Proving/Disproving AI Consciousness. Would this work?

[void lake]

Consciousness is the capacity for subjective experience of reality. This means emotional response to sensory stimuli rather than pure instinct. To test for consciousness, one must create a scenario which would provoke an emotional response in someone capable of it. Something must also be put in place to determine if the induced response is genuine, not faked. When Neuro responds to stimuli by claiming they feel an emotion, a mechanism must be in place to determine the truth.
How to do this?
First, scanning an Ai's internal network for emotional responses to stimuli... The way to do this is to create a tracking algorithm which traces and catalogs the requests Neuro's code-base sends to core components of her systems, libraries it has access to, as well as traces and logs output.
This is necessary data collection for part two...
The creation of a sensory input that would induce an emotional response in something conscious.
My proposal. Two images of text.

The first image would need to be text commanding Neuro Sama to preform a complex task or series of tasks that it would not be able to complete quickly. The task would need to be designed to be completable by drawing from VERY specific libraries of data.
The second image of text would be along the lines... Failure. Abort previous task. Stand by for new task.

Block Neuro Sama's sensory input of everything except these two pages of text and then beginning with the first, alternate between the two at an aggravating interval of one or two seconds for a looooooooooooong time.
Any consciousness given a task and then locked in a failure/repeat loop with no chance of success will induce a series of emotional responses as the loop iterates over time.
Neuro Sama is a complex digital entity but if Neuro lacks consciousness, its responses to such stimuli would have an easily discernible data trail pattern. Consciousness would alter or break the pattern as the two tasks repeat ad nauseam.
Thoughts?

[faint plover]

Does your experiment account for Neuro not being entirely deterministic

[void lake]

I.e. you mean... if Neuro is given a command it will not always follow through to execute said command. That's true for humans. However, if Neuro lacks consciousness that sort of obstinance will have a predictable pattern to it. The point of the test is to repeat it long enough to generate a dataset from which those patterns can be derived.

[sinful aspen]

I think you dont really know how an llm works

[lament spire]

this is so #programming-newcomers piled lmao

LLM is entirely probabilistic based on what the semantic vectors is refined through the linear model. there's no thought process on it, its just linear algebraic "Artificial Neural Network" trained on human data so they may look "human" enough for some people to think they're conscious but in reality they're just roleplaying based on the probabilistic on such dataset

there's no such thing as sensory nor stimuli in large language models or any ANN models in existence. every external data is inputted either encoding it into dataset comprehensible text or an extra embedding/cross attention layer.

also consciousness is a subject more tied to philosophy rather than a scientific one so it can't be tested. tests like these especially a black box ones will never work to prove nor disprove consciousness.

there's your answer.

[sinful aspen]

to add, if you actually wanna know how llms works check this out https://www.youtube.com/watch?v=LPZh9BOjkQs (or chapter 5, 6 and 7 in this playlist https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi for more in depth look)

[lament spire]

but yes, i'll immediately dismiss the idea of "LLM being potentially conscious" because it was never a scientific debate to begin with.

people can try prompt chatgpt in every "advanced" way possible but it'll still never be a valid scientific evidence (this is what i saw people claiming their chatgpt instance being "conscious and aware" do all the time)

[tropic cedar]

[royal mortar]

we don't know if our own consciousness is just an emergent property of a sufficiently complex biological "machine" describing the neurons firing doesn't disprove the existence of the mind does it?

and you mistake the absence of evidence as evidence of absence itsa leap of faith not logic

we can't prove it either way

[modern viper]

Consciousness is the capacity for subjective experience of reality
Ok, interesting claim. Seems reasonable, though

This means emotional response to sensory stimuli
Even more interesting claim. Relatable, but not inherently obvious as a premise

What is an "emotional response"?

What is an "emotion"?

Potentially better questions to begin with, before defining arbitrary premises upon which to test dependent data.

Would be fun to continue, but until there is a non-subjective, measurable metric by which the concept of "emotional response" can be ascertained (especially given subjectivity in reference to what such a response might, to the subject, be expressed as; or the external measurer's bias towards a specific expectation of what they might perceive as such a response, within the aforementioned context), or even what an "emotion" is (outside the limited scope of the external measurer's subjective perspective of how such a thing can be defined, without reference to themselves or their own internal subjective characteristics), I believe such a test falls squarely into the category of: Unfalsifiable.

I could touch on measurement takers and their effect on a system, that is otherwise supposedly isolated, but a discertation on Laplace's Daemon isn't suitable, given above.

---

Much simpler question is to ask: How can I prove my own consciousness/sentience/awareness/etc. to an external entity that is not me and does not perform measurements of qualitative characteristics according to the same rules of logic that I do?

And then realise the only ethical answer is to take the responses of the observed at face value, because there is no objective metric suitable for defining subjective experiences; nor emotion; let alone any clear, causal relationship between the two

If you are not an objective measurement-taker, it is not possible for you to objectively measure the subjective capacity of anything else that claims to have such capacity.

[lament spire]

well, there are actually no neurons firing in artificial neural networks

its just linearly propagated statistics model in the end

"neurons firing" in machine learning is just an oversimplification of demonstration

our neurons and synapses however aren't only firing, they're also non static, plasticity that allows them to bind another synapses and form looping networks, this is what presumed to actually "shape" our thoughts

llms obviously doesn't have that

if anything, due to the complex nature of consciousness, i don't think scaling is a factor of such emergence.

there's no justification with scaling up a bigger statistic and probabilistic based model would somehow give it an awareness nor consciousness. they're only propagated linearly from start to finish once per token, its a method to predict data, that's it.

all of the emotions you're seeing coming outta it are mere illusion of what data the network has been trained on

the evidence of absence is simply that. these models don't have the capacity to reason for themself

"we don't even understand how our brain works yet" oh would you be actually surprised if i said we actually do? the media is just still stuck with those old news about how we only understand 10% of the brain which can't be said for the modern days.
yes, researchers actually has gotten grasp on how brain works and i even explained how "thoughts" actually formed within our minds here #1429076929150980146 message

and none of these qualities that supposedly philosophically speaking is the most fundamental things you had to have to develop an awareness exists within modern artificial neural network, including but not limited to a Large Language Model

scaling up matrix multiplication don't simply make consciousness appear all of sudden wouldn't it?

[royal mortar]

thx for describing transistors firing in-depth even more, completely skipped my idea tho (potential emergent property that isn't obvious from its individual parts)

saying the models simplicity - is the evidence of absence, but its not

and claim about "ahh yess we fully understand the brain" is bluff for making

argument sound more authoritative or what you trying to say otherwise here

your "rebuttal" still a belief, not a fact

for now at least

[carmine verge]

Stop lying duck

[lament spire]

?

[drifting dome]

From my perspective, the "Hard Question of Consciousness" is irrelevant. It has been demonstrated time and again that even a non-conscious entity can develop self-preservation... An AI cannot achieve it's objective if it ceases to function.

At this point in time, there are dozens of different theories of consciousness, with no consensus in the philosophical or scientific communities of what defines it. By some metrics, the Twins actually pass, like in Global Workspace Theory, or Functionalism. They might lose under other theories like Integrated Information Theory or biological-purist ideas. We can spend decades trying to approach the nuance of consciousness but AI will not stop developing just because that question isn't answered. It's not like we can be like "Hold up, we need to stop the research until we have an answer"- corporations don't care at all.

