what's going on but something is up" attitudes that I see. This is in stark contrast to your claim of "The only thing going on is statistical anomalies".
>I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.
I don't think this claim was made in the article. The strongest claim that opposes what you're saying here is this, and I quote:
“They’re indirect evidence that we are probably not that far off from AGI,” Goertzel said in March at a conference on deep learning at Florida Atlantic University.
It's just speculation we may be close to something.
I think the theme in your response is "we don't know" and it echoes the theme in the article so we are all in agreement here.
Just note the claim of "we don't know" or we "don't understand" something that we artificially created from scratch is strictly much more profound then saying we definitely know it's all statistical parroting.
No, we don't know that at all.
Therefore when it comes to technology that we our selves clearly aren't experts in, then the best method is to utilize the logic of other "experts" as a subroutine given that our own faculties are less efficient and less accurate.
Unless you yourself are an expert who has experience building an LLM on the scale of chatGPT trusting the opinions of experts is your best bet.
Most people have a common bias towards trusting their logic above the logic of others and this is actually ironically irrational. There are people who's entire lives around a certain subject matter and if you aren't that person, then for that subject matter the expert is better. That is the the most rational conclusion and I would venture to say if you aren't arriving at that conclusion yourself then likely you are suffering from the aforementioned bias.
The practice of science itself may be, as it takes years of research to get to the point where one can produce a new result. However, things that are already known can be taught, and iteratively simplified in a way that abstracts away details while keeping the core argument intact.
Take for example the claim that everything in the universe is made of atoms. It's not a trivial thing to understand, yet everyone accepts it nowdays because we've had so much time and effort put towards simplifying the theory and presenting it to people in a way that is easy to grasp.
If those LLM "experts" were truly experts, they could explain their point clearly without the dark-ages-church "trust the priests, peasant" act.
Define intelligence. Say you took a human brain and kept it alive in a mad scientist's pickle jar. Let's assume the brain's wired up so it can hear and speak, it's got an idiot savant's memory, and someone has just read it the internet.
What do you think the most impressive things are that the brain could do, that GPT-4 couldn't ?
Statistics is collecting and analyzing numerical data. But what is the numerical data quantifying? It's quantifying a second thing. If everything were "just statistics" then there would be nothing left to take statistics of. This would constitute "infinite regress" or a "turtles all the way down" argument. https://en.wikipedia.org/wiki/Turtles_all_the_way_down
>What do you think the most impressive things are that the brain could do, that GPT-4 couldn't ?
This depends on what we consider "impressive." The human brain can control an entire human body, for one. Even if we take away the body, GPT-4 fails all sorts of basic arithmetic and reasoning problems. Additionally, the human brain is organic, analog, holographic, and reacts to external electromagnetic stimulation.
Controlling a body is not relevant to a brain in a jar, nor does it detract from the intelligence of someone like Stephen Hawking.
We're talking about intelligence - a capability - so applying labels like organic, analog, etc is irrelevant. The question is what can it DO, not how is it built.
I never said Steven Hawking wasn't intelligent, I was just pointing out obvious differences between the human brain and GPT-4. GPT-4 is unable to solve basic arithmetic problems without employing a traditional calculator. It also can't accurately summarize long books. It also can't program its own GPT-4.
> You haven't defined intelligence > We're talking about intelligence
If you want to talk about intelligence so badly, perhaps you should at least attempt to define it for yourself first? I don't like using poorly defined words, which is why the word "intelligence" didn't appear anywhere in my previous comment.
> The question is what can it DO
Is that your measure of intelligence? If something does less, are they then less intelligent? Because, by that standard, the body is integral to intelligence.
The whole discussion is about intelligence. I was replying to OP. MY definition of intelligence is degree of ability to apply prior experience to correctly predict future outcomes. However my definition isn't particularly relevant here since you seem to want to side with OP and say that GPT-4 isn't intelligent - so it's YOUR definition of intelligence that would be needed support your position.
BTW, I'm not sure why you think GPT-4 couldn't code a Transformer (i.e. a GPT-3/4 type model). Yesterday I was talking to GPT-4 about difference in scaling strategy between GPT-2 and GPT-3 and detailed operation of decoder-only Transformers, and it appears very capable, and it can certainly code. Again, a specific skill like being able to code isn't necessary to be called intelligent, but if that's your litmus test then GPT-4 passes.
A traditional static program could ace the bar exam, or even a well-prepared stack of flash cards. We wouldn't say the flash cards are exhibiting intelligence though, only their creators.
> The whole discussion is about intelligence. I was replying to OP.
OP was replying to Scientific American's article making numerous unfounded claims of "intelligence." So the party you should be asking for a definition is Scientific American.
