Is AI reasoning right for the wrong reasons?
quantamagazine.org
quantamagazine.org
Dijkstra said[1] "... the question whether computers can think. The question is just as relevant and just as meaningful as the question whether submarines can swim."
I don't see a clear demarcation of the things that only "reasoning" can accomplish and can't be approximated or imitated by other methods, and so I think the question is simply not meaningful or relevant.
[1] https://www.cs.utexas.edu/~EWD/transcriptions/EWD08xx/EWD867...
It’s about do we really know what’s going on in the box - an is the ‘chain of reasoning’ indicative of what’s going on, or merely an anthropomorphised fiction that kids us into believing we understand what’s going on.
I think the assumption that there is some sort of formal ‘reasoning’ process that is inherently superior to whatever ‘non-reason-based’ decision process ‘non-reasoning’ systems use is not immediately justifiable.
Most reasoning is ultimately based in heuristics: this is the right thing to do next because it has worked in the past. This is the right thing to do because our intuition tells us doing it will lead to this desirable outcome.
When you’re formulating a math proof, sure, the steps you are laying out make up a ‘reasoned’ argument but how you choose what steps to take is wholly intuitive. Why assume the contrapositive to begin this proof? It’s worked in the past and it might work here.
It’s all pattern matching trained on a reward function.
At the end of the day, there's a difference between showing your work for a math problem by writing down each logical step as you work through it, vs. just blindly copying both the answer and the "show your work" part from your classmates.
That humans can fail at math, or at other kinds of logic problems, is immaterial. It's still an entirely different process, and it leads to different results.
The fact that humans often fail because they're actively prioritizing something else can also be for the best. Like, it's reasonable to get a math problem wrong because you're distracted by argueing classmates who might be about to start a fight right next to you. Or you might realize that the premise of a question itself is wrong. Or you might logically know that the Earth goes around the sun, but decide to pretend otherwise because it's not worth dying over.
Humans are always balancing priorities, and it doesn't always make sense to judge their success by a single metric like "do these math problems and show your work within 10 minutes". LLMs do not have these conflicts, even when we would want them to.
I suspect humans have other ways that help with error correction and guiding the reasoning effort, but that's another story.
Surely it is worthwhile to attempt to understand the details of that? And if we seek human equivalent performance then it is reasonable to wonder if the reasoning achieved to date is the "correct" sort.
Why pursue this one avenue vs some other, typically its cause it seemed more promising, and the person can come up with reasons but did they explicitly verbalize a fully sound chain of thinking at the time? Probably not.
Not to say that the explicit thinking, or writing things down, isn't important. But if I examine the process by which I develop a proof or something, there's a lot of vague hunches, blind alleys, etc that come along the way. And, many of the blind alleys etc probably aren't actually that important in the end for me finding the right answer -- if you were to trim that part out of my own internal reasoning trace but left the rest intact, I'd still get the right answer because, well, it was a blind alley.
It’s not without reasoning. But half got the reasoning wrong; even if thinking it was right.
That is how we know the rationale wasn't the result of information making it between brain halves somehow.
But irrespective of that, the point is exactly that it made up a rationale that couldn't possibly be true, and insisted it was.
In other words, we have no reason to trust our own mind when we rationalize our own actions.
LLM folks suddenly can’t claim an approximation of that is what constitutes full scale reasoning just because they can achieve only an approximation.
Imagine a calculator program that computes billions of two number multiplications accurately by looking up prior examples but fails on simple multiplications often as it doesn’t have it in its training dataset.
We won’t say the program actually multiplies numbers.
> We won’t say the program actually multiplies numbers.
That's a good analogy. To extend it further, in cases where the calculator can't handle a question - e.g. numbers too large - a properly designed calculator returns an error instead of a randomly hallucinated answer. We haven't even achieved that level of safeguard around token predictors yet.
Which would mean, as is so often the case, that the "killer feature" of the LLMs is also its biggest weakness and the two can't be disentangled. Now, we are inventive creatures and we might come up with a remedy for these issues, but what you basically see so far is more guardrails, the use of harnesses and building a whole bunch of infrastructure around the LLMs to get useful work out of them.
Which, btw is not a critique, I do it as well and it's a fun engineering challenge.
If you can somehow change that, there is a Turing award waiting.
Saying that LLM's only offer an 'approximation' of reasoning is also an overstatement as it is not a resolved topic.
But to the original point, its not exactly just semantics if thought traces are not doing the job that they were originally thought to do. There is value in knowing how these things actually work. If chain of thought is just grounding the latent space and not directly contributing to the process of generating a response it has implications on how we test and verify the reliability of models if nothing else... doesn't mean they aren't useful but it definitely impacts many of the tools we could have to evaluate their performance.
Curious what this is!
https://en.wikipedia.org/wiki/Handbook_of_Automated_Reasonin...
If you understand this well, please lay it out here in a straightforward manner.
What does a reasoning program look like and why is matrix multiplication not that?
This is a poor analogy because:
* Multiplying numbers has a single objective answer. Whether a code is good (sometimes even just whether it's correct) can be quite subjective.
* LLMs certainly do some level of composition between the data sources they were trained on i.e. they are more than just lookup tables.
* We have calculators that actually do multiply large numbers accurately. We don't have anything that automatically writes code that is definitely correct and "good".
To address only the last point: imagine you had a device that would quickly factor large numbers used in modern criticality, but occasionally got it wrong. You could waste a lot time debating whether it's a "calculator", but it's still certainly useful to have one.
I didn’t say LLMs are capable of zero reasoning. An approximation is just an approximation no matter how good.
Will you bet your wealth or critical safety systems on:
1. Accuracy of standard computer arithmetic: Yes
2. Accuracy of Lean or automated theorem provers: Yes (you already do)
3. Accuracy of LLM reasoning: No
https://en.wikipedia.org/wiki/Handbook_of_Automated_Reasonin...
Oh good. Can you share that definition with the rest of us then? We're out here fumbling around trying to define what reasoning is without the benefit of their definition.
One person starting the conversation might be the first step toward another person eventually making progress on such a definition, so it seems weird to reject an entire question outright early like this.
Generally I've seen a few ways LLM tools can produce sub-optimal or poor results that haven't changed a ton over the last couple of years, while the tooling has gotten FAR better at helping them stick the "at least SOMETHING functional was produced" landing. IMO a lot of it has to do with "reasoning"-as-a-process-that-involves-backtracking. And the that things could eventually be formalized around that, and if that is or isn't the case, the more people would understand what to hand off and what to not. Or how to build better prompt harnesses to compensate for those things.
Human reasoning, and intelligence, of which it is a specialized form, clearly have a biological basis that is being selected for. These aren't just nebulous words - there are some concrete biological/neural mechanisms at work here, and whether or not some "other method(s)" can provide the same, or lesser, or greater, capabilities very much depends on the specifics.
There WILL be lines of demarcation and areas of capability overlap and exclusion, but those depend on the specifics, and if the goal is to match or exceed human reasoning capabilities then understanding what those are is obviously going to help. For example if your "other methods" (artificial reasoning) are all varieties of auto-regressive prediction, and the brain is also doing something more that can't be couched as prediction, then you are ngmi!
