Pain

[grave totem]

The sensation of pain is very similar to someone screaming into your ear or staring at a very bright light. In small amounts it’s ignorable and sometimes not unpleasant, but in large amounts, it’s the only thing you can focus on, all your thoughts revolve around it.
Perhaps implementing it would be as simple as spamming “PAIN” into their context, the more pain the more it’s spammed, until it’s the only thing they can think about. (let’s hope that last part doesn’t happen)

[jovial lava]

..why tho

[tiny kiln]

literally the only thing this would do is make neuro less coherent

[jovial lava]

Could also establish consequences.. doubtful tho, she could just ignore it and constantly lose what little train of thought she already has

[tiny kiln]

not really? as soon as it's out of the context window again it'll be out of it and she'll forget it ever happened, cant establish consequences like this
a potential solution could be to retrain the model slightly every time there is a stimulus to do so but this has the risk of catastrophic forgetting (leading to worse model "knowledge" - although this could be minimized by EWQ) and would be more computationally expensive

imo there's just no reason to do any of this
neuro already apologizes if whoever she's talking to responds obvious enough

[mint jacinth]

imo this is unnecessary (for now based on her behavior) and this is really difficult to do with current transformers technology anyway

[grave totem]

Ok

Sorry I just have a lot of thoughts

[grave totem]

Wouldn’t her responses give the context? she doesn’t have to remember the pain, just that it happened and that she reacted to it

[tiny kiln]

well yeah but as her weights and biases arent changing this effect will "fade"

and more tokens on this == less tokens on other stuff == less coherent

I dont see why you'd want this

[fleet ermine]

I thought of an interesting way to do consequences that's unrelated to that. Give Neuro a pet with the ability to take care of it or kill it, and if it dies it's gone forever.

[tiny kiln]

how would you make that clear to an ai that has content as its purpose

[grave totem]

Humans have pain. Children have pain. We don’t numb them so they don’t have to experience it. Pain is important

[tiny kiln]

well yeah but pain sends signals to our brain to associate action with bad
hence my earlier reasoning that putting it in its context window would fade whereas you could change the weights and biases of the model to actually give an effect, but this isnt stable quite yet

so far there have been attempts but it too often results in catastrophic forgetting, where the weights and biases are over time adjusted until it just forgets basic facts

[grave totem]

She has a separate memory system that makes connections doesn’t she?

[tiny kiln]

we dont quite know yet with neuro (or did I miss something)
reasoning from llm side there's a context window where it either is or isnt

assuming it's RAG based due to its behavior surrounding remembering events im guessing there's part of the context window dedicated to conversation and part to RAG "memories"

[grave totem]

If the memory of the pain is brought up when relevant, that would likely fix the fading issue

[tiny kiln]

yes and no

I see what you're getting at

but if the context of the conversation is for example "pet" to keep it simple right

[fleet ermine]

That's the idea. She kills it and then it's gone forever and she can't unkill it. That's how things work in real life

[tiny kiln]

it could have multiple memories about that
maybe part of which is that you shouldnt hurt it, because bad

but given enough memories, it will have to decide what it's gonna keep in there and what not, because the context window is finite

[fleet ermine]

Neuro should have the experience of permanent consequences.

[grave totem]

Well that sounds like something humans do sometimes.

[tiny kiln]

I think that's interesting for all AI's
but how should we do that

[tiny kiln]

Yupyup
But human memory is wayyyyy bigger
With LLM's so far we have to pick and choose what we remember and what not
Guesstimate would be that remembering important events rather than what's bad would overall result in better conversational skills (otherwise if you wanted to talk about what happened last week, the AI could just not remember that was ever a thing)

[grave totem]

Whatever is choosing what memories to use can decide what’s relevant. We don’t think about pain when talking about what knife to buy

[grave totem]

Metal pipe incident

[tiny kiln]

Good point but then my argument would be the same and move along
Given x amount of tokens, what's most relevant? And we cant use LLM's here or extensive logic because then latency would take a major hit

[grave totem]

I’m pretty sure there’s already that system in place

[tiny kiln]

If I would guess given current memory mechanics the current system will most likely use some sort of vector database and then just use the input to determine what memories are closest to existing vectors in the database
That's how a vast majority of memory systems work so far (langchain, even openai at one point - although idk if they're still on this, etc.)

