Inductive or deductive? Rethinking the fundamental reasoning abilities of LLMs
arxiv.org
arxiv.org
You cannot test reasoning when you don't know what's in the training set. You have to be able to differentiate reasoning from memorization, and that's not trivial.
Moreso, the results look to confirm that at least some memorization is going on. Do we really not think GPT has extensively been trained on arithmetic in base 10, 8, and 16? This seems like a terrible prior. Even if not explicitly, how much code has it read that performs these tasks. How many web pages, tutorials, Reddit posts cover oct and hex? They also haven't defined zero shot correctly. Arithmetic in these bases aren't 0-shot. They're explicitly in distribution...
I'm unsure about base 9 and 11. It's pretty interesting to see that GPT 4 is much better at these. Anyone know why? Did they train on these? More bases? Doesn't seem unreasonable but I don't know.
The experimentation is also extremely lacking. The arithmetic questions only have 1000 tests where they add two digits. This is certainly in the training data. I'm also unconvinced by the syntax reasoning tasks since the transformer (attention) architecture seems to be designed for this. I'm also unconvinced these tasks aren't in training. Caesar ciphers are also certainly in the training data.
The prompts are also odd and I guess that's why they're in the appendix. For example, getting GPT to be better at math or many tasks by having it write python code is not novel.
There's some stuff here but this really doesn't seem like a lot of work for 12 people from a top university and a trillion dollar company. It's odd to see that many people when the experiments can be run in a fairly short time.
Discrepancies in mathematical ability between the various bases would seem to suggest memorization as opposed to generalization.
More than that -- It's not clear that what humans do is not "just" a memorization. We can always look at human experience mechanisticly and say that we don't think -- we just memorized thinking patterns and apply them when speaking and "thinking"
> It's not clear that what humans do is not "just" a memorization.
If you doubt me I very much encourage you to dive deeper into the history of science and get doing deep deep knowledge on any subject. Because you'll find this happen all the time. But if you apply a loose enough definition to memorization (that isn't one that would be generally agreed upon if you used it's logical conclusions) then yeah, everything is memorization. But everything is foo if I define everything to be foo, so let's not.
Memorizing facts and algorithms is memorization. The rest of what you are talking about is not.
Applying existing knowledge on new data without deriving new information is generalization. An example of this is the case of a semantic segmentation model classifying a car that it has never seen. If the model was not trained on birds, it will never classify a bird as a bird.
Computation of decidable problems is a large, possibly the largest subset of reasoning. Most humans do not struggle with solving decidable problems, the problem is that they are slow and can only solve small problem sizes, but most problems encountered in practice aren't one large decidable problem, but a long chain of many small, dozens to hundreds of heterogeneous problems that are seamlessly mixed with one another. LLMs struggle with decidable problems that are out of distribution, but you can give a human instructions on how to do something they have never done before and they will follow them with no problem.
> More than that -- It's not clear that what humans do is not "just" a memorization.
I hope it is clear that I did not memorize this message I am writing here and that it is the unique result of processes inside my brain that were not captured in the training process of an LLM.
>We can always look at human experience mechanisticly and say that we don't think -- we just memorized thinking patterns and apply them when speaking and "thinking"
Again you are trying to twist this in an absurd direction. Let's come up with a teleoperated humanoid robot on Mars that is controlled by a human on Earth. The robot acts exactly like a human does. Does this mean the robot is now capable of reasoning and thinking like a human, simply because it is replaying a recording of the human's body and speech? This is the argument you are making. You're arguing that the robot's ability to replay a human's actions is equivalent to the processes that brought about that human action.
> Let's come up with a teleoperated humanoid robot on Mars
I don't care what Iilya said, mimicking a human does not make a human. It may look like a duck, swims like a duck, and quack like a duck, then it's probably a duck, but you haven't ruled out an advanced animatronic. In fact, I'm betting right now people could make an animatronic that would convince most people it is a duck because most people just don't know the nuances of duck behavior.
If todays LLMs are not really intelligent (hardly anyone is arguing that LLMs are literally humans) then by all means, devise your intelligence test that will cleanly qualify Humans (and all creatures you ascertain are capable on intelligence) and disqualify LLMs. In good faith, it should be something both or all can reasonably attempt.
I'll save you some time but feel free to prove me wrong. You will not be able to do so. Not because LLMs can solve every problem under the Sun but because you will at best create something that disqualifies LLMs and a good chunk of humans along the way.
Anybody who cannot manage this task (which should be very simple if the difference is as obvious as a lot of people claim) has no business saying LLMs aren't intelligent.
> I think your "advanced animatronic" is a duck until you can devise a test that cleanly separates it from a "real duck". A test of "duckness".
> You will not be able to do so.
River crossing puzzles are a common test. Sure, humans sometimes fail even the trivial variations, but the important part is how they fail. Humans guess the wrong answer. LLMs will tell you the wrong answer while describing steps that are correct and result in a different answer. It's the inconsistency and contradiction that's the demonstration of lack of reasoning and intelligence, not the failure itself.
That's the spirit of the popular phrase isn't it ? I genuinely never read that and thought that literally only those 3 properties were the bar.
>I am unable to because any test I make or from any other researcher in the field makes will result in an answer you don't like.
This is kind of an odd response. I mean maybe. I'll be charitable and agree wholeheartedly.
But i don't know what my disappointment has to do with anything? I mean if i could prove something i deeply believed to be True and which would also be paper worthy, i certainly wouldn't let the reactions of an internet stranger stop me from doing it.