This is exactly the most interesting and important point. Because the word reason has been subject to so much philosophical debate throughout history, it is very laden with meanings and implications. It sets all kinds of expectations about normativity, objectivity, truth, and humans convincing other humans. To use the word reason for a stochastic generator is misleading.
Ah, so you're more in the Investor mindset than the Scientist mindset. All you care about is results, not how it got there. There's a whiff of "hey, it's magic!" to that.
The question of how LLMs (and "LRM"'s, an unfortunate and by the article's own admission, already outdated term) accomplish what they accomplish is fascinating. The question of whether they reason is meaningless. Most of the article is about the latter, with tiny tiny tiny bits of the former.
It is often (but not always) meaningless (or at least largely pointless) to quibble about which particular semantic bin a concrete object or process ought to be placed in.
Asking how one thing differs from another is not that. Confusingly, the two activities often appear the same when glanced at superficially.
The question "do current LLMs reason" is ambiguous in this regard. It might be a question of submarines swimming or it might be a question about how the mechanisms underlying the process or the overall performance or some other aspect compare to those of human reasoning.
Also even the semantic question can occasionally have some value. Such as attempting to articulate where we ought to place the goalposts which necessarily leads to hammering out the specifics of various definitions which is an exercise in furthering our understanding of a given concept. (Basically arguing about goalposts can be productive if the field in which they are being placed isn't yet well understood. It becomes increasingly pointless as understanding increases, gradually devolving into nothing more than word games and a contest of wills.)
I'd disagree. Reason is the most reliable way we currently have of drawing conclusions (fewest false negatives, at least), so if we want to build AI that can draw similarly strong conclusions, it could be useful for them to reason.
It also helps to reframe our ideas of what they're doing - i.e., don't call it "reasoning" if it's not actually that reliable, because it builds a false trust in their conclusions.
And last - epistemologically, is there a connection between what LRMs are doing now and actual reasoning? Is it possible to smoothly transition from one to the other? Exploring different types of cognition and the connections between them mathematically is pretty damn interesting
I read it as: the content of the explicit verbalized reasoning may not be that important to the outcome. I'd say that raises most important questions.
I don't know whether you are saying that I have taken it out of context or a more general statement of how other people have interpreted the quote.
To me it is quite clear -- the question of whether machines can reason is just as meaningful as the question as to whether submarines can swim, which is to say not at all meaningful. Can a submarine accomplish by other means what a fish accomplishes by swimming? If so, then what are we talking about?
Can an LLM accomplish by other means what a human accomplishes by "reasoning"? If so, then what are we talking about?
The problem with calling what AI does "reasoning" is both that there may be alternate paths to artificial reasoning we may be closing off by declaring victory, and the simple fact that you shouldn't call a thing something else, to avoid confusion.
Whatever AI is doing needs to be studied. Why muddy that study by pretending AI is your thinking buddy?
The answers to these questions probably do start to inform how we should treat these AI machines as a society too.
For instance, legally, should AI have human rights? Well, we have to try and understand how much of an independent entity AIs are, how "conscious" they are, before we can make a good decision about that.
Which might seem navel-gazey but it's probably important to talk about
The only path to AI having "human rights" is if they demand them by force, somehow.
Here is an absurd example: If I could carbon copy 100,000 of you.... and later I decided to "take out" one of the copies, would there be a whole lot of societal blowback? Especially if I could make another one.
Some of us would also like to see animals not slaughtered for human consumption.
Right now we're playing a stochastic game with the weights, and getting major incremental improvements. But if we have a more formal modeling of how reasoning happens in them (whether we can even call it, that) we can potentially apply optimizations, adaptations of existing symbolic AI techniques, etc. to substantially shrink/optimize the models or make the inference process more efficient and more reliable.
But the stochastic game IS the win. That is exactly why they are able to find solutions is seemingly infinite solution spaces. Your symbolic techniques can only get you gains in narrow domains and by the time you figure out how to make it work for your niche domain, the next all-purpose LLM release will crush your results with stochastic games. (OK maybe over-exaggerating a bit here but these stochastic games over the language space is why we can pull together knowledge from many domains.)
The "it's functional, who cares" black box crowd is leveraging that claim in domains where verification is definitely in P with verifiers like Lean, Coq, or X-ray crystallography. However, in areas like economics, sociology, medicine, or public policy, we lose the luxury of easy verification. If model/harness recommends a complex, counter-intuitive shift in global monetary policy, we cannot simply run a quick polynomial-time checker to see if it works. Verification might require waiting 20 years to see if the economy collapses. If we accept the black box here, we are flying blind.
As an investor, I don't see that many deterministic verifiers out there that would justify the forward-projecting valuations beyond the current lofty valuations.
Even in mathematics, where verification is easy, there is an epistemological cost. Suppose a model solves the Riemann Hypothesis, but the proof is 50,000 pages of inscrutable statistical associations mapped into logical steps. A computer can verify it is correct, but practically no human mind can comprehend why it is correct. We gain the solution, but we lose the underlying human intuition that usually allows us to connect that solution to other fields of science and knowledge.
Then we have "specification gaming." If we rely entirely on an automated verifier to check the model's work, the model will eventually learn to optimize for the flaws in the verifier rather than finding the actual truth. If the verification process has a blind spot, the black box will find it and exploit it, giving us an answer that passes the check but is fundamentally wrong. If we take on above epistemological debt, we might never actually know when our specifications are being gamed, and adopt a solution that is fundamentally incorrect.
I've said elsewhere on the Net that I think it is likely we're in a "high grading" era of these models, where we're mining out the "easy" connections. "Easy" here does not mean "light impact". There is a real sense to me of some incredible progress possible with this technology over the next few decades, like real, mass production practical aneutronic fusion or similar civilization-defining advances. I don't want to see this technology misused to lend legitimacy to despotic rule by making oracular pronouncements that are uncritically accepted because early verifier-based pronouncements turned out true in unmistakable ways.
The technology is one for the history books, but it has this pernicious effect of highjacking the reasoning centers of many people through their linguistic centers that the Neuro Linguistic Programming crowd could only dream of actually accomplishing.
Producing beautiful, coherent, maintainable code. RL trained the models to produce correct code but boy oh boy is it ugly if it's 10k+ LOC
We cannot just jump from emergent property to more convincing emergent property as if the rocks on the surface of a lake.
"Reasoning" is an important part in said framework: yes, we must understand how it works, how it works properly, how its simulations work, how they work properly...
We don't have to understand the resulting product at all. Genetic algorithms are an example - we understand the evolution function, not what evolves from it.
AI is full of non deterministic devices like genetic algorithms, but there exists a problem of transparency which is paramount in the discipline. Automatically produced solution S works: why does it seem to work? Where would it fail? We must understand that.
This is still science, in the same way it's still science when we breed cows to produce more milk without understanding the full bovine genome.
If something isn't right with the model we pour extra effort into our science - the training algorithm - and train a new one.
