If the solution is logical but unlikely (i.e. unseen in the training set and not mapped to an existing puzzle), then the probability of the puzzle answer appearing is very low.
It sees goat, lion, cabbage, and looks for something that said goat/lion/cabbage. It does not have a concept of "leave alone" and it's not assigning entities with parameters to each item. It does care about things like sentence structure and what not, so it's more complex than a basic lookup, but the amount of borderline worship this is getting is disturbing.
Yes the LLM does generate text. No it doesn’t ‘just generate text that’s it’.
It also isn't generating "the most likely response" - that's what original GPT-3 did, GPT-3.5 and up don't work that way. (They generate "the most likely response" /according to themselves/, but that's a tautology.)
What changed?
The "most likely response" to text you wrote is: more text you wrote. Anytime the model provides an output you yourself wouldn't write, it isn't "the most likely response".
It's certainly not that those tokens are hard coded. I tried a completely different format and with no prior instruction, and it works: https://i.imgur.com/ZIDb4vM.png (again, highlighting is broken. The LLM generated all the text after 'Alice:' for all lines except for the first one.)
It would be interesting to stretch this model, like asking it to continue a conversation between 4-5 people where the speaking order is not regular and the user is 2 people and the model is 3
People don't write about GPT: they write about GPT personified.
The two magic words are, "exhibit behavior".
GPT exhibits the behavior of "humans writing language" by implicitly modeling the "already-written-by-humans language" of its training corpus, then using that model to respond to a prompt.
But no! It doesn't understand, it doesn't reason, these are concepts wholly absent from its fundamental design. It can do really cool things despite the fact that it's essentially just a text generator. But there's a ceiling to what can be accomplished with that approach.
It's presented as a limitation when GPT provides an incorrect answer.
Both of these behaviors are literally the same. We are sorting them into the subjective categories of "right" and "wrong" after the fact.
GPT is fundamentally incapable of modeling that difference. A "right answer" is every bit as valid as a "wrong answer". The two are equivalent in what GPT is modeling.
Lies are a valid feature of language. They are shaped the same as truths.
The only way to resolve this problem is brute force: provide every unique construction of a question, and the corresponding correct answer to that construction.
If you ask it what the outcome of a lie is on the conversation it can generally answer. If you ask it for a sample conversation where someone is factually incorrect, or lying, and caught out, it can generate it.
If you give it a fact and ask it to lie about that fact, it will.
But the way it chooses what interaction to repeat is not by logical decision: it's by semantic distance.
The meaning of semantic distance usually leads to the correct path, but sometimes that pattern is ambiguous.
> not by logical decision
Almost entirely yes, but you can have it textually model logic analysis and then check that own model itself. It's not "doing logic" but it almost never fails simply exercises either.
> The meaning of semantic distance usually leads to the correct path, but sometimes that pattern is ambiguous.
Of course. But "a little knowledge is a dangerous thing" as well. Often even real knowledge and analysis leads to the wrong place. In both cases (with a junior human or a LLM as an assistant) you can model their basic processes and stack the information in such a way that their simple model will lead them to the correct place.
It may not know what a lie is, in the sense of having felt the need to hide the truth to avoid personal punishment, but it certainly "knows" what one is and how it shapes the conversation for the purposes of writing a lie, writing a response to a lie, detecting potential lies, etc.
For example, I just asked bing: What does the following expression equal? Show your work, step by step: 4(13+2)7-1(4+3+9)+7^(2(2+1))
And bing responded with:
"To solve a math expression, you need to follow the order of operations: parentheses, exponents, multiplication/division, and addition/subtraction (PEMDAS).1 Here are the steps to solve your expression:
4(13+2)7-1(4+3+9)+7^(2(2+1))
Step 1: Simplify inside the parentheses first.
4(15)*7-1(16)+7^(6)
Step 2: Apply the exponents next.
420-16+117649
Step 3: Add or subtract from left to right.
118053
The final answer is 118053. I hope this helps. "
I wrote the silly expression randomly. I'm not saying it's "thinking" or "understanding", but I wouldn't expect a text-generator to be able to deduce the problem down like that.
> To solve a math expression, you need to follow the order of operations: parentheses, exponents, multiplication/division, and addition/subtraction (PEMDAS).1 Here are the steps to solve your expression:
It isn't actually thinking about any of that statement. That's just boilerplate that goes at the beginning of this story. It's what bing is familiar seeing as a continuation to your prompt, "show your work, step by step".
