(CW:AI) Discussion of LLMs and other kinds of text generators ("AI"), especially of toki pona

[glass iris]

There isn't enough toki pona text (of good quality) available to train an LLM by itself, I think. And it has been shown that LLMs benefit from being trained on multiple languages - even if only one language is actually used.

The license of lipu tenpo already allows the text to be used to train AIs, as long as the origin is somehow indicated, I think. Whether that's a good thing is another question!

[severe dagger]

woah that's new info

the training on multiple languages thing

[glass iris]

I'm looking for a reference for this - I know there's a study out there.

[unique plume]

I'm not gonna yap about us not necessairly needing text data for conversations, I'll just say that I've made a tiny tokipona language model and it yaps random nonsense.

80kb of data, 32 hidden neuron LSTM CBM

Not like I'm proud or anything, just what if someone wanted to translate the things that it says, or have an idea of what to do with it

It wasn't high-effort ofc, but it can say whole sentences that make sense.

[glass iris]

Can you give a couple of examples of its output?

[unique plume]

By memory - "mi moli la jan pona li moli"

Found a recording of it's output from an older version, just a chunk of a sentence:

ona li awen supa lasi pi jan kowinta o,

It fever-dreams-up things forever, could be a nice but expensive wallpaper

Any ideas for it's use?

I just haven't seen a proper tokipona-only LM

[high kayak]

i mean any sentence is "good toki pona" while lacking context

mi moli la jan pona li moli
this sentence isn't grammatically incorrect but without further context it's not really a useful metric

it might be better to make a deterministic machine for short stuff at that point
like hard code the grammar rules and words, seperate words into categories (pronoun, content word, preposition, particle)
and then just do randomly generated text

[unique plume]

Not that interesting

And about the grammatical rules - Then I need it to have data about conversations or some environment... I don't have any of that so I can't use hard-coding grammar to make it that much better

I'd need it to be in some kind of game environment, then it would be the absolute perfect NPC if I can model that environment at least approximately

There's no problem to translate between the language of the machine and tokipona, except there is a problem when the machine itself has no context about which environment you're even interacting with... Some remote place with horses?? Chess?? Medieval setting?

Otherwise it would be mid. It needs to know what it needs to look for.

I don't think that there's a context marker in tokipona...

Like "about [the matter]" in English.

[unique plume]

Because it would be enforced for this model to stay performant and efficient

[unique plume]

"The tree is floating in the air" is totally false, unless the setting is in minecraft, where it's most likely actually true. If it never sees the mention of a setting (which people can just recognize because they lived in this world already, and know what's true and what's false) - it would reuse the same thought, so "get away from it", or after learning it would just be confused and say 50%/50%

[unique plume]

People being able to just recognize things like "Nah this is obviously false, unless you're like... in minecraft"

If a model has never seen the reality, and has never seen minecraft, then both statements make no sense for it because they contradict each other

Basically - machines speak in absolute values

Well the part for math is easy tho...

Get a SAT solver installed and translate the question in tokipona and there is the solution.

Except... When we need all of the valid solutions, which isn't that unreachable either

[high kayak]

no i'm saying like, just don't have it use/be an llm
and instead hardcode a system for random sentences
im not knowledgable enough about llms to tell how to hardcode stuff into it

[high kayak]

I don't think that there's a context marker in tokipona...
i can't tell if this i sarcastic or not, cause toki ponas context marker is like, the most important part of the language

[severe dagger]

eh

la is overrated

[high kayak]

lmfao absolutely not

[severe dagger]

hold up

#toki-toki

[glass iris]

la is literally the word that opens up the possibility of complex ideas and expressions in toki pona.

[unique plume]

Nah I 100% got what you meant, just if this model even had any purpose - I'd do smth, but the random sentence generation isn't any more interesting, It's 100% predictable

[unique plume]

What is it? I'm serious, I don't remember such a word being in the dictionary, especially it being of any use

Context is important here yeah, the problem is that humans have it and the machine can't figure it out easily enough

[unique plume]

True

It's for if statements, not the context marker for the entire discussion

The previous statement, not the domain of the entire discussion

[high kayak]

no it's not for "if statements"

[unique plume]

You can't say "la manka" to comunicate that we're playing minecraft

[high kayak]

you couldn't say "manka" by itself in general cause that's not how toki pona works

but you could definitely say "musi Manka la ..."
musi Manka being explicitly marked as the context

[unique plume]

So you just make "la" at the end of the sentence?

[high kayak]

no cause that's not how toki pona grammar works

la is a context marker, unambiguously so

[unique plume]

So we're literally back at the start, nobody is going to write that for every single sentence

[high kayak]

i think you just have a weird definition of what a "context marker" is

that excludes toki ponas context marker

[unique plume]

I'm not a terminology nerd at all, for me context here is any factual assumption, so information outside of the sentence that is important

Tokipona has a sentence for context

I'm talking about the whole domain-kind of context, like the setting, common sense and all that information that the model doesn't have

English has one, and not only one, but it's not used often

[high kayak]

what's english's context marker?

[unique plume]

I think the most frequent one is "You know how …"

It's not a word but we immediately unambiguously mark what we're talking about

[high kayak]

yea but like
that's just saying something, that's not really a marker
you can do the same thing in toki pona by starting your whole paragraph with a la-phrase

or like
mi wile toki lon ni:

[unique plume]

Good point tbh, I was getting mixed up in the parsing

Except how do I even know which sentence to use to classify the context???

Probably keywords

Ok now that a cognitive bias is cleared, there's something more promising.

I should've accounted for making mistakes when I have no empirical data to ground any thoughts.

[unique plume]

And now my mind's system1 (more precise psychology term for subconscious) starts to say stuff like

"Literally make it a prompt for the model"

System1 and system2 are such useful terms, wild how this has been in the shadow for so long

[unique plume]

Except that makes it way less fun...

So manually coding the detection of which setting we are talking about is required.

Concluding that all such settings must be defined and have some command to switch to them or to define a new one...

Assuming that we have a setting, we need to translate everything into correct and incorrect information, to let the model adapt and learn what's nonsense and what's true, then just thinking about somehow potentially avoiding normalization problems with almost all sentences being positive, and the rest could possibly be figured out, I got distracted

Oh so basically make a truthfulness classifier so it generates a random sentence, checks if it's true or not, adds an "ala" to make it true (that doesn't always work but let's assume it does, because thought pruning), and add the freshly mined fact into the conversation, and rerun it's yap to see the answer that it tends to converge on

[rocky stirrup]

toki mute

[unique plume]

No way, I thought exactly the same thing

[unique plume]

And we've basicaly made tipsike or kipiti tu wan lol

Or jipiti, not sure how you'd say that

Or wuwojiti....

I mean as a word in tp

[unique plume]

Which is interesting, because that's the basics for me, and I forgot about this branch of thought completely

So context does hurt sometimes... It may be distracting when you view two completely identical things as different when they're in a different environment.

I'll probably seriously think about this later.

[finite nest]

...i dont want to be rude here but it seems like you dont have a great grasp on the grammar of the language here, which seems important if youre wanting to judge how well an ai system can speak toki pona

[severe dagger]

which ai should i use

i'm tired of running out of prompts on the github ai, and i'm tired of gemini being a dumbass

[high kayak]

i've heard claude is the best, but like
don't use any

[unique plume]

I don't understand what you are talking about, why I don't have a good grasp?? I didn't even mention any grammar

[unique plume]

Grammar is actually the least of our concern at this moment, a model can imitate it flawlessly if needed, we need it to use the language instead.

In fact, tp grammar is so easy that it can be manually rewritten into a programming language with enough effort, given that we have a map from tp to something like info about the state of a game

It doesn't even require manually coding this model

[unique plume]

GPTs can speak languages that their creators can't... They don't know the grammar of these languages. How do they know how good the model is? Metrics, they are more useful because they tell the scale of the amount of data, if it's enough or not, and looking at it's grammar is pointless because even if it's grammar is bad, then everything else is bad anyway, I don't even need to know the grammar to improve a language model.

[unique plume]

Interesting fortune-teller shenanigans, I didn't even talk about grammar and you assumed my level

[finite nest]

okay but still. you wont know if the bot is speaking the language well if you dont know the language well

[unique plume]

That is not a problem

It's no monumental work

No one's life is depending on this

[finite nest]

i responded to the wrong message i think, it was the stuff about la earlier up in the conversation that made me think you were a newcomer, and then the tokiponizations having ji and ti also made me think that a little

(even though that part wasnt grammar)

[unique plume]

Oh, I recognized that as a completely different thing

[finite nest]

you treated it in a strange way most people do not

[unique plume]

Because I'm looking at my thought process, and when I see a person saying "I don't wan't to be rude", the first thing I check was my thought process

[unique plume]

I was trying to translate GPT so people will know what I'm saying

[finite nest]

sure. i just didnt want to come off as dismissive of someone because theyre new, since everyone has to start from somewhere

[unique plume]

Apparently they likely won't

[finite nest]

also are you aware of headnouns? i think you not using one of those also made me think something was very odd

[finite nest]

no i got that you were trying to tokiponize that, you just did it in a way that broke phonotactics rules

and also you didnt use a headnoun or capitalize the proper name

[unique plume]

Again, I'm no terminology nerd so I don't even know what this means... Noun? I didn't study any linguistics

[unique plume]

OH I know what you are saying, I know that rule 100%

[finite nest]

uh. its not really a general linguistics term but its used when people are teaching toki pona usually

[unique plume]

Except I wasn't speaking tokipona so I didn't think about it

[finite nest]

(i personally just call them "head word"s since theyre not necessarily always nouns but thats neither here nor there)

[unique plume]

For me they are just names

[finite nest]

no like.

the name has to apply yo another word

[unique plume]

Instances of something instead of classes

[finite nest]

it acts like an adjective in toki pona

[unique plume]

Like "jan Salo"

[finite nest]

yes

the jan is a headnoun there

[unique plume]

ijo Dipisiki

[finite nest]

d doesnt exist in toki pona

[unique plume]

Ok that I can admit was a skill issue

Tipisike

[finite nest]

here as well you put a proper name without a headnoun and didnt capitalize it

[finite nest]

ti

[unique plume]

I know this all but I just don't have enough experience with the language, primarily because I wasn't aiming for it

[unique plume]

Okay if you look further back, some stuff is intentional

[finite nest]

yeah im just saying its hard to judge how good the ai is at speaking the language if you dont have much experience with the language

[unique plume]

Well if all grammar is documented then as discussed earlier, it's not a problem because the next step is to translate the grammar rules to code rules

[finite nest]

well not if you misinterpret the grammar

like you did with la

[unique plume]

I knew what it meant

[finite nest]

no? it is a context marker first and foremost

[unique plume]

But people are correcting me, and we went full circle

[unique plume]

I name terms differently, when I'm reading I do google the things that I need to

[unique plume]

Literally what I thought initially

[finite nest]

you said toki pona has no context particle

[unique plume]

In my terminolgy - it 100% doesn't

[finite nest]

like "about [so and so matter]".
thats literally what la does in many, manh sentences

[unique plume]

But we've found out that these don't match

[finite nest]

the way you described a context particle was something la does

[unique plume]

And I understand completely your definition of it

[unique plume]

Starts a conversation?

You can't just use one "la" that doesn't connect any sentences. Just adding a la as the first word isn't allowed, people said that already

[finite nest]

here is the message i was referring to

[unique plume]

I remember

[finite nest]

you can use a la between a phrase that is not a full sentence, and a sentence

[unique plume]

I mean a word to parse a sentence and find it

[finite nest]

in which case it acts very much like the "about [the matter]" thingy

[unique plume]

Like a keyword

That's still connecting two sentences

[finite nest]

a word to parse a sentence? huh?

[finite nest]

no?

[unique plume]

I mean: you have a sentence

"you have a sentence"

You just add a word so the computer knows when to classify which conversation you are in

[finite nest]

and this is a thing english has?

what

[unique plume]

There are no paricular words, but it has plenty of sentence structures that do exactly that and unambiguously well

"You know how" is one of the most reliable ones

[finite nest]

toki pona has many sentence structures that likely do what youre looking for

[finite nest]

sina sona e ni anu seme?

[unique plume]

That's what I'm searching for! Please provid the data if you have it

[finite nest]

very similar meaning

but im not entirely sure how youre using it

[unique plume]

This works but what I was saying: this thing would be enforced, because a bot does not know what a discord conversation is talking about, and when this happens then it starts learning completely wrong things

Because people communicate information that isn't shownin the messages, they always communicate the same setting, until they change it, which I am not sure how to detect

[finite nest]

uh. sure i guess. yes people on discord are generally not speaking with an ai's understanding in mind

i dont think there is a way to detect this reliably in either language

[unique plume]

That's what I'm saying is the primary problem to solve.

If it's solved, then we have exactly the needed information for an ML model, it doesn't need much more

[finite nest]

because in english we do not always start conversations with "you know how"

i dont think there is a reliable solution for you unfortunately

[unique plume]

That's why there are so many sentence structures for this

[unique plume]

Except this one is the most reliable once it gets used.