>River crossing puzzles are a common test. Sure, humans sometimes fail even the trivial variations, but the important part is how they fail. Humans guess the wrong answer. LLMs will tell you the wrong answer while describing steps that are correct and result in a different answer. It's the inconsistency and contradiction that's the demonstration of lack of reasoning and intelligence, not the failure itself.
Inconsistency and contradiction with the reasoning they say (and even believe) and the decisions they make is such a common staple of human reasoning we have a name for it...At worst, you could say these contradictions don't always take the same form but this just kind of loops back to my original point.
Let me be clear here, if you want to look at results like these and say - "This is room for improvement", then Great!, I agree. But it certainly feels like a lot of people have a standard of reasoning (for machines) that only exists in fiction or their own imaginations. This general reasoning engine that makes neither mistake nor contradiction in output or process does not exist in real life whether you believe humans are the only beings capable of reasoning or are gracious enough to extend this capability to some of our animal friends.
Also, i've seen LLMs fail trivial variations of said logic puzzles, only to get them right when you present the problem in a way that doesn't look exactly like the logic puzzle they've almost certainly memorized. Sometimes, it's as simple as changing the names involved. Isn't that fascinating ? Humans have a similar cognitive shortcoming.
Who are those "some" who say it is reasoning?
Here's a question. If you enter a command in your terminal do you expect your shell to have memorised the result of the command from some previous experience, or do you expect your shell to compute the result of your command according to its programming and your inputs? A rhetorical question: we all assume the latter.
Which one is more like what most peoples' informal conception of "reasoning"? Retrieving an answer from storage, or computing an answer from the inputs given to a program?
>> We can always look at human experience mechanisticly and say that we don't think -- we just memorized thinking patterns and apply them when speaking and "thinking"
I think this is confusing memorisation of the rules required to perform a computation, like a program stored in computer memory, with memorisation of the results of a computation. When we distinguish between memorisation and reasoning we usually make a distinction between computing a result from scratch and retrieving it from storage without having to re-compute it, like in caching or memoization, or getting data from a database.
For a real world example, we memorise our time tables but we don't memorise the result of every sum x + y = z, instead we memorise a summation algorithm that we then use to derive the sum of two numbers.
[1] Physics of LLMs: https://www.youtube.com/watch?v=yBL7J0kgldU&t=7s
This looks like an LLM doing the usual LLM thing - finding relevant items and combining them to fit. This doesn't require the level of abstraction and induction the authors impute to the LLM. Ordinary LLM behavior explains this, once you've found the relevant training data.
People often solve problems that way too, of course.
The intent was to point out that the educational approach was flawed, but I think there are interesting parallels to token processing in LLMs, which--unlike a human child--are built in such a way that crazy partial-fit rules are likely their only option.
> Benny believed that the fraction 5/10 = 1.5 and 400/400 = 8.00 because he believed the rule was to add the numerator and denominator and then divide by the number represented by the highest place value.
https://blog.mathed.net/2011/07/rysk-erlwangers-bennys-conce...
See this article on SAT test prep.[1] The requirement that only one answer can be right means that wrong answers have easily identifiable properties.
[1] https://blog.prepscholar.com/the-critical-fundamental-strate...
Yeah, it's confirmation bias. People do that sort of thing all the time in machine learning research, specially in the recent surge of LLM-poking papers. If they don't do that, they don't have a paper, so we're going to see much more of it before the trend exhausts itself.
Abductive reasoning is common in day-to-day life. It seeks the best explanation for some (often incomplete) observations, and reaches conclusions without certainty. I would have thought it would be important to assess for LLMs.
Is abductive inference synonymous with bayesian inference?
https://plato.stanford.edu/entries/abduction/#AbdVerBayConTh...
For example suppose you go outside and the streets are wet. Perhaps it rained, or perhaps someone drove a fire truck around spraying water all over the streets. You might select the former because of its higher prior probability.
They are statistical text generators, whose results are defined by their training data set. This is why the paper cited reads thusly:
Despite extensive research into the reasoning capabilities
of Large Language Models (LLMs), most studies have failed
to rigorously differentiate between inductive and deductive
reasoning ...
The authors then postulate:
This raises an essential question: In LLM reasoning, which
poses a greater challenge - deductive or inductive reasoning?
You can write programs into transformer models and run them deterministically; they aren't "statistical".
It's also still statistical if it gives you the same answer 99999/100000 times. You can definitely tune these things to be much more certain about certain problems but it tends to decrease capabilities in others.
I also don't like the term "black box." While they aren't transparent they aren't completely opaque either. A big reason I didn't like the term is that I feel it also encourages people to not research this more. While we don't completely know what's going on I see no reason we can't. Is there some proof they their calculations are an unbreakable encryption or is it just a hard problem. I'm pretty sure it's the latter, and I think it's good to differentiate "we're dumb" from "indeterminate"
I was thinking temperature=0, but what more do you need than that?
(nb there is also some randomness from non-associative FP operations happening in different orders)
> I also don't like the term "black box." While they aren't transparent they aren't completely opaque either.
A white box is a subset of a black box here. The important part is that the sampler/text generation is wrapped around the transformer model and is not actually the same thing as it.
> I was thinking temperature=0, but what more do you need than that?
So say such a thing makes something non-statistical is the same as saying prime numbers don't exist because any prime p * 0 is 0.
> A white box is a subset of a black box here
Please take your passion and use it to drive you to learn about these things deeper. It's actually the same advice I give to the LLM fanboys who think it's AGI. But trying to dispel that myth without correct information makes it harder. It muddies the waters. These topics are extremely complex and the issue with the ML field is itself a hypocritical claim that it is easy (scale is all you need while calling these machines black boxes). I mean we can talk about the elephant in the room but certainly Von Neumann has said enough.
