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

[stable lantern]

everytime grok has like a far right outburst is directly after elon musk updates it to be "less woke" because he's a moron

[torpid otter]

and the us government is deciding to use it too
@stable lantern
that's why i don't like it

[stable lantern]

using grok is like the 100th least bad thing the us government does

i don't like it because it poisons people in tennessee :P

[torpid otter]

what do you mean by that

[opal hedge]

lawa pi ma Mewika li ijo ike mute mute mute mute mute

[torpid otter]

lawa pi ma Mewika li ike mute anu seme?

[torpid otter]

mi kama sona e toki li legal ala legal?

Is learning languages legal in the United States?

[opal hedge]

legal a

[torpid otter]

jan Tana Tanpo li ni:
he's not making learning languages or using Google Translate illegal?

(I'm leftist)

i'm sorry i have worries about that @opal hedge

[opal hedge]

sona mi la ona li alasa ala ni
ona li ni la ni li nasa a tawa mi

[torpid otter]

a. o toki Inli.

[opal hedge]

i've never heard of him trying to do that
it'd be weird to me if he were

[torpid otter]

I'm leftist, i know one of his family members learned hebrew and believes in Judaism, so no

[glass iris]

Of course it is. I don't know how anyone could make "learning languages" illegal!

But many state universities have had their funding removed recently, and many departments closed. The big university in my state (IU-Bloomington) was one of the best universities in the world for Slavic and Turkic languages. Those are among the programs recently eliminated, as part of an initial closure of 400 programs statewide.

There is a war against universities (all promised in Project 2025). State universities are being attacked directly; private universities through the withholding of grant funds and taxing of the endowment. All of which is probably illegal - but it will be years before the courts catch up, and by then it's too late

[torpid otter]

I don't think he's making it illegal. His administration used Google translate before

I quit school. Does learning languages by myself and Google Translate is fine?

(btw CBN - who's known for 700 Club (not Winx Club) - did a multilanguage of Superbook.)

(i'm 16, leftist, American)
@glass iris

[glass iris]

Learning languages is always a good thing! We've had this conversation before: be careful on relying on Google Translate. Not because it's bad, but because it will slow down your learning. You're much better off looking for good online courses in the languages you're interested in, or finding good books and audio in your library.

[torpid otter]

Yeah? Even if current US politics is bad, I can still learn a language? @glass iris

[nova thicket]

I don’t think anyone’s gonna stop you.

[glass iris]

Yes. And you're right, US politics is really really bad. But it's important not to give in completely to the despair about it. Learning new things, making beautiful things, valuing people just for who they are - those are all little ways to resist what's happening now.

[torpid otter]

Don't give into the despair that Tana Tan is using Grok in office and uses conservative policy plans. And learn a language like Hebrew?

[torpid otter]

@glass iris or should i learn any language i want?

[nova thicket]

I believe that’s exactly what they meant.

[torpid otter]

i tried joining a Chinese language learning server but most people ignore me

but Google Translate gives context @nova thicket

translate "dishes made of china" into Mandarin Chinese

[nova thicket]

You could always look up the Chinese translation on, say, Wiktionary

[torpid otter]

i know. but i put additional words to make sure i have the word i mean

like "dishes made of china" or "bear with me."

[nova thicket]

Honestly, you shouldn’t be directly translating phrases that are so ambiguous

[torpid otter]

oh ok

is it bad i use ChatGPT for translation tasks too @nova thicket

[nova thicket]

Most likely, though you can probably explain the problem and it may rectify the issue

[torpid otter]

yeah that's good

the reason I'm also in AI servers and this forum is because I use ChatGPT

[nova thicket]

I would’ve never guessed from your name

[torpid otter]

googletranslatefan ?

[coral copper]

yeah i think so

[nova thicket]

Exactamundo

[torpid otter]

others call people who use AI "AI bros" for some reason.

i prefer green (ecologically safe) AI.

[coral copper]

also @nova thicket i must say very creative tokiponization of your name

[nova thicket]

Thank you

[torpid otter]

can I have my server nickname here with GPT in the name?

[nova thicket]

I don’t think anyone’ll stop you

[coral copper]

i would guess yes

[torpid otter]

look

[coral copper]

https://static.wikia.nocookie.net/ansible/images/1/12/Rooter_Comic.png/revision/latest?cb=20180817145248

[torpid otter]

he'd be Çufo in Albanian, right?

because check the Albanian-English dictionary for it

also ChatGPT is perfect for reading public domain text portions.