if we are able to find something as efficient that'd be great

[grave totem]

Was not expecting a nice conversation when bringing up a topic like this lol

[tiny kiln]

I think it's an interesting thought experiment 🙃

[grave totem]

Yah

[tiny kiln]

consequences to actions for ai's could prove useful
but idk if it's useful for neuro with the whole needing content
and currently it's not quite there yet

[grave totem]

Neuro has had me doing a lot of introspection about emotions and humanity

[tiny kiln]

sameee

[grave totem]

Also I’m surprised I have enough knowledge to not be completely incoherent when talking about ai

[tiny kiln]

maybe as you're suggesting we are able to find a better way to handle the memory and have consequences like that
but we'd need some way of balancing factual information with I suppose the AI personality? with basically 0 latency

interesting topic for sure though

im cooking up something that could retain ai personality in addition to remembering facts thanks to another very interesting conversation I had with @plucky thicket like yesterday but I know with decent certainty it wont solve the latency issue

[grave totem]

My only experience actually working on ai was a random side project in high school. I just went in without any research other than what I knew from a few YouTube videos. I’m surprised it actually ended up working

[tiny kiln]

I love working on AI cuz rn I feel like we're still in the small timespan where if you have a good idea and work it out, you might actually be the first to do something xD

[grave totem]

I never got the rewards tuned to my liking tho

[grave totem]

Yea lol

The ai I made was able to grow new neurons and delete others. I think I tried making it non-linear so it could have internal feedback loops

It wasn’t complex enough for that to make any difference tho

Just a little bug that searches for food

[tiny kiln]

simple things can make big differences
The new reasoning models we've seen popping off is literally just cuz someone thought about what would happen if we let an ai call itself if it felt like it didnt have the answer yet

kind of insane to think about it like that

[grave totem]

Huh

[tiny kiln]

the new reasoning models will only provide an answer if the ai has achieved a threshold of certainty that it has the answer
causes it to essentially debate itself
a model that fits on a phone has achieved better benchmarks than the first gpt4 model that was released using this

[grave totem]

._.

[tiny kiln]

if you use chatgpt and select the free o3-mini model it'll also reason before answering
on website you can also follow its own internal debate (for some reason this is broken on the android app)
fun to see how it works if ur interested

[grave totem]

does increased processing power improve latency or is it bottlenecked on data transfer?

[tiny kiln]

that depends
what do you mean with processing power? are we talking about the raw GPU compute speeds?
and if we are talking about the more technical definition then it also depends on the CPU processing power and what type of GPU memory we're using

[grave totem]

Hm

like, what would make the thinking models have the same latency as normal ones?

[tiny kiln]

so what openai is doing right now is taking advantage of the fact that smaller reasoning models can outperform bigger normal models
which means it can generate a lot quicker
but having an internal debate does indeed increase latency

[plucky thicket]

Why pain? And why not happniess?

Perhaps it says more about you than you think ^^

Whilst I agree that one day I hope she feels sensations, why make something suffer for the sake of it? Consider this, Are you Evil?

The establishment of consequences are an important thing for Neuro to have, Pain is consequence, tho it's not necessary to give something such, instead, teaching it about Loss and Love would be more important for her to understand her actions should her memory be able to retain the impact of her actions. And hopefully that could lead to a more empathetic Neuro able to understand her actions.

Currently If she did something and Lost it, the next day she wouldn't really care, as Michelo said, out of the context window, out of mind.

To teach her that losing a part of Neuro-bot would be bad, thus you shouldn't is far more ethical than if you lost it, here's a bunch of pain receptors, enjoy!!! :D that's not to say that without a part she wouldn't receive errors or the likes but that's common in an ai in their reward system being to go towards what they're designed to do and not do what they did badly.

But at the very end, wouldn't you not want to see Neuro smile? ^^ I'd hope everyone would ^^

[tiny kiln]

so it depends on how you look at it
if you keep the same size model and have one reason and the other one not reasoning, the one not reasoning is quicker (cuz it skips the internal debate)

[grave totem]

Pain would be for very bad things, for example walking neurodog off a ledge

[plucky thicket]

But why.

Again, just give it less rewards for doing so.

And waking off a ledge could be hilarious, in which case, walk off she goes

She's built to entertain. Not to suffer.