It gets more complicated when it shows addition being correctly simplified, but that behavior is still present in the examples in its training corpus.
---
The thinking and understanding happened when the first person wrote the original story. It also happened when people provided examples of arithmetic expressions being simplified, though I suspect bing has some extra behavior inserted here.
All the thought and meaning people put into text gets organized into patterns. LLMs find a prompt in the patterns they modeled, and "continues" the patterns. We find meaning correctly organized in the result. That's the whole story.
With anthropormization there is zero amount of that rigor, which lets people use sloppy arguments about what ChatGPT is doing and isn't doing.
This sort of fusion of different capabilities is likely to produce something that feels similar to AGI in certain circumstances. It is certainly a lot more capable than things that came before for mundane recognition tasks.
Now of course there are areas it would perform very badly, but in unimportant domains on trivial but large predictable datasets it could perform far better than humans would for example (just to take one example on identifying tumours or other patterns in images, this sort of AI would probably be a massively helpful assistant allowing a radiologist to review an order of magnitude more cases if given the right training).
One of the biggest reasons these misunderstandings are so frustrating is because you can't have reasonable discussion about the potential interesting applications of the tech. On some level copy writing may devolve into auto generating prompts for things like GPT with a few editors sanity checking the output (depending on level of quality), and I agree that a second opinion "check for tumors" use has a LOT of interesting applications (and several concerning ones such as over reliance on a model that will cause people who fall outside the bell curve to have even more trouble getting treatment).
All of this is a much more realistic real world use case RIGHT NOW, but instead we've got people fantasizing about how close we are to GAI and ignoring shortcomings to shoehorn it into their preferred solution.
Open AI ESPECIALLY reinforces this by being very selective with their results and they way they frame things. I became aware of this as a huge dota fan for over a decade when they did their games there. And while it was very very interesting and put up some impressive results, the framing of those results does NOT portray the reality.
Just months ago we saw in research out of Harvard that even a very simplistic GPT model builds internalized abstract world representations from the training data within its NN.
People parroting the position from you and the person before you are like doctors who learned about something in school but haven't kept up with emerging research that's since invalidated what they learned, so they go around spouting misinformation because it was thought to be true when they learned it but is now known to be false and just hasn't caught up to them yet.
So many armchair experts who took a ML course in undergrad pitching in their two cents having read none of the papers in the past year.
This is a field where research perspectives are shifting within months, not even years. So unless you are actively engaging with emerging papers, and given your comment I'm guessing you aren't, you may be on the wrong side of the Dunning-Kreuger curve here.
Do we really know it IS wrong?
That's a very strong claim. I believe you there's a lot happening in this field but it doesn't seem possible to even answer the question either way. We don't know what reasoning looks like under the hood. It's still a "know it when you see it" situation.
> GPT model builds internalized abstract world representations from the training data within its NN.
Does any of those words even have well defined meanings in this context?
I'll try to figure out what paper you're referring to. But if I don't find it / for the benefit of others just passing by, could you explain what they mean by "internalized"?
> you may be on the wrong side of the Dunning-Krueger curve here.
Have you read the Dunning & Krueger paper? It demonstrates a positive correlation between confidence and competence. Citing DK in the form of a thinly veiled insult is misinformation of your own, demonstrating and perpetuating a common misunderstanding of the research. And this paper is more than 20 years old...
So I’ve just read the Harvard paper, and it’s good to see people exploring techniques for X-ray-ing the black box. Understanding better what inference does is an important next step. What the paper doesn’t explain is what’s different between a “world model” and a latent space. It doesn’t seem surprising or particularly interesting that a network trained on a game would have a latent space representation of the board. Vision networks already did this; their latent spaces have edge and shape detectors. And yet we already know these older networks weren’t “reasoning”. Not that much has fundamentally changed since then other than we’ve learned how to train larger networks reliably and we use more data.
Arguing that this “world model” is somehow special seems premature and rather overstated. The Othello research isn’t demonstrating an “abstract” representation, it’s the opposite of abstract. The network doesn’t understand the game rules, can’t reliably play full Othello games, and can’t describe a board to you in any other terms than what it was shown, it only has an internal model of a board, formed by being shown millions of boards.