So it's a subset, because "understanding something" is a refinement of the state of "not understanding something" - you can just pretend you don't know what's happening in that part of the system and it is now a black box.
Any part of a program can become a black box if you decide it is one (if you can figure out an interface anyway…), and that's good because it means you can swap it out.
The next-word predictions output by the model itself are based on the patterns and statistics of the training set. The next word that is actually generated is typically a random selection per a top-k or top-p sampling scheme.
you mean like this?
People really don't understand consciousness, but it's actually quite simple when you realize that the question isn't "is this conscious?", it's "what is this conscious of?". A rock is not conscious of anything, because it has no state that represents anything. A sunflower is conscious of the sun's position, because it has state that represents that. A cat is conscious of much more, such as "this mouse is food". An LLM is conscious of much more, but less than a human.
The LLM is a model that is conscious of many things, as experienced through the medium of human text.
I think Qualia is a wave phenomena, and all wave phenomena simultaneously can be described as "existing" and "not existing" as follows:
Think of a football stadium "wave". It has a beginning, ending, location, velocity, etc., all being scientifically measurable observables. Yet I ask you "Does the wave exist, or is it merely people in various states of arm positions?" There is no right or wrong answer. This is the nature of waves.
So if you want to argue consciousness "does not exist" I say fine, that is a valid "perspective", just like the denial of the existence of any other wave phenomena is a valid perspective. However keep in mind also that the entire universe is also made of waves.
I'm not arguing that consciousness doesn't exist, I'm saying that it's not a binary thing, and trying to provide a useful, falsifiable definition of it that provides a way to actually communicate about it instead of talking past each other.
I'm not even pushing the panpsychism view that "Everything has Consciousness, because all atoms have EM fields" at all, but when you get right down to it, EM fields are more "real" than a computation is. A computation is purely theoretical. We can build computer logic circuits out of water pipes or dominoes. So it's just obvious to me Qualia is not built on computations.
Believing Qualia is made of computations is more of a naieve "Universe is Love" type position than believing that Qualia is built of something REAL like EM fields. Calculations are not even real. For example the number 0100101110110 in a computer is purely a "label" for something real. Numbers in computers are just labels. EM waves are the actual thing, and actually real. The Map is not the Territory. Computers are Maps. Qualia is a Territory.
Like for all anyone knows, they're simply present to trigger (and keep triggered) "consciousness computations".
We must be *very* careful before making authoritative statements about who and what has qualia.
Sure, it's still debatable for us humans (the term isn't well defined) but that doesn't change the direction of required proof.
There is cultural/normative truth, and then there are various stricter/technical interpretations (using forms of logic other than binary, which causes reality to appear very differently), and making such distinctions is typically culturally "discouraged" (it is rude, pedantic, not what HN is for, etc).
The burden of proof does have direction. It's in the direction of the more complicated thing. Do you think a rock is conscious? What about lightning? If you're 99.999% of people, then the answer is no. So if you want to claim AI is conscious you need make the claim where inanimate becomes animate. And you know what they say about extraordinary claims... But personally I don't believe in ghosts, even those in machines
It's not about animate vs inanimate, it's about models that represent something.
There's also some confusion because we have limited language here. A person can be "unconscious", but that doesn't mean they've lost their world model. It's just currently not executing.
EDIT: I mean neuroscientists know brain frequency is more than an accidental correlation with consciousness.
I mention "unconscious" only to help disambiguate an overloaded word. It's used for both "this model includes something in its representation" as well as "this model is currently executing".
People often just hear the phrase "brain waves" and immediately think of "woo woo spirituality" because all the spiritual crack pots are constantly talking about waves without even having a physics background.
How about picking up a physics textbook? The status quo for Quantum is Griffith's Introduction to Quantum Mechanics. Might I also suggest his E&M book (again, a standard). If you're half as passionate as you portray, I'd also pick up Jackson. That's the standard graduate level E&M book (but a warning on this: to get through Chapter 1 you're going to need to be intimately familiar with PDEs and Green's Function. To get further, I sure hope you like Bessel functions).
Take your passion and use it to drive you to prove. Don't just conjecture and don't try to derive things from first principles. You will get it wrong because you won't actually be working from first principles. Again, don't ignore humanity's greatest invention. It's so much easier to catch up than reinvent the wheel.
Also, you should become familiar with Penrose. I'll leave it to you to figure out why.
Have you done substantial reading?
> because they have no state that represents anything.
The point I'm making is that representation is not enough. Storage (or memory) is not enough even recall is not enough. If it is, then any mechanical process is. Because you can build computation machines out of nearly anything. You can build memory out of almost anything. And the hello world python program I wrote has both!
You can argue scale, but we see animals with much smaller scale than these ML models perform unambiguous reasoning tasks and critical thinking. You can then argue architecture and data, but then that's not just scale.
The problem with all these people making claims is that they don't consider base assumptions and they don't follow the consequences of the model (your mental model of what is consciousness) is and must mean! So many people argue that they are working from "first principles" but I don't see anyone writing down their fucking axioms.
Of course you should try to prove what you want, but not at the cost of not being wrong. Doing so will prevent you from achieving what you want. I'm just so tired of people shooting themselves in the foot and calling it medicine. We need to stop with all the bullshit because I actually want to build AGI
Sure then following OP's model outlined here https://news.ycombinator.com/item?id=41427796, rocks are simply conscious of the passage of time. Perhaps it can be said that is the one thing every matter in the universe is conscious of.