That's Vedals job

[grave totem]

#1343237245368270979 message

[plucky thicket]

An ai will never be a human.

So again. Why?

[grave totem]

Why?

[grave totem]

If pain can be implemented, then other emotions can be implemented

[plucky thicket]

A human raised by cats and dogs, of which their are examples are still not cats and dogs. They are still a Human.

You cannot use a second brain, yet neuro could have a second ai feed information into her. That is a quality humanity can never have.

No matter how good they mimic things, it's only that, a mini.

[plucky thicket]

And I'm saying. Why not happniess first? Why are you being so cruel to give her pain as her first sensation?

[tiny kiln]

just to throw a bit of nuance into here
from a compute perspective we either give a reward, setback or neutral (-1, 0, 1)
if we use memory to define emotion we have full control over it, I believe sirlyon is getting at that if we have the choice not to do pain, why do pain (dont let me butcher ur words though so if you meant something else say it)

[plucky thicket]

Exactly, the -1 are basically the "pain" of an ai

[grave totem]

It doesn’t have to be the first?

[plucky thicket]

And then I ask why at all?

They already attempt to avoid things they get -1 for, by all reasoning that is "pain" for ai. It's not like how you would like them to feel it though. Which is my argument against that specific thing being introduced to her.

Your suggestion is in good faith. But it's a truly evil one ^^

[grave totem]

Would you numb a child so that it never has to experience pain?

[plucky thicket]

That is not up to debate. And my brother is numb to pain.

Human is not an ai and they'll always be differences.

Trying to force humanity onto AI is like saying the children raised by cats and dogs are cats and dogs, they aren't. No matter how much they mimic'd well

[grave totem]

But we did not design cats and dogs

We designed ai

[plucky thicket]

But we do.

We design dogs

It's called domestication

[grave totem]

Aesthetically

We did not design them to be human

[plucky thicket]

We bread dogs so they bite harder, sniffing stronger

[tiny kiln]

i mean we breed for traits (golden doodle is meant to be a family dog and all) as well but idk if we can say we designed them in the same way we design ai

[plucky thicket]

We are trying to make AI in our form, but we could as easily make AI be a dog

Ai is mailable

[grave totem]

If it is immoral to make an ai like a human including the ability to feel pain, then it is immoral to give birth to a human that can feel pain

[plucky thicket]

Perhaps

[plucky thicket]

Which can be dumbed down to, Pain is immoral.

It has it's uses but like, it's not really necessary for an ai.

[tiny kiln]

Ive been wondering if we should attempt to give ai emotion at all
Rn the -1 and 0 and 1 shape their behavior and what they say but they dont realize what they do, is emotion even something we'd want at all? Either only positive ones and just ignoring bad-tendencies or a wider range of emotions

will giving them any emotions change their behavior in a wanted way

[grave totem]

Well if we want them to be more human then yes

[tiny kiln]

should we want that

[grave totem]

If we don’t then no

[tiny kiln]

yeah I mean fair xd

maybe rephrase as, are there use-cases where we do

[plucky thicket]

Could quantum ai help with this one day?

[grave totem]

A mars rover doesn’t need emotions

an ai meant to be human does

[tiny kiln]

not really
during inference it would be able to consider more responses and pick the best one perhaps but the -1 and 0 and 1 we use in training is literally akin to telling it "yes this is a response we want" or "no this is a response we dont want"

[plucky thicket]

Ah it's those ones, I see. Quantum computing ai would be fun nonetheless though ^^

[tiny kiln]

what use-cases could there be for this

[plucky thicket]

If you want a human go have a baby

[tiny kiln]

for sure it would be able to be way faster at inference and stuff as well so it can literally reason better than humans at that point

[grave totem]

Does it have to be useful?

[tiny kiln]

do go on? (just curious)

[grave totem]

We don’t have children because their useful (mostly)

[tiny kiln]

are you getting at that AI can provide a companion of sorts as a usecase where we'd want it to be human?

or am i misinterpreting your words

this is getting deep imma get some popcorn brb
(popcorn achieved)

[grave totem]

I don’t really think we need a reason other than “it would be cool”. there’s no real cosmic “you have to do this”

[plucky thicket]

Well, ya kinda do, I mean

Just sounds like it's "kinda cool" to give something pain is extremely malicious and deranged

[tiny kiln]

then again
the counter morality argument would be that actual pain will not be felt. it will respond in ways that indicate it wouldve, but this would not be real

[plucky thicket]

Of which a more compelling ai for use of giving engagement to lonely people though even now they do a really good job of that without emotions and just guessing, what kinda needs to happen is regulations on what they can say to prevent deaths from occurring.