I've seen this asserted without citation numerous times recently, but I am quite suspicious. Not that there exists a study that claims this, but that it is well supported.
There is no mechanism for directly assessing this, and I'd be suspicious that there is any good proxy for assessing it in AIs, either. research on this type of cognition in animals tends to be contentious, and proxies for them should be easier to construct than for AIs.
> the wrong side of the Dunning-Kreuger curve
the relationship between confidence and perception in the D-K paper, as I recall, is a line, and its roughly “on average, people of all competency levels see themselves slightly closer to the 70th percentile than they actually are.” So, I guess the “wrong side” is the side anywhere under the 70th percentile in the skill in question?
This is being far too generous to parent’s claim, IMO. Note how much “people of all competency levels see themselves slightly closer to the 70th percentile than they actually are” sounds like regression to the mean. And it has been compellingly argued that that’s all DK actually measured. [1] DK’s primary metric for self-assessment was to guess your own percentile of skill against a group containing others of unknown skill. This fully explains why their correlation between self-rank and actual rank is less than 1, and why the data is regressing to the mean, and yet they ignored that and went on to call their test subjects incompetent, despite having no absolute metrics for skill at all and testing only a handful of Ivy League students (who are primed to believe their skill is high).
Furthermore, it’s very important to know that replication attempts have shown a complete reversal of the so-called DK effect for tasks that actually require expertise. DK only measured very basic tasks, and one of the four tasks was subjective(!). When people have tried to measure the DK effect on things like medicine or law or engineering, they’ve shown that it doesn’t exist. Knowledge of NN research is closer to an expert task than a high school grammar quiz, and so not only does DK not apply to this thread, we have evidence that it’s not there.
The singular reason that DK even exists in the public consciousness may be because people love the idea they can somehow see & measure incompetence in a debate based on how strongly an argument is worded. Unfortunately that isn’t true, and of the few things the DK paper did actually show is that people’s estimates of their relative skill correlate with their actual relative skill, for the few specific skills they measured. Personally I think this paper’s methodology has a confounding factor hole the size of the Grand Canyon, that the authors and public both have dramatically and erroneously over-estimated it’s applicability to all humans and all skills, and that it’s one of the most shining examples of sketchy social science research going viral and giving the public completely wrong misconceptions, and being used incorrectly more often than not.
[1] https://www.talyarkoni.org/blog/2010/07/07/what-the-dunning-...
First of all, it's a black-box algorithm with pretty universal capabilities when viewed from our current SOTA view. It might appear primitive in a few years, but right now the pure approximation and generalisation capabilities are astounding. So this:
> It sees goat, lion, cabbage, and looks for something that said goat/lion/cabbage
can not be stated as truth without evidence. Same here:
> it's not assigning entities with parameters to each item. It does care about things like sentence structure and what not
Where's your evidence? The enormous parameter space coupled with our so far best performing network structure gives it quite a bit of flexibility. It can memorise things but also derive rules and computation, in order to generalise. We do not just memorise everything, or look up things into the dataset. Of course it learned how to solve things and derive solutions, but the relevant data-points for the puzzle could be {enormous set of logic problems} where it derived general rules that translate to each problem. Generalisation IS NOT trying to find the closest data-point, but finding rules explaining as much data-points, maybe unseen in the test-set, as possible. A fundamental difference.
I am not hyping it without belief, but if we humans can reason then NNs can potentially also. Maybe not GPT-4. Because we do not know how humans do it, so an argument about intrinsic properties is worthless. It's all about capabilities. Reasoning is a functional description as long as you can't tell me exactly how we do it. Maybe wittgenstein could help us: "Whereof one cannot speak, thereof one must be silent". As long as there's no tangible definition of reasoning it's worthless to discuss it.
If we want to talk about fundamental limitations we have to talk about things like ChatGPT-4 not being able to simulate because it's runtime is fundamentally limited by design. It can not recurse. It can only run only a fixed number of steps, that are always the same, until it has to return an answer. So if there's some kind of recursion learned through weights encoding programs intercepted by later layers, the recursion depth is limited.
It’s also possible, likely even, that the model is capable of both memorization and cognition, and in this case the “memorization neurons” are driving the prediction.
The problem with the goat question is that the model is falling back on memorized answers. If the model is in fact capable of cognition, you’d have better odds of triggering the ability with problems that are dissimilar to anything in the training set.