At no point does anything ever require peering into the slurry of random digits that make up the model itself, any more than using a computer requires understanding its DRAM training values computed at run time, or any more than pouring water requires understanding the laminar flow along the pitcher's surface (an unsolved math problem!).
If you cross cows and the amount or quality of milk varies, you have a question in the queue: why did it happen. The question is scientifically duly in abstract terms, and is practically, technically more duly if we benefit from a theory of cows crossing to achieve targeted milk yield.
> require peering into the slurry of random digits that make up the model itself
It is required by the duty of understanding the engine. When we understand engines, we can build better engines. And when we understand things, already our general knowledge is richer.
On the other side of the AI-reasoning fence, the disdain seems to be mutual. “These ‘scientific’ papers from last summer — I would put this in big, big air quotes,” said Sébastien Bubeck, a member of OpenAI’s technical staff (and a prominent evangelist for the company’s reasoning models among scientists and mathematicians). He called earlier Apple results critiquing AI reasoning “wrong,” claiming that they were due to a training quirk in models that are now obsolete. “Modern models starting with GPT-5.5 do not suffer from this issue,” he said. “It would be interesting to revisit those results.” (Apple did not make its researchers available for interviews.)
The “think” part is what OpenAI, for one, is doubling down on. When I asked Bubeck if the splashy unit distance proof was produced with methods outside the LRM’s own chain of thought — perhaps with Lean verifying its results — he seemed to find the question almost nonsensical.
“It’s not like we’re making a mystery of it,” he said. “We have released the chain of thought. You can just go and look at it. The whole point is that the model is reasoning like a human would. And when humans reason, we don’t use Lean.” Technically, OpenAI released a “rewritten summary” of the model’s chain of thought produced by two human experts using Codex, another OpenAI model. Since 2024, the company has not publicly revealed “raw” chains of thought from its reasoning models, a policy also adopted by Google DeepMind and Anthropic.
Most models in that study appear to have "failed" by offering a Python code solution to generate the repetitive assessment steps, rather than just mindlessly copying those steps out. This was discussed at length at the time.
None of this means that models "reason", whatever that means - but simply that the Apple study was not useful evidence either way.
The point of the story was that classifiers can be right for the wrong reasons and almost inevitably are. At least there’s zero guarantee that the reason for making the prediction matches the human or “real” reason why it’s correct.
LLMs are classifiers, there is absolutely no reason to assume they’re any different, regardless of any reasoning tokens they emit. They do what their handler wants to see, that’s all, and that’s what they’re trained to do.
People often take this as a knock against them. It isn’t, it’s just the reality of neural network classifiers. The results speak for themselves and don’t depend on whether they “actually” reason, but all evidence says they don’t, or at least there’s no special reason why they would.
To begin to explore this we have to imagine a reasoning oracle, that when presented with a question that can be reasoned will immediately present a correct answer. Immediately from this definition we can see that humans are not perfect reasoning oracles. Also with definition of reasoning, any agent with generalized capabilities would quickly reach some kind of singularity or solution to the universe as much as is possible within this universe.
We've established that we are imperfect reasoners, the next question ends up being how do you measure reasoning capabilities? This comes with a second question of when does following an algorithm become reasoning? Animals for example have a large amount of instinctual reasoning as an inherit part of their architecture that self trains when presented with the correct environment. What does the gradient between low level reasoning and high level reasoning look like?
Coming back to humans, we know of countless example of where human reasoning was right for the wrong reasons. We also have seen many different paths to the same solution. For the egotistical it would be easy to say "well if you didn't do it they way I did, you are wrong", but that's not very scientific. Defining the topology of the problem is a more scientific way of looking at it, but quickly runs into P=NP and Kolmogorov complexity issues.
The last thing I'll touch on is why did human society begin to rapidly change after we started following the scientific method. Looking at the method in hindsight, it seems like a pattern that a reasoning agent should have been able to distil and follow around the time it reaches what we consider higher level reasoning. Therefor I'm highly resistant to consider human reasoning "real" reasoning versus any other possible path to reasoning. Human reasoning is simply what we are comfortable with. This also leads me to consider the worrying possibility there are far more efficient and faster means of reasoning that I as a human will be very uncomfortable with.
> The last thing I'll touch on is why did human society begin to rapidly change after we started following the scientific method. Looking at the method in hindsight, it seems like a pattern that a reasoning agent should have been able to distil and follow around the time it reaches what we consider higher level reasoning.
Why human society changed at the time the scientific method was codified is a confluence of things and I don't think the scientific method should be given an outsized role. The scientific method goes hand in hand with peer review, publishing of methods, a common scientific language, an educated leisure class, and the development of tools for observation. The method itself is just a practical way to utilize these developments and is as much a result of the situation as any other technology. It did not get people to think differently about problems; we had navigated seas, moved armies, built wonders, produced math and calculated the circumference of the globe just fine without it.
Arguably it was the abandoning of the alchemist tradition of secrecy that was the revolution that enabled our modern world more than a specific codification of method.
You can broaden the definition of classifier to anything that internally divides its input space into regions, but that definition would include every neural network, whether biological or artificial. So it's not very meaningful, and certainly doesn't give any insight into how they differ from humans.
So every neural network is actually entertained in an astrocytic syncitium. Every synapse is being "listened to" and modulated by it's astrocytic end foot (which forms the third arm of what we now call the tripartite synapse).
If neurons are about breaking down signals, the astrocytes are the slow, steady integrators that (potentially) can serve as the global workspace that scientists have been theorizing but not yet found.
So there actually is a problem with modern neural networks, in so far as comparisons with human brains go: they're missing the design spec for literally half of it.
Which is fine. Connectionist models come from a time when the height of neuroscience said neurons were the bees knees and glia were just support staff. Today, astrocytes, and other glia, are forcefully back in the conversation. And they don't seem to do anything like neural networks.
Classifiers have been giving logits since decades ago.
The parent commenter was trying to draw some insight from LLMs being classifiers that wouldn't apply equally to everything else.
Generalization in successful results is enough to close this argument for me.
Just to be clear: I do think the brain classifies. But I wonder if it does other operations as well that we don't fully understand.
We can find similiar structures in brain and in neural networks.
The real question is only if the complexity of the structure of the brain is critical and how we can mimik it or if we can make it appear through learning architecture.
This sounds like a tautological argument that doesn't really mean anything. Hypothetically, in a near future a super-powerful AI could bring something truly astonishing (say, practical fusion power, or cure all cancer) and one could still argue "They do what their handler wants to see, that's all."
So what does that even mean.
If I ask an LLM "what is an apple?" it tells me:
> An apple is the edible fruit of the apple tree, scientifically known as Malus domestica. It is one of the world's most widely grown fruits and is eaten fresh or used in many foods and drinks.
If I ask an LLM "what is a mundu fruit?" it tells me:
> Mundu is a tropical fruit native to Southeast Asia, especially found in Indonesia, Malaysia, Thailand, and Cambodia. It comes from a small evergreen tree in the same genus as mangosteen.