>The point I'm making is that representation is not enough. Storage (or memory) is not enough even recall is not enough.
You don't know what is or isn't enough.
And by the way, what 99% of people think of a concept they can scarcely define, test for or prove doesn't mean shit. We are regularly wrong about concepts with much stronger standings.
Humans don't have any special insight to consciousness. We can't even make decisions without unknowingly altering our preferences and fabricating reasons in the process. What makes the conscious mind tick was clearly lost along the way.
>We need to stop with all the bullshit because I actually want to build AGI
No, you want to build the idea of AGI that exists in your mind. Clearly not Artificial General Intelligence seeing as what exists today checks all those boxes.
You want something more, and that's fine because I do too. Our state-of-the-art models are baby AGI, because they can accomplish a wide variety of tasks without being specifically trained for it. They can do limited reasoning on new tasks. That's intelligence that's both artificial and general. But it's just the start.
There's a few axes that need to be captured. Breadth of knowledge, ability to reason, and ability to update the model based on new information. LLMs are great at 1, possible superhuman. For 2, they're 1/10 on a scale of 0 to human. For 3, fine tuning and prompt stuffing get the job done for now, but they're a far cry from human ability.
And yes, you can build computation machines out of nearly anything. If you simulate a human brain with sticks and stones, what does it matter the substrate?
> AGI exists now
What I hate about these conversations is that people learn the basics but there's been a lot of advancements since the 50's. We've learned a lot since the 90's too! Yes, the metrics keep changing, but they are not in response to the current state of the art in a specific effort to make the systems not AGI[1]. But rather the criteria keeps changing because we keep learning. It would be ludicrous if it didn't! Because WE STILL HAVE NOT CONCRETELY DEFINED CONSCIOUSNESS, REASONING, THINKING, and many other similar terms. The criteria are generally guides created by best guesses. But until we have formal definitions, expect it to change (and when we do have formal definitions, be sure as hell to understand their assumptions!).
But I can confidently tell you that LLMs do not exhibit generalization in the way we've been using it for decades. This is easy to demonstrate but if you're a true believer then there's no convincing. And in that case, just don't bother arguing either.
I really do encourage people to take their passion and use it to drive gaining a deep expertise on these subjects. You're passionate right? I hope that passion is not about "being right" or "having answers" and is actually about "understanding." And if you can't differentiate that, I suggest starting at the beginning, since these topics are covered there.
[0] Sure, you can find examples of people stating criteria, but I can find you samples of idiots, even among scientists.
[1] There are some bad faith people who do this but they're terrible people too. You may freely disregard their opinions.
You are incorrect. Any claim of fact has a burden of proof.
> So if you want to claim AI is conscious
You made a claim that it is not. How did you determine that is True? Or did you?
```python
def main():
print("Hello gaganyaan")
if __name__ == '__main__':
main()
```
So... The burden of proof is about what level of complexity is sufficient.
Sure, no one has this answer. But also despite a few billion people believing in ghosts, including some saying a large portion of that saying they've seen ghosts, the burden of proof is to show ghosts exist, not showing ghosts don't exist. The reason being that one is falsifiable while the other isn't.
Don't confuse "the absence of proof is not proof of absence" with "the absence of evidence is not evidence of absence." If you do you might start a war in the middle east.
This is why people who believe the "Consciousness is a Computation" view never want to discuss brain waves, despite the overwhelming evidence that brain waves is precisely the thing out of which consciousness is composed.
I'd be interested in seeing any of this overwhelming evidence, but it's also not relevant, unless you're arguing that brain waves are supernatural in origin and can't be simulated. I think the crux of your argument can be boiled down to: "What's inferior about a perfect brain simulation, regardless of medium?"
But I don't think this is a relevant analogy. It seems that you're arguing a specific form of the embodiment hypothesis; that consciousness requires a body. Surely this is a sufficient condition (and is why I think roboticists have a good chance at cracking AGI), but is it a necessary condition? I'm not sure anyone has any convincing evidence that cognition requires "a body" (is a standard computer "a body"? What about a laptop that has a camera? What body parts are necessary?) and if it does, it is unclear what constitutes the minimal necessary requirements. It is quite possible that this is all "software" (e.g. a person in a vegetative state may be completely oblivious to their surroundings but still may have conscious experiences. Most people do this every night!).
I think you are erroring too far in the other direction. While I highly disagree with the person you're responding to, I do no think you are making strong arguments yourself. With so little known about what actually constitutes consciousness, I think it is difficult to make bold claims about what it is (though we've ruled out things that it isn't; e.g. a rock. Which is our typical method of doing reasoning in science; we generally disprove things, not prove. Like you said, the map is not the territory, and in this say way our models (yes, plural) of physics is not physics itself).
I think the is evidence to believe that consciousness may be a "software" but that is not proof. Similarly I think we have lots of reasons to believe that embodiment is highly beneficial to creating consciousness, as not just by nature of examples, but our understanding of consciousness is in part about responding to environments and stimuli. But that's clearly not sufficient either and pinning down where this is, is quite complex (even our definition of life is ill-defined). I'm highly convinced that embodiment is the best path to pursue to achieve AGI, but I'm not convinced it is required either.
[0] If we get pedantic, we should note that any simulation has to run on hardware and that creates magnetic fields but these are not in fact necessarily the fields being simulated. I don't think we need to pursue this avenue for this conversation.