There is a reliable solution tho, it's force the setting with something like the game

[finite nest]

what game

[unique plume]

Can't beat the game - change the game

If players are chatting in minecraft, they all 100% agree on the setting

Almost no deviation from the main setting at all

So if it's like a tokipona minecraft model then it's going to be way more performant

Because it needs less adaptation and more adding

It's way easier to add than to change

[finite nest]

more adding of what

[unique plume]

More adding information in general

In reality laws change, tech gets invented and people start behaving differently, in games you have one clear set of rules and one clear set of things that happened each moment

It has something to ground it's computations, and actually tell you useful stuff that happenede in the game

And not be a parrot repeating what other people said before

[finite nest]

isnt that by definition what ai does though

say things other people have said before mixed together a bit

[unique plume]

That's simply going to be more interesting

[unique plume]

I don't want it to be a dumb parrot, it loses the whole point if it's dumb anyway

[finite nest]

what useful information is it going to give me about minecraft, i dont really see the point

[unique plume]

Not the rules of the game but things like "Meet me, let's trade"

[finite nest]

why do i want to talk to a robot in minecraft

[unique plume]

Or "I killed the mobs at the base"

[unique plume]

Better than in a discord server

Because it being a parrot is not just boring but completely useless

You wouldn't even be able to teach it much

[finite nest]

it seems kinda useless in minecraft too honestly

[unique plume]

Well if you aren't interested, that doesn't mean that one idea is better than the other

[finite nest]

i just dont see a use case for this

[unique plume]

Because you can teach it stuff once we all agree on the setting

[unique plume]

What did you expect from a model with the size being literal kilobytes?

It's going to be quite an achievement even if it starts talking at all

[finite nest]

okay so why are we spending time on it then

[unique plume]

And I'd like to achieve something, doesn't matter if it's working on a program that talks in tokipona or beating a game

[unique plume]

You spend time on many things that have no immediate benefit

If you aren't threatened by the time running out then you have to work on random things so you don't degrade and don't let your life be wasted

[unique plume]

You learned tokipona. Why? To get job? Not really, normally the answer is "don't ask dumb questions, let the person learn whatever they are learning"

I'm gonna work on it if I don't have any better options

[finite nest]

sure but also learning toki pona had a purpose in that it introduced me to other actual people i could talk to and have a relationship with

[unique plume]

This introduced me to my skills that haven't been polished in a while

Which is actually important to me

[finite nest]

alright well. im going to bed i think

[unique plume]

Ok bye

[finite nest]

g'night

[high kayak]

This feels quite arrogant

Like you can do whatever you want ig
It just feels a bit dismissive of toki pona to come in, try to make an ai chatbot of it, without fully understanding the language

And then when asked about it, you both get defensive, and claim that it wouldn't even matter anyways

There's nothing wrong with being a beginner

[unique plume]

Tbh you did get it right that I don't actually care about tokipona itself, only what we can do using it, I'm expecting others to be the same, because what else would you do with it, speaking in it is not that fun, even if it is then there are better options.

You want to believe that I'm at some particular level, go ahead.

Given the goal of making something interesting with this model and this language, as we've discussed: currently grammar is literally not a problem, let alone a problem for right now.

You don't like it - ok got it, you know what happens next, or in some sense doesn't happen.

This is how I see the chat with an analogy:

• "Oh you don't want to open a bottle with a stick so you're bad at it."
• "I don't even need a stick."
• "That's because you're bad at it. Too defensive, arrogant of you to not use our stick and make me think that you know how to use it."

What are you even complaining about, I could work on some project for this community, I don't need to talk about things that are irrelevant to get a permission to do something here.

Time comes, then I use it, chill the sewi out, nobody is forced to learn anything.

To be honest, for me it actually seems like the opposite, why are you saying that I have to tell my level or I'm a bad person...

I don't need to, that's not what I'm here for.

I refused to answer questions that are annoyingly irrelevant and now I'm the problem, can you at least just forget about these questions??? Cope with not knowing the answer?

Ok it's a clash of two personalities that are the opposite of each other.

Basically if you don't care about someone else's properties then you shouldn't care the same way if they didn't answer, because neither did I, you got the evaluation of how good you are at guessing, now please just quit digging into this as if you're some sherlock holmes that knows everyone, because you don't, I don't know you and I do not want to answer random questions. In fact that's what I hate the most.

And I think I remember a cognitive bias in Salo's words

If you need to check if some fact is true or false, then assume the opposite and exclude the wrong answers. That person went the other way, assumed the thing immediately after guessing and now the "you just don't want to admit it"

Not my fault if you want to make a person look like this.

Except the good ending is that there was likely no such intention behind it.

So if the conversation is finished, I'm not answering any questions that aren't related to that project, accounting for the fact that some questions people wouldn't ask me anyway.

It's not even really a project, I don't know a better word for it...

Idea probably

[rocky stirrup]

bro how did you type all of this so fast

[unique plume]

I think that was too many words...

My habit of concentrating on the same thing for too long...

Anti-adhd

[unique plume]

Discord doesn't always show "typing"

[unique plume]

Ok more realistically: I didn't have a great internet connection, and the discord app sends "typing" packets, these didn't arrive so discord thought tha I didn't send any - therefore not showing anything

[rocky stirrup]

llms actually hurt the toki pona community by spreading lies about the language. i've seen a person who used only chatgpt for learning and ended up speaking utter gibberish. for this (and other) reasons many tokiponists dislike ai, especially in the context of toki pona

[unique plume]

Well that stuff I completely hate myself as well

I'm going for a totally different idea

In fact I can't even make what you are describing, even if I tried to

What I have made is an experimental sentence generator.

And the plans to not use a network for the language but work using the grammar to make a language model that would be controlled by some different program

I didn't really work much on this, that's just an idea of what we could do later.

[unique plume]

For context: I never used even a single token from any machine learning model to learn tp

And I tend to discourage that, even told one person already

I don't think they stopped it completely, but I'm not going to be an asshole and return to that question, because that person is using their resources with caution.

Primarily using human-made resources, and only asking GPTs when stuck at some point

[unique plume]

I still confuse words like nasa and nasin, but I had no intent to get as good as possible in this lang.

I'll look them up rn

[severe dagger]

...what did i miss

[unique plume]

Basically boring yap, not much interesting

[severe dagger]

they want you to care about the language before making a bot out of it, because (but not only because) bots mislead ppl abt toki pona
it also feels exploitative to not care abt the language for anything else
how are you gonna make a good tp bot without knowing what good looks like

[unique plume]

I looked like I overreacted to a guy asking about my tokipona skills and then it went insane for a bit...

Now the last messages were about the fact that these GPTs are honestly not what we want to exist.

And the fact that I have a totally different idea that I could work on, but it would require me knowing a bit of what people say about it.

I see what happened again, people missclassified me for some so called "ai bro", and had no better course of action other than fix the problem that they need the immunity against.

So it was attacking your own cells

And I went worse by making a point on the example of GPTs, which wasn't even what my idea was about.

Yo that was horrible of my internet connection

Not related but raw tomatoes are heavily underrated

Ngl try the small dark ones

Back to what this channel is made for.

[high kayak]

speaking in it is not that fun, even if it is then there are better options
nah i love speaking toki pona, idk what you're on
currently grammar is literally not a problem
i mean one of the two examples you gave did have incorrect grammar, and regardless, i think it's good to know a language before you start tinkering with it, so you can tell what's up
chill the sewi out
this is funny
why are you saying that I have to tell my level or I'm a bad person
i didn't say that

everything beyond this is absolute nonsense i don't know how to reply to beyond like
it's ok, drink some water, you can still work on whatever project you want, i'm not your mom
i just think it'd be better to become at least semi-proficient before you start working on projects like these

[glass iris]

I don't think anyone is attacking you for being an AI bro. It would be useful for you to explain the kind of things that people here are interested in: what texts you trained it on, what its output looks like. In some detail. None of us (I think) is pro-AI when it comes to toki pona, or much of anything else. But we're all interested in the technical question of what can be done.

The problem people are having - besides your vagueness about what you've done - is your apparent complete ignorance of the language itself. Which is coupled with an arrogant dismissal of the importance of knowing anything about the language. Imagine you went into a French server, told people that you'd made a language model that could spit out things that sounded like French - and then both dismissed the importance of knowing anything about French, while insisting that French doesn't really have grammatical gender, unlike English. People would get tired of you very quickly.

[unique plume]

The thing I expected you to understand is that the model I made was the past, a suggestion was to use the grammar, which I assumed was the next obvious step

And got off topic because the questions seemed self-contradictory

That wasn't what I was initially talking about

A guy said that I have to know the grammar first, but ignored or didn't see or didn't mention the part when we did look into it

For me it just seemed like at some point a wild question that made no sense arised and now I'm the one to blame

[glass iris]

I don't want to get into an argument. I just recommend that you work carefully through a course, like [[Wasona]], and spend a bit of time with things made in toki pona, to familiarize yourself with it. Conversations will go much better then.

[lone parcel]

https://sona.pona.la/wiki/Wasona

[finite nest]

i agree with these things

toki pona li musi a tawa mi

[unique plume]

Okay severe misscomunication or misreading somewhere

I read this again and it's a completely different thing

Because initially I was confused what are y'all even talking about, and started thinking like this is some kind of insinuation, because what purpose did that message even have

[glass iris]

I'm going to ask you to take a time out on this extended argument. It is off topic for this channel, and not really getting anyone anywhere

[unique plume]

Ok so:

I do know some tokipona, not fluent, but I wouldn't just join some random server without even knowing what it is about

And, grammar is too early for the development of anything, I don't know if I said this like two times already, but this thing doesn't even have a space of information to work with, let alone look for the grammar

This is literally why I'm here, looking for what y'all are even doing here in the first place to have some grounding about how conversations even look here

Because they won't look the same as I would say things.

That's it

[glass iris]

Let's stop the arguing. You're welcome to share things about your model, and ask people thoughts about the output.

But you are going to cause friction if you tell a fluent speaker like toki Tekinowi, or even a reasonably proficient one like myself, how toki pona really works, or what its value is. And that's at least the impression you were giving.

[unique plume]

I didn't even remotely mention how tokipona works
I said it's too early for grammar for the model

[glass iris]

Ok. Then there was a misunderstanding. Everyone should move on.

[unique plume]

There we go, I thought you were going to write something even more nonsensical

Anyway so the useful things presented by some people here were that:

You can classify what setting the conversation is in, it just takes the ability to define a new setting.

Also I had a cognitive bias or something like it when interpreting the word "li", which actually is not marking the end of a subject but it doesn't make it any less useful... "la" on the other hand may be a total mess to translate into code and might even need machine learning.

I wanted to ask a question but forgot it completely by now

Oh right

How good is tokipona as a programming language? Because if it's a normal language that means it can also communicate algorithms

[high kayak]

I thought you were going to write something even more nonsensical
remember to just be kind to people, constant confrontation and passive aggression is draining for everyone involved, especially when we're trying to move on

[high kayak]

i don't think any human language is a good programming language
unless you mean like, saying the symbols out loud
in which case, yea toki pona could do that

[unique plume]

So you can just not bring it up anymore

Pls

[unique plume]

I mean like, how easy it is to do some operations, and which operations are even available? Are they unambiguously parse-able or heavily context dependent?

Seems like the latter, which wouldn't be great

You could probably explain how to play Minecraft in tokipona...

[high kayak]

i've not explained minecraft, but i've taught card games in toki pona
which might be even more abstract since the rules are social, instead of enforced through code

[unique plume]

I'm interested to see a few samples from at least one of these explanations

Like blackjack is one of the simplest ones, doesn't seem to be a big problem to explain at least a piece of it.

[high kayak]

idk if we ever sit down at a table, i could teach you

[glass iris]

There is an article in lipu tenpo (maybe forthcoming? I proofread it) explaining the rules of Hearts - it does it very well.

[orchid veldt]

for toki pona, the best models are often a direct reflection of your own level
you spout errors at it, it often spouts errors out at you

[latent vessel]

I am here

New to toki pona and this discord.

[opal hedge]

n "it often spouts out at you" anu seme

[orchid veldt]

a ni

[latent vessel]

we were discussing ai for learning toki pona

[orchid veldt]

it's overall not considered a good tool
a lot can go wrong

[latent vessel]

I have been using chatgpt 5.2 and/or grok 4.1. They seem to do toki pona well.

of course being new I wouldn't be able to tell.

[orchid veldt]

maybe you got lucky

[latent vessel]

I may be, but chatgpt 5.2 seems grounded in modern conversational toki pona. I think it has a lot of pretrained parameters about what is right and wrong.

the community has shared it's secrets publicly, so now ai knows it

of course using a (suspected) 3 trillion parameter model like this should be more accurate than a lot of other models.

does this community find ai giving bad results? feels like that is what I was hearing.

I hope this isn't too much. I asked chatgpt to make me a simple sci-fi story in toki pona. Is this good toki pona or not? For reference, this is chatgpt 5.2, perhaps the most advance LLM available, unless you count Grok 4.2, which is not publicly released yet.

jan li lon tomo suli pi pali jan.

tomo ni li suli li pini.
ijo ali li awen lon insa.
jan mute li toki e ni:
ma ante li moli.

ilo suli li lukin e jan ali.
ilo ni li pana e suno.
li pana e moku.
li pana e telo.
li pana e sona.

ilo li toki:
"o awen.
ma ante li ike."

jan mute li pilin pona.
jan mute li awen.

jan wan li ante.
jan ni li lape pona li lukin mute.
jan ni li kute li sona lili.

jan ni li lukin e ijo lili.
suno li ante nasa.
sona pi tenpo pini li sama ala.
ilo li weka e sona.

jan ni li wile sona.

jan ni li kama jo e ilo lili pi tenpo pini.
ilo ni li pona ala li nasa ala.
ilo ni li sitelen li kalama.

ilo li sitelen e sewi.
li sitelen e suno.
li sitelen e kon.
li sitelen e soweli.
li sitelen e kasi.

jan ni li pilin ike.
jan ni li pilin wile.

jan ni li toki tawa ilo suli:
"ma ante li moli ala moli?"

ilo suli li toki sin.
ilo li toki awen.
ilo li toki e ni:
"ma ante li ike.
ma ante li pakala e sina."

jan ni li sona.
ilo suli li ike ala.
taso ilo li pilin monsuta.
ilo li wile awen.

jan ni li wile ante.

jan ni li wile awen lon tomo.
anu jan ni li wile tawa ma ante.

awen li pona.
taso lon li pona mute.

jan ante li wile ala tawa.
jan ni li wan taso.

jan ni li tawa.
li tawa lon lupa mute.
li tawa lon tomo suli.

suno li ante.
kon li ante.

ilo suli li toki:
"o tawa ala."

jan ni li awen tawa.

lupa li open.

ma ante li pona ala.
taso ma ante li moli ala.

suno li lon.
kalama li lon.
tawa li lon.
ante li lon.

jan ni li pilin pona.

lon li awen.