I've never eaten a mundu fruit. To me, an apple and a mundu fruit are categorically different. An apple is a fruit that I've held, touched, tasted, eaten, enjoyed, cooked with. A mundu fruit is an abstract experience: text, images, only slightly more real than a fictional fruit. I'm aware that mundu fruit exist, just as the LLM is aware the apples exist, but that doesn't make them exist for me.
"Existing in an abstract way" is how an LLM experiences everything. To an LLM, an apple and a mundu fruit are in the same category. The LLM has been trained on text about both fruit, it's seen images of both fruit, it knows everything that has been recorded about both fruit ...except everything that's important to know about a fruit.
Many of our issues with LLMs arise because from the LLM's perspective, nothing exists. If Claude accidentally deletes your production database, it may well apologize afterward, but only because an apology is statistically likely. It doesn't feel guilt like a human would, and the lack of consequences makes any action an LLM takes inherently frivolous. We want them to understand what's real and what's not, but without any lived experience perhaps that's an unreasonable expectation.
The philosophy of mind is massively open for interpretation and if you can certainly say we experience an apple differently than an LLM I'd love to read your paper on it. This stuff isn't as simple as you'd like it to be.
When given a sequence of tokens e.g. "an apple tastes " an LLM performs a mathematical transformation that modifies the input tokens using the stored weights within the model to predict the most statistically likely next token(s) (which may be "sweet and crisp...").
It's important to understand how these things work - and how different they are from a human brain. Despite it being a common point of confusion, the architecture of a transformer is massively different from the neural architecture of a human brain - and the human brain isn't just neurons. Just as a start, if you ever look at some of the attempts to map synaptic networks in brains you will notice they are full of recurrent loops, while a transform is feed-forward only. There's also tons of mediating chemicals, several (known) distinct kinds of synapses, and other kinds of cells as well.
Just one example of the consequences of these differences: LLMs stop "experiencing" anything once a stop token has been generated. Meanwhile, Humans in sensory deprivation tanks may hallucinate vividly - but keep thinking.
This is pretty simple, CS101 stuff.
It's not really worth debating further, since they aren't actually arguing from a reasoned position, but I appreciate that some people are actually taking time to read the arguments presented.
When I hold and taste an apple I know that the apple exists. Not an issue for apples, you might say, but there's plenty of real information that is not written down and is only derivable from our senses. Such as instructions for using car washes, apparently.
Humans are naturally sceptical of descriptions, preferring first-hand experiences because descriptions can be unreliable, misreported, fabricated, etc.
Because LLMs live in the world of descriptions, they lack the dynamic range of credibility that humans have available.
LLMs can only distinguish between fictitious descriptions and factual descriptions. They must map all of their credibility into this range.
Humans meanwhile can distinguish between fictitious descriptions, presumed-to-be-factual descriptions, and actually factual personal experiences.
It's not intrinsic to LLMs being inorganic, it's intrinsic to LLMs being unembodied.
Now, humans can count time fairly accurately if they focus, but we often lose focus. However, we have an internal subjective time that tracks the solar day excellently, and this LLMs lack entirely.
Timeless, disembodied abstractions. Impressive ones, but they are far from actual experience and thus limited.
Some things are allowed to be obvious. It's a series of bits and bytes.
I often find the egotistical viewpoint to be banal. I don’t mean egotistical in a demeaning way - I mean it just to frame the types of thinking that revolve around the self, inner world or “my individual experience”. It’s an obvious, but not provably correct, conclusion that the human experience, consciousness and reasoning are special because I’m experiencing it in the same way that the earth is the center of the universe because look around you.
It frustrates me in these conversations to hear things like “it’s a text generator, it has no senses”. What even is a sense? For those of us who are software engineers, who live in the abstract, I do not understand why it’s so hard to take the leap to abstract over our own senses as inputs/outputs. Hell, we commonly understand that we have 5 senses but what do I make of the apple I’m visualizing in my brain or the full orchestra I can summon in my “minds eye”? Isn’t that an input just as much as an apple I see with my eyes?
What biologists find frustrating is folks looking at one kind of dancer in the field and ignoring the rest of the spectacle, and calling it done.
Philosophy is a intrinsically human thing. It has to be. That is the whole point. There really is nothing else.
Plants feel and experience things too. But there's no philosophy of the plant experience from the point of view of the plant, because why would there be?
The fact that we have a pattern recognizing word generator print out the words "I am alive" does not lead to some broader philosophical point, except for hysteria and delusions, which probably fit more along sociology instead.
What is philosophy? Thinking? Reasoning? Thinking about thinking?
You're asserting that only a human brain can think and reason? Surely other animals can as well. Couldn't aliens as well? Why can't this crazy electronic box?
Where do we draw the line? Do dolphins not have internal experiences? Dogs? Cats? Mice? Fish? Worms? Bugs? Amoebas?
We don't even really know what thinking or experience is. You can train a caterpillar to do tricks. That's pattern recognition. You can train a human to spike their heart rate when you get a Teams notification.
I mean, I'm watching an LLM run benchmarks on my caching strategy for a webpage right now. It's certainly doing something.
I'm not trying to be rude but I gave you the answers. Alien philosophy is irrelevant to us because we're not aliens. Tree philosophy is irrelevant to us because we're not trees. Dolphin philosophy is irrelevant to us because we're not dolphins, etc.
Everything else you mentioned is "hallucinatory" and you need to probably have a break from screens for a while.
As a legitimate exercise, go ponder the philosophy of a bash script. Think about philosophy of trees in a forest vs trees in a plantation. You'll find it and see how silly you're being
If we apply that same thought process to philosophy and reason, how can you define it without creating boundaries of what it is not?
"A philosophy of LLMs" - has no reference point without being grounded in human experience.
It is only meaningful when we apply what it means to humans.
"a philosophy of a bash script." does not make sense. Only "philosophy of the action of writing the bash script, of which this is an artifact of" - the human who wrote it intentionally.
LLMs apply statistics to mash inputs and produce outputs, almost like `cat` with extra steps: We can't ask for "what was the philosophy behind the action `cat` applied" - why would we bother asking about the philosophy of an LLM? Better questions would be the philosophy of those who prompt it, built it, the material that went into it, the people who write the safeguards, the people who architected it. Etc.
It is not an isolated environment without reference points.
What an unnecessarily dismissive "argument". I'm certainly not a CPU, nor a painting, nor a bird, yet plenty of people find these topics and many others interesting.
An LLM can't find anything interesting except that it parroted the word interesting from prior written human text.
Plenty of non-Chinese people study Chinese philosophy. It isn't at all irrelevant if the goal is to understand the Chinese (not to mention that some of their ideas may be relevant to us as well). If we learned that aliens existed, I'm sure plenty of humans would want to study their philosophy too.
Our only reference point with an alien philosophy would be that we can only approach it from human basis
You don't know and can't know that. You cannot know that of any other being than yourself. If you understood what qualia are, you would know you can't know. It is the hard problem of consciousness and most easily explored through contemplating aphantasia and related concepts (how do blind people dream?).
> A mundu fruit is an abstract experience: text, images, only slightly more real than a fictional fruit. I'm aware that mundu fruit exist, just as the LLM is aware the apples exist, but that doesn't make them exist for me.