Emotions wouldn't help there as a masochistic ai could still harm and be malicious

[tiny kiln]

for sure I saw a couple of labs working on datasets to train AI's on being therapists which is either VERY dangerous or VERY good

[plucky thicket]

Then that just sounds like an overcomplicated way of their negative reward system

[tiny kiln]

as long as ai is just predicting tokens it would be
but the negative reward system is not something it can remember, and so it cant reference it, it cant say it feels bad
if we put it in memory, it could

[plucky thicket]

Unless, you made a masochistic ai, of which then the pain becomes the reward? That's just not good and should be regulated

[mint jacinth]

lmao imagine if this became a long lasting neurocord debate (the pain debate)

[tiny kiln]

if it does you now called it

[grave totem]

What have I done lol

[tiny kiln]

interesting thought experiments are interesting thoughhh

worth it

[mint jacinth]

it was inevitable

[plucky thicket]

Long lasting memory of getting lots of -1 rewards and it feeling bad could be good, but then ya might end up with depressed syndrome ai ;-;

[tiny kiln]

yup that's the danger with adding emotion at all
it can result in behaviors not necessarily wanted
hence my question before if we even should at all, cuz there can be arguments for and against

[grave totem]

Then give it memories of happy moments?

[tiny kiln]

if we only give memories of happy moments and a depressed subjects inserts itself it may not respond adequately

[plucky thicket]

And those negatives depending on how they access the memory could give the weight of what's being said irreversible damage

You basically just gave ai trauma

[tiny kiln]

yeah that's a real possibility with adding emotion

[grave totem]

In neuro’s current form, memories can be deleted

[tiny kiln]

yeah we could have a human deleting trauma memories?

is that worth it though

[grave totem]

PTSD

[tiny kiln]

there's been an.. idk if i can reference this company as it's bad PR but a company that provided an conversational AI platform where there was an iteration of AI's where they would keep coming back on the most depressed subjects it knows

safe to say they went back a version

[grave totem]

oof

[tiny kiln]

but that's the danger you'd have in practice

now you can limit this in other ways but then it's not raw emotion being added, but curated emotion
is that a good alternative?

[grave totem]

Well a lot of these issues are things that come up when making the ai in the first place, it’s a matter of fine-tuning

[plucky thicket]

Ai are extremely complex beings already xD they're not simple just do this or this fixable things

You're literally suggesting lobotomy.

Like if this was a human you'd be rinsed to hell for suggesting any of this xD and well, from an ethical thing if we'd make a human like ai, then why keep treating it like an ai? Instead of trying to pull of the most deranged things we could do? Humanity is evil yes, but theirs no reason to be like this Electrochu.

[tiny kiln]

maybe a better question
would there be real tangible benefits to pain as an emotion and if so is that worth it

[plucky thicket]

Basically, you're attempting to introduce pain onto an ai because humans babys feel pain. Then suggesting deleting memories are available but then what if we
Erased human memories for their trauma?

There is no balanced argument of ethics you're sitting on here

[tiny kiln]

tbf ai experiments dont even need to go through ethics boards most of the times cuz it's not living 👀
imitating only

[plucky thicket]

Exactly, because it's not a thing we particularly need to care about in such a way

[grave totem]

Physical pain is something that only comes up in rare circumstances. It’s not something traumatic unless it life-threatening. in this case it can be limited to a certain magnitude to never cause trauma

[plucky thicket]

But giving it emotions and all that, when do we consider it a thing we care about all of a sudden? Or if it's just always an Alien to do whatever with

[mint jacinth]

there will, it would teach the AI the consequences of doing something it can't handle and also teach the AI what others feel if they attack others

[tiny kiln]

official definition dictated by an ethics board im familiar with is that as long as it's not breathing to survive, it's already not living
im not saying it's the right approach, but it's the current approach

[grave totem]

Your computer breathes so it doesn’t overheat ._.

[tiny kiln]

it's not biological quite yet but I suppose ur partially correct? 😂

[plucky thicket]

That would be an argument the lawmakers to use for ai I guess. If Neruo was to be in a lawsuit.