I think neurons do 4 distinct types of things. They 1) gather data (sensory inputs) and route it to 2) locations where it's oriented into 3D spatial shapes that function essentially like "Fractal Radio Antenna" [for lack of better description] where the 3) waves resonate to create what we call qualia, and then 4) the waves induce a current/action going to motor neurons. It works similar to a television set working in reverse, with the screen being an eye, and everything going backwards in time, not forwards. Pixels are sensors, not lights, in the TV analogy. Thousands of pixels are input (rods/cones) and it all routs to (not from) an antenna.
You can explain all of that with Maxwellian Electrodynamics. There's no "woo woo" there. It's physics. If you look at it like this, you can see that most of the neural network wiring is I/O and signal routing, except for the antenna-like structures, the best of which being the Corpus Callosum and Hippocampus, which evolved first, which is why they're deepest in the brain. Their physical shape is critical, because they are both indeed functioning as a kind of resonator/antenna, and are getting memories (what most of consciousness is built on is memories) thru wave effects.
CORRECTION: You don't need a biological brain for qualia, but you need something that generates the same actual wave mechanics in actual physical reality. 0s and 1s in a computer ain't that.
> but definitely a brain (excluding plantlife to keep this short)
> I think neurons do 4 distinct types of things.
I ask because I suspect you're trying to solve the problem yourself and not building on the wide breadth of literature and research we have.
Brains do not operate in waves the same way a wave of water or electromagnetic wave does. Though anything can be approximated by a wave too, but signals need not be waves. Yes neurons conduct electric signals, but neuron to neuron communication (synaptic transmission) is chemical. These are discrete particles. But as I mentioned before, if we're talking about the electrical transmission inside the neuron, these are spikey[0,1]. When you look at neurons firing, there is a digital like signal. If anything, these look like Gaussian mixtures.
> There's no "woo woo" there.
So don't try to dispel people who are calling to magic by making a call to a different kind of magic. I don't think you're dumb, but rather it is easy to oversimplify and end up with something that requires magic. It's FAR easier to do this than to get the right answers. There's a reason people study very narrow topics for decades. Because it is all that complicated and trying to reinvent their work (while a good exercise) is just more likely to end up making many of the same mistakes that were then later resolved. We have the power of literature and the internet makes it highly available (though can be hard to differentiate from hogwash and it can be hard to find truth). Use the tools to your advantage instead of trying to do everything from scratch. Do the exercise, but also check and find the mistakes, improve, and do this until you catch up. Hell, you may even find mistakes the experts made along the way! The good news is that its SO much easier to catch up than it is to extend knowledge. You can learn what took thousands of years to get to in a week! It's hard, but in comparison catching up is trivial. So don't squander humanity's greatest innovation just because you worked hard at building your model. But refine it and make sure it doesn't invoke magic.
[0] https://en.wikipedia.org/wiki/Biological_neuron_model
[1] https://en.wikipedia.org/wiki/Soliton_model_in_neuroscience
> This is why people who believe the "Consciousness is a Computation" view never want to discuss brain waves, despite the overwhelming evidence that brain waves is precisely the thing out of which consciousness is composed.
But I really disagree with your argument. The difference doesn't come down to analogue vs digital and that seems like quite a bold assumption, especially considering 1) digital can get quite approximate to continuous signals and it can represent continuous signals 2) the brain doesn't function entirely continuously. Neuron signals are quite spikey. The neuromorphic people drew their inspiration from these and other aspects of the brain. But the brain is not operating on just continuous signals. There's a lot going on in there.
But if your true complaint is about how few ML people and far fewer ML enthusiasts will passionately discuss these subjects but also refuse to read literature from neuroscience, mathematical logic & reasoning, and other domains, then yeah, I also have that complaint. When you're a hammer I guess....
> Do you think a computer simulation of an EMF wave is an actual EMF wave?
> aren't you also claiming that actual EMF waves are "just computations"?
You are confusing what simulation means, which is why I explicitly put it in quotes. There's a dumb "simulation hypothesis" (the likes of which people like Elon Musk discuss) and there's a "yeah the terms are broad enough so it can be viewed as a simulation - simulation hypothesis". The latter isn't really contested because it also isn't very meaningful in the first place. But it does matter here for what you're talking about and why we brought up that a rock has "memory" since we can store information in the atomic structure. You're missing the abstraction of the stuff and how we're just talking about entropy.
Look at it this way, how does nand flash store memory? You trap electrons in states, right? This is to say you isolate charge to one side of a transistor gate from another. You can also do this with magnetic storage. There's plenty of people who build macro computers with things like water, and in those you're separating buckets of water. Physicists love to build computers out of things, and there are even thermal computers, using heat flow for computation.
That's the thing, computation is just about differentials in entropic states. So literally everything is a computer and everything is computing. Thanks emergence (another extremely misinterpreted word that people believe holds much more meaning than it does). But this does not mean there's some programmer, god, or any of that. It just means that information exists... You're making the same mistake these Musk like people talking about "living in a simulation" make, conflating the computer you're sitting in front of now that you control and have power over with just information flowing (obviously you can affect things in the physical world, and that's why experiments are physical simulations -- since they are controlled and limited settings -- but good luck programming "the universe" -- whatever the fuck that means. You're going to need more energy that it's got...)
So view physics as the instruction set of the universal computer or whatever. We're dealing with huge levels of abstraction so you can't go about understanding these things with reference to how we normally operate (this is why physics is so fucking hard and why advanced math -- well beyond calculus -- gets so crazy). Just like you can't work with high dimensional data by trying to relate to things in 2 or 3d, because the rules are very different at those abstraction levels (for example, the concept of a distance measure becomes useless. This is the curse of dimensionality that people very often misinterpret due to looking at it in a limited scope).