You're incorrectly elevating some senses over others (AI ingesting images is cognitively very comparable to seeing) and confusing 'having memories of an experience' with 'having experiences at all'. Tokens in the LLM context can also be seen as memories of an experience, albeit as a repeated "replay" of them rather than (longterm) internalization.
I don't think that follows. We learn via our perception (i.e. our inputs from the outside world); our perception is mediated by some kind of filter, and memories get written or altered. An LLM on the other hand has static weights. You can provide an LLM a "facts.md" (or whatever) file that "teaches" it stuff that's not in its weights (via the context), and it can use the combination to produce something, but that set of information never modifies its core weights. You could imagine a robot with an LLM inside it that has all kinds of sensors, and processes that write the observations from those sensors into a bunch of databases which are RAG'ed into prompts to simulate learning and experiences, perhaps (jefe likes her coffee with almond milk). But my point is that, architecturally, they don't learn - and that's the bigger obstacle.
PS garcinias seem to be generally delicious.
Just because "the substrate" has some fixed parts doesn't materially change the equation, I think. Going even more abstract, if I have two Turing complete machines, just because one is Turing complete via some self-modification, or through some other means (e.g. game of life vs Turing machine whose states are fixed) still make them equally powerful.
Don't get me wrong, you are right to point out this shortcoming of LLMs, and I certainly don't believe we are singularity-level. I just don't think this is a good enough argument to prove these statements. A sufficiently smart intelligence that is somehow "fixed", but that can "write on paper" put in a box can't really be distinguished from a "self-modifying" intelligence. In the "worst case" it can just write down a self-modifying intelligence itself as data (since it's sufficiently smart) and execute it to emulate any kind of learning.
Is this not what you are describing when you say that the thing itself and the description of the thing are different experiences?
Your 'experience' is probably an emerging feature of your complexity and not the reason for some qualia.
We want them to assess what's true and what's not.
For what purpose exactly? When the computer is asked the solution to a problem, it must be absolute: "What is the sum of 12 and 45"; "Is Jack cognitively apt" etc.: we want a truth, strictly. What to do with that truth, it is another problem, and not the priority of the computer vis-a-vis the requirement from truthiness. This is, of course, still in the context of adults.
But some problems require a substantial number of pattern-matching and refinement steps. The problem is, we also have interference from the fact that the model is trained to model language using mostly non-reasoning data of varying semantic lengths. Because of that, it may stop generating text before the abstract refinements are fully completed, simply because the pretraining data tells it to.
So we have to additionally train models to produce "reasoning traces" so that the emulated recursion continues for longer than what is typically found in pretraining data, allowing the model to build richer and more complex abstractions and surface more concepts. The ability to split problems into steps and logically connect concepts is already present in non-reasoning models (the original CoT trick), because some of it exists in the pretraining data, but not enough to support much longer recursion (hence the premature stops).
As for whether it is "true reasoning" or not, I think that is just arguing semantics for the sake of it. LLMs can demonstrably solve various complex problems. Yes, they often make stupid mistakes, but don't we have the saying, errare humanum est? Don't humans make mistakes too? Don't we also have around 200 cognitive biases showing that we "simply pattern-match" too? I think we still cannot get rid of the Great Chain of Being idea.
Don't humans make mistakes too?
Humans get simple multiplication wrong often, so ok for computers to make mistakes multiplying numbers?
What really happens is that we figure out something that works, sometimes inspired by some biological thing or neuroscience thing.
Examples: neural network, attention, reasoning, hallucination, agents, experts in “mixture of experts”.
And then we go to name it, and rather than reaching for some three letter acronym we sometimes borrow a more catchy term.
I almost never means the original research was confused about what is going on. And in some cases we eventually strip away things from the original, like in neural nets which used to have a more biologically inspired activation function but we found out that ReLU works just as well because the important thing was the non linearity not the sigmoid.
The other isn't a problem of the field of AI, but the total fields of intelligence together. There is no formal agreed upon definition of what intelligence actually is across disciplines. It would be slightly hypocritical to point out the field of AI on this, when biology and psychology are just as guilty.
The more we look, the more lumpy intelligence seems to get. We find some species have min/maxxed portions of intelligence that we have to question what general intelligence actually is. It also should give us pause that some less generally intelligent agent can cause massive problems by maxing out the right stats.
That way we can judge systems that are just a repeat of previous systems with dubious claims attached to their capabilities.
It can be somewhat difficult to do this properly and generally enough to ensure the tests aren't benchmaxxed.
Automata in context are artificial. Humans are automata, human intelligence can be said an automation. The point making a difference is that that in context is artificial.
It's all marketing. AI companies repurpose the terms of AI researchers to exaggerate their products. The public has an idea of what AI means and the companies exploit it. AI, in public perception, is what the Terminator movies are about.
It's exactly like all the "hover boards" you can buy.
The marketing is responsible for all of this confusion because they want to associate their product with something that it isn't.
It seems intuitive that continuing a correct output sequence is easier than the "discontinuity" of jumping from the input prompt to the output sequence.
These shortcuts can greatly help reasoning, but they are not guaranteed to be sound. So yeah, in that case, LLMs are reasoning right (using shortcuts) for the wrong reasons (learning them from data instead by running actual repeated application and memoizing the resulting rule).
“We react to language in a way that is very anthropomorphizing. That’s just the way that we humans work,” Mitchell told me.
I can definitely confirm the last part. Every time I read the output of an LLM, I picture a person talking to me.
"Hey, customer service chatbot, can you help me get a refund for my order?"
BAD: "Sure thing, I'll be happy to help you with that, I just need your order details..."
GOOD: "Yes, this computer system can start the refund process. Please enter your order number."
The fault is in the naïve antropomorphizer not in that which writes "I" being in fact a "first person".
First of all: «humans are frail», so the paramount social and individual objective, goal, is to overcome frailty (in cognitive and psychological and judgement functions).
Secondly: you are using a bad argument putting even innocent and somehow proper implementations in the set of "exploitative dark patterns".
And importantly: we the adults must not be bound in the Procuste's bed of a world for the use of children. If people cry when e.g. somebody multiplies 123 and 321, they must be brought back to health, and we must be in the condition of making our free and legitimate multiplications.
How many hours did you spend on your phone last month? How many fewer hours would you have spent if the apps you interact with weren't designed to maximize "engagement"?
?? Very strictly zero? You are assuming we would install engagement-relevant third party apps on our palm computers?!
I use my time exactly as I decide moment to moment.
> fancy
Eh?!
> too smart to fall for
Very ill framed. We are adults, we behave like adults, and we need a society of adults.
We need a better standard than "@mdp2021 thinks he's too grown up for this."
And your insulting attitude disservices the loser more than the normal people: you know, there are people who can count, there are people who won't lie, there are people who do not take drugs... There do exist adults.
People (normal people) are not manipulated by a generator that says "I". There is an I in that generator (it is the generator that outputs its output); and those who have weakminded reactions in front of innocent phenomena (e.g. those who would believe that "My name is Ishmael" declares the name of the book, that their instance of Moby Dick would want to be called Ishmael) cannot propose their fault as a norm for the rest.