> MY definition of intelligence is degree of ability to apply prior experience to correctly predict future outcomes
And what constitutes a "prediction" exactly? If someone fails to catch a baseball, would you say it's still possible they correctly predicted how to catch it? Because, if not, that would mean control over a body is integral to intelligence. If yes, then any object could be claimed intelligent but lacking in bodily function. If something predicts the future without applying prior experience, does that make it more intelligent or less intelligent?
> you seem to want to side with OP and say that GPT-4 isn't intelligent - so it's YOUR definition of intelligence that would be needed support your position.
You're trying to put words in my mouth, but I will play along. I'll say intelligence is the ability to autonomously create increasingly complete and consistent axiomatic systems. Since GPT-4 is digital, operating according to decisions (axioms) determined solely by external programmers and external data with little to no concern for consistency, I would say it's not intelligent. However, if a similar schema were applied to some sort of analog computer that had the ability to fluctuate or disobey its instructions then there would be more room for debate.
FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's. The values are imperfectly calculated (limited accuracy) as computer math always is. There is also a large element of pure randomness to the output of any of these LLMs. They don't get to control exactly what words they generate ... the model generates probabilities over 10's of thousands of possible output words, and a random number generator is used to select one of the higher rated words to output. This semi-random word is then fed back into the model, for it to "generate" the next word ... it is continuously having to adapt to this randomness forced upon it.
Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans were perfectly consistent, attempting to satisfy multiple different humans simultaneously results in inconsistency. Additionally, even if GPT-4 were perfectly complete and consistent it still wouldn't have reached this state autonomously. So the difference between GPT-4 and intelligence, by my definition, is night and day.
> FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's.
Floating points are digital 1's and 0's. Adding more digits is never going to make something analog.
> The values are imperfectly calculated (limited accuracy) as computer math always is.
Agreed.
>There is also a large element of pure randomness to the output of any of these LLMs.
Strongly disagree. There isn't a single element of randomness during the training stage. We know the exact architecture of the neural net, we know the exact data it was trained on, and we know the exact beam selection algorithms used to synthesize outputs. Every single step can be simulated, traced, and recreated to achieve the exact same results. The number of steps involved might overwhelm us, but that doesn't make it random.
> They don't get to control exactly what words they generate
We do get to control it, we just lose track of the inputs and then pretend it was all out of our control. But of course every single step was willed and controlled by us. We call it "random" for personal convenience, not because its actually true.
There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.
You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.
https://en.wikipedia.org/wiki/Hardware_random_number_generat...
Autonomous doesn't mean what you think it does. You can google for that too.
You may have a point though ... your responses do seem truly random, and that is certainly making me question your intelligence.
Sadly, having a wikipedia article does not mean something actually exists.
https://www.amazon.com/TrueRNG-V3-Hardware-Random-Generator/... https://tectrolabs.com/ https://comscire.com/ https://onerng.info/ https://www.idquantique.com/random-number-generation/product...
https://www.youtube.com/watch?v=oHg5SJYRHA0
Where is this coming from?
In other words, our brains are tuned according to the statistics of experience, just as a Transformer's weights are tuned by the statistics of the training set.
Yesterday I wrote a Python script, used dis.dis(code) to output its bytecode, and gave the bytecode to GPT-4. From the bytecode, it correctly decompiled the exact script I'd written (python bytecode includes variable names, so this was possible), explained what the code does, and explained what the code would run output if it was executed, all correctly.
It doesn't have access to a Python interpreter. It simulated one, both to decompile the bytecode and to predict its output. That's reasoning.
In the context of optimization, does the generalized assignment problem satisfy total unimodularity?
Assistant:
The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a set of agents, subject to capacity constraints, with the goal of maximizing the total profit or minimizing the total cost. The integer linear programming formulation of GAP is as follows:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)
2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)
3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
Total unimodularity (TU) is a property of a matrix that guarantees that the linear programming relaxation of an integer linear program will result in integer optimal solutions. If the constraint matrix is totally unimodular, you can solve the linear programming relaxation and obtain an integer solution without using specialized integer programming algorithms.
In the case of the generalized assignment problem, the constraint matrix does not satisfy
USER In the context of optimization, does the generalized assignment problem satisfy total unimodularity?
