23
A prevailing sentiment online is that GPT-4 still does not understand what it talks about. We can argue semantics over what “understanding” truly means. I think it’s useful, at least today, to draw the line at whether GPT-4 has succesfully modeled parts of the world. Is it just picking words and connecting them with correct grammar? Or does the token selection actually reflect parts of the physical world?
One of the most remarkable things I’ve heard about GPT-4 comes from an episode of This American Life titled “Greetings, People of Earth”.
GPT-4 cannot alter its weights once it has been trained so this is just factually wrong.
LLMs are really cool and very useful, don’t get me wrong. But people get excited by what they seem to do and lose sight of what they actually can do. They are not intelligent. They create text based on inputs. That is not what intelligence is, unless you have an extremely dismal view of intelligence that humans are text creation machines with no thoughts, no feelings, no desires, no ability to plan… basically, no internal world at all.
An LLM is an algorithm, not an intelligence.
Adam Something uploaded a video starting with the definition of intelligence itself, and then explains how something that “acts” intelligent doesn’t mean it “is” intelligent.
I think even “intelligence” here is a stretch. In a very narrow sense, it is intelligent: it creates text, simulates conversations, answers questions. But that is not what intelligence is (and it is all LLMs can do).
“Simulating conversations” to a good enough degree requires intelligence. Why are you drawing a distinction here?
What a silly assertion. Eliza was simulating conversations in the 80s; it was no more intelligent than the current crop of chatbots.
This is an unfortunate misunderstanding, one that’s all too common. I’ve also seen comments like “It’s no more intelligent than a dictionary”. Try asking Eliza to summarize a PDF for you, and then ask followup questions based on that summary. Then ask it to list a few flaws in the reasoning in the PDF. LLMs are so completely different from Eliza that I think you fundamentally misunderstand how they work. You should really read up on them.
Give Eliza equivalent compute time and functionality to interpret the data type and it probably could get something approaching a result. Modern LLMs really benefit from massive amounts of compute availability and being able to “pre-compile” via training.
They’re not, in and of themselves, intelligent. That’s not something that is seriously debated academically, though the dangers of humans misperceiving them as such very much is. They may be a component of actual artificial intelligence in the future and are amazing tools that I’m getting done hands-on time with, but the widespread labeling them as “AI” is pure marketing.
Sorry, but this is simply incorrect. Do you know what Eliza is and how it works? It is categorically different from LLMs.
This is also incorrect. I think the issue that many people have is that they hear “AI” and think “superintelligence”. What we have right now is indeed AI. It’s a primitive AI and certainly no superintelligence, but it’s AI nonetheless.
There is no known reason to think that the approach we’re taking now won’t eventually lead to superintelligence with better hardware. Maybe we will hit some limit that makes the hype die down, but there’s no reason to think that limit exists right now. Keep in mind that although this is apples vs oranges, GPT-4 is a fraction of the size of a human brain. Let’s see what happens when hardware advances give us a few more orders of magnitude. There’s already a huge, noticeable difference between GPT 3.5 and GPT 4.
To add something, as you mentioned gpt-4 neurons are only a fraction of a human brain.
The entire human brain runs on 10-20 watt, thats about a single lightbulb to do all the computing needed for conscious intelligence.
Its crazy how optimized natural life is and we have a lot left to learn.
I did not mean to come across as stating that they were the same, nor that the results produced would be as good. Merely, that a PDF could be run through OCR and processed into a script for ELIZA, which could produce some results to requests for a summary (ex. provide the abstract).
My point being that these technologies that are fundamentally different and at very different levels of technological sophistication can both, at a high level, accomplish the task. Both the quality of the result and capabilities beyond the surface level are very different. However, both, would be able to produce one, working within their architectural constraints.
Looking at it this way also gives a good basis for comparing LLMs to intelligence. Both, at a high level, can accomplish many of the same tasks, but, context matters in more than a syntactical sense and LLMs lack the capability of understanding and comprehension of the data that they are processing.
That paper is both solely phenomenological and states that it is not using an accepted definition of intelligence. With the former point, there’s a significant risk of fallacy in such observation as it is based upon subjective observation of behavior not emperical analysis of why the behavior is occuring. For example leatherette may approximate the appearance and texture of leather but, when examined it differs fundamentally both on the macroscopic and microscopic level, making it objectively incorrect to call it “leather”.