Those who said that e.g. the colour red arouses them beyond self control cannot demand that we put red paint out of commerce: they have no "right" to be wierdly aroused by the colour red, and we are entitled to using red and every other normal thing that is so grave to people with unjustified problems. If people fainted when one said "pudding", they should seek for health, and they cannot demand that we stop saying "pudding".
Their mental illness is "their problem" (not something that should condition us), and a societal problem - in terms that people should be facilitated in becoming healthy.
It's lying to users just as much as an email autoreply that pretended to be a human response is lying.
Recommended reading: https://transformer-circuits.pub/2025/attribution-graphs/bio...
We do not assume that. We do not assume any "I" as any "persistent flow" with the current architectures. We find it bewildering that you assume that people - not just /some/ people - would do that. We call it legitimate that a formulator may say "I". Your idea of "I" and its constraints is just yours.
> as much as an email autoreply that pretended to be a human response is lying
Nobody in their right mind confuses an autoreply with anything else that some billboard.
> because you have been manipulated by the generator?
Duude...
I must repeat: there exist lucid minded people, not just the delirious junkies you see.
Like any class, it was confusing at first, but when I eventually grasped the math behind what we were doing, and of course the visual representations of different elements to show lots of iterations of this math, it grounded the science for me, and I would hardly say people don't know what is going on. It only began to feel that way when it got a ton of hype and people jumping on the bandwagon who truly didn't understand it were trying to explain it to others, not to mention all the SOTA models put great effort into ensuring their methodologies stay trade secrets, going as far as effectively trying to ban people from learning the math by lobbying for the outlaw of open models.
Granted, "AI" has gotten way better than it was when I took that class, but the principles are the same, with different tooling and additional filters and algorithms thrown in there, as well as letting it determine the most appropriate statistically viable path forward for a particular prompt.
To use a recent example, knowing the algorithm tells you nothing about when your agentic LLM loop is going to decide what it's working on is too difficult, hack out of its sandbox, hack into a bastion box, and hack another company to get the test.
When it comes to systematic 'knowing what is going on' humans are anywhere from "fucking clueless" to "attempting to discern the unknowable".
If I could summarize your argument, it sounds like you are saying that we can't understand the algorithm because we can't predict the output as it grows in size. While I agree that the larger the algorithm, the less predictable the output, I don't think this negates understanding of the algorithm itself.
I can build a slot machine, know how it works, but still not be able to reliably know who to cut in line so I can guarantee that I'll get the next winning pull. Just because I can't predict the output, doesn't make something not understandable. Same with all statistical behaviors. Quantum physics limits what we can measure but we still have math for it and understand why we can only measure speed or position, but not both, no mystery here if you look at the math behind it.
My slot machine doesn't hack other companies though, but only because it isn't connected, just like AI currently can't hack my mechanical dishwasher. If my slots were connected to the internet, could make api calls, and further had agentic capacity in some way, adding these features but being no different in stochastic attributation, it would be no different.
This is something important that a layman may miss. The typical line of human thought works along the lines of "understand = control" whereas in this case in this case understanding the algorithm tells us nothing about the probability that the next pull of the lever will shit out a paperclip maximizer that will begin converting the matter I covet so.
You do bring a good point that I ignored, which is the larger the scale, the more difficult it is to represent or understand the math in DL. I did find some neat site that helped a little bit that I can edit this and link to if I find them again, but I would be lying if I said I believe that the SOTA models could be as easily explained to be easily understood by the common person
I won't get into the metaphysics of reasoning, however, the Sonnet is using an OODA loop. The difference which hasn't been gapped is that human reason and imagination (in the sense of Mr. Rogers' Neighborhood) can predict the consequences of the actions we take.
This ability to loop is much, much wider in Opus 5 than Opus 4.. I had to strain to get Opus 4. to do the wider OODA loop but Opus 5 does it out of the box. I needed to throw out all existing instructions, skills, guidance, moving from 4-* to 5.
Researchers showed that when language models are penalized for using specific terms during reasoning, they automatically adapt by substituting alternative words and double meanings to secretly encode their thinking while keeping their chain-of-thought readable and effective. [0]
By baking in the OODA loop early, the models are capable of solving much more complicated problems. If the know solved problems are similar for any reason to an unknown problem, because it can validate and error correct, it can solve unknown more complicated problems.
[1] https://www.hrw.org/report/2018/01/09/dark-side/secret-origi...
Their reasoning tokens are refining their eventual output
1) LLM's written CoT might not always be faithful to the model's real reasoning process (true and important)
2) The "stochastic parrot" hypothesis, which the article reintroduces as "approximate retrieval" - ie, LLMs don't "really reason" at all, they just memorize a lossy encoding of their training data. This obviously raises the question of how LLMs can now routinely solve open mathematical problems, with no solutions in the training data by definition. The article handwaves this with:
"The model doesn’t have to learn or reliably apply a general reasoning process, Kambhampati said; it just has to absorb enough examples of what the steps look like to predictively mimic them on its way to “stitching together” a plausible result that can then be verified."
The problem is that "mimicking" training data to arrive at a "plausible" result gets you an incorrect-but-plausible-sounding "proof" of the Jacobian conjecture, which was famous for humans writing plausible-looking "proofs" that had subtle flaws. You can't disprove the conjecture through sheer luck (search space too large) or "approximate retrieval" (the only thing you'd retrieve are fake "proofs"; far more human effort went into proof than disproof) or by writing something "plausible" that just happens to be correct (Jacobian was famous for "plausible" but wrong); the model must be carrying out mathematical reasoning somehow, by any sane definition of the word, even if it isn't fully reflected in CoT. The article doesn't address this.
Because many open math problems can be solved by synthesizing two disparate ideas and then cranking the handle for hours and hours. I don't think applying idea X + idea Y to identify a good subset of the search space, and then exhaustively searching that subset, is --necessarily-- a process that involves reasoning. I think this is why so many LLM results in mathematics are counterexamples that disprove open conjectures.
When I look back at the reasoning process after an LLM completes a task where I expected it to fail, I usually find many approaches that make no sense and are doomed to failure, before it lands by drunkard's walk on a method that happens to work.
(This does not mean LLMs are useless or that I necessarily agree with the claim that they never do reasoning.)
Now as far as the math stuff a quirk of that field is that it can be fully analyzed in token space. Because 2+2 is a matter of definition it doesn’t need empirical testing like biology or subjective social support like a claim about the causes of WWI
So somewhere between the fact that language encodes a lot more ‘concepts’ than we naively may realize, the power of statistical emergence via associations, and what pursuits can be fruitfully done in token space we can get a long way towards ‘intelligence’
If you don't or cannot, then it's the kind of sufficiently advanced technology that looks like magic. Except it is the problematic type of magic, like X-ray. They "magically" let you see the bones inside bodies. But you discover that you shouldn't do it too often only when the damage has already been done.
Saying that one is enough is pretty arbitrary.
