23
A prevailing sentiment online is that GPT-4 still does not understand what it talks about. We can argue semantics over what āunderstandingā truly means. I think itās 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 Iāve heard about GPT-4 comes from an episode of This American Life titled āGreetings, People of Earthā.
Yes, thatās exactly what Iām saying.
I mean. Not in the way we do, and not with any agency, but I hadnāt argued either way on thoughts because I donāt know the answer to that.
Huh? We know what they are doing but we donāt? Yes, we know the math, people wrote it. I coded my first neural network 35 years ago. I understand the math. We donāt understand how the math is able to do what LLMs do. If thatās what youāre saying then we agree on this.
āThe neurons are cells. When neurotransmitters are sent through the synapses, they become words, but they were never concepts at any point.ā
What do you mean by āthey were never conceptsā? Concepts of things are abstract. Nothing physical can ābeā an abstract concept. If you think about a chair, there isnāt suddenly a physical chair in your head. Thereās some sort of abstract representation. Thatās 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 theyāre doing conceptually, and individually, and in aggregate. Read your own sources from your previous post, thatās what theyāre 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 āthought,ā and anyway there is nothing to āthinkā 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 āchair,ā 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 ā 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.
Can you define and give examples of what you mean at each level here? Maybe weāre just not understanding each other and mean the same thing.
The Anthropic one is saying they think they have a way to figure it out, but it hasnāt been tested on large models. This is their last paragraph:
They are literally only able to do this on a small one layer transformer model. GPT 3 has 96 layers and 175 billion parameters.
Also, in their linked paper:
Under the Future Work heading:
How are you getting from that that this is a solved problem?
Again, you arenāt making sense here. Word/sentence vectors are literally a way to represent the concept of those words/sentences. Thatās what they were built for. Thatās how they are described. Letās take a step back to try to understand each other.
Are you trying to say that only human minds can understand concepts? I donāt buy the human brains are magic bit, and neither does our current understanding of physics. Are you assuming Iām saying that LLMs are sentient, conscious, have thoughts or similar? Iām not. Juryās out on the thought thing, but I certainly donāt believe the other two things. Thereās no magic with them, same with human brains. We just donāt fully understand what happens inside either. Anthropic in the work I quoted is making good progress at that, and I think they may be pretty close, but in terms of LLMs (and not Small LMs), they are still a black box. We know the math behind them, the software, etc. We have some theories. We still do not understand. If you can prove otherwise, please provide me with a source. Stuff is happening really fast in AI, and maybe I blinked and missed something.
I think youāre maybe having a hard time with using numbers to represent concepts. While a lot less abstract, we do this all the time in geometry. ((0, 0), (10, 0), (10, 10), (0, 10), (0, 0)) Whatās that? Itās a square. Word vectors work differently but have the same outcome (albeit in a more abstract way).
I was talking word vectors where the vectors DO represent words. Itās in the name. LLMs donāt specifically use word vectors, but the embeddings they do use work similarly.
You are correct tokens donāt represent the meaning of a word. However, tokens are scalars. You are conflating tokens and embeddings / word vectors here. Tokens are used to simplify converting a string into a format a neural network can understand (a vector). If we used each ascii character in the input/output string as a vector input to the network, weād have to have a lot more parameters than if we combine the characters in some way (i.e. tokens). As you said, they can be a word or a part of a word. Thereās no statistics embedded with the tokens (there are some methods of using statistics to choose what tokens to use, but thatās decided before even training the model and can not ever change [with our current approach]). You can read here for more information on tokens. Or you can play around with the gpt3 tokenizer.
If you know Python, you should grab nltk and experiment with gensim, their word vectors.
king + woman - man = queen
Seems like an abstract representation of those things as concepts using math. For the record, word vectors are actually pretty understandable/understood by people because you can visualize them easily. When you do, you find similar concepts clustered together (this is how vector search works except with text embeddings). Anyway, it just really seems like linking numbers to concepts is not clicking with you, or you somehow think itās not possible. Reading up on computational linguistics might help.
Yes, neural networks (although initially built thinking they were a computer version of a neuron), are a lot different from how actual brains work as weāve learned in however many decades since they were invented. If youāre saying that intelligence and understanding is limited to the human mind, then please point to some non-religious literature that backs up your assertion.
Iām pretty confident in my understanding, though Iām always open to new ideas that are backed with peer reviewed research. Iām not going to get into a dick waving contest here, so I guess weāll have to agree to disagree.
As a side note, going back to your definition of intelligence. That was for psychology. Iāll note that the Wikipedia page for Intelligence has this to say:
And so Iāll reiterate that we donāt have a good definition of intelligence.
Again, all your quotes indicate that what theyāve figured out is a way to inspect the interior state of models and transform the vector space into something humans can understand without analyzing the output.
I think your confusion is you believe that because we donāt know what the vector space is on the inside, we donāt know how AI works. But we actually do know how it accomplishes what it accomplishes. Simply because its interior is a black box doesnāt mean we donāt understand how we built that black box, or how it operates and functions.
For an overview of how many different kinds of LLMs function, hereās a good paper: https://arxiv.org/pdf/2307.06435.pdf Youāll note that nowhere is there any confusion about the process of how they generate input or produce output. It is all extremely well-understood. You are correct that we cannot interrogate their internals, but that is also not what I mean, at least, when I say that we can understand them and how they work.