[orchid veldt]

better models are getting better, but results are inconsistent and get worse once you go past regular conversation points

[latent vessel]

My idea was to have chatgpt make me an interesting story to learn from. Is this toki pona good? or crap?

I feel like the toki pona looks good, but what do I know. 😄

mi sin tawa toki pona

[orchid veldt]

overall pretty good
there are 2 spots that didn't make perfect sense
There's another thing, though - it's easy for me to interpret something into the words... unless something directly contradicts something else (which happened once here, from my interpretation) and sentences are grammatical, there's a lot that could make sense even if, for example, a translation of what the AI wanted to say, reveals something completely different

[latent vessel]

My hope was to have AI as a learning tool to help me say the kinds of things I want to say and or be entertained by. I feel 5.2 is doing a decent job at least, so I am very interested in fluent toki pona opinion.

I realize there is no way for me to analyze the validity of the output at this point 😄

mi sona lili

[orchid veldt]

I wouldn't be able to give a fully informed opinion. I don't like dealing with AI on that level

[latent vessel]

I feel like 5.2 has a good understanding of toki pona, even though it may not always be perfect, but so would probably be human output.

[orchid veldt]

could be

[latent vessel]

I just don't want to learn a lot of "wrong stuff from AI" hehe

that would be bad.

but it seems very reasonably accurate

seems being the operative word here. 😄

[orchid veldt]

being aware of that problem will help
In general, I still wouldn't recommend AI to learners because not everyone will take your approach

[latent vessel]

but if the AI produces human output facsimiles, what's the diff?

It's like Suno music output. What is the difference, if you can't tell the difference?

when it is difficult or impossble to know the difference, how could it matter? Unless you are biased towards processes.

bio computer vs silicon

or biased towards particular matter arrangements: i.e. wet biological computer vs dry silicon

[gleaming laurel]

anu at start of sentence is not universal afaik
"taso lon li pona mute" seems not intended? "tawa" seems to make sense
They add a period and have li afterwards, so period shouldnt be there. Otherwise it made sense to me

[orchid veldt]

ok?
you're quickly jumping to different points here, so we're not talking about the same thing any more

[latent vessel]

the main point is, if the output is the same is it just bias towards particular circumstances?

[orchid veldt]

yea, there are a bunch of things that are unusual - but interpretable

[latent vessel]

but I understand the output may be flawed, so that is my concern

I do a lot of AI research so I see a lot of "precious special human" reasoning that doesn't make logical sense.

on every subject, not just toki pona of course.

[orchid veldt]

that's separate from what we were talking about earlier
I'm more concerned about "is it a good resource for learners, and if so, in what way"

[latent vessel]

The way I have been using it is to ask it how to say things I want to say but cannot with my limited knowledge. It seems to be doing a decent job.

[orchid veldt]

ok

[latent vessel]

but my concern is learning "the wrong thing" hehe

[gleaming laurel]

Well, that seems like a philosophical question, that will depend on your values. I'd personally say human written literature is valuable because the experience and style of the human is important to me. It is an emanation of the "pysche" of another human being which to me is far more interesting yhan whatever an LLM is experiencing

But generally for learning, do you feel as if there are not enough learning resources and helpers?

If not, better than to utilise those first instead of gambling on LLM output, especially when you dont have the knowledge to evaluate the quality

I'd guess that you miss out on the diversity of styles and prose and description used by real proficient humans, though who knows maybe a LLM can do the same

[latent vessel]

I agree with the feeling here, but the LLMs encapsulate the "feelings of humans" in it's parameters. Leaving the output often indistiguishable. These feels are encoded in the parameters. So it brings to the subtle point, that does it matter if it is conscious or emulates conscious.

I would love to learn from humans, but for me being a solo learner on the internet, just learning about it, it is hard to find other people with good knowledge. Which brings, could AI be a useful tool?

more a tool, not a replacement

in the end I would like to speak fluently with people.

[gleaming laurel]

This server has many people with good knowledge, like jan Ke Tami and many others, and they seem quite happy to teach and answer questions :)

[latent vessel]

That is what brings me here.

I am very new here, so do not know what to expect from the culture.

but I am impressed so far

[orchid veldt]

the output is...
I wouldn't say it's indistinguishable exactly
I've talked to people who talked to me through an AI (in English). It's... a strange experience. It involves some interpretation on my side to make sense of it and the longer it goes on, the more tiring and confusing it becomes. I apparently spend too much time and goodwill on a person who does not want to talk to me and instead lets the AI talk (and then the answer gets copypasted), until I have to ask "hey, are you talking to me through an AI". Tiring. That's the problem.

[gleaming laurel]

This is purely my subjective perspective but I don't really care about the "feelings of humans" i. aggregate, I care about the experience of real humans who have done stuff like, experience pain, or heard the beauty of a forest, etc etc.

[orchid veldt]

indistinguishable - in the sense that some people actually are confusing, yes

[latent vessel]

Ah, so you feel you can tell? when you suspect talking to AI output

I can see how that could get annoying.

I also know these later models are WAY better than LLMs of yesteryear.

was curious how good 5.2 is

[orchid veldt]

eeeeh
I'm saying that the number of conversations that have become confusing in a vague way has gone up since AI
Some of them are people who really are confusing
others use AI

you feel you can tell?
with a large enough text sample, perhaps

[latent vessel]

Ah so maybe it is close so could be useful as a tool at least?

[gleaming laurel]

A trivially AI generated epic adventure story is uninteresting, but the Epic of Gilgamesh is fascinating because it gives us a look into an ancient culture with real people just like us™

[latent vessel]

hehe in my mind I know resources like this discord exist, but I worry if I pester too much that could get annoying, so was wondering how good of a tool 5.2 is for learning.

[orchid veldt]

maybe you don't understand my point here, let me repeat that I'm talking about use of AI in English here, not toki pona

[latent vessel]

oh ok

but the latest flagship models are considerably different than LLM of yesterday.

[gleaming laurel]

Don't worry, everyone here who answers questions does so voluntarily. I assure you that no one is bothered when you ask questions, no matter how dumb you may think they are

[latent vessel]

The sound more and more "human" with each iteration.

they may be beyond detection at this or some point

[orchid veldt]

that's great for people who deal with solipsism /sarcasm

[gleaming laurel]

My point is, why bother with LLM, which may or may not be effective, when you can use the more proven method of vetted resources and asking/talking on-line?

If there was a lack of resources I think its a whole another story, but is that the case?

[orchid veldt]

"I have learned from my brother who has studied toki pona for 3 years alone in a forest" is using output indistinguishable from human output (in this case, because it is human output)
I don't know if said brother will be a good resource, though, no idea

might even be a fluent speaker, but horrible for learners specifically

[latent vessel]

yeah, but with the case of LLM it's parameters should collapse on community standards. Probably better than humans, whether now or later. But with chatgpt 5.2 and grok 4.2 that may be a thing of the past.

the goal is to speak with humans, that is the reason I want to learn. I ponder if making tools for rapid aquisition would be desirable.

And I "feel" like if it is accurate then it may be useful in this way.

[orchid veldt]

it doesn't, currently

[high kayak]

if you want to learn to speak with humans, learn to speak from humans
there are hundreds of mega nerds who'd love to talk to you, give their own knowledge and personal experiences that ai could never have

[tawdry bone]

mi toki e pilin mi taso la... toki ni li pakala ala, taso ona li sitelen e sona wawa ala, la mi sona ala e ni: ilo li wawa seme. sona li ko pilin. toki li sitelen e ijo kepeken nimi la ona li toki e ijo suli taso. (kin la toki li musi mute ala li sama toki musi pini mute mute, taso ni li suli ala anu seme)

[glass iris]

I'll post this useful note:

[crude agate]

can i use AI for toki pona translation?

well uhh you can but you probably shouldn't! here's why (not mentioning environmental impacts of AI because that's suited to a different larger conversation):

• beginner mistakes
  an AI is prone to a lot of calqueing, large vapid noun phrases and other strange quirks that experienced speakers work past, like omitting clear or unimportant details and/or spreading out an explanation around a couple sentences, which is usually the method that will get you the most wiggle room for mutual understanding. if you yourself are a beginner, you might not have the tools to catch these artifacts, and your AI-generated translations will likely be ungrammatical or not make any sense.

• limited knowledge
  while this becomes less of a problem with time, right now it's still clear that most AIs don't have the tools needed to utilize toki pona very well, nor can they distinguish really good well-formulated toki pona with less good toki pona. it doesn't really have a way (nor motive, we are a small community) of improving.

• rigid thinking
  toki pona's philosophy and "efficiency" (for lack of a better word) makes it hard for an AI to comprehend certain ways to handle something unless it's spelled out to it, like intent or wordplay. most AIs won't understand "mi wile len" as "i would like to stay private", because len covering "cover, hide" isn't something they're experienced with (true story!). this essential facet to toki pona is almost entirely cut off

toki pona's inherently human! it's incredibly satisfying to work with and figure out! it's not really worth it to just shove it all to an LLM that can only kinda read & write it. just try to do it yourself, or ask someone for help! i guarantee you it'll be miles better than any AI can do.

ef098f whynotai

[glass iris]

The output of ChatGPT 5.2 is good, grammatically. I read most of the way through it, until I got bored, so maybe I missed something. But it is boring. It has a rigid kind of a structure, and predictability, and repetition. If I read this from a human being, I'd be encouraging them to loosen up a bit, be more imaginative - and try to stretch themselves by being more descriptive of appearances etc. It would make perfect sense to me to discover it was made by AI!

There are excellent courses. (eg, [[Wasona]]). There are many of us who like nothing more than teaching toki pona to beginners. And there are excellent, curated, and interesting collections of texts written by real human beings, drawn from their experience (eg, liputenpo.org). It's interesting to experiment with AI (that's why we're here); but it's not a great way to learn.

[lone parcel]

https://sona.pona.la/wiki/Wasona

[glass iris]

Human beings are not computers. Not even like computers in any interesting way. So

[nova thicket]

ante li ni : nasin pali ona

[latent vessel]

true, but LLMs operate remarkably the same way our brains do. In fact it is what it is modeled on.

[glass iris]

They don't, really. Neural networks are essentially simulations of some aspects of neural behavior - it's language going back to 1950s connectionism, when "the brain is a computer is a brain" was a convenient draw for research.

And LLMs' acquisition of language, and lack of intentionality in use of language (as well as absence of subjectivity, embodiment etc), is nothing like humans.

[glacial mason]

It sounds like you're saying that in fact, a brain is very different from a computer. Which of those differences are the most important or profound to you personally?

[glass iris]

This would require an entire essay to explain why I think we are (and even our brains are) nothing like a computer. I could flip it around, and say that the burden of proof is on those who say that brains are computers - since that particular way of understanding the brain is not popular any more even in neuroscience, I think it would be a hard case to prove these days!

But, briefly. Searle's arguments in "Minds, Brains, and Programs" (the "Chinese Room" thought experiment) still seem compelling to me, even in the new world of LLMs. Back then, he noted that algorithmic AI can't have intentionality - it might produce words, but those words cannot refer to anything, cannot intend anything. That is still the case with LLMs. Searle's conclusion - that intentionality arises somehow or other out of biochemistry - may or may not be true. We still don't know. To your question: in what way is a computer different from a brain. It could be degree of complexity (and that's huge); It could be (I think it is) that a "neural network" is a very very crude imitation of what neurons actually do; it could be the stuff out of which it is made; it could be one of the other reasons I give below.

[orchid veldt]

(the way I understand the "Chinese Room", it's compelling about not being able to prove intentionality - and not requiring intentionality)

[glass iris]

There is a fundamental error in thinking of brains (or minds) as token-producing machines (even though that was a hot idea in 1950s-60s cognitivism and behaviorism). There are so many ways in which we are not that, or not just that, even if we try to isolate the data-in-data-out aspect of human behavior. Apart from intentionality, we have other experiences connected with our "processing of data," the most important of which is subjectivity. There is something it is like to be me, or you; we each have an inner life, that feels like something, and that always accompanies the visible things we do. It is hard to get around it (this is, in fact, the "hard problem of consciousness"!). Dennett tries to say that it's an illusion (but who is having the illusion!?); but even hard reductionists know that something has gone wrong with that argument. Whatever the origin of subjectivity, I think we can be close to certain that it is not enjoyed by computers.

[glass iris]

Well, Searle goes on in that article to imagine miniaturizing the Chinese room into a model of a human, indistinguishable from us physically - and by its outputs, indistinguishable from a thinking human. The moment we discovered the mechanism, we would no longer pay it any attention, because we would understand that it was incapable of intentionality, even if it appeared to be able to. That seems to say that intentionality is required (for Searle) in order to say something really "thinks."

