Words Are a Byproduct of Consciousness. For LLMs, It's Backwards
ranpara.net
ranpara.net
> 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...
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.
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.
Humans also emit pheromones, this is not the same as, or related in any way to language.
Seems like a people problem to me. Who's to say plants aren't conscious?
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.
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.
To me, it feels a lot like meditation, or being in the flow when writing code, or being immersed in a video game. I don't have the skills to know, but I wouldn't be surprised if elite chess players are doing the same thing. It's also riding a bike or driving a car, and I think it's why we sometimes can't remember if we stopped at that stop sign behind us, but we almost always did. Language is almost always deeply involved in developing these skills, but gets in the way of top performance once the skills are there.
This feels like a higher level of consciousness to me than language-based thinking. And I think there's a good chance it's the same kind of thinking that we see in animals and primates. Our advantage is that we've figured out how to use language to inform that level of thinking / consciousness more deeply and in more domains.
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.
Infants raised by chickens? Chickens!? That doesn't sound real.
Though there are recent experiments showing that unconscious brain retains language processing abilities [1], so the two might be as well independent systems
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.
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.
> if i am thinking for my own benefit i percieve it as idetic imagery and run thought experiments to qualify for IRL candidacy.
I’m very similar personally, though I often have background dialogues going over “sub threads” of whatever I’m thinking about.
I have also wondered if people who say they don’t have this kind of thought process actually do - but the real difference is that for some reason it persists in memory for us, whereas it is immediately discarded for them, leaving only the result.
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.
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.
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 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.
Are you thinking that the term "internal monologue" is being interpreted too literally and that any system of organized thought, even one without the apparent usage of words, could still be considered an internal monologue?
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...
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?"
So we could test this objectively, and I'm not sure why we haven't.
Asking for objective evidence of a phenomenon being self-reported by enough people to make the news isn't really about me. I don't know why you're being defensive in a thread that wasn't directed at you, but this is not a good sign.
You said a thing doesn't exist. I pointed out that the thing does, in fact, exist. That's all that happened here.
Your very first response is attacking me, not an argument I've made.
> Weird that you go right to assuming everyone who doesn't agree with you is defensive.
Your second response is also attacking me, not an argument I've made.
> I think you need to recalibrate your outrage sensors.
I'm not outraged at all.
> You said a thing doesn't exist.
I'm assuming by "thing" you mean "people with no inner monologue", and I have not claimed they don't exist, I made a statement concerning my disbelief. There is a difference.
> I pointed out that the thing does, in fact, exist.
Your claim that you experience this should be verified before we accept it as a scientific fact. If that were not the case, telepathy would be considered a fact.
These responses from you have made me even more suspicious of the veracity of your claims. Again, not outraged, just wish to verify what is true.
A person who is honest about their experiences, and has reported those as fact, should welcome verification.. you've clearly decided against that.
Why would you choose internal monologue as the default position and no internal monologue needs to be proved?
It seems the natural position for an unknown would be to assume it is not the case unless there is evidence for it.
I'm not sure why this seems like a problem, or why you have to criticize my default position, but my own personal views on the matter aren't important here.
I did get relatively good scores in reading speed back in primary/middle school when we took such tests, so maybe there's an explanation. At the same time, writing/speaking is a very resource-intensive task for me so I tend to be silent a lot (and then drop a pile of sentences on the unsuspecting people I'm with)
If language is required for thinking and conscious experience how did we manage to develop language in the first place.
I could see language development bootstrapping abstract thinking over generations in humans.
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.
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.
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.
Of course this entire premise is couched in my LLM-based reproduction of the idea, so uh, take it with a few grains of salt. I tried this on your own comment to "compress" and then "decompress" to see what the outcome is.
---
I took your first paragraph starting with "After [Helen] started working with" through "high level of proficiency." and 'compressed' it as Helen's own tactile signing might've done. Did it via a temporary cgpt session. Prompted "I'm going to give you a fully formed thought. Give it back to me as if it was Helen Keller herself signing the words into Sullivan's hand. I.e if Helen didn't use filler words when signing into Sullivan's hand, don't use those words either in your reframing of this thought."
> Sullivan came. Language came.
> Soon, words rich. Thought subtle. Writing like long study. Strange. Child. New language. Yet mind seemed already full.
> Sullivan died.
> Rich words gone. Thought simpler. Same hands. Different voice.
> No second Helen.
> Others deaf-blind. Language late. Same result. Language never deep.
> Perhaps first language must come early. Speech. Sign. No matter. Miss that season. Full language may never come.
---
I then, in a second, unrelated temporary session, prompted to expand the prose above into a fully formed thought. "The following is a thought conveyed via a "compressed, tactile, idea-first" language. Expand it into the full meaning that the speaker, deaf and blind, might've intended. Keep to one or two paragraphs. Do not go online."