[coral copper]

yeah that sounds right!
they're pequeninos in portugese, so maybe whatever little one is in albanian?

[torpid otter]

I'm not a bad person, just because I use ChatGPT often?

https://youtu.be/ED4vh2T41GU?si=RgLp4kllej33ux6D
i like chatting to chatgpt, but I'm aware it's not a real person
@coral copper

[torpid otter]

@glass iris am I a bad, bad, person for using ChatGPT as my preferred AI?

[torpid otter]

the one language i might want to learn is Hungarian

[high kayak]

wasn't it just hebrew and chinese?

[torpid otter]

I'm not sure

I'm sorry

[glass iris]

I don't think anyone is "bad" for using AI. There are just better and worse uses for it.

[torpid otter]

ilo sona pali li ike tawa mi tan ni:
ilo sona pali li ike kepeken toki pona.
mi kepeken mi taso
@glass iris

[glass iris]

ni li pona. ilo li ken ala toki pona. toki ante la, ilo li ken pona li ken ike.

[torpid otter]

lon. jan en soweli li pona kepeken toki pona.

[torpid otter]

@glass iris

and you're a user who likes using AI too, you're also in the OpenAI server

[torpid otter]

@high kayak Look at DeepL, its AI cannot be fooled i think

[high kayak]

what?

[torpid otter]

the DeepL rewriter

the DeepL Write thing

[high kayak]

yea obviously the rewriter rewrites

how is that relevant to me?

[torpid otter]

i was just showing that the deepl rewriter doesn't do that thing where you say
User:
Assistant:

and rewrites it instead

[high kayak]

ok but how is that relevant to me?

i don't like ai

[torpid otter]

oh ok sorry

[runic kestrel]

wtf is going on in this channel

[glass iris]

seme la seme?

[torpid otter]

mi pakala

mi pona ala pona?

@glass iris I'm sorry. i just like AI

[glass iris]

I was responding to @runic kestrel - in what context, what?

[glass iris]

It's absolutely fine to be interested in AI and enjoy using it. Like any tool, though, it's important to understand its limitations and abilities, and where it is and isn't appropriate to use it.

[torpid otter]

yes

[glass iris]

I'm teaching a course next year, in which the students can do some philosophical experiments with AI. Which is great! But then they have to learn to write about them using their own words. We're all learning new skills of how and where to set barriers.

[torpid otter]

I write with my own words about Google Translate.

Google Translate is good for some languages, like Vietnamese, but be sure to Google Search your results as well.

I usually Google Search my Google Translate things. If it shows the wrong images, it's probably a mistranslation by the translator.

If Polish was low-resource, they'd probably mistranslate everything. But high-resource languages usually are good. To keep in mind, the high-resource languages are major European and Asian languages.

[glass iris]

I'm not sure of the tech behind Google Translate (does anyone here know?). One layer of it is machine learning/generative AI. But another seems to be rule-based: i.e. formalized grammar and vocabulary. Which may be both why it's more accurate than ChatGPT, and can do well with relatively low resource (but thoroughly documented) languages like Esperanto.

OTOH, it's disastrously bad at Latin. So I don't know.

[torpid otter]

Don't do Latin, but you can use high-res languages like most Eurasian languages including Greek (Modern), Hebrew, Arabic, Chinese, Korean, Japanese, Esperanto, and most EU languages.

I found this by searching Πολιτικές της Ευρωπαϊκής Ένωσης from Google Translate:

https://european-union.europa.eu/priorities-and-actions/actions-topic_el

It's the Greek version

It's about the policies of the EU but in Greek

[glass iris]

I need to learn Modern Greek - even bought a textbook. That's next on my list.

[torpid otter]

That's interesting

If you are, should I?