When someone makes a conjecture about physical reality like "Consciousness is made of EMF Waves, not Computations" (like I did) and then someone else tries to refute that with an appeal to the Simulation Theory (like you did) then that's called "Appeal to Base Reality".
Even if we're living in a simulation we still should consider EMF waves a very different phenomena than computations. Computations are not a phenomena of physics. Computations can be done on an abacus.
When run, that program will have a very simple model that includes only a few things like stdout. More complex programs will have models that represent more things. A game of tic tac toe written in Python is conscious of the state of the board, because it models it.
It's not "when does it become conscious.", it's "how much is it conscious of?". Asking the prior question is like asking "When does someone become rich?" or "When is the apple ripe?"
This isn't about ghosts. This is exactly a way of looking at it that provides a concrete, falsifiable definition, because far too many people are talking right past each other whenever "consciousness" comes up. This is trying to get past the magic words, and get at something useful.
> like "become conscious", which binarizes it.
There are countless things where we recognize they are on a spectrum yet we use shortcuts in language to specify a sufficient threshold. If you don't believe me, may I introduce to you the color blue. Or any other color for that matter.
There is still the aspect that even when we consider consciousness to lie upon a continuum this does not mean that the function yields a non-zero result (or even approximately) prior to some threshold. In this case there is no ambiguity of "becomes conscious" as you can interpret this in the most weak form of consciousness. If you believe that the consciousness is only zero at the domain minimum then please state so, as I believe you are smart enough to not misinterpret my meaning and we can actually have a productive conversation. If you believe in that and also that it is a very slow growing function until a certain point, well... if you're going to be fucking pedantic then be fucking pedantic. In either case, you don't have to force this conversation into an argument and reconstruct it so that you're smart and I'm dumb. Who knows that relationship and such definitions are typically obtuse anyways.
So please, engage with at least a little good faith here.
> This isn't about ghosts.
Don't talk about people "talking past one another" when you are egregiously misinterpreting. In any conversation you have to make sure that you are correctly decoding the other person's words. At least if you do indeed want to "get at something useful."
I'm not trying to pigeonhole you. I do believe that consciousness is only zero at the domain minimum and it's a slowly growing function, but I disagree that there's any sort of "certain point". Humans are "more conscious" than cats (and better terminology is that they are conscious of more things than cats, or their world model encompasses more than cats, or something like that), but that doesn't mean humans are conscious and cats aren't.
My entire point is that "consciousness" is entirely unlike "blue". You can ask the question "is this blue?" and have a coherent answer. You can't ask the question "is this conscious?". You need to ask the question "what is this conscious of?".
> I realize you're using ghosts as an analogy
> I'm saying this isn't about anything mystical.
> I'm not trying to pigeonhole you.
> You can ask the question "is this blue?" and have a coherent answer.
> You need to ask the question "what is this conscious of?".
The thing is, everything is fuzzy. Embrace the chaos, because if we're going to talk about brains and consciousness and abstractions, well you're not going to be able to work with things that are concrete. But, this too is fuzzy and I think you're likely to misinterpret. Just because everything is fuzzy doesn't mean there aren't things that are more wrong than others. It's more that you can't have infinite precision. So stop trying to deal in absolute answers, especially when we're in topics that people are have been unable to solve for thousands of years.
[0] https://en.wikipedia.org/wiki/Blue#History
[1] https://en.wikipedia.org/wiki/The_dress
[2] https://en.wikipedia.org/wiki/Linguistic_relativity_and_the_...
[3] https://www.reddit.com/r/colors/comments/1ao2osm/is_this_col...
[4] https://www.color-meanings.com/shades-of-blue-color-names-ht...
The real issue is establishing the limitations of their reasoning capabilities, and how to improve them.
Reducing reasoning beings to the level of AI is an affront to—and demonstrates a genuine lack of understanding of—the nature of organisms and reason itself, and the nature of AI and its capabilites.
I'm tired of reading comments from people who keep repeating that LLMs don't think, don't reason, isn't intelligence because it is not human. If your definition of the above is because it's not human, it's quite useless as a comment. We know LLMs aren't biological human brains. So what?
Define what reasoning is to you. Then tell us why LLMs don't reason and why it matters.
1. Reasoning is ability to at least carry out proofs in FOL (first-order logic). FOL can simulate Turing Mach
2. LLMs are formally equivalent to only a subset of FOL.
Why is this important? To model human mathematics, you need at least first-order logic.
These arguments have been around for decades, e.g., Penrose. I am tired of people bringing up strawmen arguments ("Not intelligent because not human!")
The type of 'reason' a cat uses is different from the 'reason' used in math
The information a cat uses is incomplete whereas the information used in math and logic is theoretically all accessible.
The reasoning a cat uses allows for inconsistencies with its model because of its use of incomplete information whereas no inconsistencies are permissible in math and logic.
Formally speaking, math uses deductive reasoning whereas the cat uses abductive reasoning.
They can, at times, and do so best when emotion is not involved.
> I'm tired of reading comments from people who keep repeating that LLMs don't think, don't reason, isn't intelligence because it is not human.
LLM's represent a category of algorithm. Quite elegant and useful in some circumstances, but an algorithm none the less. A quick search produced this[0] example discussing same.
Another reference, which may not be authoritative based on whatever most recent edit the link produces, is[1]:
A large language model (LLM) is a computational model
capable of language generation or other natural language
processing tasks. As language models, LLMs acquire these
abilities by learning statistical relationships from vast
amounts of text during a self-supervised and semi-supervised
training process.
