It is not all obvious that consciousness precedes language. Hellen Keller provides a vivid description of her pre-verbal mental state as something less than reflective conscious.
> Before my teacher came to me, I did not know that I am. I lived in a world that was a no-world. I cannot hope to describe adequately that unconscious, yet conscious time of nothingness. I did not know that I knew aught, or that I lived or acted or desired. I had neither will nor intellect.
There are real questions about the veracity of some of these Helen Keller quotes, and reason to believe that a lot of it was really coming from Anne Sullivan.
There are really three main observations that give pause. First, after she started working with Sullivan, Keller apparently developed a writing style and nuanced opinions reminiscent of a college-educated person virtually overnight. That’s pretty surprising for any child, let alone one who had only recently acquired language. Second, all of that disappeared overnight when Sullivan died - Keller’s writing style changed drastically, becoming much simpler, and all that eloquent insight also disappeared along with Sullivan. Finally, we’ve never since seen another Helen Keller. On the contrary, later observation of deaf-blind people who don’t acquire language until later in life is part of what inspired a variant of the critical period hypothesis that posits that people who don’t learn their first language (doesn’t matter whether it’s spoken or signed) in early childhood are never able to acquire any language to a particularly high level of proficiency.
That said, even if that Keller quote is now considered dubious, it may not be entirely off base. Supposedly people who don’t acquire language in early childhood also tend to show less capacity for abstract reasoning tasks. Which is still a far cry from the “no-consciousness” that Keller-maybe-Sullivan describes, but does still suggest that language and reasoning are mutually supportive.
In “The Origin of Consciousness” Julian James argues that human consciousness substantially predates language and even writing.
For example he contrasts the language in the Iliad (which he claims was written before the development of conscious ess) and the Odyssey (which came after).
I think this is different if you have vision though. I struggle to understand how someone can look at intelligent animals that don’t have language and not intuitively assume that they’re conscious. Yes this is a tricky problem without an answer but IMO the elements of consciousness can exist without language if visual abstractions are present.
I think one should bring into Chomsky's Universal Grammar the discussion, or at least some version of it.
If for no other reason, I believe it's quite reasonable to believe that cognition must be Turing-complete (okay, linear bounded automaton), and there must be some kind of support for recursive structures, no matter what the "primitives" of such consciousness is (either visual or linguistic).
I believe in the case of animal intelligence, they (we) have a pretty good hard-coded 3d/4d world model that ought to encode some level of recursion intrinsically (in thought, we very very often use spatial and temporal words), e.g. understanding the motion of another animal, e.g. a rodent running through a tree trunk requires imagination and future prediction already over an abstract agent. Humans just probably have more "stacks" available, and can offload them into either limited "brain memory", or very importantly drawings, text, etc, for unbounded memory.
But I'm just thinking out loud, I'm absolutely no expert on any of these topics.
You could redefine language to include mental abstractions and maybe it wouldn't be so weird.
Language is a shared medium. Maybe an intelligence is capable of building abstractions in their head without the need for say, a symbol system to hopefully communicate with species similar to itself.
You may argue that the agent needs persistent memory. Well, so what? If the world has stable regularities from the perspective of the agent - if the sun rises every day, and language is associating certain visuals with certain thoughts (it's not when viewed in its original social definition - but conduct the thought experiment with sociality removed), who's to say that you can't form a kind of language around that?
Of course, this is all armchair, and I'm not trying to separate out what "real language" is or whatever, just thought experiments.
But the original poster has a good point though, that's been proven in papers, that the tokens an LLM spits out, themselves, attract them to certain distributions, and at least a decent amount of the tokens in CoT are there solely to bring itself to the right distribution.
Animals don’t have language the same way we do, but many of them do have some kind of semi-structured communication. Cats for example, when domesticated, develop an entirely parallel method of communication specifically to interact with humans - one that is neotenous in genesis, but nonetheless absent from their feral counterparts.
While true, the connection between communication and consciousness is thin. Plenty of plants communicate with each other, but it’s seen as signaling rather than consciousness.
I believe that conscious is a gradient not a binary, and I believe that cats have some level of consciousness, but I’m not sure communication is the thing to hang that on.
Please give an example of an interaction with an animal that suggests that the animal is conscious, but not using language.
Challenge round: Please demonstrate that conscious animals are not using internal idiolect (language with 1 user) and lack the physical ability to develop a form of shared communication for that language.
Well, language is not really required for thinking and conscious experiences, a huge chunk of people don't even have an internal monologue, and some (that'd be me) don't have any "byproducts" of thought process.
Though there are recent experiments showing that unconscious brain retains language processing abilities [1], so the two might be as well independent systems
IMO it's far more likely that people don't agree on what it means to "have an internal monologue" than that people actually bifurcate that dramatically on cognition. Surely there are some people with an actual difference from baseline, but I don't think it's the near-equal split that seems to appear when you just ask people about internal monologue
This is in contrast with aphantasia (complete lack of mental imagery) which can be correlated with outwardly detectable responses. E.g. people who can imagine mental images will dilate pupils when imagining a bright light with eyes closed, while people with aphantasia will not. Not sure there's an analog for internal monologue but if anyone.
I don't think a bifurcation of cognition is necessary to explain this. You get both people who claim to always think with words (to the point where they don't understand how it's possible to think without one), and people who claim never to outside of linguistic tasks like writing (to the point where they need "internal monologue" explained to them). In my experience, if you get down to it -- most people will have experienced cases of the thought preceding the word (you can skip the word, or not, but either way it demonstrates a thought without a word) and of the word preceding the thought (people who claim not to have an internal monologue will usually still do this sometimes when writing).
It's just that subvocalization is one powerful technique for organizing thoughts, but not the only one. If you had a swimmer who only ever learned the crawl stroke (freestyle), they might wonder how someone can swim without it. The difference, of course, is that the breaststroke is a demonstrable physical act whereas cognitive techniques are not. The monologuers are just more practiced at subvocalization-based reasoning, and the non-monologuers are more practiced at other types.
It's not that hard to exert limited control over which technique you're using -- and I think there are multiple nonverbal techniques. I'd encourage people to explore the space a bit. Subvocalization-first approaches are very good for linear progressions, whereas subvocalization-last (or never) approaches are usually better for big-picture reasoning.
I still think it's far more likely that people aren't describing the same thing. People are very, very bad at interpreting and experiencing their mind's inner workings. Then getting them to describe and compare those inner workings is, IMO, largely a fool's errand. I'll wait until someone develops an external validation of this to believe it.
I say this as someone who pays a lot of attention to my mind's inner workings (meditation etc, near-complete aphantasic) and the more I pay attention to the inner workings, the less I know the answer to whether or not I have an inner monologue.
I'm very reluctant to believe people when they say they don't have an internal monologue. Studies on the matter all appear to rely on self-reported data, not objective measurements. It's been shown that electromyography can be used to measure myoelectric signals produced when a person subvocalizes their speech (i.e. an inner monologue). It works well enough to reconstruct and synthesize that speech.
So we could test this objectively, and I'm not sure why we haven't.
I'm not sure what you think all these people have to gain from tricking you. I have no internal monologue. None whatsoever. If I have something like an internal monologue I don't have any conscious access to it at all. When I want to think through a spatial type problem, I can visualize shapes, images, rotations, relationships, etc but it has no verbal component. If I want to think through a text or verbal heavy task, like perhaps planning some code, I type it out. I do have intrusive thoughts but they are in the form of memories of a thing that happened or simulations of something that could happen, I guess. I most surely do not have an internal narrator or anything that approaches any kind of inner verbal life.
Well, I can't say for the general population, but for me there's just nothing unless I actively try to think of words, sentences and other things. One moment you don't have an idea and the next you do, like a state switch.
I really feel like humans argue too much about how computer programs think and at the same time too little about how humans think. The topic is criminally underresearched and the existing research is all over the place.
You make it sound like there are no ideas until words arrive? For me, my normal ground state is full of thinking that is not linguistic. My mode switch is more like having to turn on the whole language subsystem and pay attention to it, often to the detriment of my native concentration.
When I do think with words, it is usually something I am manipulating with a purpose. It is not the thoughts themselves. I have a thought, and I attempt to render it into words.
Outside of recollection of past speech, it is extremely rare for me to feel like words arrive with a thought I wasn't already holding. Most often, it happens when making jokes and puns in real-time. I am in a mood to be playful, and may have a vague feeling that there is something funny to interject, but cannot separate it from the word combination that comes to mind.
But I can't really think and deliberate verbally. It feels like an act of puppetry.
No, I actually say that there are no words unless I think of them. Ideas appear out of nowhere and are made of nothing, like they just are integral part of my mind state
My mom has aphantasia and we were both shocked to discover it even exists.
She saw something about it online and then started asking everyone in the family if they could picture an apple in their mind. We all said - yes of course, what are you talking about?
And she said, "When you say the word apple, I think of the concept of an apple. It's juicy. It's crisp. It's red, etc, but I don't SEE an apple. I just know what it is."
She had no concept that other people literally saw the apple and I had no concept that people couldn't see the apple.
I asked her what she thought the phrase "in your minds eye" meant for her entire life and she said she thought it just meant exactly what she was describing and not literally seeing anything...
Yeah this was my experience too as a near-complete aphantasic. Doing the apple test around the dinner table was enlightening, and it also made a whole lot of other things make sense through my life as well: I have virtually no concrete memories from a first-person point of view (I remember the fact that events happened, I do not remember subjectively being there)
It's unfortunately very insulting when someone says, "hey remember when we did that very important thing together?" and my answer has always been "uhh... I remember that we did it?"
I scored a 68 on the VVIQ and I can watch my pupils dilate in response to me 'seeing' something in my mind. It doesn't have to be a bright light in my mind's eye, I can imagine something that would trigger a fight or flight response and make the same thing happen.
Yeah this is another test they use: imagine something frightening and "normal people" will get clammy hands and elevated heart rates while aphantasics will not
This is in the same vein to what I wanted to say here as well - I find it very likely that people with no internal monologue are executing similar mental processes, just without them being so “front and center” in the mind.
not everyone is aware of the "methodology", they just think it through, some however have an executive layer to cognition vs reflex.
imnot sure enough to form an opinion quite yet but thinking about thinking, modifying the methods of cognition, may be a learned skill, perhaps with a developmental window.
a comparison between extroverts and introverts [1] for perceived modality of cognition might be interesting.
it may very well be that a difference in neural development is the cause of this disparity.
1] this is about focus on internal conceptualization vs external perceptualization.
myself, i think in terms of a mono/dialogue if i am preparing to disseminate.
if i am thinking for my own benefit i percieve it as idetic imagery and run thought experiments to qualify for IRL candidacy.
Human level thinking and conscious experience aren't required for language. Dogs understand spoken words and gestures. They're also pretty good at making themselves understood through noises and body language. Humans have better hardware than dogs for making noises.
I could see language development bootstrapping abstract thinking over generations in humans.
