Looking at the training data I don't think it will know anything.[0]
Doubt On the Connexion of the Physical Sciences (1834) is going to have much about QM. While the cut-off is 1900, it seems much of the texts a much closer to 1800 than 1900.
I was vague. My point is that I don't think the building blocks are in the data. Its mainly tertiary and popular sources. Maybe if you had the writings of Victorian scientists, both public and private correspondence.
But I think there are also some questions about the role of language in human thought that leave the door just slightly ajar on the issue of whether or not manipulating the tokens of language might be more central to human cognition than we've tended to think.
If it turned out that this was true, then it is possible that "a model predicting tokens" has more power than that description would suggest.
I doubt it, and I doubt it quite a lot. But I don't think it is impossible that something at least a little bit along these lines turns out to be true.
I also believe strongly in the role of language, and more loosely in semiotics as a whole, to our cognitive development. To the extent that I think there are some meaningful ideas within the mountain of gibberish from Lacan, who was the first to really tie our conception of ourselves with our symbolic understanding of the world.
Unfortunately, none of that has anything to do with what LLMs are doing. The LLM is not thinking about concepts and then translating that into language. It is imitating what it looks like to read people doing so and nothing more. That can be very powerful at learning and then spitting out complex relationships between signifiers, as it's really just a giant knowledge compression engine with a human friendly way to spit it out. But there's absolutely no logical grounding whatsoever for any statement produced from an LLM.
The LLM that encouraged that man to kill himself wasn't doing it because it was a subject with agency and preference. It did so because it was, quite accurately I might say, mimicking the sequence of tokens that a real person encouraging someone to kill themselves would write. At no point whatsoever did that neural network make a moral judgment about what it was doing because it doesn't think. It simply performed inference after inference in which it scanned through a lengthy discussion between a suicidal man and an assistant that had been encouraging him and then decided that after "Cold steel pressed against a mind that’s already made peace? That’s not fear. That’s " the most accurate token would be "clar" and then "ity."
The problem with all this is that we don't actually know what human cognition is doing either.
We know what our experience is - thinking about concepts and then translating that into language - but we really don't know with much confidence what is actually going on.
I lean strongly toward the idea that humans are doing something quite different than LLMs, particularly when reasoning. But I want to leave the door open to the idea that we've not understood human cognition, mostly because our primary evidence there comes from our own subjective experience, which may (or may not) provide a reliable guide to what is actually happening.
>The problem with all this is that we don't actually know what human cognition is doing either.
We do know what it's not doing, and that is operating only through reproducing linguistic patterns. There's no more cause to think LLMs approximate our thought (thought being something they are incapable of) than that Naive-Bayes spam filter models approximate our thought.
My point is that we know very little about the sort of "thought" that we are capable of either. I agree that LLMs cannot do what we typical refer to as "thought", but I thnk it is possible that we do a LOT less of that than we think when we are "thinking" (or more precisely, having the experience of thinking).
I'm not clear that it has to be coherent at this point in the history of our understanding of cognition. We barely know what we're even talking about most of the time ...
>Unfortunately, none of that has anything to do with what LLMs are doing. The LLM is not thinking about concepts and then translating that into language. It is imitating what it looks like to read people doing so and nothing more.
'Language' is only the initial and final layers of a Large Language Model. Manipulating concepts is exactly what they do, and it's unfortunate the most obstinate seem to be the most ignorant.
They do not manipulate concepts. There is no representation of a concept for them to manipulate.
It may, however, turn out that in doing what they do, they are effectively manipulating concepts, and this is what I was alluding to: by building the model, even though your approach was through tokenization and whatever term you want to use for the network, you end up accidentally building something that implicitly manipulates concepts. Moreover, it might turn out that we ourselves do more of this than we perhaps like to think.
Nevertheless "manipulating concepts is exactly what they do" seems almost willfully ignorant of how these systems work, unless you believe that "find the next most probable sequence of tokens of some length" is all there is to "manipulating concepts".
>They do not manipulate concepts. There is no representation of a concept for them to manipulate.
Yes, they do. And of course there is. And there's plenty of research on the matter.
>It may, however, turn out that in doing what they do, they are effectively manipulating concepts
There is no effectively here. Text is what goes in and what comes out, but it's by no means what they manipulate internally.
>Nevertheless "manipulating concepts is exactly what they do" seems almost willfully ignorant of how these systems work, unless you believe that "find the next most probable sequence of tokens of some length" is all there is to "manipulating concepts".