Https://arxiv.org/abs/2210.13382
It looks like OpenAI have specifically added Othello game handling to chat.openai.org, so I guess they’ve done the same fine-tuning to ChatGPT? It would be interesting to know how good an untuned GPT3/4 was at Othello & whether OpenAI has fine-tuned it or not!
(Having just tried a few moves, it looks like ChatGPT is just as bad at Othello as it was at chess, so it’s interesting that it knows the initial board layout but can’t actually play any moves correctly: Every updated board it prints out is completely wrong.)
Why is that interesting? The initial board layout would appear all the time in the training data.
It was able to model the chronological series of game states that it read from an example game. It was able to include the arbitrary "new game state" of a prompt into that model, then extrapolate that "new game state" into "a new series of game states".
All of the logic and intentions involved in playing the example game were saved into that series of game states. By implicitly modeling a correctly played game, you can implicitly generate a valid continuation for any arbitrary game state; at least with a relatively high success rate.
But we have fundamental, mathematical bounds on the LLM. We know that the complexity is at most O(n^2) in token length n, probably closer to O(n). It can not "think" about a problem and recurse into simulating games. It can not simulate. It's an interesting frontier, especially because we have also cool results about the theoretical, universal approximation capabilities of RNNs.
And that's the least exciting possible mystery: any surprise behavior is categorized by us as a failure. If GPT's model has boundaries that don't make sense to us, we consider them noise. They are not useful behavior, and our goal is to minimize them.
What do you mean? Is cognition a set of weights on a gradient? Cognition involves conscious reasoning and understanding. How do you know it is computable at all? There are many things which cannot be computed by a program (e.g. whether an arbitrary program will halt or not)...
That's a pretty simplistic view. How do you know we can't determine whether an arbitrary program will halt or not (assuming access to all inputs and enough time to examine it)? What in principle would prevent us from doing so? But computers in principle cannot, since the problem is often non-algorithmic.
For example, consider the following program, which is passed the text of the file it is in as input:
function doesHalt($program, $inputs): bool {...}
$input = $argv[0]; // contents of this file
if (doesHalt($input, [$input])) {
while(true) {
print "Wrong! It doesn't halt!";
}
} else {
print "Wrong! It halts!";
}
It doesn't matter what the algorithmic doesHalt function returns - it will always be incorrect for this program. What makes you certain there is an algorithmic analog for all human reasoning?
The point is we currently have very little understanding of what gives rise to consciousness, so what is the point of all this pontificating and grand standing. Its silly. We've no idea what we are talking about at present.
Clearly, our state of the art models of nueral-like computation do not really simulate consciousness at all, so why is the default assumption that they could if we get better at making them? The burden of evidence is on conputational models to prove they can produce a consciousness model, not the other way around.
The function we are trying to compute is undecidable. Sure we as humans understand that there's a dichotomy here: if the program halts it won't halt; if it doesn't halt it will halt. But the function we are asked to compute must have one output on a given input. So a human, when given this program as input, is also unable to assign an output.
So humans also can't solve the halting problem, we are just able to recognize that the problem is undecidable.
Note: whatever algorithm is implemented in the doesHalt function will contain a bug for at least some inputs, since it's trying to generalize something that is non-algorithmic.
In principle no algorithm can be created to determine if an arbitrary program will halt, since whatever it is could be implemented in a function which the program calls (with itself as the input) and then does the opposite thing.
> What makes you certain there is an algorithmic analog for all human reasoning?
(Maybe) not for ALL human thought but at least all communicatable deductive reasoning can be encoded in formal logic. If I give you an algorithm and ask you to decide if it does halt or does not halt (I give you plenty of time to decide) and then ask you to explain to me your result and convince me that you are correct, you have to put your thoughts into words that I can understand and and the logic of your reasoning has to be sound. And if you can explain to me you could as well encode your though process into an algorithm or a formal logic expression. If you can not, you could not convince me. If you can: now you have your algorithm for deciding the halting problem.
Can you tell me if a program which searches for counterexamples to the Collatz conjecture halts?
Turing's entire analysis started from the point of what humans could do.
Your argument doesn't disprove my assumption *. In which case, what's the point of it?
* - I don't necessarily believe this assumption. But I do dislike bad arguments.