Sure, but that's a non-sequitur. Sight and hearing are the most important human senses that practically inform consciousness in the sense of where we live, and what we can do. The other 3 (touch, smell and taste) are at best secondary and cannot provide the groundedness in our world experience - you are the easiest prey in the animal kingdom. Language is primarily a tool for communication, not thought [0], and "thought" is a slippery concept when you're lacking a guidance that informs you of the world if you can't see or hear it.
Similarly, I'm reading "Supernormal Stimuli" now, and there's a whole section on infants raised by various animals (chimps, wolves, chickens, leopards, etc, etc). When they are found and brought back to human societies, basically none of them develop full language and integrate fully.
> Similarly, I'm reading "Supernormal Stimuli" now, and there's a whole section on infants raised by various animals (chimps, wolves, chickens, leopards, etc, etc). When they are found and brought back to human societies, basically none of them develop full language and integrate fully.
Infants raised by chickens? Chickens!? That doesn't sound real.
That humans who have grown out of their optimal brain plasticity and hence never acquire the capacity for ordinary human language are not somehow "conscious" is not supported. Rather, it means that, by its very definition, we do not have the means to communicate with them in order to even ask the sorts of questions necessary to determine such a thing. The premise is unfalsifiable.
Animals communicate, and it may be primitive, but a lot of it doesn't seem to require consciousness. Should someone insist, it then depends on what you call "a word".
It is likely that language forms a sort of quanta of reasoning. It is far easier to posit a verbal hypothesis pertaining to verbal subjects and test it than to try and do it nonverbally.
Maths is probably a great example of this. Try and describe Pythagoras' Theorem without using maths notation or words. Difficult, right? Reasoning is a House of Cards, and without understanding what the card is that exercise becomes significantly more difficult.
a^2 + b^2 = c^2 is perhaps a poor example, since there are several visual proofs that don’t rely on language. I don’t think this negates your overall point fwiw.
Math is a terrible example. Euclid's "Elements" was a core book in Western math education for millennia, and it uses drawn geometric proofs. The proofs are then described in Greek, but the rigorous definitions are the pictures. A lot of mathematical fields can be reasoned about and explored by imagining shapes and objects. The end results today are written in notation because this is a standard language that helps clarify ideas.
Geometry pictures are absolutely not "rigorous definitions". For 1 thing, pictures of lines aren't finite in extent, and pictures of points are finite in extent.
Some of Euclid's proofs are wrong, because they relied on those non-rigorous picture definitions.
* Book I, Proposition 1 (constructing an equilateral triangle)
* Book I, Proposition 4 (Side-Angle-Side triangle congruence).
* Multiple theorems throughout the Elements rely on the visual "betweenness" of points.
Helen Keller was blind and deaf from infancy. What her teacher gave her was a way of systematizing what little sensory stimulation she was capable of receiving. "Language", in the topical sense of "the textual encoding of language that LLMs ingest and emit", is certainly not required for consciousness. Rather, what you need is a way to identify and predict patterns borne out of repeated observations, which is complicated by the lack of ability to observe.
LLMs as clearly doing more than just modeling words. In order to predict word placement, they need to build some kind of model, their latent space, of the types of things they are able to predict. Not really full world models yet - but they have decent "blog space" or "github project" models. And you can see this with multimodality, or non-text modality modals such as images and audio. They map from their latent spaces to the outputs. The fact that multi-modal systems can share the interior layers shows some kind of internal representation is created.
Love the analogy.
I a lot of people had their mind blown by vector space representation of words. The idea that Female + King averaged out around "Queen". Words might foundational to concepts or just really-well-designed ways to transmit them.
The assumption of this sort of argument, if I understand it correctly, is that consciousness is just an ordinary byproduct that appears (or grows gradually) somewhere on the spectrum of complexity.
If that’s true, mastering language is basically orthogonal to being conscious (as you’d maybe expect, GPT-2 was pretty good at language, but had a relatively tiny amount of “neurons”).
The same is arguably true for world modeling: a math textbook has a very complex and coherent model of a world, and is, probably, unconscious.
Then the question is: what reason do we have to assume that these models have some form of consciousness?
I don’t know why I’m consciousness, but I know that I am. Other people are outwardly very similar to me, so I assume the same is broadly true inwardly.
A rock is outwardly pretty similar to me; in that it exists in the world, has obvious physical boundaries, and is affected by the passage of time. But it is not so outwardly similar that I assume consciousness.
An LLM is also outwardly similar to me, in that it can express itself in language, and seems to ‘contain’ notions about the world. But it is again not necessarily so similar to me, outwardly, that inward similarity is obvious (to me)
When you say "we" you're referring to something besides our minds?
Because sure, when we get knocked out, we're out. That's not an observation, it's a tautology. But there's nothing indicating that whatever remains doesn't have its own kind of consciousness.
Except that the Chinese Room shows that the existence of a mapping from input to output, however emergently it might have been devised, is not alone sufficient to demonstrate understanding.
An LLM is a big equation that we solve to get textual output. If Ai proponents already believe an equation can contain consciousness, what about the Chinese Room presents a more compelling counterargument?
I'm not bringing up the Chinese Room in its original sense of asserting that a machine can't have consciousness. Indeed, I have no reason to suspect that human consciousness is irreproducible. Rather, the point here is to emphasize that our methods of determination are insufficient. A man armed with an English-to-Chinese dictionary appears to know Chinese until he's faced with an instance of linguistic ambiguity outside the scope of what a dictionary encodes; in other words, you can, with enough probing, disprove that the room understands Chinese. But how do you prove it in an affirmative sense? Let's use a simpler example: does Hans the Clever Horse understand arithmetic? We can disprove it by throwing enough arithmetic at Hans to prove that his understanding does not generalize, but if Hans did know arithmetic, how would we prove it? There are plenty of things in this category--things that we believe to some confidence level, but cannot prove--but let's frame it for what it is: belief and faith, not proof or logic. Whatever utility I may derive from LLMs, I have every reason to be skeptical of a movement of people whose awe of LLMs echoes the steadfast furor of the religious adherent.
The man in the room is comparable to a human hand, and the magic rulebook to a human brain.
This in the sense that we can easily retain human "understanding" by stripping away almost all parts of the human body or replacing them with fairly trivially made replacements, except for the brain.
In the Chinese Room the equivalent is the magic rulebook: We have no idea how to construct/replace it, yet people somehow handwave that away whilst simultaneously confidently asserting it does not understand anything.
That's irrelevant to what we're talking about, though. The point here is not to assert whether or not the room understands, but to emphasize that our methods are insufficient to demonstrate this. You could posit that the room understands, or you could posit the reverse, and neither argument can be refuted.
If our own brains were insufficient to demonstrate the fallacy of the Chinese Room; the particles, molecules, cells, synapses, and so forth of our bodies cannot be believed to have understanding, but we recognize it in the sum of the part anyway....
... we now have LLMs to even more pointedly show the deficiencies of the argument.
It does so only in the claims of its creator. Plenty of other people have pointed out fallacies in the claim. My favorite is the Dennett/Hofstadter observation that while the man in the room may not understand chinese, the room/system certainly does.
"Understanding" is your word, it doesn't appear in the source article nor the comment you're responding to. The Chinese Room argument does not attempt to show that a mapping from input to output can't implement a latent space, or that it can't implement complex models of what the language is describing. If a person can express those things explicitly in the output, or if you have to do them in order to correctly respond how a person would, then the room by definition has those capabilities.
(What's the point of the argument if it doesn't tell us anything about the capabilities or internals of an AI? I'm not sure.)
Right - the point I was refuting is that LLMs are "just statistical models of words." More is going on. Does that imply "understanding?" I don't know, I'm not sure we have a good enough definition to say. But it does mean that the models are more complex that say, markov chain graphs with corpus frequencies. It seems we are encoding data in the latent space with much higher complexity than "just words." There is higher order semantic information being captured - probably not the same has human "thoughts" - but again - also not _just words_.
This person is just making stuff up that feels true to them. Caveat lector.
There are reams of grown up scientific research and philosophy on the relationship between language, thought, neurobiology, and inference. It's nice to hear how one person feels they think, but there is no weight to it if they don't know page one of the syllabus.
I find the comment section interesting. There is so much discussion about the dichotomy of consciousness, in support and opposition as discussed by the author. Some of the "rights" and "wrongs" feel way more combative than, simply noticing where the author is currently at.
I see the temporal disconnected strong here. The author is young-ish, and appears to be exploring ideas of many that have come before him. And, that he'll likely continue to explore on his journey. You can see his thoughts and see echos of long debated thoughts about consciousness, language, meaning, bottom-Up/top-down processing, not being our thoughts, etc..
Thanks to those who are leaving their knowledge, book references, quotes, and etc., to help myself and others who are somewhere along the continuum of understanding ourselves more thoroughly!
The author is making a big mistake by making a normative claim about 'true' intelligence and consciousness.
Is there a correct way to cognize? And although he feels cognizing works in a very specific direction, his brain is basically doing a very similar guessing game on a deep level with training of pathways that started at birth. Basing the argument is a feeling about consciousness and is not convincing.
The author thinks they are describing a unique human magic, forgetting Augustine's insight that human thought doesn't precede the word, but is brought into consciousness by it.
Maybe consciousness just does not exist. Could it be that humans invented it to feel superior to animals, but it does not really exist? I wonder if we will think of the word "consciousness" in the same way we think of the word "aether" in the future.
Consciousness exists by definition. It is a subjective experience and you are experiencing it.
Or more precisely I am experiencing it, but since you are, I think, a human just like me, with a brain and biology very similar to mine, I assume that you must experience consciousness too.
And that is the big question, and one that science may never have an answer to. What is conscious and what is not? The only consciousness I know exists is mine, the rest is just guessing based on similarities. Other humans: almost certainly conscious, higher intelligence life forms: probably conscious, plants: probably not conscious, rocks: almost certainly not conscious, LLMs: ???
LLMs are unusual in that they have a very human-like behavior, more than anything non human, which would make them a good candidate for having consciousness, but on a material level, they are more like rocks, which point them as not being conscious.
There’s certainly a debate as to whether animals are conscious or not, and if we’ve just been engaging in a kind of human chauvinism all this time, but I mean consciousness definitely exists. There’s an “I” in here that perceives, and it’s reasonable to extrapolate that I’m not particularly special, that others that have the same parts as me have the same properties (consciousness).
> Maybe consciousness just does not exist. Could it be that humans invented it to feel superior to animals, but it does not really exist? I wonder if we will think of the word "consciousness" in the same way we think of the word "aether" in the future.
Correct, in the Koine Greek, logos is the word that's actually used. Logos is a bit deeper than words, and is more akin to a universal principle of order that governs the cosmos.
"In the beginning was the logos" is like saying in the beginning were the laws of physics, and God is the laws of physics and the laws of physics are God.[1]
[1] This also entirely depends on which framework you are translating from. This would be accurate from the stoic interpretation of the logos but differs from a more mystical Johannine interpretation, in which the logos is outside the physical universe, not just operating within it. Logos is like the source code of the universe.
That words are a byproduct of consciousness is a bold claim and requires some justification. I'm not implying it's backwards also for humans, but that the statement might not be true, and something else could be. Also, defining consciousness is hard.