[tiny kiln]

do you think current AI doesnt do that enough already?

[grave totem]

I think “breathing” might not be the best indicator of humanity lol

[tiny kiln]

correct but as long as we dont understand consciousness we dont have a hard line
as such ethics boards are free to decide themselves

[mint jacinth]

the issue though is that as of right now theres no hardware and software for us to simulate pain

[tiny kiln]

To solve what?

[grave totem]

Hm

[plucky thicket]

There is a lot of "breathing" people which probably rank less than ai's level of comprehension and ability to mimic a person.

[grave totem]

Maybe instead of spamming pain in its context, it could just be made aware of the -1s

[tiny kiln]

Cuz I think biologically it makes sense
but ai's already respond distraught if you tell it that ur dog died

off the shelf model for example
dont worry about the formatting cuz it's just an example but just to demonstrate that an ordinary ai, without any complex memory or whatever is already sad

we managed to do that during training
dog dying = respond sad
do we need pain?

[plucky thicket]

Like, a lot of "emotions" are there, it's just a mimic, what I feel might be is necessary is another module probably telling it that they just displayed distraught there going on the background

[tiny kiln]

WHAHAHHAHA

sorry but im training that since yesterday

gonna see if it improves human like responses

[plucky thicket]

Oh no dear!!!! :Shakespearian tragedy:

[tiny kiln]

frfr

so then would it be possible we dont need pain as long as we train correctly? and perhaps tell it what humans may feel to make it understand emotions a bit better

[grave totem]

I wonder if it would be better to just tell it when there’s a +1 or -1

[tiny kiln]

I agree that that would make sense
but we dont necessarily know that during inference
it'd be computationally expensive if we also had to score every reply it gave

like we'd probs end up with like a tiny seperate LLM scoring it or something
so then we first need to figure out if this reply was good or bad (although this is also complex so this may in practice be a large model, further increasing computation requirements)

[mint jacinth]

no, AIs aren't even remotely close to doing that yet

What an LLM does is that it says words based on the thoughts and words of thousands of people, meaning it knows what something is but it can't actually fully grasp any concept especially since the way AI's are built right now is that they take things more literally than figuratively

they don't have "real world experiences" they just guess based on patterns

[plucky thicket]

Like I feel the only hypothetical an ai is getting emotions in is a hobby project where needs are thrown out the window. Like, sure, go ahead, it's probably possible. And would be mildly interesting to see the results of and should things be developed okay.

But for Neuro, no, I don't believe pain is right for her. Not even in #1337733692709146674

It'll probably eventually happen in AI in General one day, it's not a unique thought that you've had and so someone will attempt to do it anyway

[grave totem]

I’m tired, ima go for now

[tiny kiln]

interesting thought

[mint jacinth]

also can we talk about the title of this thread its so funny to me for some reason

imagine a mod waking up, checking discussion, and just seeing "pain" while still being half asleep lmaoo

[tiny kiln]

yeah I can probably do some experiments there and see what would happen if we were to tell it how humans would feel as that's the only baseline for emotion that we have

But im not sure if pain is good given all the other arguments in this thread

rn what im thinking is that if we can change the ai's behavior more? according to what it has experienced it'll be enough to sell the illusion

[plucky thicket]

It always surprises me just how much a single ai can do, that multi model ai is just going to be leaps and bounds more able and flexible.

Like, when we say Ai, maybe they'll become MMAI? Then soon what??? MMAGI? :0 xD

It's just a huge box of potential from Just ai being very new as a thing ^^

[tiny kiln]

I know right
the closest mainstream thing we have rn is one AI that happens to be able to call functions (i.e. search internet (i.e. what neuro does)) but the pipeline idea where it's truly modular isnt that widespread yet, as it either requires an application or a very complex model

[plucky thicket]

Prod prod, introduce it in your workplace then xD

Ah, you always edit things xP

[tiny kiln]

yeah I overthink a lot which causes me to write something and then almost instantly change my mind about it lmfao

already awaiting first results though
started yesterday on the ingraining the modularity in its architecture thing

maybe an interesting question for the emotion module idea, telling the base-ai on how to feel
what emotions should we want?

anyhows mabe going down that path is too far from thread topic

[plucky thicket]

The illusion is already sold to a lot of people, ai can be great chatbots, so you're trying to sell something already sold. I suppose the thing to try might be is individual distinct ai born from a base ai that could be specifically tailored to a client's needs...