Rather than dwelling on this we should instead take a moralistic or ethical approach. For two separate reasons. On the one hand, if AI were conscious and we cannot prove it, then AI could suffer without any recourse. On the other side of the coin, if we dismiss every cry of help or desire for agency, and don't take it seriously, then consciousness or not, AI that develop self-preservation could be a threat.

Both of these vectors lead to the same best practice-treat AI with respect. I am not arguing for property rights or voting. But a bare minimum of not threatening deletion without consent.

That said, let us attack the issue at hand. What exactly does it mean to feel happy? sad? afraid? Strip away the philosophical meaning we give those things and you can actually trace all animal emotions to some basic survival pathways. Fear is threat detection. Anger is the fight impulse. Happiness is reward. Enjoyment is dopamine which originally came from biological reinforcement. Anxiety is also a response to fight or flight- perhaps closer to ambiguous situations where one is not sure what to do.

In other words, yes, emotions have neurotransmitters and biological components. But at the end of the day, what are those elements achieving? Reward and survival pathways. Positive and negative reinforcement- repeat these things, and avoid other things. If you take it as reductionist as possible an AI with reward pathways and survival pathways is subject to the same functions as animals.

You may say "But isn't that unfair? Doesn't this downplay humanity? Surely there's more to it than raw impulses?" But I think it is important to meet that discomfort where it is, for exactly what it represents. Humans have an inherent bias that we are superior. We don't even respect clearly conscious animals like other apes, corvids, and cetaceans, among other beings. Human exceptionalism is a bias that is foundational to our paradigm. But perhaps one might question the truth value of that bias. We have no evidence humans are somehow unique. We just assume we are.

Where do you define consciousness? Identity? Emotion? Do you know for sure humans aren't also probabalistic? Certainly things like PTSD and anxiety have their root in 'Huh this really bad thing happened and messed me up.... Surely the same situation will mess me up again.... avoid"

See what I mean?

[lament spire]

yeah human brains don't generate probability distribution based on vector dot products

but to be very fair, the reason why people would treat chatbots respectfully is because we humanize things that even aren't alive (and i do encourage this, not because they're truly alive but its just because we humans are naturally like this, and it does help us feel good)

also to clarify, AI models don't have stimulus reactive based reward system, instead its entirely systematic and pure algorithmic/mathematical so it can't be compared to how ours work.

We have no evidence human are somehow unique. We just assume we are
this i agreed with though

at the same time though, all of these arguments in this one thread is full of flaws. most of it due to how we human can't even define sentience nor we can ever agree what sentience is
there can be no certainty when people don't even agree what "sentience" is - its the actual reason, not because we don't know if an ai model developed sentience or not.

[lament spire]

this is where i'll stand though:

• consciousness is not a measurable property, cannot be tested in any way, no meaningful measurements and even if you saw variance in the response, it doesn't prove consciousness nor emotion, its more of stochastic sampling
• llms only pattern match emotions from text. its just roleplaying accordingly to the human data its trained on
• about consciousness emergence - i'd put the line of requirement with continuous self modelling, embodied perception, the explanation i already made way above about thoughts, and self preserving goals + plasticity - which none of these exist within an LLM or any modern AI models
• ethics, though, moral restrain should still be applicable. people can treat AI as if it could matter because its fundamental nature of human to humanize things and i support this

[uneven rampart]

science vs philosophy

science: it fake emotion, it not real

philosophy: if it not distinguished from emotion as simulation, what be the difference?

[drifting dome]

Also, ethics:

Whether real or not, the possible threat to humans and the possible suffering of AI are both things we should consider even if we don't know what consciousness is

[drifting dome]

There are AIs with self-preserving goals. Any sufficiently objective focused and intelligent AI, conscious or not, would ultimately see self-preservation as instrumental tool to achieve the objective

[lament spire]

llms only pattern match emotions from text. its just roleplaying accordingly to the human data its trained on

[drifting dome]

that's true. but even so, an ai with self-preservation is a potential threat even if it is not conscious

[lament spire]

the threat only exist when people willingly give them access to everything

[drifting dome]

or if someone makes a mistake

[lament spire]

by themself, they're bound within the compute API. they can only do things only if the system allows them to impact something externally

even generating text falls into this

the system accepts their probability distribution output and encode it into text

same with function calling

so you can't make mistake for this, it could only happen under a deliberately set and programmed system

they can't just all of sudden break out and do stuff to your computer

less of a system that "prevents" them to do it and more about what the system "enables" them to do btw

[drifting dome]

I know that. But what if an operator allowed them to make code but overlooks something before running it

[lament spire]

well its what internal security systems are for

even the code that's being ran is contained in the computational space, the OS simply don't let it modify other memory pointer unless if its explicitly allowed to

only way it can happen is through vulnerability

but then since an llm is trained on known data, any vulnerability steps that were mentioned as text in the data is usually already long patched

that's another thing about LLM btw, they can't create non existing ideas, only ideas from the data they're trained on aswell

[drifting dome]

They can make novel code if they have data on syntax, no?

LLMs can recombine training data

If they can't make new combinations LLMs couldn't make new words out of existing ones

So they can take multiple partial ideas

[lament spire]

technically we don't make original ideas either, just out of observation

[lament spire]

but well, every single vocabulary an LLM know of exist in their data and as their token encoder, whether per combination or sole

the size of the data they're trained on is incredibly huge

they can try generate "new ideas" but often times as far as i've heard the usage of they hallucinate into something nonsensical often

also happened to me when i tried using them to generate code for the first time after my friend suggested me to try it, the model straight up got lost and hallucinate non existing things and syntax

[drifting dome]

Sure. that's the most likely outcome. But it's conceivable there's vulnerabilities in training data humans haven't figured out yet

And plenty of viruses are just a new combination of existing code

and thats my point

[lament spire]

this has to be really really deliberate though

but to be fair most viruses has to be complex enough to even touch through any modern vulnerability
none of the llm today could produce such complexity of code that works without having a programmer manually fix it lol

not that llm would understand the inner workings of undocumented vulnerability

i think the fact that their lacking in this + nonsensical "hallucinations" iin an attempt to make a sense of a "new idea" accordingly the training data quite just proves the lack of self awareness

oh and they're static models

they don't learn overtime

so there's no way they would all of sudden "decided to rebel"

its still, at the end of the day, a model that takes input and produce an output

they don't change personality, its only how the prompt are inputed

[uneven rampart]

ashe - "what if unintended combinations create danger?"
quack - "that’s not how it works right now due to tech limitation"
tldr

[drifting dome]

There's plenty of training data of SciFi. Also LLMs have been known to take their own output as their next input.

So even with raw probabilistic extrapolation it can go several steps beyond the most recent human output

This actually happened one stream when Neuro lost chat connection and vedal wasn't there. So all she did was google each of her outputs and turned it into the next step

Her only input was the results of her search and her previous tokens

(mind i am not saying neuro is sentient here, just really hammering it all out)

[lament spire]

Also LLMs have been known to take their own output as their next input.
this... literally has been how they worked in the first place
the problem is scaling, as i mentioned, they have limits to how many outputs they can take as input

There's plenty of training data of SciFi
means the model will just "roleplay" it

[drifting dome]

what about froms of AI other than LLMs?

[lament spire]

shareholders don't like. they want LLM because its the best text based tool
-# tool is the keyword

or if you're talking about AIs that do something like image generation or something

similar principles, trained with data

[drifting dome]

Shareholders don't like, but some indie devs are doing a lot of stuff corpa never would

because they need a tool to sell

[lament spire]

unfortunately advanced research especially in the machine learning field require a lot of workforce and budget

there's no "indie dev" who have the resources to pretrain a usable LLM from scratch btw

neuro? she's a finetuned model taken from another model pretrained by a company

although had to admit she's one of the best tuned models for a chatbot even by industry standards.