Have you ever reasoned through a complex coding or maths problem? Tell me, how did you do that without words?
You (and TFA) are making a false dichotomy here. Yes, of course we think in images. But we also have an inner monologue that is critically important for much of our higher-level thought. How do you even write a HN comment without thinking through it in words?
I agree, and I think that's a good example. Sometimes we can hold a concept in our minds of something and cannot find the words to attach to it. The word we eventually give that concept when it's human, as in your example, is a name. This is because the concept of a personality is too complex to describe with other, preexisting words.
Still, even though I think the author is right about consciousness/concepts preceding language, I think he's wrong about the other part. The statistical models that underlie AI very much remind me of what we call a human concept. It's a statistical model of that thing, person, or idea. However, we humans seem to have this ability to create new concepts that are wildly different than our training data (what we have experienced in life), and we call this creativity. It's missing from AI for some reason. That is the part that makes me question if AI will ever match humans.
I share your positivity, however I don’t think figuring out how to market through the noise is the important part.
LLMs don’t need to think, they just need to be a more efficient data retrieval system than what already exists (and what has already been sold to the highest bidder). With LLMs, I’ve find I’ve started reaching the corners of the internet, where the true lovers of knowledge are doing the work they have always been doing. Quality does eventually speak for itself, at least once we’re able to strip away the marketing (you mentioned Apple, do you believe they are truly a cut above, or is it just hype?)
If LLMs can filter the noise, build it and they will come might return as a valid way to be.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do. The word is never the start. The word is just the skin. The idea, the consciousness, is the thing sitting under it.
In (at least once school of) Tibetan Buddhist meditation you observe the part of the mind that is producing ideas, and then suppress it in order to explore deeper layers of consciousness. The "lights on" "you" that is observing is _not_ the idea producing part of the mind, which to me at least seems to be more mechanical.
I do agree with the author of course that a concept isn't necessarily wrapped with language from the start, and can be a skin; but equally it is possible to think in a verbal mode.
As I see it, LLMs will/do develop 'introspectiveness' if it is selected for during RL. If its not delibrativly selected for (this is probably done because it helps with alignment; if the machine can align itself that's great) its probably a generally useful skill to have when doing normal tasks, https://www.anthropic.com/research/introspection
Their introspectiveness doesn't strike me as indicative of anything, since we might be just reinforcing a "plausible word salad generator for input context that leads to better solutions" or just a "post hoc rationalizer", both of which may or may not be the key mechanism behind verbal consciousness.
Writing is a shadow of thought. The better the writing, the more clearly the shadow represents the shape of the original thought. Even the best writing can never perfectly capture the original thought. Writing is one of the best tools we have to share thoughts across space and time.
I look at culture, under which I put language as a component, as the O/S of the human brain. So yes a brain with no O/S can think but it probably will never replicate, in one life time, the accumulated thinking skills of a brain given a good O/S.
How many of us would independently reinvent language with specific tenses, the diatonic scale, the number zero, perspective drawing, geometry, calculus, atomic theory, quantum mechanics etc. if left on our own from childhood?
So to me, the question becomes, can human language encode the sum of all human thought? The answer might be no. But it might be good enough to get close.
Isn't it interesting then how much of our intelligence is captured by our words/language? I would've thought to create AI you needed to replicate neurons and synapses and learning. So it's still amazing to me that statistical modelling of our words creates a Claude Opus.
Perhaps LLM's do replicate the important features of "neurons and synapses and learning" and perhaps a "statistical modelling of our words" is pretty much what our brains are doing? Most of the counter-arguments I've seen boil down to "but humans are special" which I'm not sure I find compelling.
Not saying LLM's are conscious. Just that much of our amazement about their behavior seems to say more about us realizing the things we can do are not magic, rather than them being so amazing.
Words limit our understanding of words because they are intrinsic to a particular kind of understanding. yes? no?
In "Notes on the Synthesis of Form", Christopher Alexander talks about how we can know something is wrong even when we do not know how to make something right. He was talking about this I think
=z==
as opposed to
+|+|+
or
+|+|+|
I happened upon the word "entanglement" and it seems an interesting alternative to the order that is inherent in words.
Our language is all about Actor-verb-object. Entanglement provides a fundamentally different concept. I cannot say "I drove my car on the road" with entanglement. Or at least with entanglement I can say something equally valid like "the road moved my car with me inside."
And entanglement works for the + and | things, at least for me. There is some kind of entanglement (degree) that creates a gestalt.
Reminds me of this from "Gödel, Escher, Bach: an Eternal Golden Braid":
"Let me rephrase these last couple of sentences without using the slightly
technical term "isomorphism". When a system of "meaningless" symbols has
patterns in it that accurately track, or mirror, various phenomena in the
world, then that tracking or mirroring imbues the symbols with some degree
of meaning - indeed, such tracking or mirroring is no less and no more
than what meaning is. Depending on how complex and subtle and reliable
the tracking is, different degrees of meaningfulness arise." - P-3
EDIT: I initially wrote "G.E.B." instead of "Gödel, Escher, Bach: an Eternal Golden Braid" which was too cryptic.
While I know what Douglas Hofstadter's book "Godel Escher Bach - an eternal golden braid" is, it might be worth revisiting your apparent assumption that more or less everybody on HN knows about a book from 47 years ago (regardless of its significance to those of us who have encountered it already).
Language is just a shortcut to logic. LLMs don’t seem to have the capacity to logic, they just pattern match. LLMs seem conscious because they are pattern matching against all of human written cognition.
I am pretty sure human level intelligence is emergent from language use. The more complex the language the higher the intelligence since language is itself a tool used to not only to communicate, but also classify the human experience. Without language, we are fairly primitive relying on instincts and more primitive communication. Feral children are an example of what a base state might be.
> But one thing scares me a little. What if LLMs slowly become worse? Think about it. Everything written before 2017 was made by humans. After 2017, LLMs started filling the open web with their own content. And now that same content is being fed back to train the next LLMs. An LLM can give you perfect grammar and a rich vocabulary, but it can quietly lose the real context. Right now the share of AI content is small. But what happens when it keeps growing?
Have the last-generation LLM pre-screen new training data for the next generation LLM.
> For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
Recently, I've been thinking a lot about context and whether an LLM actually creates a true representation of a concept within its vectors. And for our current usage of LLMs, does it even matter? As models become so large, will the difference even be perceivable to us?
It’s funny because, ever since Norbert Wiener’s Cybernetics was published in 1948, we’ve been trying to replicate how the human brain works. In that sense, we’ve been building increasingly powerful machines that mimic human thought. And now, the most interesting thing is something that turns that idea on its head: LLMs think in reverse.
> Now come back to my first question. An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath.
No!! This is misinformation.
An LLM predicts the next words (tokens) based on the words before it *and* on the internal representations it has learned from training.
Processing language can lead to internal representations that capture patterns and relationships; LLMs can exhibit emergent abilities, including reasoning-like (stress on "-like") behavior, even when they weren't explicitly trained for that.
Making reductionist claims about LLMs, based on next-word prediction, is similar to making such claims about biology, based on amino acids.
"An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath." At most, and I will argue for less, that is half the story. This egregious over-simplification doesn't account for training and the massive dimensional space of thought it forms. Given that frontier models likely have trillions of parameters, the combinatorial space of ideas within approaches the infinite -- hardly zero.
As others have remarked: this is an unsubstantiated claim.
I am actually using LLMs to perform controlled experiments to determine if the opposite is true, and that consciousness is a byproduct of language as Jaynes might have agreed.
> But your brain works the other way around. First there is a concept, a feeling, an image, and then the words come out to describe it.
After this I can't take the essay seriously- this sort of blanket statement about the one true hierarchy of consciousness and knowledge is BS.
While it seems to have been disproven that words in other languages cause people to speak and think differently, that also doesn't mean that words don't have any effect on the way we think, or that the concept of a word always has to come before the word itself.
The reason why we experience the uncanny valley of LLMs is because they don't represent a true consciousness, BUT it's also clear that the architecture represents certain qualities of consciousness- as the models have scaled we can see that it has some other non-word related understanding.
The evidence points to consciousness as a set of interlocking systems- attention, long term memory, short term memory, emotions, etc.
Steve Jobs didn't invent the PC. It's reductionist and offensive to single him out and thanking him for that.
Computers didn't "figure out the maths". People did the math, the computers did the calculations.
Human knowledge is nowhere complete, it's silly to think so, and by extension it's silly to think the LLM's have anything near a complete knowledge set.
They don't even have access to all of current human knowledge, much less so future knowledge. A lot of text and information is locked away from the public internet and the internet altogether.
Not all types of useful ideas for applications have been invented, that's ridiculous, there are plenty of ideas that people haven't had yet. There are however a massive amount of shitty copies of the same shitty ideas.
I'm going to go on a limb here and challenge this assumption. As a parallel thread to the point put forward by michael1999, can we not argue words manifest as the result of the maths performed within LLMs, maths that, for all intents and purposes, is their world? Because the internal maths of how LLMs work is likely to be significantly cruder than the 'real' world that defines consciousness, perhaps we are closer to the ghost in the machine than we would like to admit.
> First there is a concept, a feeling, an image, and then the words come out to describe it.
Nope. Concepts, feelings and ideas do not exist independent of expression (words, sounds, responses). Humans have the same process as LlMs have. And that's precisely why LMs are able to succeed. Similar to how camera works based on the working of an eye, or how aeroplane works based on a birds wing.
If you still believe that concepts exist independently, imagine a creature that can't express, make sounds, or respond with actions. What exactly is a "concept or feeling" to such creature? Do trees have concepts and feelings? Infact trees could, because they can respond to stimulants.
If you can't describe or perceive a thing, that thing doesn't exist in your world. It's that simple.
What you are describing as "translating thoughts into words" is actually, refining your expression. The expression already exists at the same time as the thought.
My son is insufficiently verbal (he's of the kind that communicates with quotes, but really struggles to express or understand new sentences), and I am quite sure he has concepts, feelings and ideas.
St Augustine had it right 1500 years ago: humans learn words opportunistically according to their desires and problems and existing deep understanding of the world, and (critically) humans really can't learn language without being natively fluent in great ape facial expressions, gestures, grunts, etc.
Did I not, then, as I grew out of infancy, come next to boyhood, or rather did it not come to me and succeed my infancy? My infancy did not go away (for where would it go?). It was simply no longer present; and I was no longer an infant who could not speak, but now a chattering boy. I remember this, and I have since observed how I learned to speak. My elders did not teach me words by rote, as they taught me my letters afterward. But I myself, when I was unable to communicate all I wished to say to whomever I wished by means of whimperings and grunts and various gestures of my limbs (which I used to reinforce my demands), I myself repeated the sounds already stored in my memory by the mind which thou, O my God, hadst given me. When they called some thing by name and pointed it out while they spoke, I saw it and realized that the thing they wished to indicate was called by the name they then uttered. And what they meant was made plain by the gestures of their bodies, by a kind of natural language, common to all nations, which expresses itself through changes of countenance, glances of the eye, gestures and intonations which indicate a disposition and attitude--either to seek or to possess, to reject or to avoid. So it was that by frequently hearing words, in different phrases, I gradually identified the objects which the words stood for and, having formed my mouth to repeat these signs, I was thereby able to express my will. Thus I exchanged with those about me the verbal signs by which we express our wishes and advanced deeper into the stormy fellowship of human life, depending all the while upon the authority of my parents and the behest of my elders.