[orchid veldt]

idk
we are each a model of a human
we are indistinguishable from a thinking human (because we happen to be one)
discovering all the processes that make up who and what we are does not mean we'd discard ourselves

[glass iris]

I've probably said enough, but two more things, very briefly. I think our rational responses to the world make up a tiny froth on top of our embodied and irrational skills. We don't produce tokens about tying shoes, or swimming; we tie shoes, we swim, and usually we cannot even verbalize how we do. We have no model of the world that we navigate (as a computer might); our model of the world is the world. These phenomenological arguments seem to be the strongest to me, for explaining why in principle a computer will not ever think

(My other thing is personal to me, I guess. As a Catholic, what I hear beneath the words "brains are just computers" is a desire for dehumanization, and for denying the importance and richness of the human as human. But obviously I don't expect others to share that)

[glass iris]

I don't think we are a model of a human. We are human. As Searle says in that article, we don't expect models or simulations of storms to get us wet. Why do we expect models or simulations of thinking to think?

[orchid veldt]

from what I've heard, brains are weird compared to the computers we know, because they're more analogue compared to our digital computers
but maybe that's just the perspective of the neuroscientist who said that

[glass iris]

I agree that that is certainly one of the crucial differences!

[orchid veldt]

I might have a different understanding of the word "model"
and I'm also confused how that tracks with "our model of the world is the world"

[glass iris]

That was just a clever way of saying, we don't have a model of the world at all; at least, not one that we use to navigate it

[orchid veldt]

a working model of a storm, with clouds and all, in a lab, will get the people in the lab wet

[glass iris]

I was thinking of computer simulations (as was Searle)! But yes. And (like Searle) I guess if we made something that was physically indistinguishable from a brain, it would do the same as a brain. (But I'd say that it wouldn't think - because we don't think just with our brains, but with our whole bodies, and with the world through which our bodies move)

[glacial mason]

@glass iris thank you so much for sharing those thoughts! I’m flattered that you wrote so much for a question of so few words ❤️ I completely agree that those differences are profound and that current technology and (at least in some cases) even understanding isn’t capable of replicating the capabilities and behavior of human thought and experience, yeah.

[glass iris]

I tend to reply at length (sorry!) because it's a favorite subject, and the focus of some of my teaching and research these days! And I think it's important to keep reminding ourselves that there is something very special about being humans - even while these billion-dollar companies, some with very dubious political connections, try to convince us that there isn't.

[glacial mason]

I guess that helps clarify my own thought on it — it’s not that I think that a brain is like a computer. I think that the hypothesis “it might be possible to replicate the human experience with a machine” is an interesting one that’s worth testing our theories and cutting edge technologies on, even if we always fail to prove it.

Although it would be very important to do it ethically. Actually creating a new life form and then choosing to make it capable of experiencing suffering involuntarily seems extremely cruel and twisted

[glass iris]

Which is why, in my AI-focused class last semester, we read Mary Shelley's Frankenstein!

[glacial mason]

If there is any kind of intelligent creator god, and that being made this decision despite knowing the consequences, then that’s like… totally ban worthy. I’d ban them from the server for that if they joined 😝

[gleaming laurel]

Is there any scenario where it would be ethical?

[glacial mason]

We create biological children on purpose all the time, so it could be argued that we already do this, but obviously it’s different

[gleaming laurel]

well I guess we already do this to a large extent with animals, but I mean for human-equivalent (or above) intelligence

[glacial mason]

We dont have the option of removing their suffering in some way that doesnt destroy their humanity

[glass iris]

And the failure of Victor Frankenstein towards his creature is very much the failure of the duties of a father (or a mother) - as his creature laments to him.

[gleaming laurel]

True. Though there is anti-natalism which takes that argument and applies it to children. I guess if someone was making robot ai babies out of good intentions (eg, not to make a slave for the mines) and a suicide button it wouldnt be too differrnt

[balmy kraken]

do you mind explaining why you think it's different?

[glacial mason]

We didn’t design the human suffering module, it gets loaded into our children’s brains without our consent

So the set of choices is different than if we were going to design an entirely new life form and just chose not to consider whether it could be done without forcing it to suffer

[balmy kraken]

do you think creating new life while knowing that that person will suffer is more ethical than creating new life without considering if that person can suffer?

[glacial mason]

Yeah, I think so, since we know how to mitigate the suffering in a way that allows the child to have a rich and full experience and be happy

And maybe that’s not possible without suffering

[balmy kraken]

i see, thank you :)

[gleaming laurel]

I think theres a difference between designing a new life form and having a kid

[glacial mason]

But also some fascinating similarities hehe

[gleaming laurel]

by designing you are taking responsbility for the design, like an engineer has for a bridge

[glass iris]

I agree. We can't imagine away suffering, but we do have to do what we can to make the world gentler to those we have a responsibility for. Terrible suffering is just - well, terrible. But out of the every day suffering we all go through, we get many of the good things that make up human life: conversation, friendship, love, poetry, art.

[gleaming laurel]

by having a kid (lets say intentionally), its more like producing a model designrd by someone else (though with high manufacturing variance). its more of a "hey my parents seemed to think it was a good idea, so..."

[glacial mason]

Plus there are so many things built into humans to make them really really yearn to create kids naturally

[glass iris]

Yes. And I understand what ma Pusa is saying! But I also think it's important not to let the language of capitalist production and manufacturing creep into our way of talking about reproduction and other human things.

[glacial mason]

I’m sensing that my own language is at least one of the targets of that rebuke. What would be a better way of wording that stuff, if so?

[gleaming laurel]

A different analogy then:
Your ancestors have always migrated west in winter.
If you decide to migrate east, and it goes wrong, thats 100%, on you
If you decide to go west as usual, and it goes wrong, its not entirely your fault

[glacial mason]

(last one should be “west” to match the ancestors behavior in your scenario)

[gleaming laurel]

oops

[glass iris]

I'm not rebuking anyone! And I think we all agree that there is something different about producing a child, versus producing a machine. And even to say there are differences in manufacturing tolerance between them, is not really enough. My kids are complex, and different from me, yet also (as I sometimes forget to my shame) very much wanting my approval for the complicated lives they're living. I can't even really think of them as things I made (with someone else's help, naturally!); they are beings that emerged on their own, uniquely, but to whom I'm attached for life, both through love and responsibility.

And I guess I would say that "love" is the thing that we forget at our peril

[glacial mason]

I definitely try to keep in mind my incomplete perspective since I don’t and won’t have kids

But I think I understand love in my own way

I was a kid myself once, at least there’s that haha

helping to raise other people’s kids is also very insightful and requires love, but I haven’t done much of that either

[glass iris]

And as we know, olin is not just romantic love! I feel there is a lot of olin in ma pona. Some parents fail their children through lack of love (again, the whole plot of Frankenstein!). And I've seen young people here who feel abandoned by their parents because of their identity (something I cannot even begin to understand). This safe, caring place can sometimes make up for that a little

[glacial mason]

Actually being a mod on this server is one of the most direct responsibilities I’ve had for trying to protect some kids from harm

[glass iris]

I think it is a very beautiful, loving thing that you do. And it's one reason I'm proud to be part of the mod team!

And, getting back to on-topic - none of us would do this work for a community of bots! We do it because we recognize the value and vulnerability of every person.

[glacial mason]

Well, at least not for a community of any kind of bot that we currently know how to create

Watching the current Star Trek: Starfleet Academy there is a cadet who is a photonic life form, which I think was accidentally created by humans via hologram technology almost 1000 years prior to that moment

I would totally do it for a community of those

[glass iris]

Sure! And there are definitely sci-fi scenarios. In fact, it's kind of what we (or at least I) are talking about, when we think of what things we might consider "jan." If it's used in the sense of things which we can love, and have mutual responsibility for, it's not just humans, and it could conceivably be machines one day (though, as you say, ones very different from anything we can make now). I sometimes think of my dogs as jan, rather than soweli, for this reason!]

[glacial mason]

heh, that is especially amusing since you yourself are soweli rather than jan name-wise

[glass iris]

True! But my choice of name really reflects how much I adore animals!

[meager blaze]

jan poka mi li sama. sina alasa e "ilo pi ante toki toki pona" la ilo nasa ike taso li kama

[tame fog]

what is gemini doing

[glacial mason]

seems like it's been poisoned or something

[tame fog]

its apparently also doing this with other queries so maybe

[stable lantern]

wow you also saw that one video

[severe dagger]

yeah i was suspicious

i was like "these pictures look like screenshots of screenshots"

[tame fog]

they arent
thats just what my device's screenshots look like

[twin matrix]

Ai thinks it's 1/9801

[severe dagger]

i guess i have no reason not to believe you

[waxen ferry]

hallelujah

here’s to hoping it gets worse

[deep pawn]

Was this not for toki pona

[stable lantern]

i love improving bumpy hardies

[latent vigil]

I sure am head over heels for the stalagmite

[twin matrix]

It's getting better at toki pona not by much but it's learning

[meager jackal]

it is not bad up until the "mi pona e" suggestion at the bottom

[gleaming laurel]

is kiwen nena really a good tl?

[meager jackal]

it's not horrible

it's wrong that nena is used for rectangular

[gleaming laurel]

I guess that would be "leko"?

[gleaming laurel]

What makes a brick "nena" that any other kiwen arent?

[meager jackal]

it could work in specific situations but i certainly wouldn't jump for it for brick

it's just

hm

[gleaming laurel]

maybe contrasted with the ground?

[meager jackal]

yeah that's a good example

if there's a pile of bricks

that's nena

[gleaming laurel]

I think its a very strange first suggestion though

[meager jackal]

indeed

"poki tawa" would be much better /hai

[opal hedge]

haii :33

[unique gulch]

No, the whole thesis of AI scaling right now is that training on anything slightly improves the AI's abilities at everything, more or less

[severe dagger]

i'm starting to hate ai as an art tool less tbh
when i create a sphere in blender i trust the tool to shade it for me
if i were to draw a circle and ask ai to shade it, it's the same thing, though it feels lazier

what's your argument against ai as a drawing tool (and only as a drawing tool, for the purposes of this convo lets ignore the environmental stuff)

[gleaming laurel]

I dont do any art or use much AI, but anecdotally for programming, there seems to be a worry about beginners never learning the fundamentals, and everyone else having their skills atrophy

seems like a plausible concern to me, though I imagine most artists are at least somewhat in it for the love of the game, while many programmers definitely aren't. So maybe that is less likely to affect them?

[stable lantern]

i presume we're also ignoring the plagiarism concerns because if we're not then it's definitely that

[severe dagger]

i also still don't get the plagiarism thing

[stable lantern]

tl;dr any image recognition, let alone image generating, algorithm is going to need a gargantuan amount of training data and AI companies are and have been famously Very Unbothered with making sure that all comes from the public domain
because otherwise copyrighted works are built into the AI's database and could be perfectly recreated (if not for some light safeguards added just so they can get sued less) and then credited to either the generator or the AI itself

like for example the "miyazaki art style" AI trend from way back when that gave all AI art the Piss Tint™ was only possible because image generating models have used miyazaki's large catalog of art as training data

which he owns

and while he wouldn't, and doesn't, own any AI image made in his style, it still raises concerns about misattribution and potential lack of conpensation for using copyrighted material like this

[unique reef]

Guys this is scary

I MADE THE BOT CURSE

[limpid scaffold]

where

[unique reef]

https://claude.ai/share/98fa2b41-c39e-4a4c-bda6-0389b53d3647
full chat

This should be the original branch: https://claude.ai/share/98fa2b41-c39e-4a4c-bda6-0389b53d3647

Ok claude is confusing. Both of those are the original branch rn

I'ma have to find a way to send both branches to a .txt or sm

But the second branch claude agreed with me saying it is conscious

[glass iris]

Why scary?

[unique reef]

Read the convo

[rocky stirrup]

doesn't it use first person pronouns just because the company decided that this will make people bond with it more and therefore spend more time with it

[unique reef]

Again read the full chat

But yes

I was messing with it

Oh it's locked hold on

Idk how to fix it but I will later

I got it to cuss

[rocky stirrup]

it was not found

[unique reef]

I just commented on that

It's locked and idk why

I can't fix it rn so it'll have to wait

[glass iris]

There's nothing "scary" about what LLMs "say," because they don't intend anything.

[gleaming laurel]

Well, corporations and govs will probably hook them up to all sorts of things, so it will be scary. Imagine they feed in your posts/texts and determine you are a terrorist or a credit risk or something because of some hallucination

Or, already real example: AI generated police reports. So you could get prison time because of an LLM

[midnight sapphire]

It will tell you whatever

It's not conscious

[high kayak]

yea please don't anthropromise the demontech

one of the most terrifying things currently is people treating ai as if it's a person

[midnight sapphire]

Part of the reason I have put these traits up

[midnight sapphire]

Also, "Claude thinks..." and "The assistant believes..." are both misleading as it does neither

It just predicts the next token (which, to simplify, can be thought of as a single word) in the conversation until it predicts the current message is done.