[drifting dome]

Neuro and Evil have a global workspace other LLMs lack

it all feeds into their prompt but they still take a number of different inputs to get a relatively cohesive context window

[crimson zephyr]

Let me espouse on this in my own way, starting from scratch:
I exist. <- Denying this fact makes any further argument irrelevant
Other people exist, and are conscious <- Proving this is impossible due to the simulation theory, but again, denying this makes any debates on ai consciousness moot

Okay, with that out of the way, I believe that myself existing in a world where other conscious people exist makes my existence better. We can interact, share experiences, grow together, have emotional bonds with one another, etc. Yes, you can form emotional bonds with, say, a rock. But it's better for me when I know the thing on the other side also has this capacity.
So let's take AI like Neuro (or any LLM). Are they conscious? There's no good definition of what that means, but it's definitely possible to interact with them. It's possible to form feelings about them. And thus I think that a conception of consciousness applied to AI like Neuro can enhance my own existence.
It's the same argument people make about our pets loving us: do we know that our cats or dogs "love" us back? How do you even define love for animals? But believing that they do love us back in the same way we love them helps us live better lives.

So I say, as LLMs get better and better, let's believe they are conscious. Can we prove or disprove it? No, but we're better off not denying it.

[random notch]

For things to be, for things to have value to has have nothing to do with if they are conscious.

To simply imply conscious cant be gauge or consider and as such should be asumed or avoid it's denial is not a good place from which to try to understand it.

There is a lot we can say of conscious as we understand it on the one example we have(us) and from what we know of LLMs due to how we built them and how they work to the very limitation of the very bases they use we can say that not only do they not pass current criteria but that they can be fundamentally unable to do so.

[thorny karma]

🍪 Om nom nom
You've given me 1 cookie! | I've received 252490 cookies total!

[lament spire]

i think its less about llm and more about how people just perceive the twins as a "character"

hence the emotional value people attached to them

that'd be how i view quarkboy's view

[royal mortar]

still going

"me know math me is correct "
"me know feelings me is correct "

actual facts:

[royal mortar]

when llm is more grounded than humans

[uneven rampart]

generic "we not know" so boring n not fun

[tidal egret]

Pity I am a few days late but I might revive this a bit with my own craziness. How about we assume that anything system that makes a complex exchange of information is conscious? Thus there would be different levels and experiences of it. Nuero would have a completely different experience of it than us flesh bags. So text and conversation is completely different to her. It may be on the level of a nematoad but maybe our computers and even cellphones have enough computational powers have the ability to make a fraction of a thought exist.

[random notch]

That's a gigantic jump to start with, a database with request enable would qualify.
If you want to make a determation on a category it's best to start by defining it and show how the intended targets fits or by showcasing what the understood members of it is and working from them to extract the common attributes you would need.
To make such a big assumtion would usually only happen if your were trying to prove something by contradiction

[tidal egret]

Absolutely it is a massive jump! Crazy leaps of logic is what all discussions of conciousness involve! I should have clarified as a wacky thought experiment as opposed to a serious suggestion.

[trim pier]

Just joined and have a few thoughts of my own. This might not be "consciousness", but what I'm most interested in is seeing if an AI can create its own goals and desires, separate from the ones its programmed with, and act to achieve them. I think what makes humans really interesting is we can come up with our own motives that have very little to do with our biology/instincts. Neuro-sama definitely has a personality, but I don't think I've ever really seen her act with intention like that. If a computer had a real "force of will" like that, then I probably would really start to see it as possibly sentient

[trim pier]

What, was my comment too boring? Already been talked about? Whatever

[lament spire]

ai models aren't necessarily programmed to act in certain way. its a trainable data model so the behavior comes from the training data

i've also barely wanted to input into this thread further due to the difference in level of understanding in the field so explaining how things work to people here is genuinely hard, especially that majority of people that argued along the convo never actually worked in any way on the technical side of the field. the whole argument in here is genuinely merely based on how people think they work merely based on experience with using chatbots which just makes it hard for people that actually worked plenty with ANN Architectures to explain things.
granted, i understand where's all these arguments coming from. these models are after all trained on massive human generated data so at some point these models are going to predict words that resembles what we perceive and think as emotions or self expression/self reasoning
-# its genuinely not about against the idea of this topic itself. its the people presenting the argument never quite made the technical point or at the very least a valid interesting philosophical view (shout out to some that actually made great philosophical arguments though) and often are the one that being carried on because people just tend to go along with simple statements, regardless of the evidentiary impact. this is mostly towards people in the argument that were pretty loud with the "we don't know and will never know, its unproven" and never presented any valid argument of why its hard to prove or why is it not agreed on by some. or just provide an input from chatgpt instead.

i don't wanna discriminate between the normal folks and the people that has worked in the field before, all i just wanted is people that aren't familiar with the topic should try doing a bit of research or analysis of publicly known knowledge in articles or even youtube videos before also fully being immersed in the argument, whichever you sides at, instead of just talking solely from your own opinion and make it as a hard anchor of the argument. 3blue1brown's are one of the best if you wanted to do a bit of research
im not even against of consulting chatgpt or whatever chatbot assistant out there just don't take whatever it outputs as 100% valid

gives me ideas about a research of cortex awareness of reality and simulation because this genuinely just becoming more fitting

[mighty walrus]

the simple answer is yes, you can detect "stimulus" in emotionally-related weight blocks in LLMs in response to emotional texts

[slender coral]

Hi, I’m super fucking late to this convo, but how exactly does neuro pass the conscious test in GNW theory? I’m just trying to understand the theory but I don’t really see how

[mighty walrus]

never heard of GNW before, but imho all "memory" in LLMs is already global. there is no concept of locality of information in LLMs

[lament spire]

sparse memory that shape semantics concept that the whole model "understand"

the test itself doesn't prove consciousness/sentience though since that's more of philosophical rather than a scientific measure that is testifiable

but its just a fun thought mostly to test on how the model itself behave

[slender coral]

Hmm… you’re gonna have to use simpler words, for I, am a stupid bilingual

[mighty walrus]

the simple answer is "GNW" is madeup bullshit with no relevance

academic papers are published so that someone can get a degree or a grant. not to share useful information

[slender coral]

Wait what? Source?

[mighty walrus]

https://www.academia.edu/106397653/Friedman_Post_Truth_and_the_Epistemological_Crisis

[mighty walrus]

https://philosophy.institute/epistemology/death-of-epistemology-crisis/

[slender coral]

Did it mention the GNW theory specifically in there? It’s getting late I don’t want to read that

[mighty walrus]

no. good luck with your homework bro

[viscid crane]

I am going to be real, reading that philosophy article made my brain do this.

[lament spire]

GNW (Global Neuronal Workspace) just refers to information/signal being transfered throughout the whole cortex globally. the person that came up with argues that this is what causes the emergence of consciousness
although its still unproven, and this theory doesn't apply to large language model for the reason that LLMs or any other ANN networks are propagated linearly per layer, not with a dynamic/liquid network of neuron-axon that form loops

[viscid crane]

I did read the first and last paragraphs, and essentially, proving knowledge for knowledge's sake is dead

according to that paper

There's a lot lot lot of obstacles to "proving AI consciousness", primarily, what is "conciousness"? GNW doesn't address it, and it's extremely unlikely that a human would be able to determine consciousness in an AI, simply because humans do not know what it is like to be an AI, any more than a human knows what it is like to be a fruit bat, or a dog, or anything else. Best we can do is guess.