Maybe it is. If you think about it hard, you will find that you cannot separate the idea of dog from the word dog. The object dog can only be understood as a perception linked to that idea.
Stop for one second and ask yourself a simple question. Where do your words come from?
When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do. The word is never the start. The word is just the skin. The idea, the consciousness, is the thing sitting under it.
Is it really? LLMs don't have words inside, for the most part they operate by applying transformations to a vector that does not contain any words at all. Words that come out of an LLM are just what sampler gives us by looking at the vector that is the result of those transformations.
Does this vector contain a world model? Some form of thoughts or reasoning to arrive at the result? That's an open question really.
Sure, rerunning that whole process for each token might not be the best solution, although that's an open question too. But saying that LLMs operate on words first is too big of an oversimplification
>Does this vector contain a world model? Some form of thoughts or reasoning to arrive at the result? That's an open question really.
It is, but it has been explored in various forms. Anthropic have some interesting papers where they can map out concepts to some extent as they exist inside the vectors of the LLM, and they can see e.g. that vectors relating to some concept can appear there some iterations before the LLM emits the tokens that spell it out. (though also, quite frankly, this shouldn't be all that surprising. If they were just picking the next token with no representation of what might follow it the result would be more like the 'one-word-at-a-time' party game than coherent text. They used poetry with a rhyming scheme as an example because it's something where you need to have some idea of what might rhyme with the line before and aim at it at the start of the next line in order to have any degree of success).
My favourite way of thinking about this is by going back to the End Poem at the end of Minecraft. The fact that words are an "interface", "very flexible" and "less terrifying than staring at the reality beyond the screen".
Maybe that reality is the true understanding of the neurobiology that defines our thoughts. To be reminded that the magic of experience can be reduced to neuronic hallucination is, frankly, horrifying. Maybe we dislike LLMs because we see their vectors and numbers as crude reminders of our own intellectual banality.
Or maybe I'm just feeling hungry and should really go eat dinner.
I don't know about you, but I think with words all the time when reasoning through complex ideas. Where do you think the phrase "thinking out loud" comes from?
> Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it
There are enough neuroscience experiments demonstrating how we create post hoc explanations of our actions that I wouldn't trust this intuition prima facie.
I've noticed that more and more of my mental attention while speaking is observing the words that come out of my mouth rather than 'generating' them. I split time between rough pathfinding in my mind to fit a conceptual framework then kind of sit and listen as it comes out to think about how they sound. If I had to guess this happens on a 1-2 sentence chunk at a time.
The fact that we're even capable of doing a study like that and introspecting is probably a good indication of our cognitive abilities as humans
Consider someone designing a skyscraper. They're thinking about material strengths and such while they draft it, they aren't just rationalizing their choice later.
What is reasoning except applying another pattern on top of existing thought? Personally it seems like I and everyone else is simply pattern matching, albeit at a higher level than current LLMs. There is no difference in the process as far as I can tell, just different inputs.
No it's down voted because we didn't stochastic parrot ourselves into landing on the moon and building the internet and modern highways and skyscrapers
In some meditation practices, the task is to observe the stochastic parrot in your head. You will soon realise that you cannot control this parrot. If you cannot control your thoughts, could it be that you are not your thoughts? And if you are not your thoughts--the word stream in your head--what are you then?
But LLMs aren't just shuffling words around. Words (tokens) are the "human interface" at the edges. Text input gets embedded, then there's this huge latent-space computation in the middle, and only at the tail end does that get converted back into word/token probabilities. So just saying that "words are the source" isn't entirely correct and feels misleading.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it?
Let's take a feral human, do they achieve the same with language powered humans? No, obviously. It means language does something we can't replace with "consciousness". Maybe something life preserving, putting consciousness downstream of language use.
But more generally - how can 8B humans make a living on this planet? Not without language, that is for sure! We long passed the stage where we could exist without language at our consumption rate. We can't even exist without math, population collapse would be the outcome.
Minds are downstream from language & math. For example, is music downstream of piano or piano downstream of music? I think you can't cut this cleanly. They developed in relation to the other, recursion between tool and art form is already old.
"At its core, it is a big pile of words that predicts the next word, using some maths the computers figured out."
"An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath. The words are everything."
This is just flat out wrong.
Words are used as training data, to build a system of vector embeddings. The LLM contains no words. That was the training data long discarded.
Vector embeddings are groupings of meanings derived from the relationship between the words in its training data. This entire system is modelled after human neural mechanisms in a way that machines can emulate.
"But your brain works the other way around. First there is a concept, a feeling, an image, and then the words come out to describe it. (At least, this is how I feel my own brain working.) For us, words are the byproduct of consciousness."
It's working the same way around (you are not saying Ai has consciousness that is derived from words!). And you make a huge jump from the concept that 'words come from concepts' to 'words are the byproduct of consciousness'. Because 'concepts' are not equal to 'consciousness'.
In fact you do not define consciousness at all, making it hard to determine what argument you are actually making at all. You seem to think humans have this trait of 'consciousness' but can't explain or evidence it beyond 'feeling it'.
Consciousness it defined by many as the ability to experience events and process thoughts or qualia. It's hard to test this, and we can see why that is with an AI - it may claim to be conscious, or even claim not to be, but how can we trust either answer? Philosophers aren't even sure to trust another human who claims to be conscious, and we struggle constantly to determine at what point creatures the animal kingdom are conscious or not. Cats? Lobsters? Snails? or even across the plants?
Before you can refute the consciousness of AI, first establish the consciousness of humans. A better question to ask is what does the uncanny emergent ability of an AI to mimic a human say about consciousness?
Heck, they even ignore the studies that show that conscious thought is a lagging indicator of 'thinking'. Much like you don't tell your heart to beat or your stomach to digest there is a fair amount of neural activity that occurs either before you're aware of it, or outside of your consciousness at all.
With any articles like this I take any claims with a grain of salt. We don't have the needed information to make any bold claims on what consciousness is other than a bunch of electrical signals we can see occur in the brain.
An approach that might shed some light is instead to define what consciousness ISN'T. Naively let us say consciousness is NOT a large list of weights (i.e. an LLM).
The uncanny emergent ability depends entirely on training data. A mathematical model is used to match output against training data (via loss functions etc). The training data contains all the human ingenuity, logic, rational, patterns and features.
Try giving an LLM model the alphabet ALONE and see what it comes up with?
I think small children learning language learn in a way similar to LLM's, just saying words that seem appropriate to the situation without any understanding of the meaning. But a key difference in my mind, is they use the reactions to this text prediction to help them construct a picture of the world, which then can interact with their thinking of what words to say.
>An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath. The words are everything. For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
I don't think that's right. What makes LLMs (and other forms of "AI") work is the very fact that they hold a statistical model underneath and put words on top of it. That statistical model, at least to me, appears to be somewhat akin to what a human would call a "concept".
Like a stable diffusion model has a statistical model of what a "chair" is. Not one exact chair, but all chairs. They all have backs, 4 legs, a flat area to sit on. It can then carve an image of a chair out of noise using this concept. Take a music model like Suno as another example. If you tell it to make a rock song, it has a statistical idea of what this would looks like. The harmonic progression, the tempo, the types of instruments used. It then attaches sounds to this statistical model and you get a "song". LLMs also appear to do this. If you ask it for a paragraph of a specific flavor of prose, it has that as a statistical model, and it attaches words to it.
I am not a big believer in LLMs delivering on even half the hype they've generated. However, this very idea of human concepts vs AI's statistical models is the one thing makes me wonder, sometimes, if they're on the right track, even if we are a long way off from AGI. It's kind of funny that the author of the above article landed on the opposite conclusion.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do.
Most people don't actively think when they are talking or writing most of the time. Only during period of thinking that we do think. Most of the time, we just talk words. There is evidence is that we just observe ourselves talking and that the thinking itself is not conscious but it's still a theory.
> Now ask the same question about an LLM. For an LLM, it is exactly the opposite.
No one has any idea how an LLM come to correct words. LLMs produce numbers (tokens) that end up to coherent words.
> So what is an LLM, really? At its core, it is a big pile of words that predicts the next word, using some maths the computers figured out.
No, not really.
> The words are everything. For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
The author provides no proof for this and there are very good reasons to think that LLMs have inherent thinking.
The rest of the article was just some random thoughts the OP jotted down? I just don't see their coherence to the beginning of the article.
No, I don't think so. Language is intimately intertwined with our understanding of the world. Many people make the mistake to think that languages are about the spelling or the sound or maybe the grammar. But in reality they are about how words are defined only on relation to each other. A symbol is defined by what it is not. This relationship between symbols structures how we think about the world and how we fantasize and desire.
You can perhaps have a consciousness without words, but not without language.
It is not all obvious that consciousness precedes language. Hellen Keller provides a vivid description of her pre-verbal mental state as something less than reflective conscious.
> Before my teacher came to me, I did not know that I am. I lived in a world that was a no-world. I cannot hope to describe adequately that unconscious, yet conscious time of nothingness. I did not know that I knew aught, or that I lived or acted or desired. I had neither will nor intellect.
https://scentofdawn.blogspot.com/2011/07/before-soul-dawn-he...
There are real questions about the veracity of some of these Helen Keller quotes, and reason to believe that a lot of it was really coming from Anne Sullivan.
There are really three main observations that give pause. First, after she started working with Sullivan, Keller apparently developed a writing style and nuanced opinions reminiscent of a college-educated person virtually overnight. That’s pretty surprising for any child, let alone one who had only recently acquired language. Second, all of that disappeared overnight when Sullivan died - Keller’s writing style changed drastically, becoming much simpler, and all that eloquent insight also disappeared along with Sullivan. Finally, we’ve never since seen another Helen Keller. On the contrary, later observation of deaf-blind people who don’t acquire language until later in life is part of what inspired a variant of the critical period hypothesis that posits that people who don’t learn their first language (doesn’t matter whether it’s spoken or signed) in early childhood are never able to acquire any language to a particularly high level of proficiency.
That said, even if that Keller quote is now considered dubious, it may not be entirely off base. Supposedly people who don’t acquire language in early childhood also tend to show less capacity for abstract reasoning tasks. Which is still a far cry from the “no-consciousness” that Keller-maybe-Sullivan describes, but does still suggest that language and reasoning are mutually supportive.
In “The Origin of Consciousness” Julian James argues that human consciousness substantially predates language and even writing.
For example he contrasts the language in the Iliad (which he claims was written before the development of conscious ess) and the Odyssey (which came after).
I think this is different if you have vision though. I struggle to understand how someone can look at intelligent animals that don’t have language and not intuitively assume that they’re conscious. Yes this is a tricky problem without an answer but IMO the elements of consciousness can exist without language if visual abstractions are present.