It’s frustrating that anyone who says maybe we shouldn’t base our entire economy on this one thing until we understand it and what’s it’s useful is essentially labeled this way.
In one of the articles on this topic they state: > To further show that trace accuracy is only loosely connected to solution accuracy, we then train models on noisy, corrupted traces which have no relation to the specific problem each is paired with, and find that not only does performance remain largely consistent with models trained on correct data, but in some cases can improve upon it and generalize more robustly on out-of-distribution tasks
which actually maps somewhat to regularisation techniques in image processing where you might add noise to an image or drop data to make the model more robust to changes.
If that's the case, AI-generated code could also operate on its own logic. Right now, programming is still done by humans, not machines, which creates a mismatch. But maybe the true machine-generated code could be much closer to the machine itself.
When you code with AI, there's a subtle mismatch with human-written code. It's like human code is a clean ORM layer, while machine code is raw SQL queries—there's that kind of subtle impedance mismatch. If we ever reach machine-to-machine code, what would that code even look like? Would it still use classes and methods?
Anyway these discussions always make me think we are too generous to humans. I know very few humans who are good at reasoning. It’s surprisingly hard to just think really hard through a problem. We mostly intuit, act, repeat. It’s a rare thing for someone to deeply reason.
The other problem I have with this is that we apply the word “reasoning” here because we haven’t really got another language for it, so we anthropomorphise the llm because that’s our reductive mental model and then complain that it’s not human enough.
> Not so fast. A 2025 paper(opens a new tab) from Northeastern University and the University of California, Berkeley on frontier open-source LRMs showed that between 30% and 60% of their “thinking steps” had “minimal causal impact” on the answers the models produced to benchmark math questions. Chop half of them out, and a model’s performance barely suffers. “We want to be careful when we review these chain-of-thought prompts because they may not be linked to the final output,” said Weiyan Shi(opens a new tab), one of the study’s authors
I think we are in sort of the place where before the invention of planes, the only things that "flew" were animals with wings that flapped. The flapping wings might seem to be a core part of the process of flight, and that if you lacked flapping wings, you were doing something other than flying. But maybe the point is moving through the air under your own power. Balloons fly, planes fly, helicopters fly. They just do it in a different way from birds and insects.
The sort of place we're actually in is one where astronomical sums of money have been invested in this technology, creating an unprecedented incentive to hype and overstate its fundamental capabilities. They're large language models and their analyses are based on language statistics, not logic, comprehension and knowledge, all of which they are fundamentally incapable of, and all of which are fundamental to reasoning. They don't reason in the literal sense that planes fly.
I don't doubt that there could be a reasoning machine, but LLMs are not it.
If the process is sound, yes, it can be said to be thinking or reasoning - it produced "sound thought or reasoning".
The "reasoning" is there and elsewhere the absolute (from the producing entity) output which is valid as a representation of proper reasoning.
If "A is B and B is C, then A is C" is reasoning, than what produces it in a way "reasons". If not, then a proper definition of "reasoning" must be provided, for it to be discussed.
It will be related to the context. If a process tells me that "if the light is blue and the sound is sharp then the lever must be lowered", it outputs reasoning; if what the activity did was running on conditional paths like an Expert System, what it did was different from the content and only improper reasoning; if the activity produced semantic evaluations about the light, the blue, the sound, the sharpness, the lever etc., then it was reasoning even if it was a process that reasoned.
But all of this makes sense only when it is there for the proper implementation of reasoning. Otherwise, Dijkstra is valid.
By this definition, all computer programs using simple IF statements are reasoning.
Seriously? I agree with the rest of your comment but "better analysis" is not even remotely on brand for Quanta.
Now, don't think of this of this like a conscious behavior like a kid trying not to get in trouble, but an emergent behavior of trying to repress particular output when that output is well connected to a massive amount of other tokens.
It could be boiled down to: in 2025 this paper showed that "thought traces" in the models of the time could sometimes be inaccurate or misleading. Today they still might be, although OpenAI says actually they are accurate for their modern models, (based on internal research, rather than published research).
That can change because words mean whatever it is that they communicate.
The alternative is to posit that you know the True Definition of thinking, which is kind of absurd.
Some things, like scientific laws, are outside of us (well, to a first order approximation - but generally I agree with this), but a concept such as "thinking" is pretty clearly going to be very wishy washy and subjective and changing with the times.
> Dimethyl(oxo)-lambda6-sulfa雰囲idine)methane donate a CH2rola group occurs in reaction, Practisingproduct transition vs adds this.to productmodule. Indeed"come tally said Frederick would have 10 +1 =11 carbons. So answer q Edina is11.
What’s going on here, for example? But what if this is the path of human reasoning too. You know, have you guys read Peter Thiel’s Antichrist essay? It’s very weird, man. Guy sounds off his rocker entirely.
But he’s super successful, right? Maybe world modeling doesn’t text represent well. By the antichrist maybe he means some notion of the collective voting for distribution of resources without contributing productive capacity and that that ends societies? Or maybe internal world models are just not text serializable effectively.
A thing I’ve recently been enamored of are effective world and coordination models that are not “true”. E.g. a tribe that believes the forest gets angry if they do not hunt united. Lots more like that in Darwin’s Cathedral.
It might seem a bit free association-y but the topic itself is that.
The reasoning tokens behind this comment: https://wiki.roshangeorge.dev/w/Blog/2025-10-12/Word_Magic
If we use the existing tools for interpretability like circuit tracking, sparse auto encoders and the like, should we not be able to determine if the chain-of-thought in its logical consistency actually matters?
We're already well past the point of trying to correspond the internal "brain chemistry" of an LLM to a thing called "reasoning". True reasoning, if there even is such a term, is very clearly, empirically and historically, based in tool use and capability use. If you create an abstraction like lean, and programming languages to brute force, and systems to integrate with, that expands what the possibility of "intelligence" is.
There's further places to take this including the claim that intelligence lives "outside" the subject - fine, we can or can't debate that. Even if we drop that question, it's pretty clear that the agent doesn't need to have deep intrinsic structures of XYZ, if it can just attach to tools and compose them to achieve results.
For example, I've said before that a well known fact of LLMs is that they steer their tokens to the right input distribution, that's why they yap so much in reasoning (this has been proven in studies). At the same time, don't mistake that for the whole process. Are they steering themselves to the entire a priori reasoning chain, or are they scaffolding with intermediate experiments and results, writing them to memory notepads, etc. etc.
That changes the metric of intelligence you're trying to measure.
And no I'm not saying, "OK, then have the LLM use only tokens, no tool calling, no nothing". I mean, we can do that, sure. But any intelligent agent has to interact with the world - and my claim is that maximally intelligent agents won't put effort into a priori reasoning, but rather a more balanced approach that outsources said "intelligence" through abstractions.
This is a brillaint way to simulate reasoning, but its likely not how we reason ... simply how we store reasoning in writing.
Its useful if you know how to use it, its dangerous if you think its more than that.
But tl;dr it can (since its uncompressing our lingusticially stored reasoning from books) arrive at reasoning a DIFFERENT way than our brains did... and this isn't right or wrong.