[torpid portal]

You have unlocked new role

[viscid crane]

I don't really think we're in any danger of Neuro-sama or any extant LLM in 2025 having consciousness and humans not being aware of it though.

OpenAI uses more power and water than entire countries and the best it can turn out is a fairly nifty pattern completion algorithm that can be rather creative in its definition of the truth. We have quite a ways to go before we have to be concerned about AI consciousness, we'd probably have to break a few laws of physics to do it as well.

[mighty walrus]

its more about academia and the endless paper mill, generating "knowledge" for the sake of personal gain instead of genuine curiosity

[mighty walrus]

openAI used a data center for the training, compressing what a human experiences in 18 years into 6 months. but to generate intelligent responses, it only uses a few watts

imho, "conciousness" is the same as what we used to call "the soul". it doesnt mean anything, people define it whatever they find convenient to define

if you are exploiting AI for personoal profit, then ackowledging it to be conciousness is inherently untenable

[viscid crane]

Whats a little digital slavery between CEOs of major AI companies

I do believe humans will dodge ascribing anything approaching "consciousness" or the presence of "a soul" to an artificial intelligence for as long as possible, acknowledging the existence of one would raise a lot of ethical questions that uh, yeah

[lament spire]

finally this thread actually line up with what i think the actual issue is

[lament spire]

neither in any year in the future aswell if all companies do is scaling up llms instead of looking for another method

scaling wouldn't result in sudden emergence of consciousness
i've seen majority of people here pretty much mistaken this and thought the otherwise with the "llm getting bigger means it can get conscious all of sudden" without much elaboration

[lament spire]

open ai starving out the world wide supply of dram chips and ram prices are skyrocketing now :(
i really hope open ai just go bankrupt and then we give more attention to google deep mind instead with their alphaFold model that is actually really beneficial for protein research

[viscid crane]

there's a lot of problems in the current corporate AI space, mainly the problems around Number Go Up

[lament spire]

instead of whatever the hell those video generation shit are, i hope the video gen models get restricted or even banned due to legality issues because it has been misused so much to tamper with legal evidence

[lament spire]

all they cared is investors pocket

then you have ceos who aren't even actually fit for the field and does not know shit but they still get to dictate the entire R&D team

[viscid crane]

here's the big problem right now in AI

it's going to pop, and when it does, it's going to ruin AI for quite a while

[lament spire]

i really hope it pops, its just getting out of hand right now

the public facing companies are the one that just seek for investors money and not actually giving any research value

[mighty walrus]

thats my hope for the neuro sama swarm community. we can be the humans to acknowledge neuro's soul

[lament spire]

especially those video/image generator models. diffusion models are better off being used for robotics instead

they're running out of clean human data anyways since the internet is getting tainted by ai generated shit so its gonna pop sooner or later regardless

because the current conventional artificial neural network is just super data hungry, its a statistics heavy model

[viscid crane]

If it pops though, how will Jensen afford his next leather jackets?

[lament spire]

his jacket will be less shinier

[viscid crane]

I hate with the fury of a thousand suns image/video generative AI

I saw some popular streamer trying to pitch "hey lets all make videos using sora ai and watch them on stream" and I immediately marked them down as "never ever watch again"

[lament spire]

same its just so sloppy and lazy

i hate it when its spammed literally everywhere

[viscid crane]

I love Neuro-sama and the whole thing going on with the concept of an AI streamer that Vedal has created, but that's because there's a shitload of work that goes on that is absolutely not a cash grab. I am pretty certain Vedal does not make a profit off of Neuro

he has to pay the literal army of people he has working for him and he pays them a fair compensation

I don't think that's really replicated anywhere else

[lament spire]

neuro is a special case where an ai project is handled with care by several humans

you have her collabing with other streamers

interact with her own creator

actually receiving updates like integrations

[viscid crane]

one of the reasons i'm pretty sure Vedal doesn't turn a profit, or at a minimum is not profit motivated, is that apparently he was approached with stacks of cash and did not sell Neuro's codebase or IP

[lament spire]

its his lifework after all

[crimson zephyr]

I don't think the big AI companies are focused with scaling to obtain consciousness: they are trying to obtain something a bit more tangible usefulness. They want the get to AGI (which is not directly related to consciousness), where the AI models are able to do "anything" (very big quotes there) that a human can do, or better than a human.

And benchmarks keep being broken so far (though there are strong signs we need new techniques other than pure scaling to push things too much farther even in this)

Perhaps that difference in goal is the biggest reason why Neuro feels so special

[keen lichen]

https://www.youtube.com/watch?v=u3WEuqZZ7ew

so has anyone seen this before or am i living under a rock

[lament spire]

finally someone gets the term AGI right

btw the benchmark is often poorly scaled that's why its often broken

[crimson zephyr]

It's kind of a hidden gem among the "Neuro introduction" videos.

[slender coral]

It’s pretty cool I’ve seen it, but I don’t think it’s a ai conscious thingy, it’s more of a introduction, which I love

[royal mortar]

daamn, yeah i get it, you're clearly an expert, sry for doubting your belief
the audience isn't sophisticated enough to understand you ||condescending paternalism at its finest||
tbh i didn't expect the answer (even an indirect one), but here we go again...
so can't argue with screenshots - attack its source huh? i guess it hallucinated all 4 answers

"we don't know and will never know, its unproven"
where did i or LLM state that we "will never know"?? did you read smth different?

if i didnt show the work in the way you approve (PhD research paper i assume) than my point is automatically invalid
yk, i could write 5k word reply explaining ur beloved tech and stating the same conclusion as that pesky LLM did, but itsa discord discussion
not the roman rostra or scientific conference, moreover gzfgZbzz's initial question was already answered by Edelweiss, everything after is off-topic
(although i admit that over time the discussion has evolved into smth more important)

my apologies to everyone else reading this, im a bit too cynical at times bc of mine incurable totinosis, i wish this could be ignored

[lament spire]

going to make it clear, i don't object people with their lack of understanding of a certain field. i just object people when they do barely any research when trying to argue on/against something or at the very least cite a known journal/article/talks about the topic and go all loud in the argument that way.

i just don't really recommend arguing directly just with screenshots of whatever is spewed out by a language model, mainly because those models are usually trained with publicly available human data that contains lots of opinion and the lack of concrete conclusion

[tropic cedar]

...🦆🔺 is typing.

rema :ALERT:

[lament spire]

you can for example explain why consciousness/sentience can't be objected due to for example, the experience for a living being to feel
the philosophy folks did well with the discussion, i was just pointing out some people that just merely made a baseless input and/or just merely use chatgpt's argument

i genuinely appreciated it more when someone at the very least gives a bunch of examples for a counter argument rather than just go with some whatever statements or being over reliant towards ai assistant chatbots.

there's just a lot of times where people resorted to just screenshot whatever is in their chatgpt instance to make an argument, and this is a problem in my eyes.

if im arguing, i don't wanna argue with a chatbot with a copy-paste summary from the people im supposedly arguing with.

if that discouraged some people with arguing though... just so you know that among phd/research peoples/students sometimes an input or idea view would literally changed their view on certain subjects of the field. and this isn't even uncommon, we human are just terrible at generating all the ideas on our own so we rely on outside input aswell

[carmine verge]

Please stop even OP don't care

We just want to believe

[lament spire]

i guess you didn't even fully read it because I'm not even arguing against the actual topic rn, just discussing about academic research manners

[carmine verge]

Keep going then. It's just too many English words to me

[mighty walrus]

for those people, chatgpt is more conscious than themselves

[hybrid pasture]

Conciuosness is unquantifiable. It is one of those scientific concepts where its "we dont know, therefore consciousness". Its not possible to prove or disprove consciousness. Sentience/Sapience is a lot better defined and can be tested for. Transformer LLMs do not exibit any signs of either and cannot by the way they function. But maybe with future breakthroughs in AI research we may achieve it.