I think one should bring into Chomsky's Universal Grammar the discussion, or at least some version of it.
If for no other reason, I believe it's quite reasonable to believe that cognition must be Turing-complete (okay, linear bounded automaton), and there must be some kind of support for recursive structures, no matter what the "primitives" of such consciousness is (either visual or linguistic).
I believe in the case of animal intelligence, they (we) have a pretty good hard-coded 3d/4d world model that ought to encode some level of recursion intrinsically (in thought, we very very often use spatial and temporal words), e.g. understanding the motion of another animal, e.g. a rodent running through a tree trunk requires imagination and future prediction already over an abstract agent. Humans just probably have more "stacks" available, and can offload them into either limited "brain memory", or very importantly drawings, text, etc, for unbounded memory.
But I'm just thinking out loud, I'm absolutely no expert on any of these topics.
You could redefine language to include mental abstractions and maybe it wouldn't be so weird.
Language is a shared medium. Maybe an intelligence is capable of building abstractions in their head without the need for say, a symbol system to hopefully communicate with species similar to itself.
You may argue that the agent needs persistent memory. Well, so what? If the world has stable regularities from the perspective of the agent - if the sun rises every day, and language is associating certain visuals with certain thoughts (it's not when viewed in its original social definition - but conduct the thought experiment with sociality removed), who's to say that you can't form a kind of language around that?
Of course, this is all armchair, and I'm not trying to separate out what "real language" is or whatever, just thought experiments.
But the original poster has a good point though, that's been proven in papers, that the tokens an LLM spits out, themselves, attract them to certain distributions, and at least a decent amount of the tokens in CoT are there solely to bring itself to the right distribution.
Animals don’t have language the same way we do, but many of them do have some kind of semi-structured communication. Cats for example, when domesticated, develop an entirely parallel method of communication specifically to interact with humans - one that is neotenous in genesis, but nonetheless absent from their feral counterparts.
While true, the connection between communication and consciousness is thin. Plenty of plants communicate with each other, but it’s seen as signaling rather than consciousness.
I believe that conscious is a gradient not a binary, and I believe that cats have some level of consciousness, but I’m not sure communication is the thing to hang that on.
Please define "conscious".
Please give an example of an interaction with an animal that suggests that the animal is conscious, but not using language.
Challenge round: Please demonstrate that conscious animals are not using internal idiolect (language with 1 user) and lack the physical ability to develop a form of shared communication for that language.
Well, language is not really required for thinking and conscious experiences, a huge chunk of people don't even have an internal monologue, and some (that'd be me) don't have any "byproducts" of thought process.
Though there are recent experiments showing that unconscious brain retains language processing abilities [1], so the two might be as well independent systems
[1]: https://www.nature.com/articles/s41586-026-10448-0
IMO it's far more likely that people don't agree on what it means to "have an internal monologue" than that people actually bifurcate that dramatically on cognition. Surely there are some people with an actual difference from baseline, but I don't think it's the near-equal split that seems to appear when you just ask people about internal monologue
This is in contrast with aphantasia (complete lack of mental imagery) which can be correlated with outwardly detectable responses. E.g. people who can imagine mental images will dilate pupils when imagining a bright light with eyes closed, while people with aphantasia will not. Not sure there's an analog for internal monologue but if anyone.
I don't think a bifurcation of cognition is necessary to explain this. You get both people who claim to always think with words (to the point where they don't understand how it's possible to think without one), and people who claim never to outside of linguistic tasks like writing (to the point where they need "internal monologue" explained to them). In my experience, if you get down to it -- most people will have experienced cases of the thought preceding the word (you can skip the word, or not, but either way it demonstrates a thought without a word) and of the word preceding the thought (people who claim not to have an internal monologue will usually still do this sometimes when writing).
It's just that subvocalization is one powerful technique for organizing thoughts, but not the only one. If you had a swimmer who only ever learned the crawl stroke (freestyle), they might wonder how someone can swim without it. The difference, of course, is that the breaststroke is a demonstrable physical act whereas cognitive techniques are not. The monologuers are just more practiced at subvocalization-based reasoning, and the non-monologuers are more practiced at other types.
It's not that hard to exert limited control over which technique you're using -- and I think there are multiple nonverbal techniques. I'd encourage people to explore the space a bit. Subvocalization-first approaches are very good for linear progressions, whereas subvocalization-last (or never) approaches are usually better for big-picture reasoning.
I still think it's far more likely that people aren't describing the same thing. People are very, very bad at interpreting and experiencing their mind's inner workings. Then getting them to describe and compare those inner workings is, IMO, largely a fool's errand. I'll wait until someone develops an external validation of this to believe it.
I say this as someone who pays a lot of attention to my mind's inner workings (meditation etc, near-complete aphantasic) and the more I pay attention to the inner workings, the less I know the answer to whether or not I have an inner monologue.
I'm very reluctant to believe people when they say they don't have an internal monologue. Studies on the matter all appear to rely on self-reported data, not objective measurements. It's been shown that electromyography can be used to measure myoelectric signals produced when a person subvocalizes their speech (i.e. an inner monologue). It works well enough to reconstruct and synthesize that speech.
So we could test this objectively, and I'm not sure why we haven't.
I'm not sure what you think all these people have to gain from tricking you. I have no internal monologue. None whatsoever. If I have something like an internal monologue I don't have any conscious access to it at all. When I want to think through a spatial type problem, I can visualize shapes, images, rotations, relationships, etc but it has no verbal component. If I want to think through a text or verbal heavy task, like perhaps planning some code, I type it out. I do have intrusive thoughts but they are in the form of memories of a thing that happened or simulations of something that could happen, I guess. I most surely do not have an internal narrator or anything that approaches any kind of inner verbal life.
Well, I can't say for the general population, but for me there's just nothing unless I actively try to think of words, sentences and other things. One moment you don't have an idea and the next you do, like a state switch.
I really feel like humans argue too much about how computer programs think and at the same time too little about how humans think. The topic is criminally underresearched and the existing research is all over the place.
You make it sound like there are no ideas until words arrive? For me, my normal ground state is full of thinking that is not linguistic. My mode switch is more like having to turn on the whole language subsystem and pay attention to it, often to the detriment of my native concentration.
When I do think with words, it is usually something I am manipulating with a purpose. It is not the thoughts themselves. I have a thought, and I attempt to render it into words.
Outside of recollection of past speech, it is extremely rare for me to feel like words arrive with a thought I wasn't already holding. Most often, it happens when making jokes and puns in real-time. I am in a mood to be playful, and may have a vague feeling that there is something funny to interject, but cannot separate it from the word combination that comes to mind.
But I can't really think and deliberate verbally. It feels like an act of puppetry.
No, I actually say that there are no words unless I think of them. Ideas appear out of nowhere and are made of nothing, like they just are integral part of my mind state
My mom has aphantasia and we were both shocked to discover it even exists.
She saw something about it online and then started asking everyone in the family if they could picture an apple in their mind. We all said - yes of course, what are you talking about?
And she said, "When you say the word apple, I think of the concept of an apple. It's juicy. It's crisp. It's red, etc, but I don't SEE an apple. I just know what it is."
She had no concept that other people literally saw the apple and I had no concept that people couldn't see the apple.
I asked her what she thought the phrase "in your minds eye" meant for her entire life and she said she thought it just meant exactly what she was describing and not literally seeing anything...
Yeah this was my experience too as a near-complete aphantasic. Doing the apple test around the dinner table was enlightening, and it also made a whole lot of other things make sense through my life as well: I have virtually no concrete memories from a first-person point of view (I remember the fact that events happened, I do not remember subjectively being there)
It's unfortunately very insulting when someone says, "hey remember when we did that very important thing together?" and my answer has always been "uhh... I remember that we did it?"
Do you have a source for the pupil dilation thing? Sounds fascinating!
I scored a 68 on the VVIQ and I can watch my pupils dilate in response to me 'seeing' something in my mind. It doesn't have to be a bright light in my mind's eye, I can imagine something that would trigger a fight or flight response and make the same thing happen.
Yeah this is another test they use: imagine something frightening and "normal people" will get clammy hands and elevated heart rates while aphantasics will not
This is in the same vein to what I wanted to say here as well - I find it very likely that people with no internal monologue are executing similar mental processes, just without them being so “front and center” in the mind.
Metacognition
https://oecs.mit.edu/pub/zjuzickv
not everyone is aware of the "methodology", they just think it through, some however have an executive layer to cognition vs reflex.
imnot sure enough to form an opinion quite yet but thinking about thinking, modifying the methods of cognition, may be a learned skill, perhaps with a developmental window.
a comparison between extroverts and introverts [1] for perceived modality of cognition might be interesting.
it may very well be that a difference in neural development is the cause of this disparity.
1] this is about focus on internal conceptualization vs external perceptualization.
myself, i think in terms of a mono/dialogue if i am preparing to disseminate.
if i am thinking for my own benefit i percieve it as idetic imagery and run thought experiments to qualify for IRL candidacy.
There is also a rather thorny chicken and egg problem.
If language is required for thinking and conscious experience how did we manage to develop language in the first place.
Human level thinking and conscious experience aren't required for language. Dogs understand spoken words and gestures. They're also pretty good at making themselves understood through noises and body language. Humans have better hardware than dogs for making noises.
I could see language development bootstrapping abstract thinking over generations in humans.
Sure, but that's a non-sequitur. Sight and hearing are the most important human senses that practically inform consciousness in the sense of where we live, and what we can do. The other 3 (touch, smell and taste) are at best secondary and cannot provide the groundedness in our world experience - you are the easiest prey in the animal kingdom. Language is primarily a tool for communication, not thought [0], and "thought" is a slippery concept when you're lacking a guidance that informs you of the world if you can't see or hear it.
Super interesting quote and link though!
[0] Fedorenko et al, https://www.nature.com/articles/s41586-024-07522-w
Similarly, I'm reading "Supernormal Stimuli" now, and there's a whole section on infants raised by various animals (chimps, wolves, chickens, leopards, etc, etc). When they are found and brought back to human societies, basically none of them develop full language and integrate fully.
There's also the grim history of language deprivation experiments: https://en.wikipedia.org/wiki/Language_deprivation_experimen...
It really does seem like language is a key part of how we develop our human intelligence.
> Similarly, I'm reading "Supernormal Stimuli" now, and there's a whole section on infants raised by various animals (chimps, wolves, chickens, leopards, etc, etc). When they are found and brought back to human societies, basically none of them develop full language and integrate fully.
Infants raised by chickens? Chickens!? That doesn't sound real.
Reading it again, the book is referencing this case: https://www.theguardian.com/world/2004/jul/11/fiji.jennyfors...
So not "raised by chickens," but rather abused and isolated in the presence of chickens. That's more plausible.
That humans who have grown out of their optimal brain plasticity and hence never acquire the capacity for ordinary human language are not somehow "conscious" is not supported. Rather, it means that, by its very definition, we do not have the means to communicate with them in order to even ask the sorts of questions necessary to determine such a thing. The premise is unfalsifiable.