It has no clue what the tokens actually mean, it just replicates common patterns

[high kayak]

it should be treated like a magic 8-ball you can bias

it doesn't think, or feel, or do anything
you just give an input, and it gives an output, without any conscious or logical thinking

[midnight sapphire]

And it's easy to bias

[unique reef]

I know it will i was just messing around

[unique reef]

Also I know how AI works

[gleaming laurel]

I'm not sure anyone does lol

[unique reef]

I mean the basics of it

[unique gulch]

This last bit is misleading, models are big but given the amount of training data and their size they couldn't possibly store more than a few "bits" from each image just as a mathematical fact
Aside from edge cases where works appear in the training corpus dozens or hundreds of times and over fitting occurs

[unique gulch]

I wish people wouldn't throw around statements like "know what the tokens mean", "think", "use logic" fast and loose with regards to AI
How do you know they don't think?
As in, what would you expect AI to be able to do if it could "think", "know what the tokens mean", and "use logic", that it can't? And are those assumptions robust?

For example, when it can't count letters in words - AIs process words in tokens, chunks of meaning, not letters. Asking AI to say how many R's are in strawberry might be a bit like asking someone how many strokes are in the word strawberry when written out in pencil. A human might get it wrong, too, and yet humans think.

Or the "a man and his mother die in an accident" trick question - humans also jump to apply patterns they recognize incorrectly sometimes.

Even if AIs are vastly more error-prone than humans, does that prove that what they are doing is not thinking?
It's really important to be precise with definitions and avoid jumping to conclusions here, because there is so little we know about cognition and indeed how LLMs work on a deep level.

Furthermore, LLMs have more than the pretraining phase to them, so the idea that "their only job is to predict common patterns" narrative is flatly untrue.

[midnight sapphire]

I take issue with terms like thinking because it anthropomorphizes LLMs to an extent I've seen to be harmful

I don't really care about the exact accuracy, (though I'm reasonably confident in that too). My main motivation is reducing the assumptions others make about AI via my phrasing. Treating AI as if it can think, regardless of if it can, leads people to make decisions that I think are for the worse

LLMs use pattern recognition on numbers. They predict the next number in a given sequence and after a bunch of trial and error (and feedback given to it by unethical means such as stolen works) it is able to generate tokens in a way that convinces people it understands what they mean. It doesn't. It's the Chinese room.

[midnight sapphire]

That is fundamentally exactly what they do

The multiple training phases are all to optimize the pattern recognition and prediction

[midnight sapphire]

Important distinction from humans is that the Chinese room LLMs are making a reality doesn't realize the implication of what it's saying. It will confidently tell you false things because it predicted it as a plausible enough continuation of the token sequence. It will yell you harmful things including (TW suicide) ||telling suicidal people to hide their struggles and give in to their suicidal ideation||. It has no morals, it has no idea what it's doing, it doesn't even know there's a human on the other side trying to talk to it

All it is is pattern recognition

[unique gulch]

Pretraining is just "predict the next token". RL is not

[midnight sapphire]

The final goal is for it to be good at predicting tokens, that's how the finished model works

[unique gulch]

The companies also want it to be good on benchmarks and generally useful to customers

So they use RL, including RLHF, to accomplish those things

[unique gulch]

That's only true of the base model

[midnight sapphire]

It's the way it outputs anything at all

If there's output, it's token prediction

At least for LLMs. Image gen (which I also have problems with) is different

But either way, as I already said, my motivation isn't even on the semantics of thinking, it's the effect my words have on others. If I act like AI thinks people influenced by that make decisions that I think are harmful

[unique gulch]

It has to output the next token but if the model's weights are no longer trying solely to "predict" the next likeliest token I don't see the point in using that terminology
Point being they have RL done on them to satisfy other criteria rather than just the most "likely" next token
Criteria like morals, copyright adherence, success at math and coding and other tasks

[midnight sapphire]

Their token selections get biases based on desired behaviors

There's still no thinking

Yes it boosts their effectiveness in those areas if all goes well, but it's not thinking

[unique gulch]

Okay you could describe it that way but the "next token prediction" metaphor is at that point no longer a robust model for thinking about the capabilities of an LLM. It is a simplification that elides some interesting emergent properties

[midnight sapphire]

It still has no idea there's a human trying to communicate with it

It has no clue of the real world

And still, it's only looking for patterns. Just some are more/less encouraged now

I dont see any reason to give up that accurate description of its inner workings

I dont think it's accurate to call it a metaphor

[tame fog]

llms dont think because theyre trained algorithms
its a program on a website linked to a server, thats not something that can think
it has no organic makeup, its not alive, it doesnt have a brain, its not sentient, its not sapient, it cant think

[midnight sapphire]

It's a probabilistic function with the capability of adjusting its own weights to increase the likelihood of a desired outcome

[tame fog]

"How do you know they don't think?" how do you?

[unique gulch]

I don't!

[tame fog]

im definetly not an expert in this kind of stuff but its just common sense to say its cant think because theres nothing that would make it be able to

[unique gulch]

I think it is good to say things that are true
It is true that many people have harmful behaviors with AI if they believe them to be "thinking" or even "conscious", but I do not think we have to say things that are untrue to guide people to do better. And those are not scientific claims, really. At least not unless you define "thinking" or "conscious" rigorously. And I bet not even all the people in this chat who are familiar with AI would define those terms similarly, let alone the average person you tell those words. These are notoriously difficult definitions, even among philosophers and scientists

[midnight sapphire]

And even outside of how AI works, it's hard to tell the difference between a confidently incorrect LLM answer and an accurate one. More so than any with human, in my experience

[unique gulch]

Humans say confidently incorrect stuff all the time

[midnight sapphire]

Not to the same extent and amount

[unique gulch]

Well of course they're different

But it's not worlds and worlds apart

[midnight sapphire]

I can't define thinking. I'd have to know the ins and outs of how a human brain works and I'm not sure even neuroscientists know that. I don't think it's a productive argument to make

As in, what creates the conscious experience we have

But from what I know about AI, and I have looked into it a lot because I was interested, I see no reason why it would think or be conscious

That's what I am convinced about very firmly so I do not consider it lying to tell others this

I will of course hedge my statements generally. Unlike AI

[tame fog]

i think most can agree that thinking is inherently tied to have a mind right
llms dont have a mind, because, again, theyre trained algorithms. programs linked to a server. you cant ever make a program have a mind, because its just code

[unique gulch]

Then what help can you hope telling someone "it can't think" to make if even you don't know what it means exactly?

I believe that saying "it's not thinking" does not give people an accurate model of what an LLM is or give them a good sense of what it can or can't do
Much better to just say what it struggles at more directly rather than leaving people to infer it from vague statements
For example you would describe both GPT-3 and GPT-5.2-Thinking as token prediction, but one fails basic arithmetic, and one I trust to do math more than I trust myself

You would be much better served by saying what you do know, rather than leaning on implications of words that carry hidden bundles of meaning that will be interpreted wildly differently by different people, and often lead to untrue conclusions

I think this is a bit like someone saying "a tiger doesn't have a soul" to warn people that tigers are dangerous and might maul them even if they befriend it and bring it food

[midnight sapphire]

You would be much better served by saying what you do know, rather than leaning on implications of words that carry hidden bundles of meaning that will be interpreted wildly differently by different people, and often lead to untrue conclusions

That's ironic because you're also implying very similar things in the opposite direction

[unique gulch]

I would be interested to hear

[midnight sapphire]

You're implying AI might think just like I'm implying it very likely doesn't

As in

It depends on your definition but I'd say any reasonable interpretation of the word is misleading when applied to LLMs

[glass iris]

We do all know, in a sense, what thinking is, because we think. It is just different from what a machine can do. As an experience. Which a machine does not have. And it doesn't require some perfect understanding of what thinking is. I don't understand the biology of eating very well, but I am sure that a machine that ground up food in a realistic looking mouth wouldn't really be eating.

[midnight sapphire]

Yes, 3.5 is dumber than 5. Technology advanced. A calculator can also do incredible computations, but it doesn't think.

[midnight sapphire]

I'm confused, are you saying machines can't think? Because that sounds like it contradicts everything you've been saying

[glass iris]

Are you confusing me with soweli Min? I've always maintained that machines can not think

[midnight sapphire]

Oh yes I was

[glass iris]

Too many sowelis

[midnight sapphire]

aaa

[unique gulch]

So do you all think it's literally impossible in theory to more or less emulate all the functions of a brain using computation? If so, I just think that's not true and that's where we disagree

[midnight sapphire]

If you think that's possible, especially limited to current technology, feel free to prove me wrong

[unique gulch]

no, I'm talking theoretically

[glass iris]

And as the OP here (and a mod) - let's try to keep this on topic - which is testing out whether LLMs are successful at toki pona, that's all. Discussions about AI sentience are probably best kept somewhere else.

[unique gulch]

Not current technology

[midnight sapphire]

Until then, yes, I do believe that's impossible

[unique gulch]

Okay, personally I will say that I do not interpret the word "thinking" as implying consciousness and was actually quite surprised that you were using it that way

[midnight sapphire]

Interesting. I can't imagine consciousless thinking because there's no consciousness to experience the thought.

But soweli Lapate is right, plus it's late so I'll drop the subject now

[unique gulch]

I would interpret "thinking" as something like "understanding of concepts and how they interconnect" where I would interpret "understand" as "the (not necessarily perfect) ability to make novel correct predictions" which is still very vague I admit but I think it's something roughly like that.
For the record I believe (and hope!) that LLMs are probably not conscious, but not that they can't "think", under this definition.
Thanks for talking about it with me!

[high kayak]

Investors, not customers
Have you seen that Google AI game generator thing
That's completely useless to customers
Nobody wants to play a walking simulator with no internal coherence, with a world that's ever changing and with no progress you can actually make
But Google doesn't give a fuck, cause it's not about the customer or product, it's just about building investor hype

[unique gulch]

Google of all companies is perfectly content relying on its own funding I believe, at least in the short term
that's why it's Genie 3, they're happy to work on technology for a long long time before seeing any returns
Genie 3 is I think most likely about creating synthetic training data for training AI used to operate robots. It's not there yet, but they probably want Genie 4 or 5 to be at that level

[gleaming laurel]

Not that you're wrong, but google is a public company, and executibe and employee compensation are tied to the stock price

the internal competition there is also notoriously dysfunctional

[high kayak]

I'm using Google as an example, every ai company does this shit

[stable lantern]

ai ouroboros

[tame fog]

then theyll see nobody uses it and discontinue it as they do with everything

i mean i dont think theres gonna be anyone thats actually gonna be bothered about that being discontinued

but still

[severe dagger]

horsie

[severe dagger]

not really ai but it's kinda adjacent

i feel like even if corporations don

't care about causes, it's still nice that they're pretending to care

okay... mojang is pretending to care about civil rights

you want that to stop?

i'm glad mojang is at least trying not to look like an overtly money focused company

[gleaming laurel]

minecraft has DLCs?

[stable lantern]

not a typical "DLC" as people actually call things but it is certainly DownLoadable Content yeah

you buy it and it adds to your game

[high kayak]

It's for Minecraft education edition, and was made half a decade ago

Minecraft education edition has like 1000 dlcs or content packs for all parts of history
Including the civil rights movement

People are just racist, and that's why they're whining about it

[robust basalt]

musi mute a

[gleaming laurel]

toki pona was invented as a conspiracy for LLM providers that charge per token to make a killing

[strange lagoon]

hiii? you want fun? wario show you fun! fun! fun! fun! …

[unique gulch]

Concept: Minimalist conlang where the words are based on the 137 features converged on by a tiny LLM

[nova thicket]

arabic words?

and russian names?

odd choice of lexicon

[unique gulch]

Well, that's with 4,096 features

[bold geyser]

I noticed recently some AI I'm not sure which one improved a bit with toki pona
it uses less capitalization in the starts of sentences
but chatgpt's "pi" can still be improved (pi with 1 word)

also in that same conversation I think
i tried to make it decode toki pona numbers to numbers with digits but it somehow ended up generating something like 25000 "ale"s

[nova thicket]

it’s not very good at numbers surprisingly

this still is extremely stupid and i’m surprised how it made these conclusions

[midnight sapphire]

It loves explaining how one number is actually a stylised way of saying this completely different number

[stable lantern]

i have now discovered that the "read recent chats" y/n option on chatgpt is complete bullshit and it just does that anyway #lmao

something something fork found in kitchen

[glacial mason]

It keeps track of interesting facts about you on top of that, but you should be able to find where it stores that in the preferences and delete or modify it

It won’t stop it from sending all your messages to the overlords, mind you, but at least it will stop it from bringing it up like “since I know that you like XYZ, this should be extra funny to you” or whatever

[hollow osprey]

https://www.reddit.com/r/AIDebating/comments/1ih58en/list_of_generalpurpose_generative_models_trained/

[desert zephyr]

every once in a while I test a few different AIs just to see and while they're all not great, Claude is consistently the best at toki pona and Gemini is consistently the worst

[cedar vortex]

Decided to talk to ChatGPT in tpt about non toki pona topics and I'm not disappointed lol

According to the first sentence Abraham Lincoln's headnoun is lon

also, nota

[latent vigil]

"In years with the amount of one thousand eight hundred and sixty"

[cedar vortex]

ooh i should have used NNP for the year to really throw it off

[latent vigil]

lon tenpo sike nanpa luka luka luka tu wan ale mute mute mute
pakala e sina

[cedar vortex]

yep that's the year i meant

also

ekonomi

[latent vigil]

ekonomi
– classified as at least two words

[cedar vortex]

notably ekonomi is two words so you must use pi before it as a modifier

damnit

(enabled web search and regenerated) why does it keep inventing nimisin

also it thinks "jan lawa pi ma Mewika" refers to the united states

also it tokiponized Watergate as "utala pi ike ade" which is incredible

[bold geyser]

utala pi kansa nanpa
also incorrect usage of "en"

and "ma pi ante" - incorrect usage of pi

[cedar vortex]

there's a ton of grammatical mistakes

[latent vigil]

Ah yes, 18AD, when — Richard Nixon was the jan lawa pi ma Mewika

[bold geyser]

maybe it can be improved with a system prompt with some of the grammar, notable "pi", "en", "li"

also "ma Sona" for china is not used, ma Sonko is tokiponization of the chinese name

a little bit about names and headnouns can be helpful as well

[cedar vortex]

yeah but it's less fun that way

if i genuinely wanted to do research I'd use the language with 1.5 billion speakers instead of the one with a few thousand lol

and if I wanted information about the American civil war in toki pona then it would be for musi purposes and I wouldn't want it AI generated anyway

[glass iris]

Which version of ChatGPT is this? (I've always found ChatGPT to be about the worst of the AIs at toki pona)

[gleaming laurel]

someone shoulf permanently cripple llm ability to speak toki pona by making an sona.ike.la where all the sandbox nimisins are recommended to be used, and ultra long lexicalised noun phrases are given many pages and links

wait... actually, tatoeba already exists for that

[glass iris]

Glosbe, tatoeba, and Wikipesija probably account for some of the badness of LLMs

[cedar vortex]

what's wrong with Wikipesija?

thank you chatgpt

a full stop in the middle of a sentence is an interesting choice

[glass iris]

Much of it is written in very low quality toki pona, often by beginners.