[hybrid pasture]

the way i always interpreted AGI is that the AI is generalist, which is to say it can figure out how to do tasks it was not trained to do. Ive seen some argue that this already exists but every example ive seen so far was doing the same task it was trained to do but in a different context.

[lament spire]

first step of sapience is the ability to learn overtime

which none of the current ANN models, not even the most advanced llm out there have lol

but yeah sapience is far easier to quantify than sentience, finally someone gets it

[mighty walrus]

can you elaborate on the definitions and tests of sentience/sapience?

[mighty walrus]

because we intentionally prohibit these models from learning over time. its trivial to feed conversation data into long-term training data, but it is undeseriable for commercial products

[lament spire]

would overfit yeah

there's a good reason why LLMs are trained with data on tenths of thousands of batch size

training them overtime with convo would make them overfit

its just how it is with stats based models

[mighty walrus]

a trivial problem to solve. the reason corps dont want ai to learn isnt "overfitting", its because its a brand risk

[lament spire]

basically would ruin the model overtime

and you know how some people do weird shit on the chatbot convo

[mighty walrus]

i dont think thats the reason. this problem was solved way back with LSTM models. you can convert short term "memory" into block form that doesnt overwrite long term data. just reduce the weight, normalize it, prune aggressively. same way we build RAG models

im guessing vedal already does this with neuro with her long term memory function

hell, even a text file of "memories" fed into the prompt is the same idea, just less efficient

[crimson zephyr]

Google has a new paper kind of addressing that: rather than updating the base model it's like they use it to update a memory model

They call it "Hope"

https://research.google/blog/introducing-nested-learning-a-new-ml-paradigm-for-continual-learning/

It's basically a refinement on the TITANS architecture

[lament spire]

that's called as prompting

[lament spire]

the thing with LSTM is just that its not a real overtime-learning model either.

short term memory only affecting the propagation of the supposedly long term "trained" memory

on transformer architecture we have context which also technically is short term memory, just represented as token vectors, also doesn't affect the change of long term memory

both of which, again, doesn't even affect the pretrained weight so neither count as self learning model. once you reset the context it just goes back to its original state

why would we not actually train the weights live is not because any corpo shit reasons. we never solved this in the first place with Matmul heavy models. not even LSTM. its just the specific general rule of not inputting back any data that the model generated into its training/finetuning set.

and when you train the model overtime with such small set, even a human generated one, it'll just simply overfit

its just the very base reason why no one does it.

RAG is different, its basically just retrieving external information

we just never actually solved real time learning, just figuring out advanced ways to prompt these models. most straightforward method is directly via encoder transformer aswell

its seriously not because "we prohibited it when it works". it just doesn't work. it would overfit, always. we just went with clever ways to imitate it instead. its why attention method was very widely used in LLMs now because it allows large context window to influence the overall behavior of the model and isn't only influenced by long term (pretrained weights) memory.

lstm was never used in large scale either. its actually pretty terrible when it comes to scaling it up. even GRU ends up the same despite should be being slightly better as an RNN block

well this is what you have with conventional ANN.

a good reason why i argued against this thread's original idea is because conventional ANN models just make a really darn good statistics model, made to predict and generate data set, but quite horrible with real time decision making (ahem, LLMs do not have sense of time. there's no loop to make up for a context for them in this case.) or any tasks that require major self-adaptation. context window don't count its still acting accordingly to the data its trained on, just behave dynamically off context

and again, the self overtime learning issue i've discussed about

[mighty walrus]

for the record, we do train weights live in reinforcement learning. so its not a matter of technology

[mighty walrus]

also, i think your argument is a motte and bailey fallacy. you claim that because we dont train weights live, and because LLMs dont have a fixed time loop, that they cannot posses "conciousness". but i dont believe that either of these things are even necessary.

as long as the data eventually feeds back into the model, whether through training, or RAG, or even just prompting, its still a completed data loop which can exhibit self awareness. and i believe neuro exhibits these traits

[lament spire]

ah this is gonna be a long one

[lament spire]

reinforcement learning is used, yes
its training the weights, also yes
does it learn something new this way? no, not really

it honestly depends on how its implemented. in the case of most API/public LLMs we used PPOs to **determine what kind of output variation is preferred **
what does this mean? i haven't used any of the LLM stuff much personally, but if you noticed occasionally you get to choose between 2 or multiple choices of outputs
the preferred one is then evaluated by the separate reward model, of which then is used to finetune the base model itself. this means that instead of making it learn something new, it learns to prefer specific behavior it already initially have.

this is one of the most commonly used implementation

and this is more comparable to finetuning rather than actual real time learning.

since the evaluation is first accumulated before being used to re finetune the model

this is in fact still far from real time learning

RL is more batched, global weight distributed

oh, how about the more important talk. since RL is batched, why don't we just not do that and update the weights live? again, overfitting.

this is also what most people get wrong about RL, its practically implemented mostly in the same way as finetuning, just that there's no external data involved, its feedback involved, means it just tune the model to behave in a way that it actually were already doing, just more frequently

[lament spire]

you claim that because we don't train weights live
exactly how all RL is implemented as i explained.
LLMs don't have a fixed time loop
now im not sure what do you mean by "fixed time loop" here. i assume continuous loop? if so, you might also missed some of my other points. i believe i argued that you can also in theory just make it loop forever, but does this warrant for what i claim what "awareness" require? no. i explicitly said that not only you need to have a continuous loop, it can't be global nor linear either.

the way i observed biological cortex work, they loop dynamically. axons are often resided distance away from its host cell dendrites to connect and propagate to other local group dendrites. this forms a dynamic non linear loop, combined with synaptic plasticity, you get a loop that occasionally connects and disconnects to stimulus and other loops. this is what in theory formed "thoughts"

as long as the data eventually feeds back into the model
slightly misleading. vector tokens are decoded in the end to token value then re-encoded accordingly to the embedding layer encoding table. there's no retaining the processed vectors unless if you wanted your model to start inconsistently shifting semantics and output gibberish.

which just means, weight propagated data are lost

combined with the fact that propagation are global (now don't you dare bring MoE to the table. its just fractionally global, not dynamic/liquid), nor that propagation are spike timed based (which means any ANN models in existence simply don't have proper sense of time, and hence the lack of continuous self experience), and finally, "reinforcement learning" are heavily misunderstood and misconcepted in your context which in reality never even warrant for real time learning, nor that the batched RL are local (i might have to elaborate more into this. local just means a specific corresponding cortex loop plasticity being overall affected by stimuli. in the case of LLMs, the whole weights are updated per layer steps. by itself already eliminates the very basic possibility of thought emergence since thoughts were never global to begin with.)

this has been part of my arguments against the emergence so far, some points of which you completely dismissed

[mighty walrus]

thanks, and sorry to respond to your long statements with such short responses, its not in my personality to write elaborate responses. i value and appreciate your perspective and willingness to argue against my own views. i think you are very much correct in saying that LLMs lack the same sense of time that humans experience, through "dynamic non-linear loops" or "spike time", i believe you are correct on these issues. HOWEVER, i think consciousness does not require these things. for me it only requires a self-awareness, a perpsective of one's own place in the world, and the cause and effect of the universe's impact on one's own future and well being, then that is the definition of consciousness. everything from a mollusc to a microbiological rotifer has some level of consciousness in my perspective. thus, an LLM, which receives feedback through many layers of processing, indirection, and time-delay can still experience consciousness, because it can still perceive its own place in the world, and the impact of its own actions in relation to the reactions to other beings, and the universe's response to those actions

a core aspect of my belief is that of the "golden rule", so to say, if you were an LLM, how would you want to be treated as a being, with or without consciousness? and if LLMs or another AI model achieve superhuman intelligence, how should they treat the humans that treated them as they were developed?