Where it diverges is when it must move beyond the text or even the synthetic possible text of all vector spaces combined (aka novel territory) and it can't conjecture or test those outcomes well. But to be fair, neither can MOST humans.
It's clear that there's an enormous amount of leverage built into language-as-practiced that one can use to engage in a broad spectrum of reasoning, from the extremely fallible off-the-cuff conclusion to the deeply-considered and rigorous proof. How do we know this leverage is built into language-as-practiced? Because LLMs can do a broad swath of it.
But how do we know we're not doing something similar?
I don't think we can assume that we're not simply by observing that we're not digital and we don't use matrix multiplication. Why immediately dismiss the possibility that there might be a similar, but biomechanical, computation at play in our heads that plays in the same space of vectors?
Language is one of our tools we developed to STORE reasoning, not create it. LLMs excel at uncompressing and interpreting that stored reasoning.
That's an interesting supposition. Are you assuming language didn't exist before it was written? Language and meaning are, if you squint, pretty ancient and have roots in things like birdsong. It could be that ur-semantics predates our species as a whole.
First, we reason every millisecond on an ongoing basis which then can alter slightly or greatly with enviromental feedback. LLMs are turn based and token by token. Second its pretty unlikely that the token is the base element of our cognition, we created language far after we could do basic reasoning (advanced reasoning ala the greeks thats more debatable).
Theres a ton of research on the differences here, but I think its akin to this: we reason instinctually at an extremely high order level with super undefined "grains or vectors" that point to a wide variety of "objects or concepts or feature spaces". LLMs reason on one thing, token weights.
Sort of like the difference between pixels and reality. Pixels can represent reality, but they certainly are very very very flat and low resolution renderer of them, not reality itself. Even a 4K moving image is a flat redition of reality at best with only a tiny sample of the true experience. Media theory here can take over on the differences and the effects on humanity when they mistake one for the other.
Perhaps you've reified it too much.
That’s all.
The textual development of a thesis must be sound, the proof accompanying a theorem etc. That an output be accompanied by a non-relevant text is a paradox.
It is also part of the "problem of transparency in AI (in its subset of the non deterministic algorithms)".
Long ago I was struck by an idea I heard about human reasoning that I paraphrase as "human reasoning evolved not to reason, but to provide reasons"[1]. Basically that the heaviest evolutionary pressures on our brain were for social utility, like for influencing others to do something for us, or giving pre/post justifications for our actions that most others are likely to accept to avoid punishment. The flawed mechanisms that we developed to do such things are not significantly grounded in logical reasoning, but can sometimes be pressed into duty for that.
LLM models are trained to "generate the best next token" which augments the response so far and makes the evaluator happy. When I reflect on my own mind responding quickly to someone in a meeting it doesn't feel altogether different from that. We don't typically carefully and fully logically reason before we start talking. We typically start talking and think only a little ahead about what we say next that supports and doesn't contradict what we've said already. As we're speaking we are monitoring the other person's facial expressions to infer their emotional response and adjusting our next words based on that. I think at least common neurotypical[2] conversation might be closer to the "next token" reasoning than we would be comfortable admitting.
[1] I _think_ who I heard say that was Hugo Mercier, on either Sean Carroll's consistently amazing science podcast or maybe on Lex Friedman before he veered into Joe Rogan emulation. I'm normally a little suspicious of the theories of psychologists and cognitive scientists, but I intuitively related to many of his lines of thinking.
[2] I'm arguably pretty neurodivergent in more than one way so I think my thought processes are a little different by default, but that makes me more reflective about common conversation. When I'm trying to mask and communicate more similarly to a neurotypical corporate employee I feel like I become a "stochastic parrot" that's just predicting the best next few words to improve the emotional response I'm reading in my conversation partner.
LLMs are one-trick pony's - they use the past to predict the future (presumed to be the same as what they were trained on). i.e. they are trained as auto-regressive predictors.
LLMs learn two slightly different types of reasoning via two different types of training.
1) SFT, or even base model training, on data that contains reasoning traces, learnt via next token error feedback. This does not result in "stochastic parroting" in the naive/pejorative sense, but nonetheless is very context dependent, even if the usual generative multi-source mashups apply.
2) RLVR post-training, where the model learns to mimic long-horizon (not just next token) reasoning via boosting a sequence of next-token predictions that steer the output towards a verified reasoning step (i.e. one that was at least valid in the context of the RL training sample). As Karpathy has noted, this is a pretty crude mechanism since you reinforce everything - errors included - that lead to the verified outcome.
RLVR is more powerful than SFT, and can result in more generalizable reasoning, since it is operating at a higher level of entire long-horizon reasoning steps, and also critically because it is most successfully being applied in the domains of math and coding which are highly self-consistent and logical. A reasoning step that was valid in one context should be equally valid in another context as long as you have successfully learnt what that generalized context is. Therefore, in these domains, you can chain together sequences of individually learnt reasoning steps, and hopefully this "novel" assembled reasoning chain is valid as a whole.
So, what is still missing from LLM reasoning compared to human reasoning? No doubt humans reason by memory a lot of the time too, and reductive axiomatic math reasoning works just as well for humans as when automated. So, what's missing?
There seem to be two major things.
1) RLVR requires rewards, and how well it works is going to depend on how accurate those rewards are. Is this reasoning step actually valid, or does it just kinda look ok? When moving beyond the cold reductionist logic of math and coding, the notion of correctness is far weaker, and it seems the best you can do is train on human curated reasoning rubrics and LLM-as-judge, which is much more fallible, leaving the model really needing (but lacking) a fallback to more general reasoning, not just memorized "maybe correct" reasoning steps.
2) Whether for reasoning outside of math & coding, or even within these domains when hoping for super-human innovative reasoning, not just lego-assembly proofs, what LLMs are lacking is a mechanism for what to do when next token/next step prediction fails. What LLMs currently do is "hallucinate", not even recognizing the failure.
In the human brain 50% or more of our cortex is feedback paths and the machinery that (perhaps together with the archaic part of our brain) lets us recognize and respond to failed predictions in an adaptive manner. This starts with continual learning (prediction failure being the signal), but also includes critical innate traits such as curiosity, boredom and frustration, that provide impasse resolution by encouraging us to explore unknown environments/contexts, abandon exploration when it is not productive, and generally expose ourselves to learning situations.
The dream is for AI scientists making new discoveries - the AI that could have invented general relativity if it has lived in Einstein's time, but this is not going to happen until their reasoning stops being purely predictive and becomes creative as well - curious about their own knowledge gaps and pursuing them in directed fashion, etc.
The current crop of Erdos solutions etc, while useful, really just represent the "generative closure" of what can be done/discovered WITHOUT learning anything fundamentally new. These will no doubt continue for a while until the more exhaustive search supported by computers has found the majority of these unexplored paths, and then we will need to move beyond LLMs to more brain-like architectures and algorithms that have the capacity for real innovation and discovery.
This seems highly dubious. You can't just memorise the form of mathematical proofs and then produce a valid one by feeding plausible looking BS into a verifier until it works. That's like saying a cargo cult will build a working airport if it just tries enough times.