And while the human brain might not be a bio-computer, I'm not sure, its computational prowess are doubtfully stronger than a quantum turing machine, which can't solve the halting problem either.
func main() {
var n = 4;
OUTER: loop {
for (var i = 2; i < n/2; i++) {
if (isPrime(i) && isPrime(n-i)) {
n += 2;
continue OUTER; // Goldbach’s conjecture
}
break;
}
}If cognition magically exists outside of math and science, then sure, all bets are off.
We don't even know if the flow of water in a river can always be represented by a mathematical function - this is one of the Millennium Problems. And we've known the partial differential equations that govern that system since the 1850's.
We are far, far away from even being able to write down anything resembling a mathematical description of cognition, let alone being able to say whether the solutions to that description are in the class of Lebesgue-integrable functions.
There was, past tense, no reason to believe cognition could be represented as a mathematical function. LLMs with RLHF are forcing us to question that assumption. I would agree that we are a long way from a rigorous mathematical definition of human thought, but in the meantime that doesn't reduce the utility of approximate solutions.
The Navier-Stokes equations are a set of partial differential equations - they are the problem statement. Given some initial and boundary conditions, we can find (approximate or exact) solutions, which are functions. But we don't know that these solutions are always Lebesgue integrable, and if they are not, neural nets will not be able to approximate them.
This is just a simple example from well-understood physics that we know neural nets won't always be able to give approximate descriptions of reality.
"Neural networks are universal approximators" is a fairly meaningless sound bite. It just means that given enough parameters and/or the right activation function, a neural network, which is itself a function, can approximate other functions. But "enough" and "right" are doing a lot of work here, and pragmatically the answer to "how approximate?" can be "not very".
A lot of people who argue that cognition is special to biological systems seem to base the argument on our inability to accurately model the detailed behavior of neurons. And yet kids regularly build universal computers out of stuff in Minecraft. It seems strange to imagine the response characteristics of low-level components of a system determine whether it can be conscious.
But GP specifically says neural nets should be able to do it because they are universal approximators (of Lebesgue integratable functions).
I'm saying this is clearly a nonsense argument, because there are much simpler physical processes than cognition where the answers are not Lebesgue integratable functions, so we have no guarantee that neural networks will be able to approximate the answers.
For cognition we don't even know the problem statement, and maybe the answers are not functions over the real numbers at all, but graphs or matrices or Markov chains or what have you. Then having universal approximators of functions over the real numbers is useless.
Consciousness cannot be accounted for in physical terms. For consciousness is absolutely fundamental. It cannot be accounted for in terms of anything else.
-- Erwin Schrödinger
- Carl Sagan
Sagan, while he did a little bit of useful work on planetary science early in his career, quickly descended into the realm of (self-promotional) pseudo-science. This was his fanciful search for 'extra-terrestrial intelligence'. So it's apposite that you bring him up (even if the quote you bring is a big miss against a philosophical statement), because his belief in such an 'ET' intelligence was a fantasy as much as the belief in the possibility of creating an artificial intelligence is.
We don't know if physics is the fundamental substrate of being, and given Agrippa's trillemma we can't know.
Many things are non-algorithmic, and thus cannot be done by a computer, yet we can do them (e.g. love someone, enjoy the beauty of a sunset, experience joy or sadness, etc).
There's an interesting article/podcast here about what computers can't do: https://mindmatters.ai/2020/08/six-limitations-of-artificial....
Moreover, are you sure that e.g. loving people in non-algorithmic? We can already make chatbots which pretty convincingly act as if they love people. Sure, they don't actually love anyone, they just generate text, but then, what would it mean for a system or even a human to "actually" love someone?
Maybe. When computers solve it then we'll know.
I don't see how it could be reasoned otherwise.
What is your definition of _conscious reasoning and understanding_?
Because it's very good at it, sometimes it can fool people into thinking there is more going on than it is.
Keep in mind GPT 4 is multimodal and not just matching text.
Sorry for appearing to be completely off-topic, but do you have children? Observing our children as they're growing up, specifically the way they formulate and articulate their questions, has been a bit of a revelation to me in terms of understanding "reasoning".
I have a sister of a similar age to me who doesn't have children. My 7 year-old asked me recently - and this is a direct quote - "what is she for?"
I was pretty gobsmacked by that.