Reasoning was the process I went through to formulate this response, doing so with intent to convey meaning as best as I can, and understand as best as possible the message to which I am replying.
> Then tell us why LLMs don't reason and why it matters.
LLM's do not possess the ability to perform the process detailed above.
This is why it matters.
That's not a useful definition for judging whether LLMs reason or not. It's not something we can measure on an objective level and introduces another concept of intent which is just as vague as reasoning.
Specifically, an LLM can produce a similar message to what you posted. Everything else about that process is not defined well enough to differentiate you.
> That's not a useful definition for judging whether LLMs reason or not.
I formulated this portion of my post specifically to induce contemplation as to what it is when we, as humans, reason. My hope was that this exercise could provide perspective regarding the difference between our existence and really cool, emergent, novel mathematical algorithms.
What irritates me when I see comments like yours is that precise knowledge of weaknesses of LLMs is necessary to either improve LLMs, so most of the people who claim LLMs reason or are already AGI basically deny the ability to improve them, since they are already perfect. Research into studying the limitations of the current generation of AI is unwanted and by extension so is the next generation of AI.
> They are statistical text generators, whose results are defined by their training data set
Beyond that, the experiments are almost certainly in the training data. Like come on! I feel like I'm going crazy here. How can any coder not think LLMs are training on oct and hex‽
The question is would the results be largely valid if this was done on a human who had learnt how to perform base 8, 9, 11 etc arithmetic instead ?
I mean, they're clearly not trying to test the ability to derive base arithmetic from scratch.
My belief is that what LLMs do is best described as approximating the outputs of reasoning. This is different from reasoning itself and also different from simply regurgitating the most similar training examples.
I also think it's extremely challenging for people to understand the significance of LLM output because language is intertwined with meaning. So if you construct a valid and seemingly responsive sequence of words in response to a prompt, there is a strong illusion of purpose behind those words, even if they were assembled in a completely naive fashion.
So the fallacy of the whole LLM field is the belief that language has some intrinsic meaning. Or if you mix the artifacts of language in some very smart way, the meaning will emerge. But it doesn't work if meaning occurs before the word. The text in books has no reasoning, it was authors. The machine shuffling the text fragments does not have a meaningful thought. The engineer which devised a shuffling machine had some meaningful thought, the users of the machine have same thoughts, but not the machine itself. To put it another way, if there was an artificial system capable of producing meaningful thoughts, it is not a presence of language which produces a proof, it's communication. Communication requires an agent (as in "agency") and an intent. We have neither in LLM. As to the argument that we ourselves are mere stochastic parrots - of course we can produce word salads, or fake mimics of coherent text, it is not a proof that LLM IS the way our minds work. It is just a witness to the fact language is a flexible medium for the meanings behind - it can just as well be used for cheating, pretending, etc.
The text from this point on seems to have lost contextual awareness of what preceded it (which was excellent imho).
> One simple argument is that language is NOT a magical reasoning substance in itself, but a communication medium.
The only reason it seems like it is reasoning is because it’s probably stuffing a lot of reasoning in its context, and regurgitating that out in ways that are statically weighted with other things in the context on what is being reasoned about.
Frankly, even most commenters on HN don’t get how LLMs operate, thinking the model itself is what knows about different bases like hex and oct, when really, it searched up a bunch of material on different bases to include in the context before the model was ever activated.
I noticed that if I ask it to tell me how good cryptocurrencies are, it'll do it, and then if I say I disagree and they're wrong, it'll simply switch and agree with me as well. The thing has no care for truth, no opinion of its own, no ability to insist, and just feeds you whatever is statistically close to your own questions.
Lagniappe: https://www.youtube.com/watch?v=9B2ww3fiX30
But any mention of LLM reasoning ability ought to address the obvious confound: the LLM is trained on examples of deductive reasoning, inductive reasoning, abductive reasoning, SAT-solver reasoning, geniuses' musings, etc. If they replicate one of those examples, then should that be called "reasoning" of any sort or not? Regurgitating those examples may even involve some generalization, if the original topics of an example are swapped out (perhaps by a nearby topic in latent space).
Given that it appears they're training and testing on synthetic problems, this objection probably does not apply to their actual results. But given the fuzziness it creates for the definition of "reasoning" of any sort, I would have expected some working definition of reasoning in the paper's abstract.
Training on Moby Dick and thus being able to regurgitate text from Moby Dick does not mean the LLM is capable of writing a new Moby Dick-like book. (Thankfully; one is more than enough!)
Because they are not.
LLM's are a better form of Bayesian inference[0] in that the generated tokens are statistically a better fit within a larger context.
A popular LLM joke at the moment is "How many Rs in strawberry?"
You, a human being, understand this question. The purpose of the activity is to count the letter R in that word and give the one correct answer.
LLMs don't understand any of that. They don't know what words are, what letters are, what numbers are, what counting is, that a question can have one definite correct answer or many answers, what a question is, nor what an answer is.
They break your input into tokens and then look at the most likely set of output tokens given your input. That's all.
If you train a model on enough sensible input tokens and sensible responses then the output will mostly seem human and sensible. But it's never reasoning.
"How many Rs in strawberry?" is not a trick question. That's about as straight and factual as a question could be.
To a system that sees letters and words, sure. To a system that doesn't, you're just asking a blind man to count how many apples are on the table and patting yourself on the back for his failure. It just doesn't make any sense to begin with.
And this is before the fact that human intelligence is rife with absurd seeming failure modes and cognitive biases.
It's frankly very telling that these token related questions are the most popular kind of questions for these discussions.