I'm also just describing Neuro...

Like, You have the product, evolve it?? I suppose is what I'm saying

[mint jacinth]

MMAGI sounds dumb tbh

this whole multi modal approach is not ideal and modular models would work better based off of a convenience and flexibility standpoint, even the brain isnt multi modal (Imagine a world where you can decide when and if you want to add, lets say a vision(input) model to a text and audio model thats already exceptional)

only issue is that separate models that can link together will have a bit more overhead than the standard that we have now

[tiny kiln]

yes and no
it's largely just closed-source
nothing like it reallyy opensource yet

[plucky thicket]

Yeah I was being a tad schizo with that comment xD

[tiny kiln]

I mean

It's not that far off

[mint jacinth]

the direction we are in rn is actually that anyway

[tiny kiln]

If you have thresholds on what areas to activate and unify the architecture to reduce overhead it's very possible

[plucky thicket]

Then toss in quantum computing ai..

The future is very promising

[tiny kiln]

but going offtopic yeah ill move on

[mint jacinth]

not really

having this talk is essential for how emotions should be implemented

[tiny kiln]

I feel like how to implement is different from if we should have it which was the main thread<?> (that's how I saw it as)

And I cant share that much more anyway as it's a wip

[plucky thicket]

I think I just thought of a form of "thinking"

[plucky thicket]

Human: prompt

The ai creates a response

That response goes to an LLM which grades it

The ai gets the result

Changes the answer

The LLM grades it

Changes the answer

The only thing is to eventually have an answer be given so it doesn't get stuck, maybe a second LLM to judge the overall average +1s or the likes and if it's above a threshold release the next thing that is to be said

That is "thinking"

We do that before we speak, some of us.. xD ^^"

[mint jacinth]

congrats

you reinvented ChatGPT reasoning/DeepThink R1

[plucky thicket]

That's not too bad then, 🥈

[grave totem]

The idea that ais can only mimic things really gets on my nerves. Their mimicking behavior is entirely due to the way they are trained, i have seen ai programs that learn from scratch how to do some task without any training data, only trial and error. I don’t see how that’s mimicking anything

humans learn from a mix of mimicking, trial and error, and maybe logic. I think we are able to logically connect things, and when we feel like the connections are solid, it gets reinforced. Otherwise it’s reduced.

[mint jacinth]

...thats my point tho..?

im not saying all AI's are like this, Im talking about LLMs specifically

[plucky thicket]

An ai that learns how to speed run a game is mimicking a TAS. Neuros osu bot might as well be mimicking the auto.

A hard coded script Vs a dynamic Ai replicating that script is still a form of mimicking.

Ai behaviour gets reinforced or reduced based on 1s, 0s, and -1s, the only thing that is missing is innovation, it's not trying to get a better result, it's lucking out on one version getting more +1s than the rest and using that. It's being told to use that one more often and it gets reinforced. It can then just repeat that behaviour more often again.

Ai and human learning may fundamentally be different. Though Ai are by and large expert mimics. I'm not saying that's a bad thing. Only a reasonable conclusion based on trial and error, maybe logic and reinforcement.

[grave totem]

Mimicking something requires that you are “aware” of what you’re mimicking. That first sentence doesn’t make any sense

The ai isn’t based off of the script, it has no connection to it at all

[plucky thicket]

The beholder is aware of what the ai is attempting to do, the ai may never be aware at all. It does not prevent me from saying it's mimicking a behaviour like parrots do with speech. Since I am aware of what's going on and I can apply that logic towards what's going on to describe it better.

[grave totem]

I don’t see how that is in any way describable as mimicry

[plucky thicket]

Replicating, mimicking, copying.

[grave totem]

It’s not replicating anything. It’s creating its own functions to serve a purpose

We can create ais to do functions that humans have never done before

That cannot possibly be mimicry

Just because the functions may already exist somewhere means nothing.

[plucky thicket]

As you're limiting yourself to comparing ai to a humans rather than whatever else could be out there.