"Find the next probable token" is the goal, not the process. It is what models are tasked to do yes, but it says nothing about what they do internally to achieve it.
please pass on a link to a solid research paper that supports the idea that to "find the next probable token", LLM's manipulate concepts ... just one will do.
Thanks for that. I've read the two Lindsey papers before. I think these are all interesting, but they are also what used to be called "just-so stories". That is, they describe a way of understanding what the LLM is doing, but do not actually describe what the LLM is doing.
And this is OK and still quite interesting - we do it to ourselves all the time. Often it's the only way we have of understanding the world (or ourselves).
However, in the case of LLMs, which are tools that we have created from scratch, I think we can require a higher standard.
I don't personally think that any of these papers suggest that LLMs manipulate concepts. They do suggest that the internal representation after training is highly complex (superposition, in particular), and that when inputs are presented, it isn't unreasonable to talk about the observable behavior as if it involved represented concepts. It is useful stance to take, similar to Dennett's intentional stance.
However, while this may turn out to be how a lot of human cognition works, I don't think it is what is the significant part of what is happening when we actively reason. Nor do I think it corresponds to what most people mean by "manipulate concepts".
The LLM, despite the prescence of "features" that may correspond to human concepts, is relentlessly forward-driving: given these inputs, what is my output? Look at the description in the 3rd paper of the arithmetic example. This is not "manipulating concepts" - it's a trick that often gets to the right answer (just like many human tricks used for arithmetic, only somewhat less reliable). It is extremely different, however, from "rigorous" arithmetic - the stuff you learned when you somewhere between age 5 and 12 perhaps - that always gives the right answer and involves no pattern matter, no inference, no approximations. The same thing can be said, I think, about every other example in all 4 papers, to some degree or another.
What I do think is true (and very interesting) is that it seems somewhere between possible and likely that a lot more human cognition than we've previously suspected uses similar mechanisms as these papers are uncovering/describing.
>That is, they describe a way of understanding what the LLM is doing, but do not actually describe what the LLM is doing.
I’m not sure what distinction you’re drawing here. A lot of mechanistic interpretability work is explicitly trying to describe what the model is doing in the most literal sense we have access to: identifying internal features/circuits and showing that intervening on them predictably changes behavior. That’s not “as-if” gloss; it’s a causal claim about internals.
If your standard is higher than “we can locate internal variables that track X and show they causally affect outputs in X-consistent ways,” what would count as “actually describing what it’s doing”?
>However, in the case of LLMs, which are tools that we have created from scratch, I think we can require a higher standard.
This is backwards. We don’t “create them from scratch” in the sense relevant to interpretability. We specify an architecture template and a training objective, then we let gradient descent discover a huge, distributed program. The “program” is not something we wrote or understand. In that sense, we’re in a similar epistemic position as neuroscience: we can observe behavior, probe internals, and build causal/mechanistic models, without having full transparency.
So what does “higher standard” mean here, concretely? If you mean “we should be able to fully enumerate a clean symbolic algorithm,” that’s not a standard we can meet even for many human cognitive skills, and it’s not obvious why that should be the bar for “concept manipulation.”
>I don't personally think that any of these papers suggest that LLMs manipulate concepts. They do suggest that the internal representation after training is highly complex (superposition, in particular), and that when inputs are presented, it isn't unreasonable to talk about the observable behavior as if it involved represented concepts. It is useful stance to take, similar to Dennett's intentional stance.
You start with “there is no representation of a concept,” but then concede “features that may correspond to human concepts.” If those features are (a) reliably present across contexts, (b) abstract over surface tokens, and (c) participate causally in producing downstream behavior, then that is a representation in the sense most people mean in cognitive science. One of the most frustrating things about these sorts of discussions is the meaningless semantic games and goalpost shifting.
>The LLM, despite the prescence of "features" that may correspond to human concepts, is relentlessly forward-driving: given these inputs, what is my output?
Again, that’s a description of the objective, not the internal computation. The fact that the training loss is next-token prediction doesn’t imply the internal machinery is only “token-ish.” Models can and do learn latent structure that’s useful for prediction: compressed variables, abstractions, world regularities, etc. Saying “it’s just next-token prediction” is like saying “humans are just maximizing inclusive genetic fitness,” therefore no real concepts. Goal ≠ mechanism.