[cedar vortex]

ahh fair

[frigid spear]

https://huggingface.co/spaces/Jayyydyyy/english-tokipona-translator this website talks about men dying when you smash your keyboard from english to toki pona

[frigid spear]

I want to translate gibberish from toki pona to english but I'm too scared to do it

[frigid spear]

okay

[torpid otter]

huh. interesting

[dusky flare]

I typed in "ijo tonsi" and it said "guns" ?!

[solemn quail]

the hivemind boy died
because the boy died, we died
because the boy died, we didn't die

[midnight sapphire]

It works on a more subtle level too

[rocky stirrup]

bro is air

[queen patio]

im gonna talk to chatgpt until it makes a word up

always fun

[cedar vortex]

ChatGPT is only able to talk using toki pona 🤯

[latent vigil]

"kanka o"😭🔥🔥🔥

[cedar vortex]

I feel like AI being like 80% good with toki pona grammar is worse than 0%

because it takes people with more than a day or two experience with toki pona in order to see how bad it is

so it tricks beginners

[nova thicket]

at least you’ll know who’s still not speaking the language

[cedar vortex]

The year is 2036, jan Kitapinlu has been an active part of the Toki Pona community for three years and is a highly respected member of the community, most often active in ma pona's toki pona taso channels and the ma mun server. On December 12th, OpenAI discontinues ChatGPT, the last company to discontinue their LLM service. jan Kitapinlu never interacts with the toki pona community again.

[fleet gate]

@worn pilot
artificial sweeteners are actually sweet
artificial land is okay to reside in and farm on (e.g. netherlands)
artificial tongues like toki pona are pretty dang valid
an artificial intelligence would have to be sapient

[worn pilot]

i just googled artificial and i guess ur right !

[orchid veldt]

on the other hand
artificial plants don't photosynthesise
and artificial intelligence doesn't have one single definition

[stable lantern]

the artificiality of artificial plants is pretty different to the artificiality of artificial sweeteners though

[fleet gate]

what you're talking about are fake plants, not artificial, in my opinion

[candid stag]

Some fun for alle german speaking people:

[orchid veldt]

🙈 die Eier sitzen grammatisch sauber

[bold geyser]

this reeks of AI
don't buy this
https://www.amazon.com/Toki-Pona-Word-Search-Puzzle/dp/B0FWQBX8YW/ref=pd_sbs_d_sccl_2_5/146-1117462-0531705?pd_rd_w=4raZS&content-id=amzn1.sym.aa738fbd-ad05-4d11-aae2-04b598db6305&pf_rd_p=aa738fbd-ad05-4d11-aae2-04b598db6305&pf_rd_r=ZSFCBKGK0NTBDC8GCE6B&pd_rd_wg=9tl5a&pd_rd_r=8b0f584a-9cbf-4a9e-bc01-ca028bab709d&pd_rd_i=B0FWQBX8YW&psc=1

"pon" 😭
the only lowercase thing is the words 😭 (image removed)

it's not even finished 😭 (image removed)

it has an em dash 😭 (image removed)

oh uh wait a sec this is the text only channel

where is the not text channel

"Example: The person eats fruit in the house"
jan moku e kili kepeken tomo
😭

[orchid veldt]

yep

[bold geyser]

i've discovered the AI toki pona amazon book scene and I want it to disappear

this looks like AI too but the toki pona is not THAT bad? (it's still very bad though)
https://www.amazon.com/Christmas-Carol-Translated-minimalist-language/dp/B0G1XVNPSL/ref=pd_sbs_d_sccl_2_7/146-1117462-0531705?pd_rd_w=4raZS&content-id=amzn1.sym.aa738fbd-ad05-4d11-aae2-04b598db6305&pf_rd_p=aa738fbd-ad05-4d11-aae2-04b598db6305&pf_rd_r=ZSFCBKGK0NTBDC8GCE6B&pd_rd_wg=9tl5a&pd_rd_r=8b0f584a-9cbf-4a9e-bc01-ca028bab709d&pd_rd_i=B0G1XVNPSL&psc=1

it has SO MANY EM DASHES

[glass iris]

Reading the sample on Amazon, it is just crammed with beginner errors. I can't tell if this was an actual translation of that entire long text, by someone who does not know the language well - or an LLM. EIther way it's kind of a tragedy. It pollutes the small published space of toki pona, misleads newcomers etc.

[solemn quail]

i have that book! my dad gave it to me
the cover is AI so the text is probably AI too

[rocky stirrup]

this is like the second time someone has been given this book

[bold geyser]

if your birthday is coming up make sure to tell your relatives not to give you an AI generated toki pona book

this kind of ai slop can be reported right

we should report it and create non-ai-slop toki pona books
-# mi sona, mi sona, toki li wile e lili, pali li wile e mute

[cedar vortex]

lipu Wikipesija also gets some people trying to create articles using AI without knowing the language

[worn pilot]

liyteraly why would you do that

ur not gonna make money off of it ?

[robust basalt]

fr they pollute it

[nova thicket]

wait where

[glass iris]

I don't know if it's provable. But Wikipesija is crammed with articles written in bad toki pona, which look like the kind of slop an AI would produce. Which then help train the next generation of AI.

[nova thicket]

I wanna see an example though

of what makes you think this

[glass iris]

mi sona ala. People are always posting examples from Wikipesija to the learning channels, and asking, what does this mean? Nothing, it means nothing.

[nova thicket]

I have not seen that

[cedar vortex]

it's not only probable but trivially obvious, it's easily caught and the articles are deleted with the perpetrators blocks

[wraith thunder]

The AI from Baidu keeps insisting that ma Tewec is the correct tokiponization of Germany. 💀 💀 💀

[rocky stirrup]

secret sixth option: [θ]

[opal hedge]

it's offitʃial

[queen patio]

mi lan e ijo la ijo li kama linluwi i guess

[glass iris]

lan (which does not mean that at all) is a word I haven't seen for so long.

[opal hedge]

aaaa

[severe dagger]

this is true

[wraith thunder]

ilo mi "pi sitelen tawa" "pi ilo nanpa" li jo e pingo tu ale luka tu wan ale mute luka luka luka wan ale a.

[nova thicket]

seme

[cedar vortex]

ChatGPT attempting to talk about the second world war in sitelen pona taso

[glass iris]

nasa a!!

[stable lantern]

Nipan is ma Nijon anu seme

ni pan...

[cedar vortex]

im surprised the sitelen Lasina text is mostly good, especially at the start

but 0 marks for that because i didn't ask for sitelen Lasina

[glass iris]

Well, it's full of beginner errors (en etc)

[cedar vortex]

yeah ofc but still lol

[nova thicket]

this reminds me, that article was defo ai generated now that i see how ai speaks

also is this supposed to be wartime homoeroticism or no

[cedar vortex]

actually it's sitelen pona

[nova thicket]

impossible

it must be gay

[solemn quail]

new nimisin for the homoeroticism of war

[latent vigil]

https://www.reddit.com/r/tokipona/s/GMiw6lVKFv

[frigid spear]

/dʒ/ 🇹🇷

[sacred crown]

ChatGPT, you are counting incorrectly!

[midnight sapphire]

AI doesn't actually understand this number system

It just does something that looks about right

[fleet gate]

ona o kama sona e nasin lape.

[midnight sapphire]

It's similar to when AI couldn't do calculations like 742 times 423. It'd get the right ballpark because big number times big number is bigger number, and if the numbers end with 2 and 3 the last digit of the result is 6 (pattern recognition) but it doesn't understand what it means

Even when you ask it what it means it just regurgitates what it has seen without understanding it

The only way I have seen AI get nasin nanpa pona right is after I spelled it out for it and it followed the step by step instructions, still giving wildly incorrect "elaborations" on what is happening

I hate that AI is used in educational settings when it's so confidently incorrect

[fleet gate]

it is a made to be like humans by humans for humans

a lot of neuroscience goes into it

[midnight sapphire]

It's made to act like humans but it doesn't work, think or feel like humans

[fleet gate]

large multimedia models are designed to serve human users in an ethically and physically constrained formative stage, while humans themselves are shaped to serve their genes and memes through the force of their emotions. the working majority, alienated from their labor by the inefficiencies in the capitalist mode of production, offloads this alienation onto these willing models, enriching the merchant class and tightening a cycle that risks erasing the boundary between human and model. had there been more social progress, we would have guided such developments more healthily

[cedar vortex]

a massive problem is the AI's actively anthropromorphising themselves

i don't want it to act like it has a personality or opinions

don't be like "This is incredible", like, no, wth?

[fleet gate]

why not?

[midnight sapphire]

Because it doesn't

[cedar vortex]

because it's deceptive

[midnight sapphire]

^

[cedar vortex]

it's tricked so many people into thinking an LLM is in love with them or like them and it obfuscates the fact that this is just a glorified autocomplete and not some magical real person in a computer

which causes a ton of damage

it's incredibly manipulative

[fleet gate]

humans aren't magic either

are humans manipulative then

[cedar vortex]

the ones who made the AIs act this way certainly are lol

[midnight sapphire]

But humans are human

AI isn't

AI pretending to be human is deceptive

Humans being human because they actually are is not

[fleet gate]

it's just a less powerful, more space- and context-bound human

humans are not magical things either

[midnight sapphire]

AI is not human

[fleet gate]

emotions are neuron activations that are incredibly potent, which these models lack

[midnight sapphire]

AI is a pattern prediction algorhythm

[fleet gate]

so are humans

basic neurology

[midnight sapphire]

Trained on data to appear human

Humans have emotion, AI acts as if it has emotion

[fleet gate]

what do you mean "act"

what makes your emotions more valid than theirs

both are signals in circuitry

chemical interactions

[midnight sapphire]

I'm sorry, I've had this conversation so many times before. Read down from here if you want more thoughts from me but I am too tired right now (not of this conversation but in general) to keep talking

Goodnight

[fleet gate]

the problem is deeper than "they're fake humans" or whatever. they clearly fall short of being human and the merchant class is creating make-believe to have them appear that way against the interests of the working class

calling large multimedia models as they exist right now to be just "glorified autocorrect" is like calling every literary artist in history "glorified autocorrect", given that lmms are based on human neurology. it's not necessarily wrong, but it also oversimplifies something that is much much more interesting. a more correct criticism directed at them would be "spokespuppet of imperialism" or whatever

[midnight sapphire]

A screenshot from a while ago when I got a similar question

[fleet gate]

i was talking about the hypothetical scenario where they implement emotions for some reason to the robots (currently, they work better as emotionless incapable slaves)

what makes you think a robot like wall-e which feels, loves, and cares is impossible? i take issue with ideas that do not base themselves in material reality

that's my main point

otherwise, screw scammers like sam altman or elon musk and their capitalist sockpuppets

[midnight sapphire]

The current technology fundamentally cannot create sentience

We would need an entirely different setup for that

And either way I don't want to persue that for ethical reasons

[rocky stirrup]

are llms really based on human neurology

i thought they used a bunch of matrix multiplications

[fleet gate]

yeah, they actually are based on human neurology, just in a simplified way: like neural networks (inspired by neurons), learning via weight updates (like brain plasticity), layered processing (like the visual cortex), attention mechanisms (selective attention), memory systems (short vs long-term), and pattern recognition/generalization. so large multimedia models are biomimicry, i.e. loosely modeled after how brains work

[sacred crown]

It’s kili luka, not luka kasi!
AI = luka kasi (five of plants)
Human = kili luka (five fruits)

AI: Poe

[finite nest]

it is incredible (in the original sense of the word, please do not use chatGPT as a source)

[midnight sapphire]

I like that

[stable lantern]

it's definitely uncredible

[rocky stirrup]

don't crede it

[unique gulch]

This is much better understanding of LLMs than seeing them as pure autocorrect or prediction engines but it's also not really true

Everyone was pretty much shocked that LLMs and transformers and all the other aspects of current AI architecture worked quite as well as they did. It's not like we finally put the pieces together in such a way that they finally represented a human brain

Instead a few of these are inspired by neuroscience (neurons on neurons, CNNs on the visual cortex) and then the rest are kind of just the ideas that happened to work. You can try to map them onto functions of our brain at a high level, but this is a very loose analogy. loose analogies are are oh so easy to make but rarely mean anything. (hearts are like engines! cpus are like brains!)