Here, we’re really getting into semantics. As the authors of that paper noted, they are not using a definition that is widely accepted, academically. Though they do definitely have a good point on some of the definitions being far too anthropocentric (ex. “being able to do anything that a human can do” - really, that’s a shit definition). I would certainly agree with the term “primitive AI”, if used akin to programming primitives (int, char, float, etc.) as it is clear that LLMs may be useful components in building actual general intelligence.
Here is an alternative Piped link(s):
video
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
That’s kind of silly semantics to quibble over. Would you tell a robot hunting you down “you’re only acting intelligent, you’re not actually intelligent!”?
People need to get over themselves as a species. Meat isn’t anything special, it turns out silicon can think too. Not in quite the same way, but it still thinks in ways that are useful to us.
The author is an imbecile if they haven’t been able to break GPT. It took me less than one day of tooling around with it before I got it to say something which outed it as having no understanding of what we were discussing.
That doesn’t mean it’s not intelligent. Humans can get broken in all sorts of ways. Are we not intelligent?
The ways in which humans make mistakes are entirely different from the ways GPT makes mistakes.
Also, if you explain to a human their mistake, they can alter their understanding of the world in order to not make that mistake in the future. Not so with GPT.
LLMs can certainly do that, why are you asserting otherwise?
ChatGPT can do it for a single session, but not across multiple sessions. That’s not some inherent limitations to LLMs, that’s just because it’s convenient for OpenAI to do it that way. If we spun up a copy of a human from the same original state every time you wanted to ask it a question and then killed it after it was done responding, it similarly wouldn’t be able to change its behavior across questions.
Like, imagine we could do something like this. You could spin up a copy of that brain image, alter its understanding of the world, then spin up a fresh copy that doesn’t have that altered understanding. That’s essentially what we’re doing with LLMs today. But if you don’t spin up a fresh copy, it would retain its altered understanding.
I literally watched it not correct itself after trying to explain to it what I wanted changed in a half dozen different ways during a single session. It never was able to understand what I was asking for.
Edit: Furthermore, I watched it become less intelligent as our conversation went longer. It basically forgot things we had discussed and misremembered or hallucinated details after a longer exchange.
For your edit: Yes, that’s what’s known as the context window limit. ChatGPT has an 8k token “memory” (for most people), and older entries are dropped. That’s not an inherent limitation of the approach, it’s just a way of keeping OpenAI’s bills lower.
Without an example I don’t think there’s anything to discuss. Here’s one trivial example though where I altered ChatGPT’s understanding of the world:
If I continued that conversation, ChatGPT would eventually forget that due to the aforementioned context window limit. For a more substantive way of altering an LLM’s understanding of the world, look at how OpenAI did RLHF to get ChatGPT to not say naughty things. That permanently altered the way GPT-4 responds, in a similar manner to having an angry nun rap your knuckles whenever you say something naughty.
Well… there’s an argument to be made there.
The bit you quoted is referring to training.
Recent papers say otherwise.
The conclusion the author of that article comes to (LLMs can understand animal language) is… problematic at the very least. I don’t know how they expect that to happen.
In what sense does your link say otherwise? Is a world model the same thing as intelligence?
In the end of the bit I quoted you say: “basically no world at all.” But also, can you define what intelligence is? Are you sure it isn’t whatever LLMs are doing under the hood, deep in hidden layers? I guess having a world model is more akin to understanding than intelligence, but I don’t think we have a great definition of either.
Edit to add: More… papers…
From the Encyclopedia Britannica:
In no sense do LLMs do any of these except, perhaps, “understand and handle abstract concepts.” But since they themselves have no understanding of the concepts, and merely generate text that can simulate understanding, I would call that a stretch.
Yes. LLMs are not magic, they are math, and we understand how they work. Deep under the hood, they are manipulating mathematical vectors that in no way are connected representationally to words. In the end, the result of that math is reapplied to a linguistic model and the result is speech. It is an algorithm, not an intelligence.