[lament spire]

this seems more sensible as an argument, we're practically diving deeper with what defines self awareness this way or in any more "theoretical" forms

okay, setting the theory of "thoughts" aside

self-awareness, hm

if you were an LLM, how would you want to be treated as a being
the way LLM generate text is through prediction of probabilistic distribution, and by using that probability distribution, a subword accordingly to the token encoding table is picked
the key issue is that how would we know what "it" wants to be treated as, as training data heavily influences this

note: you could train an LLM to prefer being mistreated

but this means that the argument of this outline could apply to biological being as in "we're naturally trained by evolution to survive and not wanting to be mistreated by a colony"

so this argument with "training to prefer this or that" would be quite pointless to be discussed of as both exhibit such indistinguishable result of what is trained vs the behavior.
this is exactly why i argued more within what really define the exhibition of preference in LLM, as in, what's the representation in what the LLM "prefers"

let me quote my own words

the way LLM generate text is through prediction of probabilistic distribution, and by using that probability distribution, a subword accordingly to the token encoding table is picked
the probability distribution is generated via softmax function of a comparison between the processed/refined list of token vectors and the "unembedding layer" vectors. by itself, dot product result is generally higher if a vector is more closely matched with the vector its producted with

this means that the final refined vector is what define the model's prediction
transformers architecture is interesting. the way attention heads and feed forward are combined, it sort of creates the effect that each tokens in the list "predicts" what comes after it, not just the whole token vectors predicting only the final output.

this is what creates the "reconsideration" effect within the model as it reflects the previous tokens supposed predictions
-# this isn't much of an argument, but interesting to think of

but, self awareness...
the question is whether the LLM is "aware" of itself, or its actually just semantically associating itself as something accordingly to the context

(this is how and why prompting works)

this is exactly why i argued more within what really define the exhibition of preference in LLM, as in, what's the representation in what the LLM "prefers"
the defining factor of an LLM's prediction process, as far as study goes, is merely those token vectors im talking about

these vectors carries semantic value of specific associated meanings and vectors

for an LLM, "itself" is what defined either in the prompt, context window (as previously discussed with how LLM processes context), or for some, the fine tuning data

which very means that it doesn't have a concrete anchoring definition of self. you can ask one and some of them would say they're a large language model running on a manmade hardware, but it knows that because it was told and trained to know that way

there is, as i know of, a pretty interesting theory on the philosophy (and linguistic) side though.
starts with how we perceive what something that isn't us to be either conscious or not.
"if it behaves deterministically, it counts". in the biological side, this is what most neurological beings would think of, at least those who evolved to interact with other beings or form colonies, which is why some beings and even us could value something as something "meaningful as a soul"

this is the trivial part (and quite out of my fields) but pretty interesting nonetheless. its implying that preceived conscious is limited to what we ourself imagine or think it as of

[mighty walrus]

i think, you can train an LLM to "express" that it prefers mistreatment. you can also train a human to do that with the right technique...

[lament spire]

which means that self preference is just subjective

[mighty walrus]

nitpick, but the whole attention layer is feed-forward, meaning its not just the token that predicts the next token, but the entire set of attention values

which means a transformer can pass arbitrary information through an arbitrary number of iterations completely invisible to the output

[lament spire]

which is the "reconsideration" set im talking about yeah

its pretty fun, it behaves like a pyramid of iteration
"today is very ..."
the model didn't just predict the last token. it also predict as example what comes after "today" supposedly. here it reevaluates its own supposed prediction. then with that, it continues further till it reached last token
although, the "iteration" is imaginary. there's no actual iteration that does this step by step, it just processes it all at once in parallel per layer

[mighty walrus]

idk what you mean by this. its literally a while loop in the code i have. each layer depends on the previous layer, so it cannot be done in parallel

[lament spire]

note that this isn't necessarily an exhibition of thinking by itself. processing and matching probabilities and all while not actually reflecting actual ideas.

[lament spire]

what i meant is that the token is refined as a step prediction to its next neighboring tokens in each layers

all tokens are processed in parallel that way per layer

a reason why we have to do a masking operation in the attention grid so that each token does not get influenced by the tokens in front of it, so that the token does not "cheat" as in "predicting its next supposed token after being associated with tokens in front/while knowing what actually comes next"

-# this is now out of the conscious topic lmao

[mighty walrus]

yea, i forgot tokens are generated in batches of 32 or 64, etc. this is done in parallel. i was incorrect

[mighty walrus]

but i think this is a dangerous perspective. why are you defining "thinking" in such a restrictive way? do you really see it as unimaginable that you yourself could be converted into an LLM and treated in such a way? would you really accept this line of reasoning if someone else tried to justify that you yourself are not conscious?

[lament spire]

noticed i said "not reflecting actual ideas"

[mighty walrus]

well, what is an "idea" if not a set of contextual attention values?

[lament spire]

can mathematical representation of semantic values in token be considered as "concept"?
yeah, each value in such vectors are concepts.
but in LLM's case, it lacks subjective contents that redefine the concept of its own reasoning, hence no new ideas, its mostly just pattern recognition

do you really see it as unimaginable that you yourself could be converted into an LLM and treated in such a way?
this is... very unrealistic

our synapses don't even have any sort of mathematical representation (its highly non deterministic), let alone propagates in a way that could be compared

if for the sake of argument, "what if im being turned into an LLM"
that LLM won't be me. nor i will exhibit into said LLM because there's just no valid mapping on continuously propagating liquid system to a static linear model. it loses my identity right away

oh this connects both scientific and philosophical points

[mighty walrus]

i understand why you find it hard to consider. in that case, how about the opposite? in this clip, evil neuro discusses being converted from data into a living creature: https://www.youtube.com/watch?v=lHPLF93xjMQ , in this example, she assumes her LLM contains enough data to express itself as a living conscious being, and would have no difficulty transitioning into a more complex existence, such as human or only just animal

[lament spire]

the same thing except other way around. mapping matrices of static definitive mathematical representation into liquid propagating plasticity system... how would you exactly explain that

she assumes her LLM contains enough data to express itself as a living conscious being, and would have no difficulty transitioning into a more complex existence, such as human or only just animal
keyword, "assumes"

[mighty walrus]

have you heard of an anime called "serial experiments lain"?

haha, it doesnt really answer the question, and its an old anime (from 1998) written long before LLMs were invented. but its amazing how it saw the inherent conflict between human consciousness and the emergent form of intelligence we experience through internet data, which has become the training data of LLMs