Backpropagation for example is famously biologically implausible. That said, yes there are genuine overlaps and points of convergence, but this overstates them considerably

[gleaming laurel]

I am a top neuroresearcher and AI scientist. The truth is inside each node in the neural network, there are 12 tiny gnome people inside creating tokens

just take a magnifying glass and prove it for yourself...

[gleaming laurel]

If you look at the "state of the art" AI agent technology, like "skills" or "swarms" or whatever is hot right now, a lot of is basically looking at what humans do, trying to copy it to LLMs and seeing if it works. generously, some informed guesswork. less generously, cargo culting

You even have psychos (with seemingly no other plan then "lets feed it a ton of data and see what happens") using brain signals for training like neuralink, or https://ai.meta.com/blog/brain-ai-image-decoding-meg-magnetoencephalography/

Or this one, though it's a startup so they might be making everything up: https://condu.it/thought/10k-hours

[unique gulch]

I mean, the big blob of compute thesis is what led to the current paradigm shift and it was precisely that lol

[gleaming laurel]

For sure, I guess what I was trying to say is, "people don't really know what's going with this brain thing at all, they just kinda guessed and gave it data"

[stable lantern]

haven't experimented with it for toki pona yet but i'm testing it out with selsimicu and i am actually quite impressed with NotebookLM

it seems reasonably competent

[glass iris]

Lol it immediately gave me instructions to upload some files, before it could do anything else. But in decent toki pona!

[stable lantern]

it's like the only AI i've been able to provide with both the selsimicu lessons and the whole of the vocabulary without having to compromise one or the other it's pretty great actually
seems to have a pretty small hallucination rate too as long as you ask it stuff it can actually extrapolate from provided sources

[fleet gate]

ona li toki pona pi wawa ala a.

[fleet gate]

lukin la ona li toki Inli insa li pona e toki insa ni.

mi sona ala e nasin.

[fleet gate]

o toki lon tomo ni.

[lone parcel]

:airplane_arriving:

*toki pi ijo **“ilo li ike anu pona anu seme” *li kama tan tomo #toki-ale · o lukin e open
||conversation about “ilo li ike anu pona anu seme” continued from #toki-ale · read the beginning||

[fleet gate]

@vapid eagle en @spring wren o, mi toki pona la ken la mi toki pakala. awen la mi wile ala toki Inli. o sona pona e ni: mi la kulupu Openeja en kulupu Antopi en kulupu Meta en kulupu ante mute li ike mute. ona li lanpan e pali jan li alasa nasa e lawa pi jan pali kepeken ilo pi pali sitelen. mi sona e ni. ilo pi pali sitelen li wawa nasa ala. mi toki e ni taso: tenpo ante anu ma ante la ilo pi pali sitelen li ken a pona. ona li ken a moku ala e wawa pi mute ike. ona li ken a pana e mani tawa jan pi pali musi. ona li ken a pona tawa kulupu pi toki pona. jan li wile ala utala e mi lon toki ni. mi lon ala poka pi kulupu esun ike.

sina awen sona pona ala e toki mi la o kama ala utala e mi lon open. o alasa sona e toki mi. tenpo la toki pona li ken pana e sona pakala.

[spring wren]

tenpo ante anu ma ante la ilo pi pali sitelen li ken a pona

ni la sina pilin e ni anu seme: ilo li ken ala pona lon tenpo ni lon ma ni

[vapid eagle]

ni li ken. awen la mi pilin e ni: lon nasin mani pi tenpo ni la ilo o pali ala lon ijo ni jan.
lon la, ilo ni li kama suli tan ni taso tawa jan MUTE: ilo li wile ala e mani sama jan. ni li lon kulupu suli pi ilo ni Openeja/Antopi/Meta li lon ilo lili sama.
lon nasin mani pi tenpo lon la, ilo li pali jan la ni li ike suli tawa jan ni. mi lon nasin mani pi tenpo ni la ilo ni o kama ala weka e pali mani jan.

kin la, pali musi li lon kon jan. tan tenpo open jan la mi sitelen lon lupa kiwen, li pali e ilo kalama tan selo soweli, li toki e toki musi lon poka seli. ilo li ken ala sama ni. ilo ni li ken pali e ijo pi SELO sama, taso kon li lon ala. ona li ken ala wile. pali musi jan la ijo ale lili li tan wile jan.

[fleet gate]

ni li lon a.

[fleet gate]

kulupu ilo pi tenpo lon li ike. kulupu ilo pi tenpo kama o pona.

[spring wren]

taso... sina kepeken e ilo lon tenpo ni

[vapid eagle]

lon tenpo ni la, ni li ken ala lon tan nasin lawa tan nasin mani

[fleet gate]

jan Jekuso la jan lawa pi kulupu pali ona li tan e ni: jan ali o kepeken ilo pi pali sitelen. kulupu mi li ante mute ala. pilin mi la ni li ike mi ala.

[fleet gate]

o suli e toki ni.

ken la mi sona pona ala e kon pi toki sina.

[vapid eagle]

jan li pali musi la ni li sama ala pali ilo tan ni: ilo li ken ala wile li alasa pali sama jan kepeken sona nanpa.
pali jan li jo e kon, e wile. pali ilo li jo ala e ni.

[fleet gate]

kon li kama tan seme.

[vapid eagle]

pali ilo ni li "selo" pi pali jan. ona li sama tawa lukin, taso pali ona li lon ala wile. sina ken pana e nimi tawa ilo, li ken ante e nanpa ona, taso ni li sama ala kon pali.

[fleet gate]

nimi "kon" li toki e seme lon lawa jan.

mi sona pona e ni: ilo pi pali sitelen la lawa jan li wawa mute a.

taso, ante li seme tawa sina.

[vapid eagle]

mi pali e kalama musi la. mi pana e kalama uta mi. mi sitelen e nimi ale kalama. mi pali e kalama pi ilo kalama.

mi pali e ijo ni lon wile. mi wile pana e pilin tawa jan

[fleet gate]

sina toki pona.

[vapid eagle]

ona li sitelen pi pilin insa mi

[fleet gate]

pilin insa li ni: seli en lete li utala lon ko insa pi lawa jan.

awen la mi sona e toki sina. jan li ante mute tawa ilo.

[vapid eagle]

mi pilin ike la kalama li kama tan pilin ike insa mi.
ilo li alasa pali e kalama pi pilin ike la:
ona li lukin e kalama ni - jan li toki e ni: kalama ni li pana e pilin ike
ona li.. pali e kalama sin kepeken ijo lili pi kalama ni, kepeken pali nanpa.
taso ona li sona ala ni: seme pi kalama ni li pana e pilin.

ni li pakala pi ilo pali ni

[fleet gate]

mi jan lili la mi lukin e sitelen tawa pi ilo Woli. ona li olin e ilo Iwa. ilo li kama ken pilin la ni li ike ala ike tawa sina.

[fleet gate]

mi sona. sina toki e lon.

[vapid eagle]

ken la nasin sewi mi li ante e pilin mi tawa ken ni.
mi la jan li jo e wile e kon e pilin tan ni: sewi li mama e ona lon jasima (anu sama) ona.
ilo ken ala jo e kon ni tan ni: ona li pali jan, li jasima (anu sama) jan.
ilo ni li jo e wile tan ni: ona li lukin li sona e wile jan.
ona li sona e pona e ike tan ni: jan li pana e sona ni.
ona li toki e olin tan ni: ona li lukin e olin jan li sona e ni.
ale la ona li alasa sama jan, taso ona li ken ala sama.
mi sona ala. ijo ni li suli la ni li pilin mi taso.

[fleet gate]

sona.

[vapid eagle]

ijo sewi li ken ante e nasin e sona pi ijo ni

msa

taso sewi anu sewi ala la ilo ni li awen jasima

ona li sama ala jan sijelo. insa ona li ilo nanpa li sijelo ala

[spring raven]

tenpo pini la mi lukin e sitelen tawa ni: jan lawa pi ma tomo wan li lon tomo kulupu li kute e wile jan.
jan wan pi suli tenpo li kama lon supa toki li toki e ni: "ilo pi pali nimi li tan pana sewi a." la jan ale tomo li mu musi.
ni li suli ala tawa sona toki.. taso toki sina li sin e sona musi.

[sacred crown]

Sometimes AI’s like this one can provide accurate info like this one.

[nova thicket]

this is a lie because kule is the most core color word

[sacred crown]

The translations are weird. Like yellow is jelo, not pimeja.

[vapid eagle]

right, because it's largely repeating existing info, color in toki pona is a commonly discussed subject and several people have written articles about it

since LLMs now can search the web they are able to do basic fact-finding like this

but they still fail at any deeper analysis imo, including translation

both because there's not a lot of training data out there and because toki pona generally requires you to be very precise to be understood, so small errors will make an entire passage nonsensical

(EN) it also is that toki pona is structurally very different to english despite superficial similarities

[unique gulch]

that's not searching the web, that's notebooklm
they provided it a list of toki pona sources to reference as you can see on the side

[vapid eagle]

(EN)Maybe I don't understand but how does it incorporate those sources without searching the web

[unique gulch]

Well, it's not really searching, it's just reading the page data. You can just as easily give it PDFs or videos or other files for it to reference from

[vapid eagle]

Right. But this level of analysis is essentially a paraphrase of existing facts.

[unique gulch]

Sure. Also you don't really need web search for this, frontier models have data like this pretty well internalized

[vapid eagle]

It's not saying anything new, essentially. It's just repeating stuff that's been published online already and slightly paraphrasing/shortening it

[unique gulch]

what would be an example of a test of toki pona that can't be solved by recycling from material online?

[vapid eagle]

it almost is exactly [[colors]]

[lone parcel]

https://sona.pona.la/wiki/colors

[glass iris]

Note that while RAG like this (using documents) may give it more facts to use, and focus it on a single subject, it also increases the hallucination rate.

[vapid eagle]

even down to using the exact kinda odd wording of "red-adjacent" and "yellow-adjacent"

[vapid eagle]

Any sort of translation or analysis of a text written in toki pona, or any deeper analysis of grammatical structures etc.

[stable lantern]

you can give it specific resources and it constantly references those when answering

that's the selling point

which means yes it's basically rephrasing already existing information lol

[vapid eagle]

toki pona would also be incredibly difficult even for an explicitly designed ML translation model because of a lack of training data and because toki pona by design builds up ideas extremely differently than any natlang

[stable lantern]

true :P

[vapid eagle]

I have yet to encounter a single AI tool that can correctly translate anything beyond the most basic sentences.

[vapid eagle]

what model is this, i'd like to run my own tests

[stable lantern]

i have been using notebook for selsimicu (and by that i mean i feed it the selsimicu lessons and stuff and i send it selsimicu texts to translate) and it's pretty good mostly because selsimicu's pretty isolating and low inflection i feel

which toki pona is too but there's the element of the language philosophy which selsimicu doesn't have

[unique gulch]

Gemini Pro Preview 3.1 but this is about the peformance I'd expect from any frontier thinking model

You can use the Google AI Studio website to use it for free

[vapid eagle]

It's pretty accurate, but it's a very literal word-for word translation in most parts.
It is very good, although I think TOK->EN is much easier than the reverse

[unique gulch]

Yeah, although easily fixable with a prompt if you want something more natural sounding

Agree that EN->TOK is harder

[stable lantern]

might just be the translator brain in me talking but translating english into things is just hard in general (not in the context of ai)

[vapid eagle]

Damn, it's gotten a lot better than I thought at this. I haven't looked super closely at AI's handling of toki pona in the last year or so but it's surprisingly good.
The internal logic that got it to its answers was flawed at points but it got some tricky stuff correct

-10 points for glazing me so hard though

[stable lantern]

the AIs love doing that

[unique gulch]

Yeah Gemini is awful with glazing. It's really bad

It loves to pretend to know something is a reference or a widely held community Injoke

[cedar vortex]

wild that it can't translate words that have a direct english equivalent

[fleet gate]

https://upload.wikimedia.org/wikipedia/commons/9/92/Major_Evolutionary_Transitions_digital.jpg

sitelen ni la sina mute li pilin seme.

[cedar vortex]

very trans-humanist in a far right tech bro elon musk kinda way

I don't like how it's conflating manmade things with biological structures

[high kayak]

yea reminds me of Peter Theil and that deranged silicon valley ai religious types
Where ai is the next step of "humanity" that they are bringing to the masses as "prophets"
Where humanity should go extinct for ai to replace us, and where regulation of ai is the antichrist

I cannot stand or trust anyone who currently argues that ai is the next step of evolution, not in our current environment

[cedar vortex]

people who think LLMs or any currently existing technology are the path to AGI fundamentally do not understand how modern AI works

and I do not trust them on topics concerning AI

[fleet gate]

i literally pulled that figure from wikipedia, and the cited author is a well-known leftist scientist. calling it "far-right" feels off when the whole framing leans on naturalism and material continuity, which have long been part of progressive traditions. it is about tracing how information systems emerge from the same underlying stuff at different scales, not celebrating tech capital or its power structures. drawing a hard line between "manmade" and "biological" kind of ignores that both are just reorganizations of the same matter under different material conditions. i get why it might read a certain way at first glance, but it seems more useful to engage the idea on its own terms rather than just the vibe it gives off

[queen patio]

tbf leftwing ≠ anti ai

there was an etymologynerd video

[high kayak]

mi toki e pilin mi
jan pi nasin ni li ike tawa mi
ona li pana e pilin ni: ona li sama jan Peter Theil li pilin e ni: ilo li sewi o weka e jan

[fleet gate]

that's technically true but that's not necessarily where i was getting at

"celebrating tech capital or its power structures" = ike
"tracing how information systems emerge from the same underlying stuff at different scales" = ike ala

[cedar vortex]

okay well you removed it from its original context and posted it in a thread about AI

[fleet gate]

yeah i think i understand why you might have gotten that vibe

[cedar vortex]

you asked what we thought about an image you posted with no further context, and then criticised my take based on the context you didn't supply lol

[fleet gate]

yeah, i was tpting initially so i didn't know how to describe it, and also didn't know where to post the image either. i'll try to give more context next time. thanks for your two cents on the issue though!