Well, we can surely communicate without language. I certainly communicate via body language with other humans, my dog, etc.
Animals communicate, and it may be primitive, but a lot of it doesn't seem to require consciousness. Should someone insist, it then depends on what you call "a word".
Words are like atoms. We also have sentences (molecules) and paragraphs (structures) and more. Complexity matters.
It is likely that language forms a sort of quanta of reasoning. It is far easier to posit a verbal hypothesis pertaining to verbal subjects and test it than to try and do it nonverbally.
Maths is probably a great example of this. Try and describe Pythagoras' Theorem without using maths notation or words. Difficult, right? Reasoning is a House of Cards, and without understanding what the card is that exercise becomes significantly more difficult.
a^2 + b^2 = c^2 is perhaps a poor example, since there are several visual proofs that don’t rely on language. I don’t think this negates your overall point fwiw.
Math is a terrible example. Euclid's "Elements" was a core book in Western math education for millennia, and it uses drawn geometric proofs. The proofs are then described in Greek, but the rigorous definitions are the pictures. A lot of mathematical fields can be reasoned about and explored by imagining shapes and objects. The end results today are written in notation because this is a standard language that helps clarify ideas.
Geometry pictures are absolutely not "rigorous definitions". For 1 thing, pictures of lines aren't finite in extent, and pictures of points are finite in extent.
Some of Euclid's proofs are wrong, because they relied on those non-rigorous picture definitions.
* Book I, Proposition 1 (constructing an equilateral triangle)
* Book I, Proposition 4 (Side-Angle-Side triangle congruence).
* Multiple theorems throughout the Elements rely on the visual "betweenness" of points.
Helen Keller was blind and deaf from infancy. What her teacher gave her was a way of systematizing what little sensory stimulation she was capable of receiving. "Language", in the topical sense of "the textual encoding of language that LLMs ingest and emit", is certainly not required for consciousness. Rather, what you need is a way to identify and predict patterns borne out of repeated observations, which is complicated by the lack of ability to observe.
this was absolutely fascinating to read
LLMs as clearly doing more than just modeling words. In order to predict word placement, they need to build some kind of model, their latent space, of the types of things they are able to predict. Not really full world models yet - but they have decent "blog space" or "github project" models. And you can see this with multimodality, or non-text modality modals such as images and audio. They map from their latent spaces to the outputs. The fact that multi-modal systems can share the interior layers shows some kind of internal representation is created.
Or.. the LLMs are just "telescope" sufficiently powerful to detect the aligned internal representation that already exists.
A more powerful telescope does not "create" new galaxies.
Love the analogy. I a lot of people had their mind blown by vector space representation of words. The idea that Female + King averaged out around "Queen". Words might foundational to concepts or just really-well-designed ways to transmit them.
You have the word 'bot' in your name and judging by your comment history love using em-dashes. Are you a machine or not?
I doubt I used emdashes, but I do us "-". I dont have emdash unicode bound to my keyboard yet. (edit: Hyphenation is not emdash, of course)
you can view my profile details freely and decide for yourself, and a quick look might even explain the "bot" suffix.
Oh I figured there's a real human being named Josh Vander and youre his openclaw or something. why do you have the `bot` suffix?
Because I spent 15 years working in robotics, and I'm a nerd who over indexed identity on that :)
I can guarantee you that the comment from jvanderbot was not AI generated.
The assumption of this sort of argument, if I understand it correctly, is that consciousness is just an ordinary byproduct that appears (or grows gradually) somewhere on the spectrum of complexity.
If that’s true, mastering language is basically orthogonal to being conscious (as you’d maybe expect, GPT-2 was pretty good at language, but had a relatively tiny amount of “neurons”).
The same is arguably true for world modeling: a math textbook has a very complex and coherent model of a world, and is, probably, unconscious.
Then the question is: what reason do we have to assume that these models have some form of consciousness?
Likewise, what reason do we have to assume that any given human is conscious?
I don’t know why I’m consciousness, but I know that I am. Other people are outwardly very similar to me, so I assume the same is broadly true inwardly.
A rock is outwardly pretty similar to me; in that it exists in the world, has obvious physical boundaries, and is affected by the passage of time. But it is not so outwardly similar that I assume consciousness.
An LLM is also outwardly similar to me, in that it can express itself in language, and seems to ‘contain’ notions about the world. But it is again not necessarily so similar to me, outwardly, that inward similarity is obvious (to me)
Or that anything at all is not conscious?
Because we are, in our life, not always conscious.
When you say "we" you're referring to something besides our minds?
Because sure, when we get knocked out, we're out. That's not an observation, it's a tautology. But there's nothing indicating that whatever remains doesn't have its own kind of consciousness.
And they do have internal latent vectors for their own states, so there's some kind of recursion/introspectiveness dynamic.
Exactly. The translation back to words is the final step, so in a way very similar to what the post describes.
Improvements in model performance have been made exactly by having intermediate steps stay in the form of internal representations rather than words.
Except that the Chinese Room shows that the existence of a mapping from input to output, however emergently it might have been devised, is not alone sufficient to demonstrate understanding.
An LLM is a big equation that we solve to get textual output. If Ai proponents already believe an equation can contain consciousness, what about the Chinese Room presents a more compelling counterargument?
I'm not bringing up the Chinese Room in its original sense of asserting that a machine can't have consciousness. Indeed, I have no reason to suspect that human consciousness is irreproducible. Rather, the point here is to emphasize that our methods of determination are insufficient. A man armed with an English-to-Chinese dictionary appears to know Chinese until he's faced with an instance of linguistic ambiguity outside the scope of what a dictionary encodes; in other words, you can, with enough probing, disprove that the room understands Chinese. But how do you prove it in an affirmative sense? Let's use a simpler example: does Hans the Clever Horse understand arithmetic? We can disprove it by throwing enough arithmetic at Hans to prove that his understanding does not generalize, but if Hans did know arithmetic, how would we prove it? There are plenty of things in this category--things that we believe to some confidence level, but cannot prove--but let's frame it for what it is: belief and faith, not proof or logic. Whatever utility I may derive from LLMs, I have every reason to be skeptical of a movement of people whose awe of LLMs echoes the steadfast furor of the religious adherent.
The Chinese Room as a whole understands.
The man in the room is comparable to a human hand, and the magic rulebook to a human brain.
This in the sense that we can easily retain human "understanding" by stripping away almost all parts of the human body or replacing them with fairly trivially made replacements, except for the brain.
In the Chinese Room the equivalent is the magic rulebook: We have no idea how to construct/replace it, yet people somehow handwave that away whilst simultaneously confidently asserting it does not understand anything.
That's irrelevant to what we're talking about, though. The point here is not to assert whether or not the room understands, but to emphasize that our methods are insufficient to demonstrate this. You could posit that the room understands, or you could posit the reverse, and neither argument can be refuted.
If our own brains were insufficient to demonstrate the fallacy of the Chinese Room; the particles, molecules, cells, synapses, and so forth of our bodies cannot be believed to have understanding, but we recognize it in the sum of the part anyway....
... we now have LLMs to even more pointedly show the deficiencies of the argument.
It does so only in the claims of its creator. Plenty of other people have pointed out fallacies in the claim. My favorite is the Dennett/Hofstadter observation that while the man in the room may not understand chinese, the room/system certainly does.
"Understanding" is your word, it doesn't appear in the source article nor the comment you're responding to. The Chinese Room argument does not attempt to show that a mapping from input to output can't implement a latent space, or that it can't implement complex models of what the language is describing. If a person can express those things explicitly in the output, or if you have to do them in order to correctly respond how a person would, then the room by definition has those capabilities.
(What's the point of the argument if it doesn't tell us anything about the capabilities or internals of an AI? I'm not sure.)
Right - the point I was refuting is that LLMs are "just statistical models of words." More is going on. Does that imply "understanding?" I don't know, I'm not sure we have a good enough definition to say. But it does mean that the models are more complex that say, markov chain graphs with corpus frequencies. It seems we are encoding data in the latent space with much higher complexity than "just words." There is higher order semantic information being captured - probably not the same has human "thoughts" - but again - also not _just words_.
This person is just making stuff up that feels true to them. Caveat lector.
There are reams of grown up scientific research and philosophy on the relationship between language, thought, neurobiology, and inference. It's nice to hear how one person feels they think, but there is no weight to it if they don't know page one of the syllabus.
would second you on this.
> just making stuff up that feels true to them
Hallucinating it?
I find the comment section interesting. There is so much discussion about the dichotomy of consciousness, in support and opposition as discussed by the author. Some of the "rights" and "wrongs" feel way more combative than, simply noticing where the author is currently at.
I see the temporal disconnected strong here. The author is young-ish, and appears to be exploring ideas of many that have come before him. And, that he'll likely continue to explore on his journey. You can see his thoughts and see echos of long debated thoughts about consciousness, language, meaning, bottom-Up/top-down processing, not being our thoughts, etc..
Thanks to those who are leaving their knowledge, book references, quotes, and etc., to help myself and others who are somewhere along the continuum of understanding ourselves more thoroughly!
The author is making a big mistake by making a normative claim about 'true' intelligence and consciousness.
Is there a correct way to cognize? And although he feels cognizing works in a very specific direction, his brain is basically doing a very similar guessing game on a deep level with training of pathways that started at birth. Basing the argument is a feeling about consciousness and is not convincing.
The author thinks they are describing a unique human magic, forgetting Augustine's insight that human thought doesn't precede the word, but is brought into consciousness by it.
Maybe consciousness just does not exist. Could it be that humans invented it to feel superior to animals, but it does not really exist? I wonder if we will think of the word "consciousness" in the same way we think of the word "aether" in the future.
Consciousness exists by definition. It is a subjective experience and you are experiencing it.
Or more precisely I am experiencing it, but since you are, I think, a human just like me, with a brain and biology very similar to mine, I assume that you must experience consciousness too.
And that is the big question, and one that science may never have an answer to. What is conscious and what is not? The only consciousness I know exists is mine, the rest is just guessing based on similarities. Other humans: almost certainly conscious, higher intelligence life forms: probably conscious, plants: probably not conscious, rocks: almost certainly not conscious, LLMs: ???
LLMs are unusual in that they have a very human-like behavior, more than anything non human, which would make them a good candidate for having consciousness, but on a material level, they are more like rocks, which point them as not being conscious.
There’s certainly a debate as to whether animals are conscious or not, and if we’ve just been engaging in a kind of human chauvinism all this time, but I mean consciousness definitely exists. There’s an “I” in here that perceives, and it’s reasonable to extrapolate that I’m not particularly special, that others that have the same parts as me have the same properties (consciousness).
> Maybe consciousness just does not exist. Could it be that humans invented it to feel superior to animals, but it does not really exist? I wonder if we will think of the word "consciousness" in the same way we think of the word "aether" in the future.
That's just lame, motivated denialism.
If consciousness didn’t exist, you wouldn’t exist! you’d have no interior experience.
How do you know they have one ?
"In the beginning was the Word, and the Word was with God, and the Word was God."