This is no testable definition of reasoning or intelligence that will cleanly separate LLMs and Humans. That is the reality today. And it should make anybody pause.
Which input is it you think an LLM doesn’t have to be able to count the number of letters in a word you literally provide it
Imagine the original question is posed in English but it is translated to Chinese and then the LLM has to answer the original question based on the Chinese translation.
It's a flaw of the tokenization we choose. We can train an LLM using letters instead of tokens as the base units but that would be inefficient.
The way we tokenize is just a design choice. Character level models(e.g. karpathy's nanoGPT) exist and are used for educational purpose. You can train it to count number of 'r' in a word.
The peephole will expand soon, as multimodal models come into their own, and as the models start getting mixed with robotics, allowing them to go and interact with the world more directly, instead of through the medium of human-written text.
We are talking about a computer program whose operation we understand, can observe, and can debug. Modern LLMs certainly take a lot of human effort to do this, but it can be done.
"Please spell the word strawberry with a phonetic representation of each letter then tell me how many 'r's are in the word strawberry?"
It's gotten the answer right every time I've tried.
> ChatGPT answer: There are two "R"s in the word "strawberry."
Given enough instances of the "LLM joke" in a training data set, the joke itself having a consistent form (sequence of tokens) and likely followed by the answer having a similarly consistent form (sequence of tokens), the probability of the latter being produced as quoted is high.
That's humans as well, possibly. This description gets repeated over and over and is factually correct, but we still don't know if brains do anything more. The "it's not reasoning" may be true, but it doesn't follow from just that description.
That isn’t right: the pre processor provides a lot of material on strawberries and counting r’s, which is then pretending to the question…and then they predict the next sentence as an answer to the question. The model by itself doesn’t know anything, it it just a statistical processor of context, just tokenizing the question and using the model to predict the answer would actually give you less than a wrong answer, it would be gibberish. It messes up on the question because the context it retrieves based on the question text isn’t useful in producing the correct answer.
"retrieves" is the wrong word. Each token (in GQA a small tuple of tokens is summarized into a single token), becomes an element in the KV cache. The dot product of every token with every other token is taken (matrix multiplication) and then a new intermediate token is produced using softmax() and multiplication by the V matrix. What the attention mechanism is supposed to do is combine two tokens and form a new token. In other words, it is supposed to perform the computation of a function f(a,b) = c. The attention layer is supposed to see "count r" and "strawberry" and determine the answer 3.
Well, at least in theory. Given the combination "count r" and "r", it is practically guaranteed that the attention mechanism succeeds. What this tells us is that the tokenization of the word "strawberry" is causing the model to fail, since it doesn't actually see the letters on the character level. So it is correct to say that the attention mechanism does not have the correct context to produce the correct answer, but it is wrong to say that "retrieval" is necessary here.
The reason why it doesn't make sense to label what is happening as "reasoning" is that the model does not consider its own limitations and plans around them. Most of the work so far has been to brute force more data and more FLOPS, with the hope that it will just work out anyway. This isn't necessarily a bad strategy as it follows the harsh truth learned from the bitter lesson, but the bitter lesson never told us that we can't improve LLMs through human ingenuity, just that human ingenuity must scale with compute and training data. For example, the human ingenuity of "self play" training (as opposed to synthetic data) works just fine, precisely because it scales so well.
Instead of complaining so much about humans trying to "gotcha" the LLMs, what we really ought to build is an adversarial model that learns to "gotcha" LLMs automatically and include it in the training process.
I love seeing Victor Taelin experimenting with parallizing these programs (with HVM and other experiments with proof languages), but it's sometimes a bit sad how much time researchers take in making papers about existing things instead of trying to improve the state-of-the art in something that's most probably missing from the current models.
I don't know about "typical" but every source that classifies reasoning (or, more appropriately, logical inference) as deductive and inductive, also includes the abductive category. This categorisation scheme goes all the way back to Charles Sanders Peirce:
'[Abduction] is logical inference( ... ) having a perfectly definite logical form. ( ... ) Namely, the hypothesis cannot be admitted, even as a hypothesis, unless it be supposed that it would account for the facts or some of them. The form of inference, therefore, is this:
The surprising fact, C, is observed;
But if A were true, C would be a matter of course,
Hence, there is reason to suspect that A is true.' (Collected Papers of Charles Sanders Peirce. Peirce, 1958)
(Quote copied from Abduction and Induction, Essays in their Relation and Integration, Peter Flach and Antonis Kakas eds. 200)
Consider a logical theory, formed of rules in the form of implications like A -> B (premise A implies conclusion B). Abduction is the inference of the premises after observation of the conclusions, i.e. if A -> B AND B is observed, then A may be inferred.
That's a different inference mode than both deduction: inferring a conclusion from a premise, e.g. if A -> B AND A, then B may be inferred; and induction: inferring a rule from an observation, e.g. inferring A -> B after observing A and B. Note that this is a simplification: induction assumes a background theory of more rules A1 -> A2, .... An -> A that can be applied to the observation A and B to infer A -> B.
Anyway, abduction is generally associated with probabilistic reasoning, albeit informally so. That probably means that we should categorise LLM inference as abductive, since it guesses the next token according to a model of probabilities of token sequences. But that's just a, er, guess.
But, they clearly struggle with generalization and rule following. This failure to generalize (extrapolate, deduce, compute) is why we still can't fire all of our DBAs.
Has anyone encountered an LLM-based text-to-SQL engine that actually gets the job done? I think that's your best canary. I stopped caring somewhere around "transpose these 2 letters of the alphabet" not working consistently.