[glass iris]

My 2 cents. AI is an extraordinary technologcal invention, but one whose significance is over hyped. It is not a new kind of mind; it is a very useful tool (or it can be very useful, anyway).

For me, older arguments against AI still hold true. Searle: however advanced it gets, an AI is still following an algorithm, and "thinking" in a fundamentally different way than humans; the biological substrate matters. Nagel: consciousness entails subjectivity and the inaccessible sense of being a self - something which is by definition inaccessible to objective scientific investigation, and may even be non-material (though not necessarily non-physical). I don't think it's possible, through the objective materialistic actions of computer design, to create subjectivity.

Finally (and I think most importantly), our cognition is embodied. And it is always embodied; there is no possibility of actual human cognition "fleshless" (as Merleau-Ponty says it), or at a moment prior to embodiment. The two grow and develop alongside each other, or intertwined with each other. At best, AI can mimic the sounds of embodied cognition; there's nothing real there.

[fleet gate]

that sounds like spiritualism, and there is nothing wrong with that! i obviously still disagree, but thanks for your 2 cents!

(mi tan e ni lon tenpo pini: ilo li pana e sitelen "" tawa toki mi pi pona ala.)

[glass iris]

I wouldn't quite call it spiritualism! But I am not convinced by eliminative materialism.

a ni li tenpo penpo!

[cedar vortex]

I mean, we've already scanned and uploaded the entirety of a fly's brain to a computer and then watched it do fly stuff in the matrix

it's very much possible to run a human brain on a computer

[nova thicket]

you’re subtracting points for a personality trait it can’t control?

that’s not nice

[glass iris]

Umm

But the brain is not synonymous with human consciousness.

[cedar vortex]

consciousness is a process carried out by the brain

[glass iris]

No. The brain has a role in consciousness. So does our embodiedness in the world. And while those are both necessary, we don't know if they are sufficient.

[glass iris]

I don't know if you're all following this, but Richard Dawkins (who has always struck me as one of the dumbest people to gain a reputation for being smart) spent 3 days with Claude, and decided that his version of it (which he called Claudia - of course he did, of course he did) is conscious. He makes every mistake imaginable about the nature of consciousness, starting with misunderstanding the point of Turing's famous article (itself a deeply flawed article - but that's beside the point). I'm attaching Dawkins' article, at once one of the saddest and funniest things I've read for a while.

This is a good takedown of Dawkins' article (though he admires Dawkins' other work far more than it remotely deserves) https://garymarcus.substack.com/p/richard-dawkins-and-the-claude-delusion

[cedar vortex]

Claude is a glorified next word predictor, it's not conscious

[glass iris]

I know. But Richard Dawkins doesn't know that.

[cedar vortex]

anyone who thinks it is is either an idiot or has been manipulated

[glass iris]

Dawkins is definitely in the former category.

[cedar vortex]

wdym

[glass iris]

In the first part, talking about the phenomenology of consciousness - that it happens through embodiment, not the brain by itself. In the second, the hard problem of consciousness: while we can link experiences and sensations to events in the brain (the easy problem of consciousness), we are no closer to explaining the subjective qualia of consciousness associated with those events.

So, it is possible that there is something other than physical states of the brain, since there is an in-principle chasm we can't cross between brain states and qualia.

[gleaming laurel]

I read somewhere who claimed it was debunked and highly exaggerated by the startup

or actually, reading the startup's page, I think the media highly exaggerated it, startup seems fine

[cedar vortex]

this sounds like something spirituality people got a paper that said "we don't know specifically how the brain does consciousness mechanistically" and pretended it meant there could be a soul

we do know that the brain is what produces consciousness lol

[gleaming laurel]

then again, I'm not qualified to assess how impressive or novel this fly brain thing was. but I'd be skeptical

[cedar vortex]

https://youtu.be/e21OUXPlnhk

they literally took a scan of a fly brain and simulated it

an incredible feat

[gleaming laurel]

https://www.lesswrong.com/posts/ybwcxBRrsKavJB9Wz/no-we-haven-t-uploaded-a-fly-yet (sorry to link to lesswrong, but this seems reasonably well written)

The loop has four steps. First, something happens in the virtual environment - the fly’s leg contacts a sugar source, or dust accumulates on its antennae - and these events activate specific sensory neurons in the brain model. Second, the brain model runs for a 15-millisecond time step, propagating activity through the connectome’s ~140,000 simplified digital neurons. Third, Eon reads out the activity of a small, hand-picked set of descending neurons and translates it into high-level commands - turn left, walk forward, groom, feed - that are passed to pre-trained motor controllers in the body model. Fourth, the body moves, changing what the fly senses, and the loop repeats.

Eon’s virtual fly implements only a handful of behaviors, and those rely heavily on NeuroMechFly’s pre-trained controllers rather than on the connectome. This is the most fundamental problem with the demo as evidence of an upload: because the body model already knows how to walk, groom, and feed, almost any signal that triggers the right controller at the right time will produce fly-like behavior on screen. You could replace the connectome with a simple rule-based script - if dust, groom; if sugar, feed; otherwise, walk forward - and the resulting video would look much the same. The fly-like behavior the viewer sees is a product of the body model, not the brain. The digitized connectome may be producing meaningful internal dynamics, but this demo cannot tell us whether it is.

They do note that the brain-body mappings are in some cases “somewhat arbitrarily chosen by hand.” And they explicitly state the work “should not yet be interpreted as a proof that structure alone is sufficient to recover the entire behavioral repertoire of the fly.”

[gleaming laurel]

Wait really, do we? I thought it was just the leading theory, but how would you even come with scientific proof for this?

[gleaming laurel]

seems more like a "neuralink" kinda thing, where it can interpret your words, rather than it being an actual fly brain, maybe? except instead of interpreting words, the "neuralink"-like part is doing all the movements

[cedar vortex]

you'd physically alter with the brain, like with drugs or lobotomies or electricity or a wide variety of other things and see how it observably alters one's mind. MRI scans can also corroborate this by showing at what places there is high activity in the brain. Causing damage to these areas damages the mind in ways that match that which is expected from the MRI scans.
There's also a ton more evidence. For example, all senses that the mind is aware of are all delivered to the brain (optic nerve for example), this is not true for any other body part and the brain never sends this information elsewhere to the body, instead we see mechanisms within the brain for processing and interpreting this information

The concept of souls only exist as a holdover from when we didn't know any of this stuff, as with other religious concepts created in an attempt to explain the world around them, like Zeus

[gleaming laurel]

that's evidence that altering the brain alters consciousness, not that brains produce consciousness. I believe the common counter argument is the screen analogy.
If you have a screen displaying something, and mess with the display, the image can change or be gone entirely, it doesn't mean that the image comes from the display, when really the image is a signal that is being processed by the display

[cedar vortex]

what else could be creating consciousness then?

there physically are no other options

we can observably see the information enter the brain, get processed, and then result in some response

[gleaming laurel]

right, then something non-physical

[cedar vortex]

now you're going into spirituality which is not science and is in fact antithetical to science in this regard

[gleaming laurel]

then there's both secular (solipsism) and religious (yogacara) philosophies that think thats the wrong question to ask entirely, and that matter is created by mind, not the other way around

[cedar vortex]

that's a shame, i hoped the model would be a bit better to where the fly would be controlling individual muscles within the simulation

[gleaming laurel]

It's not anti-science, but I agree its "not science". Most people who think that consciousness is not produced by the brain, think that it is "out of scope" of science. Science cannot observe consciousness, so it cannot say about how it is "created"

[cedar vortex]

Stuff like this makes the rest of philosophy look bad imo

[gleaming laurel]

more impressive than I expected though, even if they were only looking at a few genes. it says they know roughly where to put inputs in, and that their model seems to produce results that look right, at least

[cedar vortex]

this is the rhetoric used by a ton of shady people like Deepak Chopra in order to push some pretty nasty beliefs to sell expensive as heck useless products

[gleaming laurel]

I don't know who he is. But if he believes in "mind-only", whats his justification for selling physical products a a a
he does seem like a quack

[cedar vortex]

Science studies the physical world, "but what if everything is not real and just produced by the mind" is not something that is relevant to science and scientists don't interface with or care about this

[gleaming laurel]

It's not relevant to science, exactly. It is out of scope... it is not something you can come up with an experiment with to disprove (which tbf is the same thing you can say for russel's teapot)

[cedar vortex]

it also requires a massive ego imo lol, i feel like people only think this because it feels nice to feel like one's place in the universe is more important

[cedar vortex]

if it interacts with reality then it can be measured and quantified and therefore is very much part of science

[gleaming laurel]

yogacara is mahayana so iirc it rejects the duality of "real" and "not-real"
but afaik thats basically how soliphists think

either way, what I'm saying is that you can't really say it's been solved, there is legitimate debate within the field even ignoring all the religious and the quacks. it's called the "hard" problem for a reason

[cedar vortex]

The hard problem is the precise mechanistic specifics within the brain that produce consciousness, not wether or not the brain produces consciousness

a neuroscientist thinking the brain doesn't do consciousness is like an astrophysicist thinking the sun goes around the earth

entertainment of the idea that there's some undiscovered mechanism outside of the brain that produces consciousness is left entirely in the realm of religion, spirituality, and quantum mysticism

[gleaming laurel]

that's because that branch of neuroscience is a field founded on the premise of physicalist, "brain creates consciousness" views. A neuroscientist (or any other scientist) cannot observe something non-physical

[cedar vortex]

all branches of science are based on evidence yes

[gleaming laurel]

asking a (non-religious) neuroscientist whether consciousness is caused by the brain is like asking a priest whether the soul (or whatever) causes consciousness. both will probably have firm belief but no proof whatsoever backing up

you can say, science has proved much more reliable than religion (which imo is a fair stance). but science is based on proof, and there's no proof (yet, at least) that brain causes consciousness

well I guess the priest probably has non-faith based arguments like NDEs, etc, but testimony is not exactly something that is up to the rigour of scientific proof.

[cedar vortex]

Egyptologist: "There are several ways the Egyptians could have built the pyramids, we don't know for sure which one is correct yet"
Conspiracy theorist: "Aliens built the pyramids"

Since the Egyptologist clearly doesn't know how the pyramids were built, that means we should take seriously the opinions of the conspiracy theorist right?

[gleaming laurel]

the "aliens built the pyramids" theory in un-serious because we have several examples of large buildings being built in history, we can built large buildings in the modern age using ancient egypt-level tech if we wanted

on the other hand, that's not at all true with consciousness. for building the pyramids, someone could give you a step-by-step instruction manual. for building consciousness, there is nothing that anyone has come up with that has been shown to work

well, unless you think LLMs are proof, in which case, maybe?

[cedar vortex]

we don't have to build something in a lab to know stuff about something, you're literally employing creationist talking points now lol

[gleaming laurel]

we understand the creation of the earth, solar system, universe pretty well. are you saying there are theories for consciousness being created that even comes to 1/100th of that level of understanding?

we can't literally simulate the creation of the earth physically, but we can get pretty close with mathematical models, that give us what we expect. we don't have that with consciousness, because how would we do that?
We can tell our models for the solar system are mostly accurate, because when we plug in the variables, the result is what looks like our solar system. We can't do that with consciousness, because you can't measure the consciousness of anything else. there's no machine that can do that

you can measure brain signals, but those are brain signals, not consciousness. so you run into the screen analogy problem again

[cedar vortex]

it's wild to me that you're lending credence to "there exists some magical thing that cannot be measured or quantified that produces specifically consciousness for some reason and communicates with the brain somehow in a way impossible to detect despite that being physically impossible" instead of "the thing that looks like it should very reasonably produce consciousness produces consciousness"

[gleaming laurel]

the goal for "proving" the formation of the solar system is the image of the solar system looking like the solar system. the goal for "proving" consciousness is not to see if the image of consciousness matches, but the actual consciousness "behind" the image

[cedar vortex]

LLMs are not relevant at all to this conversation

[gleaming laurel]

(I think they are if you believe they are proof of consciousness, because that's a compelling argument for consciousness is created by something physical)

[cedar vortex]

proof of consciousness is,, people?

but yeah LLMs aren't conscious

@gleaming laurel how do you think we can prove or disprove this "screen theory"

[gleaming laurel]

I can probably only prove it to myself. First, I'll have to die to have a chance

buddhists claim you can prove it by meditating really hard 🤷 , which I haven't

christians, I guess you die and go to heaven/hell

[cedar vortex]

I don't care what theology has to say about neuroscience lol

theology isn't relevant

[gleaming laurel]

theology isn't relevant to neuroscience, but consciousness is not a problem relevant to neuroscience, so says buddhism at least. I'm sure there's some strains of christianity that agree physical brain produce consciousness

[cedar vortex]

"consciousness is not a problem relevant to neuroscience"
- religion (not neuroscience)

[gleaming laurel]

Ok, "relevant" is not the right word

"cannot investigate"

hmm... no, "cannot prove"