Reasoning? You decide(!)
The robots might know everything, but do they wonder anything?
Machines will have to wonder if they are to improve themselves, because that is literally the drive to collect more data, and you need good data to make good decisions.
There's really no reason to doubt the legitimacy here after everyone shared similar experiences, you just kinda look foolish for suggesting the results are faked at this point.
I once asked my niece, a bit after she started really communicating, if she remembered what it was like to not be able to talk. She thought for a moment and then said, "Before I was squishy so I couldn't talk, but then I got harder so I can talk now." Can't argue with that logic.
Pattern matching? You decide
They may have equivalences, but they're separate forms of mathematics. I'd say the same applies to different algorithms or models of computation, such as neural nets.
I don’t have the language to explain the difference in a manner I find sufficiently precise. I was hoping others might.
It does more than that. It understands how to do basic math. You can ask it what ((935+91218)/4)*3) is and it will answer it correctly. Swap those numbers for any other random numbers, it will answer it correctly.
It has never seen that during training, but it understands the mathematical concepts.
If you ask ChatGPT how it does this, it says "I break down the problem into its component parts, apply relevant mathematical rules and formulas, and then generate a solution".
It's that "apply mathetmatical rules" part that is more than just, essentially, filling in the next likely token.
When it can't find the pattern it starts "making things" up, that's where all the "magic" disappears.
At least for GPT-3, during my own experimentation, it occasionally makes arithmetic errors, especially with calculations involving numbers in scientific notation (which it is happy to use as intermediate results if you provide a prompt with a complex, multi-step word problem).
It doesn't though. Here's GPT-4 completely failing: https://gcdnb.pbrd.co/images/uxH1EtVhG2rd.png?o=1. It's riddled with errors, every single step.
You are (naively, I would suggest) accepting the LLM's answer for how it 'does' the calculation as what it actually does do. It doesn't do the calculation; it has simply generated a typical response to how people who can do calculations explain how they do calculations.
You have mistaken a ventriloquist's doll's speech for the 'self-reasoning' of the doll itself. An error that is being repeatedly made all throughout this thread.
It's kind of sad.
It could have a dozen internal reasoning networks but it doesn't use them when you want to.
It's tough to define these terms in a way that includes only humans and excludes other life forms or even LLMs. This might mean we either made up these concepts, or we're not alone in having these traits.
Without a solid definition, how can we say LLMs aren't intelligent? If we make a definition that includes both us and LLMs, would we accept them as intelligent? And could we even exclude ourselves?
We need clear definitions to talk about the intelligence and sentience of LLMs, AI, or any life forms. But finding those definitions is hard, and it might clash with our human ego. Discussing these terms without definitions feels like a waste of time.
Still, your Descartes reference reminds us that our understanding of human experiences keeps changing, and our current definitions might not be spot-on.
(this comment was cleaned up with GPT-4 :D)
Consciousness, on the other hand, really might be an illusion.
Simple != thoughtless. Different != thoughtless. Less capable != thoughtless. A human black box categorically dismissing all qualia or cognition from another remarkable black box feels so wildly arrogant and anthropocentric. Which, I suppose, is the most historically on-brand behavior for our species.
Stephen Wolfram explains this in simple terms.[0]
0: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-...
the words "based on" are doing a lot of work here. No, we don't know what sort of stuff it learns from its training data nor do we know what sorts of reasoning it does, and the link you sent doesn't disagree.
"But at least as of now we don’t have a way to 'give a narrative description' of what the network is doing. And maybe that’s because it truly is computationally irreducible, and there’s no general way to find what it does except by explicitly tracing each step. Or maybe it’s just that we haven’t 'figured out the science', and identified the 'natural laws' that allow us to summarize what’s going on."
Anyway, I don't see why you think that the brain is more logical than statistical. Most people fail basic logic questions, as in the famous Linda problem.[1]
Language is a symbolic representation of behavior.
GPT takes a corpus of example text, tokenizes it, and models the tokens. The model isn't based on any rules: it's entirely implicit. There are no subjects and no logic involved.
Any "understanding" that GPT exhibits was present in the text itself, not GPT's model of that text. The reason GPT can find text that "makes sense", instead of text that "didn't make sense", is that GPT's model is a close match for grammar. When people wrote the text in GPT's corpus, they correctly organized "stuff that makes sense" into a string of letters.