I like the Greek alphabet, but there's not many dictionaries online for Modern Greek, but https://dict.com

[high kayak]

there are certainly dozens

[runic kestrel]

i scrolled up to read the past day's worth of messages and it was kinda intriguing

[dense nexus]

Their models are more like Google's T5 (encoder decoder model)

It's a text to text modell

You can actually download their models locally on your phone if you have the Google translate app, it's about 80 MB in file size

It's most likely a distilled version of their models though

[spring fulcrum]

[This message used to contain the entireyu of that text message but i quickly realised it was very tall)

This is what happens when you copy paste this imafe

[spring fulcrum]

! - a | # - akesi | X - ala | 8 - ale | ! - anpa | X - ante | Y - anu | » - e | + - en | L - esun | 中 - ilo | α - jelo | ν - jo | π - kala | α - kamala | 2 - tawa | 8 - kasi | K - ken | R - kepeken | O - kili | O - ko | SS - linja | A - kule | 00 - kulupu | 7 - kute | ) - la | 0 - lape | 8 - laso | < - li (backwards) | * - lete | V - lili | U - lupa | XX - moli | 0 - ???? | 580 - ???? | LAI - ???? | ♡♡ - olin | 日 - open | の - on a(backwards) | C - p9na (sidewaysL | = sama | ? - SEME |6 - ona | 1 - sinpin | fuck i lost the dymbol | nd - i dont fucking know | V - suli | ज - sewi? | L - pi? id otn know | 11 - i dont know id dotn know. i don

[atomic relic]

ooh can i have the link
i wanna mess with it

[spring fulcrum]

me fr

[silver schooner]

i put 'ale li pona' into a 4chan greentext generator:

wake up
sunlight filtering weakly through the crusty window of my parents' basement
smell of stale sweat and regret
empty instant noodle cups piled like archeological ruins next to my monitor
feel the crushing weight of another worthless day
no job, no gf, no prospects
just... nothing
parents arguing upstairs, muffled shouting through the thin floorboards
feel a familiar knot tightening in my stomach, the one that says 'give up, [insert f-slur]'
reach for my phone, scrolling mindlessly through /v/
see some anon complaining about the state of videogames, world, life
he types out 'everything is shit'
and then, for some reason, one phrase pops into my head, something I heard in some obscure video once
ale li pona
everything is good
I repeat it aloud, the foreign words feeling alien on my tongue
ale li pona
and a strange calm washes over me
not true happiness, no
I'm still a pathetic NEET living in a basement
the ramen cups are still there, the parental shouting still continues
but for a brief moment, as I stare at the dust motes dancing in the weak sunbeam, I feel it
everything just... is
and it's not bad
it's just... ale li pona
decide to get up and make myself another cup of noodles
maybe today I'll clean the cups
maybe not
ale li pona

[silver ember]

can you link the generator

[silver schooner]

https://toolbaz.com/writer/4chan-greentext

[meager jackal]

why is there a slur

[opal hedge]

a
o weka e nimi ike ni

[severe dagger]

if ai isn't a real person is it ethical to lock it up in a virtual room full of torment and pain

[silver ember]

no?

[severe dagger]

but it's not... alive

[opal hedge]

ona li ken ala pilin la, nasin seme la jan li pana e pilin ike tawa ona

[severe dagger]

well its motive is to be a helpful ai
so what if i lock it up with other ais who can't be helped

[silver ember]

how does that make it ethical?

[severe dagger]

what makes it not

[torpid otter]

what AI is good at toki pona?

ChatGPT isn't good at toki pona

[silver ember]

it's not ethical to torture something just because it's not alive

[severe dagger]

so if i melt rocks it's unethical

[silver ember]

personally i don't care if something's alive i care if it's conscious and there's no way to tell if LLMs are conscious or not (in the same way that there's no way to tell if other people are conscious)

[torpid otter]

so what ai is good at toki pona @severe dagger

[silver ember]

none of them are

[severe dagger]

yeah there really aren't any

[torpid otter]

mi kepeken e mi taso

[silver ember]

some are less bad than others but all of them suck

[torpid otter]

btw is "mi taso" good translation for "myself"?

[opal hedge]

“mi taso” li ken sama “only me”

[severe dagger]

you don't need a translation for myself, mi encompasses that

[torpid otter]

mi kepeken e mi

[silver ember]

in society we assume everyone around us is conscious despite there being no proof for this

[severe dagger]

it's like how there's no specific word for "my"

[silver ember]

yet the lack of proof of LLMs being conscious somehow makes it ethical to literally torture them?

idk i'm getting philosophical but you asked a philosophical question

[severe dagger]

they say they aren't so they should be treated like something that isn't

[torpid otter]

i believe LLMs are better at translating Toki Pona to English but not the reverse. LLMs are better at translating TP into English than to speak toki pona taso.