And Islam the first verse was ‘read’
The first time I heard about emergent behavior in llms I thought about that.
Maybe thought is too much because I don’t have anything coherent in mind but it feels so captivating like the premise of sci-fi story.
Word in the Bible is not meaning the actual word (like sematic array of characters) but more like greek word logos
Correct, in the Koine Greek, logos is the word that's actually used. Logos is a bit deeper than words, and is more akin to a universal principle of order that governs the cosmos.
"In the beginning was the logos" is like saying in the beginning were the laws of physics, and God is the laws of physics and the laws of physics are God.[1]
[1] This also entirely depends on which framework you are translating from. This would be accurate from the stoic interpretation of the logos but differs from a more mystical Johannine interpretation, in which the logos is outside the physical universe, not just operating within it. Logos is like the source code of the universe.
It's a hard word to translate. https://en.wikipedia.org/wiki/Logos has a good overview.
Some folks interpret that sentence in a mystical way - the first word was Om, from which all the universe was created.
That words are a byproduct of consciousness is a bold claim and requires some justification. I'm not implying it's backwards also for humans, but that the statement might not be true, and something else could be. Also, defining consciousness is hard.
For instance, you can dream words.
Have you ever had impressions about something complicated, e.g. someone's personality which you cannot put into words?
I'm thinking of my kid - her personality when she was tiny. I could see what it was but not explain it.
Hence I think ideas are unquestionably first and words are the inadequate and inaccurate description of thoughts that exist without them.
Have you ever reasoned through a complex coding or maths problem? Tell me, how did you do that without words?
You (and TFA) are making a false dichotomy here. Yes, of course we think in images. But we also have an inner monologue that is critically important for much of our higher-level thought. How do you even write a HN comment without thinking through it in words?
That may be true without implying that words are a always a byproduct of consciousness. You can dream words.
I agree, and I think that's a good example. Sometimes we can hold a concept in our minds of something and cannot find the words to attach to it. The word we eventually give that concept when it's human, as in your example, is a name. This is because the concept of a personality is too complex to describe with other, preexisting words.
Still, even though I think the author is right about consciousness/concepts preceding language, I think he's wrong about the other part. The statistical models that underlie AI very much remind me of what we call a human concept. It's a statistical model of that thing, person, or idea. However, we humans seem to have this ability to create new concepts that are wildly different than our training data (what we have experienced in life), and we call this creativity. It's missing from AI for some reason. That is the part that makes me question if AI will ever match humans.
how animals perceive world? They don't have language, but they are conscious for sure.
Eh? What else could communication be there for?
Survival.
I share your positivity, however I don’t think figuring out how to market through the noise is the important part.
LLMs don’t need to think, they just need to be a more efficient data retrieval system than what already exists (and what has already been sold to the highest bidder). With LLMs, I’ve find I’ve started reaching the corners of the internet, where the true lovers of knowledge are doing the work they have always been doing. Quality does eventually speak for itself, at least once we’re able to strip away the marketing (you mentioned Apple, do you believe they are truly a cut above, or is it just hype?)
If LLMs can filter the noise, build it and they will come might return as a valid way to be.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do. The word is never the start. The word is just the skin. The idea, the consciousness, is the thing sitting under it.
In (at least once school of) Tibetan Buddhist meditation you observe the part of the mind that is producing ideas, and then suppress it in order to explore deeper layers of consciousness. The "lights on" "you" that is observing is _not_ the idea producing part of the mind, which to me at least seems to be more mechanical.
I do agree with the author of course that a concept isn't necessarily wrapped with language from the start, and can be a skin; but equally it is possible to think in a verbal mode.
As I see it, LLMs will/do develop 'introspectiveness' if it is selected for during RL. If its not delibrativly selected for (this is probably done because it helps with alignment; if the machine can align itself that's great) its probably a generally useful skill to have when doing normal tasks, https://www.anthropic.com/research/introspection
Their introspectiveness doesn't strike me as indicative of anything, since we might be just reinforcing a "plausible word salad generator for input context that leads to better solutions" or just a "post hoc rationalizer", both of which may or may not be the key mechanism behind verbal consciousness.
Forgive my quoting my own blog post:
Writing is a shadow of thought. The better the writing, the more clearly the shadow represents the shape of the original thought. Even the best writing can never perfectly capture the original thought. Writing is one of the best tools we have to share thoughts across space and time.
I look at culture, under which I put language as a component, as the O/S of the human brain. So yes a brain with no O/S can think but it probably will never replicate, in one life time, the accumulated thinking skills of a brain given a good O/S.
How many of us would independently reinvent language with specific tenses, the diatonic scale, the number zero, perspective drawing, geometry, calculus, atomic theory, quantum mechanics etc. if left on our own from childhood?
So to me, the question becomes, can human language encode the sum of all human thought? The answer might be no. But it might be good enough to get close.
Isn't it interesting then how much of our intelligence is captured by our words/language? I would've thought to create AI you needed to replicate neurons and synapses and learning. So it's still amazing to me that statistical modelling of our words creates a Claude Opus.
Perhaps LLM's do replicate the important features of "neurons and synapses and learning" and perhaps a "statistical modelling of our words" is pretty much what our brains are doing? Most of the counter-arguments I've seen boil down to "but humans are special" which I'm not sure I find compelling.
Not saying LLM's are conscious. Just that much of our amazement about their behavior seems to say more about us realizing the things we can do are not magic, rather than them being so amazing.
Words limit our understanding of words because they are intrinsic to a particular kind of understanding. yes? no?
In "Notes on the Synthesis of Form", Christopher Alexander talks about how we can know something is wrong even when we do not know how to make something right. He was talking about this I think
=z==
as opposed to
+|+|+
or
+|+|+|
I happened upon the word "entanglement" and it seems an interesting alternative to the order that is inherent in words.
Our language is all about Actor-verb-object. Entanglement provides a fundamentally different concept. I cannot say "I drove my car on the road" with entanglement. Or at least with entanglement I can say something equally valid like "the road moved my car with me inside."
And entanglement works for the + and | things, at least for me. There is some kind of entanglement (degree) that creates a gestalt.
At least that is the best I can come up with.
Reminds me of this from "Gödel, Escher, Bach: an Eternal Golden Braid":
"Let me rephrase these last couple of sentences without using the slightly technical term "isomorphism". When a system of "meaningless" symbols has patterns in it that accurately track, or mirror, various phenomena in the world, then that tracking or mirroring imbues the symbols with some degree of meaning - indeed, such tracking or mirroring is no less and no more than what meaning is. Depending on how complex and subtle and reliable the tracking is, different degrees of meaningfulness arise." - P-3
EDIT: I initially wrote "G.E.B." instead of "Gödel, Escher, Bach: an Eternal Golden Braid" which was too cryptic.
While I know what Douglas Hofstadter's book "Godel Escher Bach - an eternal golden braid" is, it might be worth revisiting your apparent assumption that more or less everybody on HN knows about a book from 47 years ago (regardless of its significance to those of us who have encountered it already).
LOL, I happen to have a "Hofstadter_GEB.pdf" open on my desktop; been there for months.
Language is just a shortcut to logic. LLMs don’t seem to have the capacity to logic, they just pattern match. LLMs seem conscious because they are pattern matching against all of human written cognition.
I am pretty sure human level intelligence is emergent from language use. The more complex the language the higher the intelligence since language is itself a tool used to not only to communicate, but also classify the human experience. Without language, we are fairly primitive relying on instincts and more primitive communication. Feral children are an example of what a base state might be.
> But one thing scares me a little. What if LLMs slowly become worse? Think about it. Everything written before 2017 was made by humans. After 2017, LLMs started filling the open web with their own content. And now that same content is being fed back to train the next LLMs. An LLM can give you perfect grammar and a rich vocabulary, but it can quietly lose the real context. Right now the share of AI content is small. But what happens when it keeps growing?
Have the last-generation LLM pre-screen new training data for the next generation LLM.
We can't define consciousness well enough to make claims like this.
> For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
Recently, I've been thinking a lot about context and whether an LLM actually creates a true representation of a concept within its vectors. And for our current usage of LLMs, does it even matter? As models become so large, will the difference even be perceivable to us?
It’s funny because, ever since Norbert Wiener’s Cybernetics was published in 1948, we’ve been trying to replicate how the human brain works. In that sense, we’ve been building increasingly powerful machines that mimic human thought. And now, the most interesting thing is something that turns that idea on its head: LLMs think in reverse.
> Now come back to my first question. An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath.
No!! This is misinformation.
An LLM predicts the next words (tokens) based on the words before it *and* on the internal representations it has learned from training.
Processing language can lead to internal representations that capture patterns and relationships; LLMs can exhibit emergent abilities, including reasoning-like (stress on "-like") behavior, even when they weren't explicitly trained for that.
Making reductionist claims about LLMs, based on next-word prediction, is similar to making such claims about biology, based on amino acids.
"An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath." At most, and I will argue for less, that is half the story. This egregious over-simplification doesn't account for training and the massive dimensional space of thought it forms. Given that frontier models likely have trillions of parameters, the combinatorial space of ideas within approaches the infinite -- hardly zero.
As others have remarked: this is an unsubstantiated claim.
I am actually using LLMs to perform controlled experiments to determine if the opposite is true, and that consciousness is a byproduct of language as Jaynes might have agreed.
I want to add one thing, Please think about how animals perceive world? They don't have language, but they are conscious for sure.
> But your brain works the other way around. First there is a concept, a feeling, an image, and then the words come out to describe it.
After this I can't take the essay seriously- this sort of blanket statement about the one true hierarchy of consciousness and knowledge is BS.
While it seems to have been disproven that words in other languages cause people to speak and think differently, that also doesn't mean that words don't have any effect on the way we think, or that the concept of a word always has to come before the word itself.
The reason why we experience the uncanny valley of LLMs is because they don't represent a true consciousness, BUT it's also clear that the architecture represents certain qualities of consciousness- as the models have scaled we can see that it has some other non-word related understanding.
The evidence points to consciousness as a set of interlocking systems- attention, long term memory, short term memory, emotions, etc.
Steve Jobs didn't invent the PC. It's reductionist and offensive to single him out and thanking him for that.
Computers didn't "figure out the maths". People did the math, the computers did the calculations.
Human knowledge is nowhere complete, it's silly to think so, and by extension it's silly to think the LLM's have anything near a complete knowledge set.
They don't even have access to all of current human knowledge, much less so future knowledge. A lot of text and information is locked away from the public internet and the internet altogether.
Not all types of useful ideas for applications have been invented, that's ridiculous, there are plenty of ideas that people haven't had yet. There are however a massive amount of shitty copies of the same shitty ideas.
I'm going to go on a limb here and challenge this assumption. As a parallel thread to the point put forward by michael1999, can we not argue words manifest as the result of the maths performed within LLMs, maths that, for all intents and purposes, is their world? Because the internal maths of how LLMs work is likely to be significantly cruder than the 'real' world that defines consciousness, perhaps we are closer to the ghost in the machine than we would like to admit.
> First there is a concept, a feeling, an image, and then the words come out to describe it.