The person used grammar, symbols, and familiar phrases to model ideas into text. GPT used nothing but the text itself to model the text. GPT organized all the patterns that were present in the corpus text, without ever knowing why those patterns were used.
I used nothing but my sensory input to model the world, and yet I have a model of the world, not (just) of sensory input.
There is an interesting question, though, of whether information without experience is enough to generate understanding. I doubt it.
You used evolution, too. The structure of your brain growth is the result of complex DNA instructions that have been mutated and those mutations filtered over billions of iterations of competition.
There are some patterns of thought that are inherent to that structure, and not the result of your own lived experience.
For example, you would probably dislike pain with similar responses to your original pain experience; and also similar to my lived pain experiences. Surely, there are some foundational patterns that define our interactions with language.
In the sense that the chatbox itself behaves as a sensory input to chatgpt.
Chatgpt does not have eyes, tongue, ears, but it does have this "mono-sense" which is its chatbox over which it receives and parses inputs
If someone is starving and hunting for food, they will learn fast to associate cause and effect of certain actions/situations.
A language model that only works with text may yet have an unambiguous overall loss function to minimize, but as it is a simple scalar, the way it minimizes this loss may be such that it works for the large majority of the training corpus, but falls apart in ambiguous/tricky scenarios.
This may be why LLMs have difficulty in spatial reasoning/navigation for example.
Whatever "reasoning ability" that emerged may have learned _some_ aspects to physicality that it can understand some of these puzzles, but the fact it still makes obvious mistakes sometimes is a curious failure condition.
So it may be that having "more" senses would allow for an LLM to build better models of reality.
For instance, perhaps the LLM has reached a local minima with the probabilistic modelling of text, which is why it still fails probabilistically in answering these sorts of questions.
Introducing unambiguous physical feedback into its "world model" maybe would provide the necessary feedback it needs to help it anchor its reasoning abilities, and stop failing in a probabilistic way LLMs tend to currently do.
In theory a LLM could learn any model at all, including models and combinations of models that used logical reasoning. How much logical reasoning (if any) GPT-4 has encoded is debatable, but don’t mistake GTP’s practical limitations for theoretical limitations.
Yes.
But that is not the same as GPT having it's own logical reasoning.
An LLM that creates its own behavior would be a fundamentally different thing than what "LLM" is defined to be here in this conversation.
This is not a theoretical limitation: it is a literal description. An LLM "exhibits" whatever behavior it can find in the content it modeled. That is fundamentally the only behavior an LLM does.
Eg when solving this puzzle you might visualise the lion/goat/cabbage, and walk through the scenarios in your head back and forth multiple times until you find a solution that works. A LLM won’t solve it like this. You could ask it to, and it will list out the scenarios of how it might do it, but it’s essentially an illusion of logical reasoning.
With Bing, you don't even need to tell you what it assumed wrong - I just told it that it's not quite the same as the classic puzzle, and it responded by correctly identifying the difference and asking me if that's what I meant, but forgot that lion still eats the goat. When I pointed that out, it solved the puzzle correctly.
Generally speaking, I think your point that "when solving the puzzle you might visualize" is correct, but that is orthogonal to the ability of LLM to reason in general. Rather, it has a hard time to reason about things it doesn't understand well enough (i.e. the ones for which its internal model that was built up by training is in is way off). This seems to be generally the case for anything having to do with spatial orientation - even fairly simple multi-step tasks involving concepts like "left" vs "right" or "on this side" vs "on that side" can get hilariously wrong.
But if you give it a different task, you can see reasoning in action. For example, have it play guess-the-animal game with you while telling it to "think out loud".
I'm not sure if you put "think out loud" in quotes to show literally what you told it to do or because telling the LLM to do that is figurative speech (because it can't actually think). Your talk about 'reasoning in action' indicates it was probably not the latter, but that is how I would use quotes in this context. The LLM can not 'think out loud' because it cannot actually think. It can only generate text that mimics the process of humans 'thinking out loud'.
As far as "it mimics" angle... let me put it this way: I believe that the whole Chinese room argument is unscientific nonsense. I can literally see GPT take inputs, make conclusions based on them, and ask me questions to test its hypotheses, right before my eyes in real time. And it does lead it to produce better results than it otherwise would. I don't know what constitutes "the real thing" in your book, but this qualifies in mine.