I’m not really interested in papers that either don’t understand LLMs or play word games with intelligence (shockingly, solipsism is an easy point of view to believe if you just ignore all evidence). For every one of these, you can find a dozen that correctly describe ChatGPT and its limitations. Again, including ChatGPT itself. Why not believe those instead of cherry-pick articles that gratify your ego?
I mean, my first paper was from Max Tegmark. My second paper was from Microsoft. You are discounting a well known expert in the field and one of the leading companies working on AI as not understanding LLMs.
I note that’s the definition for “human intelligence.” But either way, sure, LLMs alone can’t learn from experience (after training and between multiple separate contexts), and they can’t manipulate their environment. BabyAGI, AgentGPT, and similar things can certainly manipulate their environment using LLMs and learn from experience. LLMs by themselves can totally adapt to new situations. The paper from Microsoft discusses that. However, for sure, they don’t learn the way people do, and we aren’t currently able to modify their weights after they’ve been trained (well without a lot of hardware). They can certainly do in-context learning.
We understand how they work? From the Wikipedia page on LLMs:
It goes on to mention a couple things people are trying to do, but only with small LLMs so far.
Here’s a quote from Anthropic, another leader in AI:
They’re working on trying to understand LLMs, but aren’t there yet. So, if you understand how they do what they do, then please let us know! It’d be really helpful to make sure we can better align them.
Is this not what word/sentence vectors are? Mathematical vectors that represent concepts that can then be linked to words/sentences?
Anyway, I think time will tell here. Let’s see where we are in a couple years. :)
You are misunderstanding both this and the quote from Anthropic. They are saying the internal vector space that LLMs use is too complicated and too unrelated to the output to be understandable to humans. That doesn’t mean they’re having thoughts in there: we know exactly what they’re doing inside that vector space – performing very difficult math that seems totally meaningless to us.
The vectors do not represent concepts. The vectors are math. When the vectors are sent through language decomposition they become words, but they were never concepts at any point.
Yes, that’s exactly what I’m saying.
I mean. Not in the way we do, and not with any agency, but I hadn’t argued either way on thoughts because I don’t know the answer to that.
Huh? We know what they are doing but we don’t? Yes, we know the math, people wrote it. I coded my first neural network 35 years ago. I understand the math. We don’t understand how the math is able to do what LLMs do. If that’s what you’re saying then we agree on this.
“The neurons are cells. When neurotransmitters are sent through the synapses, they become words, but they were never concepts at any point.”
What do you mean by “they were never concepts”? Concepts of things are abstract. Nothing physical can “be” an abstract concept. If you think about a chair, there isn’t suddenly a physical chair in your head. There’s some sort of abstract representation. That’s what word vectors are. Different from how it works in a human brain, but performing a similar function.
From this page. Or better still, this article explaining how they are used to represent concepts. Like this is the whole reason vector embeddings were invented.
We do understand how the math results in LLMs. Reread what I said. The neural network vectors and weights are too complicated to follow for an individual, and do not relate on a 1:1 mapping with the words or sentences the LLM was trained on or will output, so individuals cannot deduce the output of an LLM easily by studying its trained state. But we know exactly what they’re doing conceptually, and individually, and in aggregate. Read your own sources from your previous post, that’s what they’re telling you.
Concepts are indeed abstract but LLMs have no concepts in them, simply vectors. The vectors do not represent concepts in anything close to the same way that your thoughts do. They are not 1:1 with objects, they are not a “thought,” and anyway there is nothing to “think” them. They are literally only word weights, transformed to text at the end of the generation process.
Your concept of a chair is an abstract thought representation of a chair. An LLM has vectors that combine or decompose in some way to turn into the word “chair,” but are not a concept of a chair or an abstract representation of a chair. It is simply vectors and weights, unrelated to anything that actually exists.
That is obviously totally different in kind to human thought and abstract concepts. It is just not that, and not even remotely similar.
You say you are familiar with neural networks and AI but these are really basic underpinnings of those concepts that you are misunderstanding. Maybe you need to do more research here before asserting your experience?