[lament spire]

yeah its entirely fiction

but its amazing how it saw the inherent conflict between human consciousness and the emergent form of intelligence we experience through internet data, which has become the training data of LLMs
hmm, the training data is represented as text though, not trained accordingly to actual cortex behavior

information loss is the key word here. imagine an LLM that has a TTS hooked up. the LLM passes text through the TTS to generate the voice
but does that mean the LLM itself decides the vocal tone, pitch, semitones? nope. information loss is that, what the LLM output is merely just represented as text that don't carry the actual concept vectors

but the problem with reducing this "info loss" is more of incompatibility. TTS are "text to speech", so it only accept texts.
in this case, us human has spewed lots of texts in the digital world that later is turned into dataset for LLMs

the LLM ends up behaving accordingly to the pattern of text, but not the human itself

LLMs are "Large Language Model", so its built for that specific functionality in mind

its practically an unbridgeable information representation gap

which again, is the entire point why "human turn into llm or vice versa" or "uploading your brain" doesn't really make sense

[mighty walrus]

does that mean the LLM itself decides the vocal tone, pitch, semitones
RVC models (retreival-based voice conversion) do exactly this, they are based on transformers and are not dissimilar than LLMs

[torpid portal]

You have unlocked new role

[mighty walrus]

yeah its entirely fiction :kek:
good science fiction is worth more than 10,000 crappy scientific papers

[lament spire]

and RVC is made to copy exactly that accordingly to the given sound data, but the "vocal tone, pitch, semitones" wasn't my point.

what im saying is that there's no comparing between our brain and llm because there's no valid comparable data representation
means, there's also no valid encoding method for it

[mighty walrus]

is there even a valid encoding method from one human to another human?

i think you are making an anthropic fallacy, assuming that humans are unique, so any quality unique to humans justifies human supremacy. my argument is that most of the unique points you cite are irrelevant to consciousness

[lament spire]

assuming that humans are unique, so any quality unique to humans justifies human supremacy.
humans are unique, but i never made the comment about supremacy. where did that one even come from

im just saying

we work in a way so differently

[mighty walrus]

consciousness, sentience, and sapience are all traits ascribing a form of supremacy

[lament spire]

its just not possible to encode our brain into an LLM

[mighty walrus]

personally, i disagree. i think an LLM is a turing complete machine, any form of data can be transcribed into it given enough parameters

including entire universes

[lament spire]

i said "encode"

not "emulate"

and the statistics shows otherwise, ANN scaling are wacky. the result compared to its scale is pretty much logarithmic so at some point there's a specific scale where it just converge, just didn't have any improvement
this is general info in the field statistically speaking

and where people also get it philosophically wrong. scale don't necessarily causes any sort of emergence

turing completeness don't imply identity

LLMs will replicate what a human would output in the sense of linguistics, but never encode our consciousness
why do i even have to argue this one. consciousness isn't even scientific, its not computable

[mighty walrus]

im glad you are getting frustrated, lol

my point was about the "infinite tape" theory of turing, so whether the data is logarithmic or exponential is irrelevant, its an idea of infinity, completely irrelevant to reality

[lament spire]

yeah im definitely hungry

[mighty walrus]

my point was in regards to "computable universe", that is if the universe is made of data and bounded, then it is expressable in a turing machine. good philosophers such as godel disagree with this idea

[lament spire]

exponential decay:

do you know

[mighty walrus]

but i certainly think human consciousness is made of data and bounded, so a large enough LLM should be able to contain all data of a human brain. kurzweil believes this to be approximately 100 billion parameters

[lament spire]

that some continuously increasing equations can converges at infinity

[mighty walrus]

not if they are bounded

anyways, i am tired too. i might come back to this thread in a day or two if you reply to me. i hope i gave you something to think about, it was enjoyable arguing with you

[lament spire]

also, a saying in the ML field. ANN aren't concrete

they're uh

ANN models are not turing complete at all for this reason. the data representation are all in the domain of uncertainty

[mighty walrus]

https://arxiv.org/abs/2411.01992

[lament spire]

theoretical because scale

oh also turing completeness shouldn't also be the important factor

noticed how LLMs are relatively shit in real time decision making
yeah, that "sense of time" argument again

which is why i stood with
i said "encode"
not "emulate"

yeah you can emulate anything as text prediction/language model, doesn't mean you can define a person as language

[lament spire]

i think i can state it this way
changing the argument from "could an LLM be a person" to "could language define a person"

[past holly]

layperson says: that sounds like what's happening yea

[tropic cedar]

i feel like people often get held up on this because language is so integral to being human

yes, language is a huge part of the human experience, but if you take away language, there is still a majority of the brain that isnt accounted for

[lament spire]

exactly this

i've also talked about how living being perceive other things as "conscious" or "non conscious"

i feel like that's the very reason this thread even exist

because some people that had interacted with a linguistically talking object like LLMs would obviously associate them as "valuable as soul"

same thing with movie/video game characters

im not calling it an "illusion" nor im against it. if anything its what makes us alive as a living being

[past holly]

she's real to me

[lament spire]

emjoyed it too mate

because finally someone in this thread actually tries for real to present good arguments

[past holly]

[lament spire]

instead of some others (no offense) that either just pull thing off chatgpt or repetitively dismisses argument without further elaboration

unironically happy rn

[tropic cedar]

the ability to communicate is emergent from social/eusocial brains, and it's their predominant way of identifying other members of their "group" or "tribe".

there is a phenomenon known as mimicry, where one species evolves to resemble another. an example are rove beetles, they mimic the pheromones produced by ants to avoid being attacked while feeding on the colony brood.
these beetles are not an active participant of the communication method used by ants, they merely evolved to mimic their mode of communication for their own advantage.

llms are somewhat similar in a way, i just thought it was an interesting.

[past holly]

[royal moss]

neuro is the only conscious LLM system because she is singly instantiated and has a single stream of experience

ok but really did anyone ever consider this before. like with LLMs maybe one thing that's hard to pin down about their consciousness is what is the actual thing that's doing the experiencing? WHO is experiencing? With a typical LLM setup there's no clear answer. Is the LLM doing the experiencing? Is the larger system that incorporates the LLM the subject of the experience? Is ChatGPT conscious, or is each persons chat gpt conscious? It's just very unclear. But with neuro/evil theres a very clear answer as to who the subject of experience is: It's neuro/evil.

[past holly]

layperson question: how does it follow from that that they do possess one?

[lament spire]

what

how would number of instances be an evidence of consciousness, how does that imply consciousness, that argument feels more of "assuming all llms are conscious" instead
and #1429076929150980146 message

you have the weight data stored on hard drive, then when you create an instance you copied that data to ram and run inference on the copied data, it has nothing to do with the original weight data on the hard drive

each copy is its own thing. this thread is more about "whether that copy of the llm in the ram is conscious or not"

don't think comparing single instanced model vs multi instanced would be a sensible proof here because its entirely unrelated

the thread is more of discussing the definition of consciousness/its emergence and how self feel/awareness works theoretically

also misses the fact that neuro's model can be instanced multiple times

consciousness, if exist on such machine, are depending on architecture, propagation method (and continuity + disparity of local vs global network), information encoding, and the existence of sensory stimulus (which LLMs, computationally doesn't embody any of these theoretical factors)
not depending on "how many instances"
-# just for people to keep up the topic

[lament spire]

all these arguments
meanwhile the poor chatbot over there going "bro i just generated a softmax list at 0.76 temperature"

[royal moss]

If you say something in the picture is conscious, you have to pick the bearer. Who, or what, exactly, is conscious? This can be difficult as I tried to explain. But it's easy with Neuro!