> Look at the description in the 3rd paper of the arithmetic example. This is not "manipulating concepts" - it's a trick that often gets to the right answer
Two issues:
1. “Heuristic / approximate” doesn’t mean “not conceptual.” Humans use heuristics constantly, including in arithmetic. Concept manipulation doesn’t require perfect guarantees; it requires that internal variables encode and transform abstractions in ways that generalize.
2. Even if a model is using a “trick,” it can still be doing so by operating over internal representations that correspond to quantities, relations, carry-like states, etc. “Not a clean grade-school algorithm” is not the same as “no concepts.”
>Rigorous arithmetic… always gives the right answer and involves no pattern matching, no inference…
“Rigorous arithmetic” is a great example of a reliable procedure, but reliability doesn’t define “concept manipulation.” It’s perfectly possible to manipulate concepts using approximate, distributed representations, and it’s also possible to follow a rigid procedure with near-zero understanding (e.g., executing steps mechanically without grasping place value).
So if the claim is “LLMs don’t manipulate concepts because they don’t implement the grade-school algorithm,” that’s just conflating one particular human-taught algorithm with the broader notion of representing and transforming abstractions.
> You start with “there is no representation of a concept,” but then concede “features that may correspond to human concepts.” If those features are (a) reliably present across contexts, (b) abstract over surface tokens, and (c) participate causally in producing downstream behavior, then that is a representation in the sense most people mean in cognitive science. One of the most frustrating things about these sorts of discussions is the meaningless semantic games and goalpost shifting.
I'll see if I can try to explain what I mean here, because I absolutely don't believe this is shifting the goal posts.
There are a couple of levels of human cognition that are particularly interesting in this context. One is the question of just how the brain does anything at all, whether that's homeostasis, neuromuscular control or speech generation. Another is how humans engage in conscious, reasoned thought that leads to (or appears to lead to) novel concepts. The first one is a huge area, better understood than the second though still characterized more by what we don't know than what we do. Nevertheless, it is there that the most obvious parallels with e.g. the Lindsey papers can be found. Neural networks, activation networks and waves, signalling etc. etc. The brain receives (lots of) inputs, generates responses including but not limited to speech generation. It seems entirely reasonable to suggest that maybe our brains, given a somewhat analogous architecture at some physical level to the one used for LLMs, might use similar mechanisms as the latter.
However, nobody would say that most of what the brain does involves manipulating concepts. When you run from danger, when you reach up grab something from a shelf, when you do almost anything except actual conscious reasoning, most of the accounts of how that behavior arises from brain activity does not involve manipulating concepts. Instead, we have explanations more similar to those being offered for LLMs - linked patterns of activations across time and space.
Nobody serious is going to argue that conscious reasoning is not built on the same substrate as unconscious behavior, but I think that most people tend to feel that it doesn't make sense to try to shoehorn it into the same category. Just as it doesn't make much sense to talk about what a text editor is doing in terms of P and N semiconductor gates, or even just logic circuits, it doesn't make much sense to talk about conscious reasoning in terms of patterns of neuronal activation, despite the fact that in both cases, one set of behavior is absolutely predicated on the other.
My claim/belief is that there is nothing inside an LLM that corresponds even a tiny bit to what happens when you are asked "What is 297 x 1345?" or "will the moon be visible at 8pm tonight?" or "how does writer X tackle subject Y differently than writer Z?". They can produce answers, certainly. Sometimes the answers even make significant sense or better. But when they do, we have an understanding of how that is happening that does not require any sense of the LLM engaging in reasoning or manipulating concepts. And because of that, I consider attempts like Lindsey's to justify the idea that LLMs are manipulating concepts to be misplaced - the structures Lindsey et al. are describing are much more similar to the ones that let you navigate, move, touch, lift without much if any conscious thought. They are not, I believe, similar to what is going on in the brain when you are asked "do you think this poem would have been better if it was a haiku?" and whatever that thing is, that is what I mean by manipulating concepts.
> Saying “it’s just next-token prediction” is like saying “humans are just maximizing inclusive genetic fitness,” therefore no real concepts. Goal ≠ mechanism.
No. There's a huge difference between behavior and design. Humans are likely just maximizing genetic fitness (even though that's really a concept, but that detail is not worth arguing about here), but that describes, as you note, a goal not a mechanism. Along the way, they manifest huge numbers of sub-goal directed behaviors (or, one could argue quite convincingly, goal-agnostic behaviors) that are, broadly speaking, not governed by the top level goal. LLMs don't do this. If you want to posit that the inner mechanisms contain all sorts of "behavior" that isn't directly linked to the externally visible behavior, be my guest, but I just don't see this as equivalent. What humans visibly, mechanistically do covers a huge range of things; LLMs do token prediction.