I also canāt inspect the electrons moving through my computerās CPU. Does that mean we donāt understand how computers work? Is there intelligence in there?
No, that is not my main objection. It is your anthropomorphization of data and LLMs ā your claim that they āhave intelligence.ā From your initial post:
I think youāre getting caught up in trying to define what intelligence is; but I am simply stating what it is not. It is not a complex statistical model with no self-awareness, no semantic understanding, no ability to learn, no emotional or ethical dimensionality, no qualiaā¦
((0, 0), (10, 0), (10, 10), (0, 10), (0, 0))
is a square to humans. This is the crux of the problem: it is not a āsquareā to a computer because a āsquareā is a human classification. Your thoughts about squares are not just more robust than GPTās, they are a different kind of thing altogether. For GPT, a square is a token that it has been trained to use in a context-appropriate manner with no idea of what it represents. It lacks semantic understanding of squares. As do all computers.Iām disappointed that youāre asking me to prove a negative. The burden of proof is on you to show that GPT4 is actually intelligent. I donāt believe intelligence and understanding are for humans only; animals clearly show it too. But GPT4 does not.
Wait a sec. I think weāre saying the same thing here. I guess depending on what you mean by how it operates and functions. Iāve said multiple times we understand the math and the code. We understand how values propagate through it because again, thatās all the math and code people wrote. What we donāt understand is how it uses that math and code to actually do thinks that seem intelligent (putting aside the point of whether it is or is not intelligent). If thatās what youāre arguing then great, weāre on the same page!
Well, I donāt have the equipment to look at electrons either (I donāt think that tech exists), but I can take a logic probe and get some information that I could probably understand, or someone who designs CPUs could look at the gates and whatever and tell you what they did and how they relate to whatever higher level operations. Youāre bringing up something completely different here. Computers are not a black box at all. LLMs are-- you just said that yourself.
Iām not anthropomorphisizing them. What are you talking about? I keep saying they donāt work like human brains. I just said I donāt think theyāre sentient or conscious. I said they donāt have agency.
How do you know what itās not if we canāt define what it is?
Juryās still out on whether human brains are complex statistical models. I mean (from here)ā¦
I donāt make any claim to understanding neuroscience, and I donāt think that article is saying for sure we know that.
Anyway, in-context learning is a thing for LLMs. Maybe one day weāll figure out how to have them adjust their weights after training, but thatās not happening now (well people are experimenting with it).
New research is showing they do have semantic understanding.
They donāt by themselves have self-awareness, but a software framework built up around them can generally do that to some extent.
They do understand emotions and ethics. Someone built a fun GPTrolley web site a while ago. I think it died pretty quickly because it was too expensive for them, but it had GPT 3(?) answering Trolley Problem questions. It did (in my memory of it) like to save any āAGIā on one track over humans, which was amusing. They donāt have emotions, no. Does something have to have emotions to be intelligent?
And no, Iāve said all along they arenāt conscious, so no qualia. Again, is that required for intelligence?
No. A square to GPTs is not just a token. Itās associated with some meaning. Iām not going to re-hash embedding and word vectors and whatever since I feel like Iāve explained that to death.
Iām literally not. āIntelligence is limited to the human mindā is not a negative.
I feel like Iāve laid out my argument for that mostly through the Microsoft and Max Tegmark papers. Are you saying intelligence is only the domain of biological life?
Hereās a question-- are you conflating āintelligenceā with āgeneral intelligenceā like AGI? I find a lot of people think āAIā means āAGI.ā It doesnāt help that some people do say those things interchangeably. I was just reading a recent argument between Yann LeCun and Yoshua Bengio and they were both totally doing that. Anyway, I donāt at all believe GPT4 is AGI or that LLMs could even be AGI.
Looks like a great paper-- I hadnāt seen it yet. I know how LLMs are constructed (generally-- while I could go and write some code for a multi-layer neural network with back propagation without looking anything up, I couldnāt do that for an LLM without looking at a diagram of the layers or whatnot).
Just so incredibly wrong. Fortunately, Iāll have save myself time arguing with such a misunderstanding. GPT-4 is here to help:
Are you kidding me? I sourced GPT4 itself disagreeing with you that it is intelligent and you told me itās lying. And here you are, using it to try to reinforce your point? Are you for real or is this some kind of complicated game?
Well, is it wrong here?
Here, letās ask GPT4 itself since youāve decided itās suddenly an okay source:
I accept itās an overstatement. But it is neither āincredibly wrong,ā nor is it thought. (Or intelligence.)
Iād just like to step in here and mention that asking an LLM is probably not a good proof (and this is directed at both of you). Its understanding of AI is from before it was trained, so it is wildly out of date at this point given how much has happened in the space since.
GPT4 has knowledge of its own training since it was trained in 2022.
Care to provide some proof of that? They did update their system prompt to include a few things like it is now GPT4 (it used to always say GPT3). Other than that, I donāt think it knows anything. But in general, I was more talking about developments in AI since it was trained which it certainly does not know.
Edit: hmm I just reviewed our discussion and I note you only provided one link which was to the psychological definition of intelligence. You otherwise are providing no sources to back up your claims while my responses are full of them. Please start backing up your assertions, or provide some evidence you are an expert in the field.
So you admit that you were wrong?
I was in this case ā but the overall point I made is still correct. If winning this minor battle is what you were seeking, congratulations. You are no closer to understanding the truth of this or what we were actually talking about. Not that that was either your point or within your capabilities.