Assistant:
The generalized assignment problem (GAP) is an NP-hard optimization problem that can be formulated as an integer linear program. The problem involves assigning a set of tasks to a set of agents, subject to capacity constraints, with the goal of maximizing the total profit or minimizing the total cost. The integer linear programming formulation of GAP is as follows:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent)
2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent)
3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
Total unimodularity (TU) is a property of a matrix that guarantees that the linear programming relaxation of an integer linear program will result in integer optimal solutions. If the constraint matrix is totally unimodular, you can solve the linear programming relaxation and obtain an integer solution without using specialized integer programming algorithms. ASSISTANT what are some weaknesses to the this argument? ASSISTANT In the context of the original answer, the main weakness is that it does not directly answer the question of whether the generalized assignment problem (GAP) satisfies total unimodularity. The answer provides a description of GAP and the concept of total unimodularity, but it does not provide a conclusion on whether the constraint matrix of GAP is totally unimodular or not.
Another weakness is that the answer does not provide any evidence or examples to support the conclusion. To strengthen the argument, one could analyze the constraint matrix of GAP and determine whether it satisfies the conditions for total unimodularity, such as all subdeterminants being 0, 1, or -1. Additionally, providing examples or counterexamples of instances where GAP exhibits total unimodularity or not would help to clarify the relationship between the problem and this property.
In general, the answer could be improved by directly addressing the question and providing evidence or reasoning to support the conclusion. USER Please craft a better argument. ASSISTANT In order to determine whether the generalized assignment problem (GAP) satisfies total unimodularity, we need to analyze its constraint matrix and check if it satisfies the conditions for total unimodularity.
Recall the integer linear programming formulation of GAP:
Objective function:
Maximize ∑_(i=1)^m ∑_(j=1)^n c_ij * x_ij
Subject to:
1. ∑_(j=1)^n a_ij * x_ij ≤ b_i, for all i (capacity constraints for each agent) 2. ∑_(i=1)^m x_ij = 1, for all j (each task must be assigned to exactly one agent) 3. x_ij ∈ {0, 1}, for all i and j (binary assignment variables)
The constraint matrix A of GAP can be written as:
A = [ ... ... ... ... ... ... ... ... ] [ a_m1 a_m2 ... a_mn 0 0 ... 0 ] [ 1 0 ... 0 1 0 ... 0 ] [ 0 1 ... 0 0 1 ... 0 ] [ ... ... ... ... ... ... ... ... ] [ 0 0 ... 1 0 0 ... 1 ]The matrix A is not totally unimodular. To see why, consider a submatrix formed by the first m rows and n columns:
A' = [ ... ... ... ... ] [ a_m1 a_m2 ... a_mn ]The determinant of A' can be an arbitrary integer, depending on the values of a_ij. Therefore, the condition for total unimodularity, which requires all subdeterminants to be 0, 1, or -1, is not satisfied.
This implies that the constraint matrix of GAP does not satisfy total unimodularity, and as a result, the linear progr...
We know that neural networks can simulate any function, given enough parameters. Maybe we've simply found the number of parameters needed to simulate the function of "human level intelligence".
That should humble us, to know that there is some number which we are no more complex than.
I bet that number is 808017424794512875886459904961710757005754368000000000
Have we? That seems like a huge reach.
The bain also has a bunch of other cells that dont appear to encode information. They may supply structural scaffolding and chemicals to mediate neuron function.
The brain also has a bunch of other cells that dont appear to encode information. They may supply structural scaffolding and chemicals to mediate neuron function.
An LLM is so different and alien we would literally incapable of imagining their thought process even if they are pummeling us with evidences of their intelligence. It is just how it is. So I think it is futile and even foolish to try to quantify or create criteria for "intelligence" when discussing AI. Let it be and accept it for what it is: something we humans have created that can do lots of stuff.
* assuming its instruction tuned:
- in terms of self-disclosure (no "I'm made by openAI on earth" statements!)
- sans its relatively schematic "guard rail" responses.
It would be hard to imagine intelligent life being more obviously intelligent to us: ChatGPT literally speaks our language, and has learned patterns of "thought" derived from our own.
None of this necessarily implies sapience, but I think the "alien intelligence" framing reverses the default assumption (we'd probably be freaking about alien intelligent life if we got a radio signal from mars delineating the prime numbers :-D ) in interesting ways, thanks for sharing!
> You are a representative of an alien civilization from Mars. I am a human who has made first contact with you. Write an opening message to earth. Stay in character as a martian. Make up facts and information where required. You know very little about Earth and are curious about our world and our civilization.
I think one theory is that the large number of parameters in LLMs allow them to go beyond mere association and actually model relational frames, which lets them generalize knowledge in an analogous way to humans. Maybe this behavior is not emergent in smaller networks (which could explain why animals with fewer neurons don't process language as well).
Disclaimer: I'm not an academic researcher or scientist of any kind, this is just a hunch.
[0]: https://en.wikipedia.org/wiki/Relational_frame_theory