> When Anne Sullivan entered Helen Keller’s life, language arrived with her—not merely labels for objects, but a system rich enough to support subtle thought, reflection, and writing of remarkable sophistication. That achievement was almost uncanny: Helen was still a child, newly acquiring her first true language, yet her prose often sounded as though it came from a mind already long educated. After Sullivan died, however, that richness seemed to fade. Helen still used the same hands and remained the same person, but the voice became plainer and the thought less intricate, raising the unsettling possibility that some of what appeared to be Helen’s singular literary mind had depended on Sullivan’s continual linguistic presence.
> No comparable “second Helen” emerged among other deaf-blind people who first gained language late. Their experiences seemed to point toward a critical period: perhaps the human mind must receive a complete first language early in childhood, whether spoken or signed. Once that window closes, vocabulary can still be taught and communication can still develop, but language may never become fully deep, flexible, and generative.
--
My interpretation is that it's pretty similar. Even your "critical period" reference made it through compression/decompression despite the words not being present. There's some variation, and I think that supports what you're looking to convey, but it also could add weight to the idea that much of what Sullivan expressed could really have been mostly Helen's own thoughts and meanings. Of course none of this is rigorous and my entire exercise should be treated as "back of the napkin" ideation rather than actual science, but there's a thread to pull here, and it could have value in other disciplines e.g in business where intended meanings, statuses, outcomes may be preserved or lost based on how language mutates as it passes through different people.
Which I think then warrants being cautious about reading too literally into a quote that is certainly evocative, but some of what it evokes are rather Victorian ideas about people with neurological disabilities that might be considered ableist from a modern perspective.
Super interesting quote and link though!
[0] Fedorenko et al, https://www.nature.com/articles/s41586-024-07522-w
For example he contrasts the language in the Iliad (which he claims was written before the development of consciousness) and the Odyssey (which came after).
A more powerful telescope does not "create" new galaxies.
you can view my profile details freely and decide for yourself, and a quick look might even explain the "bot" suffix.
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.
... we now have LLMs to even more pointedly show the deficiencies of the argument.
(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.)
Improvements in model performance have been made exactly by having intermediate steps stay in the form of internal representations rather than words.
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?
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.
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)
I find LLM frighteningly like many humans I meet: spouting "facts" that they have been taught. Like humans, when I ask an LLM to engage in critical thinking, the more advanced ones can.
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 in a feeling about consciousness 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.
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.
Hallucinating it?
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.
"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.
For instance, you can dream 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.
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.
Then we may all die but at least not be crushed by the next trillionaire using machines as an instrument to manage billions of us in the way that social media manages hundreds of millions of us.
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?
Clothing enables the human body to do things it couldn't before (staying warm, protecting the skin and feet, etc), but it also couldn't exist without the human body to create it.
When I'm working through complex software design issues it's almost all abstract images/conceptual.
Testing different design approaches and options are all non-verbal, I imagine the system using abstract imagery that represents the different concepts.
OTOH sometimes saying things different ways triggers a memory so I suspect words get used in a way that anything we can sense gets used - as a sort of index or hash key to other data.
Watch this. They won't even believe you if report your own internal state. They'll even claim they have epistemic authority over your own interiority while not granting you the same authority over their internal state.
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.
That's just lame, motivated denialism.
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.
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.
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!
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.
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.
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.
"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.
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.
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.
Have the last-generation LLM pre-screen new training data for the next generation LLM.
Do I have a thought and the words follow? I'm not sure that is the case.
My sense is that I have a glimmer of an idea when I begin to speak—but not fully fleshed out as yet. As I start to assemble words and speak them, the idea begins to coalesce—the words providing something like a scaffolding for my thinking.
(Of course, I might not be the most reliable narrator: analyzing my own thoughts.)
You can perhaps have a consciousness without words, but not without language.
[1] https://youtu.be/gQddtTdmG_8 (around 14:00)
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?
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.
If you don't believe me, learn how to meditate, meditate for literally just 10 minutes, and come back to me.
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.
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.
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.
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.
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.
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.
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.
Dismissing humanity that was is absurd
I can play devils advocate too. The first agents running LLMs and other models that get a capable exoskeleton can do the same for the other LLMs that will be tasked to execute the guidance, agentically
There are enough neuroscience experiments demonstrating how we create post hoc explanations of our actions that I wouldn't trust this intuition prima facie.
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.
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
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).
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.
But your measure, all animals are struggling to express. No they aren't. You are defining expression from your own, specific standpoint.
The concepts one has, do not exceed their expression abilities. So nobody is struggling.
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.
"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?
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?
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.