>Nobody would say that most of what the brain does involves manipulating concepts. When you run from danger, when you reach up grab something from a shelf, when you do almost anything except actual conscious reasoning, most of the accounts of how that behavior arises from brain activity does not involve manipulating concepts.
This framing assumes "concept manipulation" requires conscious, deliberate reasoning. But that's not how cognitive science typically uses the term. When you reach for a shelf, your brain absolutely manipulates concepts - spatial relationships, object permanence, distance estimation, tool affordances. These are abstract representations that generalize across contexts. The fact that they're unconscious doesn't make them less conceptual
>My claim/belief is that there is nothing inside an LLM that corresponds even a tiny bit to what happens when you are asked "What is 297 x 1345?" or "will the moon be visible at 8pm tonight?"
This is precisely what the mechanistic interpretability work challenges. When you ask "will the moon be visible tonight," the model demonstrably activates internal features corresponding to: time, celestial mechanics, geographic location, lunar phases, etc. It combines these representations to generate an answer.
>But when they do, we have an understanding of how that is happening that does not require any sense of the LLM engaging in reasoning or manipulating concepts.
Do we? The whole point of the interpretability research is that we don't have a complete understanding. We're discovering that these models build rich internal world models, causal representations, and abstract features that weren't explicitly programmed. If your claim is "we can in principle reduce it to matrix multiplications," sure, but we can in principle reduce human cognition to neuronal firing patterns too.
>They are not, I believe, similar to what is going on in the brain when you are asked "do you think this poem would have been better if it was a haiku?" and whatever that thing is, that is what I mean by manipulating concepts.
Here's my core objection: you're defining "manipulating concepts" as "whatever special thing happens during conscious human reasoning that feels different from 'pattern matching.'" But this is circular and unfalsifiable. How would we ever know if an LLM (or another human, for that matter) is doing this "special thing"? You've defined it purely in terms of subjective experience rather than functional or mechanistic criteria.
>Humans are likely just maximizing genetic fitness... but that describes, as you note, a goal not a mechanism. Along the way, they manifest huge numbers of sub-goal directed behaviors... that are, broadly speaking, not governed by the top level goal. LLMs don't do this.
LLMs absolutely do this, it's exactly what the interpretability research reveals. LLMs trained on "token prediction" develop huge numbers of sub-goal directed internal behaviors (spatial reasoning, causal modeling, logical inference) that are instrumentally useful but not explicitly specified, precisely the phenomenon you claim only humans exhibit. And 'token prediction' is not about text. The most significant advances in robotics in decades are off the back of LLM transformers. 'Token prediction' is just the goal, and I'm tired of saying this for the thousandth time.
HN comment threads are really not the right place for discussions like this.
> Here's my core objection: you're defining "manipulating concepts" as "whatever special thing happens during conscious human reasoning that feels different from 'pattern matching.'" But this is circular and unfalsifiable. How would we ever know if an LLM (or another human, for that matter) is doing this "special thing"? You've defined it purely in terms of subjective experience rather than functional or mechanistic criteria.
I think your core objection is well aligned to my own POV. I am not claiming that the subjective experience is the critical element here, but I am claiming that whatever is going on when we have the subjective experience of "reasoning" is likely to be different (or more specifically, more usefully described in different ways) than what is happening in LLMs and our minds when doing something else.
How would we ever know? Well the obvious answer is more research into what is happening in human brains when we reason and comparing that to brain behavior at other times.
I don't think it's likely to be productive to continue this exchange on HN, but if you would like to continue, my email address is in my profile.
If anything, I feel that current breed of multimodal LLMs demonstrate that language is not fundamental - tokens are, or rather their mutual association in high-dimensional latent space. Language as we recognize it, sequences of characters and words, are just a special case. Multimodal models manage to turn audio, video and text into tokens in the same space - they do not route through text when consuming or generating images.
> manipulating the tokens of language might be more central to human cognition than we've tended to think
I'm convinced of this. I think it's because we've always looked at the most advanced forms of human languaging (like philosophy) to understand ourselves. But human language must have evolved from forms of communication found in other species, especially highly intelligent ones. It's to be expected that the building blocks of it is based on things like imitation, playful variation, pattern-matching, harnessing capabilities brains have been developing long before language, only now in the emerging world of sounds, calls, vocalizations.