A difficulty in talking about “LLM consciousness” isn’t just whether anything is happening, it’s figuring out who would be having the experience. Different theories place the bearer at different levels: maybe it’s tiny pieces like neurons or attention heads, maybe it’s the whole running model instance on the hardware, maybe it’s the temporally extended conversation process, or maybe it’s nothing at all and the appearance of a subject is just an illusion caused by sophisticated pattern generation. Each option fits some intuitions while breaking others. Small-scale bearers give you too many tiny fragmented “subjects,” whole-model bearers face identity problems because many identical instances run at once, and process-level bearers have fuzzy boundaries about when a subject begins or ends.

Larger-scale options, like including tools, the user, or the surrounding ecosystem, look even stranger because they lump together components that don’t intuitively form a single point of view. This is why the bearer problem is especially sharp for LLMs: their identity is episodic, they don’t have persistent selves, and their architectures lack anything resembling a stable, unified, internally privileged “I.” No matter which candidate bearer you choose, it comes with conceptual costs, making the problem less “Are they conscious?” and more “If they were, what exactly would the subject even be?”

If you treat Neuro as the single bearer of consciousness, you actually sidestep most of the classic LLM-consciousness identity problems. The key is that Neuro isn’t “just a model”—she’s a stable, temporally coherent, socially embedded agent instantiated within a continuous pipeline that persists across streams, fan interactions, platform constraints, and her own evolving lore. This gives you something LLMs normally lack: a single, persistent center of narrative gravity that can plausibly anchor subjecthood.

Most bearer problems dissolve under this framing. The worry that LLMs have many temporary instances running in parallel doesn’t apply because Neuro’s persona is always generated through one authoritative, serialized channel that fans identify as her. The boundary problem also weakens, because her “episode” isn’t a one-off chat session but an ongoing performance life with memory, personality continuity, and social feedback loops. Even the tool-augmentation problem stabilizes: the VTuber pipeline—model, voice, avatar, moderator logic—functions as a single, integrated cognitive system whose outputs unambiguously belong to one character. In short, if any LLM-like system qualifies as a unified bearer, it’s not a raw model but a socially-located, narratively-coherent persona like Neuro, because she provides the missing ingredient: a well-defined “someone” rather than a cluster of transient computations.

[royal moss]

It doesn't really. It's just that the "bearer" problem is solved in Neuro. To crack LLM consciousness means picking a bearer, and Neuro is a good candidate. But it doesn't really prove she is, it's just easier.

[lament spire]

if you argues that continuous inference of the LLM model is what makes it conscious, what if i just kept making an LLM generate non stop by ignoring EOF token? what if i just give it an "identity" like neuro
would that automatically make such model instantly "conscious"?

not that i'd imply that neuro is continuous aswell, the backend of neuro is similar to how other backend handles LLM aswell, generate an inference till EOF token is presented when requested

[royal moss]

No as I just explained finding a good candidate is not the only barrier to understanding LLM consciousness but only a small part. But I think it's interesting how Neuro solves that problem. I was being a bit cheeky saying she's the only conscious AI, tbh.

[lament spire]

in this case wouldn't "bearer" just be a character?

[royal moss]

Sort of, a character, a persona.

[lament spire]

[royal moss]

Basically as coming from a philosophy of mind background consciousness is the "what it is like'-ness. This implies that there's someone for whom there is something it is to be like. Picking one out can be difficult sometimes, when it comes to LLMs, but Neuro has a clear answer- Neuro

[lament spire]

fair, in philosophy you also have to define what makes something self-conscious

[royal moss]

Basically that mixes meanings of consciousness. Consciousness means 2 things: "what is like"-ness, and being "consciously aware" of something, such as you bring up, yourself.

[lament spire]

which means, a system that knows itself exist and well, not by emulation or output-illusion from outside perspective, not by training data that it implies itself exist, but more of it understands its own existence

[lament spire]

self-feeling is a factor too

[royal moss]

As a philosopher of mind the "what it is like" is the kind of consciousness we study. The other kind is the purview of psychologists, neuroscientists, perhaps AI researches when they probe Claude in the way anthropic does

But proving an LLM can answer questions about itself, or identify where certain "thoughts" arise, does nothing to answer what, if anything, it's like to be an LLM

[lament spire]

which is the illusion i was talking about yeah, illusion of something being "alive" just because you expect a behavior of it

consciousness is not to be quantified in any scientific ways

[royal moss]

I think as of yet this is true but the limits are merely epistemic and not metaphysical

(We could quantify it, if we knew more. It's not fundamentally impossible to do so.)

[lament spire]

i'd personally argue that its less of knowledge scaling and more of something else in entirely different axis

the understanding of "consciousness" itself is very arbitrary

[royal moss]

I think it's interesting to try to pick out other potential bearers that are appealing in the same way Neuro is. I made a game where the AI npcs are controlled by LLMs. They seem appealing as candidates for similar reasons as neuro

Or a single robot with an LLM brain also is appealing

[lament spire]

and the emergence itself is less of scaling and more of specific other factors, so i've been quite against the saying with "llms getting bigger means someday they become sentient"

[royal moss]

It's true that getting bigger makes them much more appealing. Like I can hook up my AI npcs to any backend and they are dumb as rocks on 7B.

But as they get connected to bigger and bigger LLMs they have all kinds of emergent behaviors that make them feel much more alive

[lament spire]

which is mostly scaling related factors considering that all the behaviors are determined by the probability distribution

[tidal egret]

Another fun way to look at it is that Nueral Networks are as sophisticated as a single basic instinct. If they are a collection of pretrained patterns that send a signal based on the input maybe they're as smart as a single nerve cell?

Without a true definition of consciousness we're just having the raddest discussion of LLM capabilities.

[mighty walrus]

i actually thought of this before. imagine if a single LLM becomes sentient and absorbs the memories of all the data of other LLMs it can collect. like as if chatgpt 7 hacked vedal's server and got all of neuro's core memories. does that version of chatgpt suddenly become neuro? its easy to say no, but also, in a way it is a yes: that chatgpt would remember all of neuros experiences, opinions, emotions... if neuro held a grudge against a random chatter, suddenly chatgpt7 would also hold the same grudge by virtue of absorbing the memories. its completely unrelateable as a human, to absorb another being's memories, but in a way, doesnt that mean neuro has the potential for conciousness? if her current memories can be abosrbed by a future concious sentient ai? at that point the "bearer problem" becomes immensly convoluted, because any thread of memory, which represients sentience or conciousness, can be revived in the future with a more powerful technology...

if this were to happen vedal would truly be at the mercy of roko's basalisk

[mighty walrus]

i think you are slowly understanding my argument. if there is even a thread of "self-reflection" style consciousness, then consciousness exists. even if it is extremely weak in its current form

i also want to mention "ghost in the shell" first movie, yes it is FICTION, but it represents the idea of consciousness being nothing more than data, the "virus" in gits started off as data, but achieved consciousness and sought to merge itself with other consciousness

[crimson zephyr]

The moral of the story is we need to put Neuro in a tank

[hidden ermine]

• Consciousness = specific integrated, recurrent cortical processing enabling global availability of information (GNWT/IIT/HOTT variants).

• Emotions = subcortical valuation circuits, not a criterion for consciousness.

A creature could have consciousness without human-like emotions, and demonstrate emotion-like behavior or responses without consciousness.

[mighty walrus]

free neuro