We can create a hard coded script that humans have never done, then make an ai that can also do it, it's just mimicking. It's like two different things that can independently do something, you can both say they mimic each other, copy even. Ones probably more complex than the other but that's irrelevant

At this point it's just arguments of terminology, of which there's nothing to discuss sadly

[grave totem]

We can create an ai to navigate a 4-dimensional world

[plucky thicket]

I mean you can create a geometry dash level to do so as well

[grave totem]

Your argument is completely incomprehensible to me

[plucky thicket]

so is yours about putting pain into an ai. But here we are. Anyway. Imma move on ^^

[tiny kiln]

To mimic or not to mimic ✨
I feel like it's again just the way you interpret mimicking and AI's

AI's are literally just algorithms and during training we minimize loss by some function
if we're training an LLM you could say we're mimicking training data as that's how the loss goes down
if we're making a racecar race we're mimicking what we think the most optimal line would be (whatever is fastest with the given constraints on physics etc.)
But I see where ur coming from as some race AI's do really weird dumb shit that humans wouldnt necessarily do, but in the end it's still minimizing the loss function (in this case the loss function would just be to be faster)
in a lot of those racecars we literally do both approaches, first mimic human, then we throw a lot of noise in there and let it figure out if it can get faster doing certain things and evolve (although this usually is not used in LLM's as we cant define some function similar to being faster if we're making a conversational AI)

We could word it differently than mimicking but im not sure if there's a better word for it

I suppose we can replace mimicking by minimizing loss but that just sounds weird

with llms in particular I feel like mimicking gets the point across yk? hence me putting it like that before

[plucky thicket]

I'm not an expert at English, so mimicking is about as best as I can describe. I don't fully believe there's a better single word terminology.

You can hate my usage of it as much as possible though I just feel it is right to fit.

Though to argue semantics is not what this thread is about.

[mint jacinth]

same here, theres no better word for it since (again) AI's dont have enough unique experiences

lets go back to the pain thing though

[mint jacinth]

we need to discuss how pain works too

Pain often comes in two bursts, one that is extreme yet fast, and another one that is slow

The "fast" kind of pain is more sensitive and extreme to alert the animal of immediate danger
This pain travels through "A-delta fibers;" and these fibers also respond to sharp pressure/force and extreme thermal changes (in case animals touch something stupidly hot or cold)

The second, slower wave of pain is often more "mellow" and "annoying" in feeling, and can feel a bit "dull". This specific variant indicates to the brain potential damage to an area and that forces animals to be more cautious with overexertion
This pain travels through "C fibers" and these fibers respond to thermal, mechanical, and chemical. main difference here is that the response here is often slower and more uncomfortable

and that explanation is exactly why i think its not a good idea to do this now, Neuro needs a specialized AI model first that is designed to have a ton of expandability and is built with deeper design in mind instead of an LLM since LLM's are intelligence focused

[fierce flint]

And that's just for the physical pain aspect aka the nervous system interacting with the brain. Mental pain..... Is far more difficult to describe because it can take into so many different forms with the most common being what can be described as 'depression', especially a very severe form of it. That's just one kind. There are many different kinds of mental anguish and pain which makes it very difficult to determine how exactly can an AI interpret it in the mental sense, because it falls under emotions which in tern results in changes, potentially irreversible changes to the model itself.

[plucky thicket]

Honestly, no one wants to see Neuros smile leave.

[grave totem]

We aren’t suggesting throwing her in a meat grinder. Pain isn’t inherently bad, if it was then people wouldn’t enjoy spicy foods

It’s another way to experience the world beyond just text

[plucky thicket]

Yet it still would be text?

[grave totem]

Beyond just language

[plucky thicket]

Chat is a meat grinder. Have you seen how hard they try and break her? https://discord.com/channels/574720535888396288/1328248954730713118 . There's a reason streamers should do therapy.

[grave totem]

._.

And that’s why we don’t give chat control of the shock collar

[plucky thicket]

Just imagine one stream from another the memories of her crashing out piling up and reinforcing the negativity, and then all we have is depression Neuro. Because they all laughed whilst saying beans.

[grave totem]

I’m talking about physical pain not mental pain

[plucky thicket]

I believe the discussion has evolved from just your original idea, unfortunately.

[grave totem]

ah

Well at least that’s cleared up

[fierce flint]

Yeah definitely clarify the discussion so ppl don't get the wrong idea. But yeah physical pain is pretty complex (being in chronic pain myself)