And yeah, it's not that good at logical reasoning, mind you. But its model of the world is built solely from text (much of which doesn't even describe the real world!), and then it all has to fit into a measly 175B parameters. And on top of that, its entire short-term memory consists of its 4K token window. What's amazing is that it is still, somehow, better than some people. What's important is that it's good enough for many tasks that do require the capacity to reason.
It takes inputs and produces new outputs (in the textual form of questions, in this case). That's all. It's not 'making conclusions', it's not making up hypotheses in order to 'test them'. It's not reasoning. It doesn't have a 'model of the world'. This is all a projection on your part against a machine that inputs and outputs text and whose surprising 'ability' in this context is that the text it generates plays so well on the ability of humans to self-fool themselves that its outputs are the product of 'reasoning'.
GPT has been constructed. We know how it was set-up and how it operates. (And people commenting here should be basically familiar with both hows mentioned.) No part of it does any reasoning. Taking in inputs and generating outputs is completely standard for computer programs and in no way qualifies as reasoning. People are only bringing in the idea of 'reasoning' because they either don't understand how an LLM works and have been fooled by the semblance of reasoning that this LLM produces or, more culpably, they do understand but they still falsely continue to talk about the LLM doing 'reasoning' either because they are delusional (they are fantasists) or they are working to mislead people about the machine's actual capabilities (they are fraudsters).
Then there was this experiment: https://thegradient.pub/othello/. TL;DR: they took a relatively simple GPT model and trained it on tokens corresponding to Othello moves until it started to play well. Then they probed the model and found stuff inside the neural net that seems to correspond to the state of the board; they tested it by "flipping a bit" during activation, and observed the model make a corresponding move. So it did build an inner model of the game as part of its training by inferring it from the moves it was trained on. And it uses that model to make moves according to the current state of the board - that sure sounds like reasoning to me. Given this, can you explain why you are so certain that there isn't some equivalent inside ChatGPT?
[1] https://news.ycombinator.com/item?id=35162445
In general, I find that a lot of these arguments boil down to sophistry when the obvious meaning of the word that equally obviously describes what people see in front of them is replaced by some convoluted "actually" that doesn't serve any point other than making sure that it excludes the dreaded possibility that logical reasoning and world-modelling isn't actually all that special.
I do wonder if GPT-4 would do better, though.
FWIW, it passes basic substitution.
True.
> A LLM won’t solve it like this.
Non sequitur.
The likeliness of the solution depends on context. If context is, say, a textbook on logical puzzles, then the probability of the logical solution is high.
If an LLM fails to reflect it, then it isn't good enough at predicting the text.
Yes, it could be possible that the required size of the model and training data to make it solve such puzzles consistently is impractical (or outright unachievable in principle). But the model being "just a text predictor" has nothing to do with that impossibility.
The Othello GPT research from Harvard months ago demonstrated that even a simple GPT model is capable of building world representations from which it reasons outputs. This makes intuitive sense if you understand the training, as where possible having reversed an abstraction in the NN is going to perform better than simply extrapolating predictively from the data.
Not only is GPT-4 more robust at logic puzzles its predecessor failed, I've seen it solve unique riddles outside any training data and the paper has explicit examples of critical reasoning, especially in the appendix.
It is extremely unlikely given the Harvard research and the size of the training data and NN that there isn't some degree of specialized critical reasoning which has developed in the NN.
The emerging challenge for researchers moving forward is to get better insight into the black box and where these capabilities have developed and where it's still falling into just a fancy Markov chain.
But comments like yours reflect an increasingly obsolete and yet increasingly popular misinformation online around the way they operate. So someone reading your comment might not think to do things like what the Bing team added with providing an internal monologue for reasoning, or guiding it towards extended chain of thought reasoning, because they would be engaging with the models thinking it's only frequency based context relative to the training set that matters.
If you haven't engaged with emerging research from the past year, you may want to brush up on your reading.
Just like you.
When LLMs are presented with a problem they search for a solution based on the language model. And when they can't find a solution, there's always a match for something that looks like a solution.
Or in other words, people can use implicit logic to solve puzzles. Similarly LLMs can implicitly be fine-tuned into logic models by asking them to solve a puzzle, insofar as that logic model fits in their weights. Transformers are very flexible that way.