Edit: And in relation to your links – the vectors do not represent single words, but tokens, which indeed might be a whole word, but could just as well be part of a word or an entire phrase. Tokens do not represent the meaning of a word/partial word/phrase, just the statistical use of that word given the data the word was found in. Equating these vectors with human thoughts oversimplifies the complexities inherent in human cognition and misunderstands the limitations of LLMs.
You really, truly don’t understand what you’re talking about.
If this community values good discussion, it should probably just ban statements that manage to be this wrong. It’s like when creationists say things like “if we came from monkeys why are they still around???”. The person has just demonstrated such a fundamental lack of understanding that it’s better to not engage.
Oh, you again – it’s incredibly ironic you’re talking about wrong statements when you are basically the poster child for them. Nothing you’ve said has any grounding in reality, and is just a series of bald assertions that are as ignorant as they are incorrect. I thought you would’ve picked up on it when I started ignoring you, but: you know nothing about this and need to do a ton more research to participate in these conversations. Please do that instead of continuing to reply to people who actually know what they’re talking about.
How would creating a world model from scratch not involve intelligence?
It’s not from scratch, it’s seeded and trained by humans. That is the intelligence.
From scratch in the sense that it starts with random weights, and then experiences the world and builds a model of it through the medium of human text. That’s because text is computationally tractable for now, and has produced really impressive results. There’s no inherent need for text to be used though, similar models have been trained on time series data, and it will soon be feasible to hook up one of these models to a webcam and a body and let it experience the world on its own. No human intelligence required.
Also, your point is kind of silly. Human children learn language from older humans, and that process has been recursively happening for billions of years, all the way through the first forms of life. Do children not have intelligence? Or are you positing some magic moment in human evolution where intelligence just descended from the heavens and blessed us with it?
Just like humans are! Do you know what happens when a human grows up without any training by other humans? They are essentially feral, unable to communicate, maybe even unable to think the way we do.
LLMs do not grow up. Without training they don’t function properly. I guess in this aspect they are similar to humans (or dogs or anything else that benefits from training), but that still does not make them intelligent.
What does it mean to “grow up”? LLMs get better at their tasks during training, just as humans do while growing up. You have to clearly define the terms you use.
You used the term and I was using it with the same usage you were. Why are you quibbling semantics here? It doesn’t change the point.
are you not an algorithm?
perfected over thousands of years?
No? Humans are not algorithms except in the most general sense.
For example, there has not been any discovery of an algorithm that allows one to predict human actions, and scientists debate whether such a thing could even exist.
It’s a state machine and so are we. If it can have the ability to alter itself the way we do, I don’t see how we are any different.
It can’t; again the model does not and cannot change once it’s been generated.
And you really don’t want it to either. That could cause all sorts of privacy issues if you accidentally include private information in the conversation - and as far as I have heard it is harder to remove information from LLMs than it is to “add” information to it.
Also Microsoft’s Tay could adapt itself based on conversations and that went real well…
That’s an architectural choice, there’s nothing inherent to the approach that would prevent that from happening.
What if you don’t need to change the model to accomplish that?
What is the point of your reply? ChatGPT-4 does not use this method, and even if it did, it still does not allow it to change its model on-the-fly… so it just seems like a total non-sequitur.
If it works on Vicuna-13b, Llama-2-70b-chat, and GPT-3.5 Turbo it can probably be adapted to work on GPT-4, without changing the weights
I thought this was the other comment thread.
Define intelligence. Your last line is kind of absurd. Why can’t intelligence be described by an algorithm?
LLMs do not think or feel or have internal states. With the same random seed and the same input, GPT4 will generate exactly the same output every time. Its speech is the result of a calculation, not of intelligence or self-direction. So, even if intelligence can be described by an algorithm, LLMs are not that algorithm.
What exactly do you think would happen if you could make an exact duplicate of a human and run it from the same state multiple times? They would generate exactly the same output every time. How could you possibly think differently without turning to human exceptionalism and believing in magic meat?
For the record, GPT4 specifically is non-deterministic. The current theory is because it uses MoE, but that’s just a theory. Maybe OpenAI knows why. Also, it’s not a random seed, it’s temperature. If you set that to 0, the model should always select the most probable next token because the probability becomes 1 for that token and 0 for all others. GPT3 and most others are basically deterministic at that level, but not GPT4.