Nope. Concepts, feelings and ideas do not exist independent of expression (words, sounds, responses). Humans have the same process as LlMs have. And that's precisely why LMs are able to succeed. Similar to how camera works based on the working of an eye, or how aeroplane works based on a birds wing.
If you still believe that concepts exist independently, imagine a creature that can't express, make sounds, or respond with actions. What exactly is a "concept or feeling" to such creature? Do trees have concepts and feelings? Infact trees could, because they can respond to stimulants.
If you can't describe or perceive a thing, that thing doesn't exist in your world. It's that simple.
What you are describing as "translating thoughts into words" is actually, refining your expression. The expression already exists at the same time as the thought.
My son is insufficiently verbal (he's of the kind that communicates with quotes, but really struggles to express or understand new sentences), and I am quite sure he has concepts, feelings and ideas.
“The relationship of thought to word is not a thing but a process, a movement from thought to word and from word to thought.” Vygotsky
St Augustine had it right 1500 years ago: humans learn words opportunistically according to their desires and problems and existing deep understanding of the world, and (critically) humans really can't learn language without being natively fluent in great ape facial expressions, gestures, grunts, etc.
[https://faculty.georgetown.edu/jod/augustine/conf.pdf]I do not think the written description of a dog is the same as a dog itself…
Nothing is the same as the dog but we probably store a lot of sensory information, emotions and images relating to dogs in our heads.
Maybe it is. If you think about it hard, you will find that you cannot separate the idea of dog from the word dog. The object dog can only be understood as a perception linked to that idea.
"Ceci n'est pas un chien"
Stop for one second and ask yourself a simple question. Where do your words come from?
When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do. The word is never the start. The word is just the skin. The idea, the consciousness, is the thing sitting under it.
"For an LLM, it is exactly the opposite".
Is it really? LLMs don't have words inside, for the most part they operate by applying transformations to a vector that does not contain any words at all. Words that come out of an LLM are just what sampler gives us by looking at the vector that is the result of those transformations.
Does this vector contain a world model? Some form of thoughts or reasoning to arrive at the result? That's an open question really.
Sure, rerunning that whole process for each token might not be the best solution, although that's an open question too. But saying that LLMs operate on words first is too big of an oversimplification
>Does this vector contain a world model? Some form of thoughts or reasoning to arrive at the result? That's an open question really.
It is, but it has been explored in various forms. Anthropic have some interesting papers where they can map out concepts to some extent as they exist inside the vectors of the LLM, and they can see e.g. that vectors relating to some concept can appear there some iterations before the LLM emits the tokens that spell it out. (though also, quite frankly, this shouldn't be all that surprising. If they were just picking the next token with no representation of what might follow it the result would be more like the 'one-word-at-a-time' party game than coherent text. They used poetry with a rhyming scheme as an example because it's something where you need to have some idea of what might rhyme with the line before and aim at it at the start of the next line in order to have any degree of success).
There's plenty of people who speak before they think
Word is just the skin is the load-bearing idea. The real unlock, if you will.
My favourite way of thinking about this is by going back to the End Poem at the end of Minecraft. The fact that words are an "interface", "very flexible" and "less terrifying than staring at the reality beyond the screen".
Maybe that reality is the true understanding of the neurobiology that defines our thoughts. To be reminded that the magic of experience can be reduced to neuronic hallucination is, frankly, horrifying. Maybe we dislike LLMs because we see their vectors and numbers as crude reminders of our own intellectual banality.
Or maybe I'm just feeling hungry and should really go eat dinner.
I don't know about you, but I think with words all the time when reasoning through complex ideas. Where do you think the phrase "thinking out loud" comes from?
> Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it
There are enough neuroscience experiments demonstrating how we create post hoc explanations of our actions that I wouldn't trust this intuition prima facie.
I've noticed that more and more of my mental attention while speaking is observing the words that come out of my mouth rather than 'generating' them. I split time between rough pathfinding in my mind to fit a conceptual framework then kind of sit and listen as it comes out to think about how they sound. If I had to guess this happens on a 1-2 sentence chunk at a time.
The fact that we're even capable of doing a study like that and introspecting is probably a good indication of our cognitive abilities as humans
Consider someone designing a skyscraper. They're thinking about material strengths and such while they draft it, they aren't just rationalizing their choice later.
We’re stochastic parrots most of the time, with a varying ability to reason but opting for pattern conformity usually
What is reasoning except applying another pattern on top of existing thought? Personally it seems like I and everyone else is simply pattern matching, albeit at a higher level than current LLMs. There is no difference in the process as far as I can tell, just different inputs.
This is getting voted down because people just hate the idea that it's true.
No it's down voted because we didn't stochastic parrot ourselves into landing on the moon and building the internet and modern highways and skyscrapers
Dismissing humanity that was is absurd
In some meditation practices, the task is to observe the stochastic parrot in your head. You will soon realise that you cannot control this parrot. If you cannot control your thoughts, could it be that you are not your thoughts? And if you are not your thoughts--the word stream in your head--what are you then?
But LLMs aren't just shuffling words around. Words (tokens) are the "human interface" at the edges. Text input gets embedded, then there's this huge latent-space computation in the middle, and only at the tail end does that get converted back into word/token probabilities. So just saying that "words are the source" isn't entirely correct and feels misleading.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it?
Let's take a feral human, do they achieve the same with language powered humans? No, obviously. It means language does something we can't replace with "consciousness". Maybe something life preserving, putting consciousness downstream of language use.
But more generally - how can 8B humans make a living on this planet? Not without language, that is for sure! We long passed the stage where we could exist without language at our consumption rate. We can't even exist without math, population collapse would be the outcome.
Minds are downstream from language & math. For example, is music downstream of piano or piano downstream of music? I think you can't cut this cleanly. They developed in relation to the other, recursion between tool and art form is already old.
Can you please point me to the proof of the first claim?
"At its core, it is a big pile of words that predicts the next word, using some maths the computers figured out."
"An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath. The words are everything."
This is just flat out wrong.
Words are used as training data, to build a system of vector embeddings. The LLM contains no words. That was the training data long discarded.
Vector embeddings are groupings of meanings derived from the relationship between the words in its training data. This entire system is modelled after human neural mechanisms in a way that machines can emulate.
"But your brain works the other way around. First there is a concept, a feeling, an image, and then the words come out to describe it. (At least, this is how I feel my own brain working.) For us, words are the byproduct of consciousness."
It's working the same way around (you are not saying Ai has consciousness that is derived from words!). And you make a huge jump from the concept that 'words come from concepts' to 'words are the byproduct of consciousness'. Because 'concepts' are not equal to 'consciousness'.
In fact you do not define consciousness at all, making it hard to determine what argument you are actually making at all. You seem to think humans have this trait of 'consciousness' but can't explain or evidence it beyond 'feeling it'.
Consciousness it defined by many as the ability to experience events and process thoughts or qualia. It's hard to test this, and we can see why that is with an AI - it may claim to be conscious, or even claim not to be, but how can we trust either answer? Philosophers aren't even sure to trust another human who claims to be conscious, and we struggle constantly to determine at what point creatures the animal kingdom are conscious or not. Cats? Lobsters? Snails? or even across the plants?
Before you can refute the consciousness of AI, first establish the consciousness of humans. A better question to ask is what does the uncanny emergent ability of an AI to mimic a human say about consciousness?
Heck, they even ignore the studies that show that conscious thought is a lagging indicator of 'thinking'. Much like you don't tell your heart to beat or your stomach to digest there is a fair amount of neural activity that occurs either before you're aware of it, or outside of your consciousness at all.
With any articles like this I take any claims with a grain of salt. We don't have the needed information to make any bold claims on what consciousness is other than a bunch of electrical signals we can see occur in the brain.
Defining consciousness is one of the most difficult outstanding problems in the whole of science please see: https://en.wikipedia.org/wiki/Hard_problem_of_consciousness
An approach that might shed some light is instead to define what consciousness ISN'T. Naively let us say consciousness is NOT a large list of weights (i.e. an LLM).
The uncanny emergent ability depends entirely on training data. A mathematical model is used to match output against training data (via loss functions etc). The training data contains all the human ingenuity, logic, rational, patterns and features.
Try giving an LLM model the alphabet ALONE and see what it comes up with?
I think small children learning language learn in a way similar to LLM's, just saying words that seem appropriate to the situation without any understanding of the meaning. But a key difference in my mind, is they use the reactions to this text prediction to help them construct a picture of the world, which then can interact with their thinking of what words to say.
For LLMs there is no consciousness.
We might quibble what to call it, but they use language productively.
https://en.wikipedia.org/wiki/Model_collapse
>An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath. The words are everything. For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
I don't think that's right. What makes LLMs (and other forms of "AI") work is the very fact that they hold a statistical model underneath and put words on top of it. That statistical model, at least to me, appears to be somewhat akin to what a human would call a "concept".
Like a stable diffusion model has a statistical model of what a "chair" is. Not one exact chair, but all chairs. They all have backs, 4 legs, a flat area to sit on. It can then carve an image of a chair out of noise using this concept. Take a music model like Suno as another example. If you tell it to make a rock song, it has a statistical idea of what this would looks like. The harmonic progression, the tempo, the types of instruments used. It then attaches sounds to this statistical model and you get a "song". LLMs also appear to do this. If you ask it for a paragraph of a specific flavor of prose, it has that as a statistical model, and it attaches words to it.
I am not a big believer in LLMs delivering on even half the hype they've generated. However, this very idea of human concepts vs AI's statistical models is the one thing makes me wonder, sometimes, if they're on the right track, even if we are a long way off from AGI. It's kind of funny that the author of the above article landed on the opposite conclusion.
> When you speak, what comes first, the idea or the word? Do you first feel a thought inside you, and only after that go searching for the right word to wrap around it? I think we all do.
Most people don't actively think when they are talking or writing most of the time. Only during period of thinking that we do think. Most of the time, we just talk words. There is evidence is that we just observe ourselves talking and that the thinking itself is not conscious but it's still a theory.
> Now ask the same question about an LLM. For an LLM, it is exactly the opposite.
No one has any idea how an LLM come to correct words. LLMs produce numbers (tokens) that end up to coherent words.
> So what is an LLM, really? At its core, it is a big pile of words that predicts the next word, using some maths the computers figured out.
No, not really.
> The words are everything. For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.
The author provides no proof for this and there are very good reasons to think that LLMs have inherent thinking.
The rest of the article was just some random thoughts the OP jotted down? I just don't see their coherence to the beginning of the article.
No, I don't think so. Language is intimately intertwined with our understanding of the world. Many people make the mistake to think that languages are about the spelling or the sound or maybe the grammar. But in reality they are about how words are defined only on relation to each other. A symbol is defined by what it is not. This relationship between symbols structures how we think about the world and how we fantasize and desire.
You can perhaps have a consciousness without words, but not without language.
Meditative practices support this. Anyone who's spent even a moderate amount of time with meditation knows we are not our thoughts.
If you don't believe me, learn how to meditate, meditate for literally just 10 minutes, and come back to me.
this doesn't imply that words are a byproduct of consciousness.