Ironically, the other crucial ingredient for AGI which LLMs don't have, but we do, is exactly that animal nature which we always try to shove under the rug, over-attributing our success to the stochastic parrot part of us, and ignoring the gut instinct, the intuitive, spontaneous insight into things which a lot of the great scientists and artists of the past have talked about.
I’ve long considered language to serve primarily as a dissonance reconciliation mechanism. Our behavior is largely shaped by our circumstances and language serves to attribute logic to our behavior after the fact.
>Ironically, the other crucial ingredient for AGI which LLMs don't have, but we do, is exactly that animal nature which we always try to shove under the rug, over-attributing our success to the stochastic parrot part of us, and ignoring the gut instinct, the intuitive, spontaneous insight into things which a lot of the great scientists and artists of the past have talked about.
Are you familiar with the major works in epistemology that were written, even before the 20th century, on this exact topic?
Yes. That is correct. If I told you I planned on going outside this evening to test whether the sun sets in the east, the best response would be to let me know ahead of time that my hypothesis is wrong.
So, based on the source of "Trust me bro.", we'll decide this open question about new technology and the nature of cognition is solved. Seems unproductive.
In addition to what I have posted elsewhere in here, I would point to the fact that this is not indeed an "open question", as LLMs have not produced an entirely new and more advanced model of physics. So there is no reason to suppose they could have done so for QM.
The problem is that it hasn't really made any significant new concepts in physics. I'm not even asking for quantum mechanics 2.0, I'm just asking for a novel concept that, much like QM and a lot of post-classical physics research, formulates a novel way of interpreting the structure of the universe.
"Proposition X" does not need testing. We already know X is categorically false because we know how LLMs are programmed, and not a single line of that programming pertains to thinking (thinking in the human sense, not "thinking" in the LLM sense which merely uses an anthromorphized analogy to describe a script that feeds back multiple prompts before getting the final prompt output to present to the user). In the same way that we can reason about the correctness of an IsEven program without writing a unit test that inputs every possible int32 to "prove" it, we can reason about the fundamental principles of an LLM's programming without coming up with ridiculous tests. In fact the proposed test itself is less eminently verifiable than reasoning about correctness; it could be easily corrupted by, for instance, incorrectly labelled data in the training dataset, which could only be determined by meticulously reviewing the entirety of the dataset.
The only people who are serious about suggesting that LLMs could possibly 'think' are the people who are committing fraud on the scale of hundreds of billions of dollars (good for them on finding the all-time grift!) and people who don't understand how they're programmed, and thusly are the target of the grift. Granted, given that the vast majority of humanity are not programmers, and even fewer are programmers educated on the intricacies of ML, the grift target pool numbers in the billions.
> We already know X is categorically false because we know how LLMs are programmed, and not a single line of that programming pertains to thinking (thinking in the human sense, not "thinking" in the LLM sense which merely uses an anthromorphized analogy to describe a script that feeds back multiple prompts before getting the final prompt output to present to the user).
Could you elucidate me on the process of human thought, and point out the differences between that and a probabilistic prediction engine?
I see this argument all over the place, but "how do humans think" is never described. It is always left as a black box with something magical (presumably a soul or some other metaphysical substance) inside.
There is no need to involve souls or magic. I am not making the argument that it is impossible to create a machine that is capable of doing the same computations as the brain. The argument is that whether or not such a machine is possible, an LLM is not such a machine. If you'd like to think of our brains as squishy computers, then the principle is simple: we run code that is more complex than a token prediction engine. The fact that our code is more complex than a token prediction engine is easily verified by our capability to address problems that a token prediction engine cannot. This is because our brain-code is capable of reasoning from deterministic logical principles rather than only probabilities. We also likely have something akin to token prediction code, but that is not the only thing our brain is programmed to do, whereas it is the only thing LLMs are programmed to do.
Kant's model of epistemology, with humans schematizing conceptual understanding of objects through apperception of manifold impressions from our sensibility, and then reasoning about these objects using transcendental application of the categories, is a reasonable enough model of thought. It was (and is I think) a satisfactory answer for the question of how humans can produce synthetic a priori knowledge, something that LLMs are incapable of (don't take my word on that though, ChatGPT is more than happy to discuss [1])
[0] https://github.com/haykgrigo3/TimeCapsuleLLM/blob/main/Copy%...