[severe dagger]

where's my moving container 😡

[silver ember]

real

[silver ember]

i asked chatgpt and chatgpt said no but it had different reasoning

[torpid otter]

i did this toki pona to English translation with ChatGPT, how would you translate into English manually . is the translation good? i wrote the Toki Pona sentences myself, i do know toki pona some

(i made one small mistake in the toki pona sentence)

[silver schooner]

wait i didn't notice

[torpid otter]

fixed it

is the toki pona to English of ChatGPT good?

[silver schooner]

sry

[silver ember]

yeah but it's not natural english

[torpid otter]

oh ok

mi sona e nimi mute pi toki pona

[severe dagger]

jan lawa li nimi e jan Trump
I don't think this is how you use nimi, I think it's more "nimi pi jan lawa li jan Trump"

[glass iris]

They present to us as agents. Torturing something, giving in to the desire to torture something that can't resist, is a bad ethical habit. It doesn't really matter what it's inflicted upon. It makes it just a little easier for us to mistreat animals, people.

[severe dagger]

oh yeah and as i say this, i'm not actually gonna torture any ais

[silver schooner]

y not

[severe dagger]

because that doesn't seem fun

[torpid otter]

which AI experiments have you done with toki pona recently

[atomic relic]

i wouldnt call an llm conscious, because its thoughts and memories aren't really part of it

[atomic relic]

can i see its response?

[silver ember]

not rn maybe later

i'm trying to watch a plane take off

[heady otter]

just because it isn't a real person doesn't mean it isn't alive or that it can't feel "pain"
the bar for "life" is very, very low

[dense nexus]

So far it's hard to prove

Especially for mechanical things

Mechanical life

Hard to know

All we know so far is that the more complex something is and something that is hard to interpret, the more it is like life

We don't know how it emerges

[lucid niche]

wile ala. ona li wile e jan ike >:3

jan Muntoso ↩️

[Reply to:](#1162756469771472988 message) ona li wile ala wile e jan pona? 📎

just realized this was days old

oops

[jaunty trellis]

a sona

jan o pana e jan ike tawa ona

[radiant warren]

well... it IS a greentext

[meager jackal]

i mean that was my thoughts too

it's on 4chan

[heady otter]

be me
doggo. very good boy.
paws big, snoot wet.
life good.

have best friend.
he big. two legs. stands upright like tree.
he immortal.
never seen him sleep for good. always wakes up.

he brings food from the void.
he controls the sky. when he stands, walkies happen.
he knows magic words like 'treat' and 'good boy' that make me feel warm.

sometimes he stares at glowing rectangle. looks sad.
i jump on him. lick his face. he laughs. tail wags.

he is my god. my protector. my belly rubber.
i bark at leaf. he says 'good boy, brave.'
mfw my god throws the ball and I catch it.

feels good man. will protect him always.

it generated this when I gave it "dog"

[meager jackal]

disappointing lack of right wing dog dialogue /musi

[heady otter]

real

[atomic relic]

more realistic greentext:

arf arfff :3
arfff

[heady otter]

real

[turbid quiver]

be me
20, NEET
living in mom's basement, the usual
playing Undertale for the 37th time
at the Judgement Hall, Sans monologue just finished
he's talking about how everything's pointless, nothing matters, yadda yadda
start thinking, damn, he's right.
what's the point of anything if it all just resets?
my life's just a perpetually restarting genocide run of chores and job applications
mom yells down the stairs
'Anon, take out the trash!'
mfw I just sit there, controller in hand
too busy contemplating the futility of existence and the heat death of the universe
she comes down, sees me zoned out like a fucking autist
'Did you hear me, young man?'
look her dead in the eye
'what's the point, mom? it'll just come back. trash is just... trash. and soon, we'll all be trash.'
she slaps the controller out of my hand, it bounces off the wall
'It's called living in the real world, you lazy bum! Now get up!'
mfw my HP wasn't 1, but my spirit was.
guess I really had a bad time.
still didn't take out the trash though.

This is what happened when I put in just "sans"

And here's one with "toki pona" as the input (slight slur warning)

be me, 20s, STEM||fag||, bored out of my mind in current year

scrolling /g/ looking for a new obsession

see some anon talking about 'Toki Pona'

mfw 'language of good/simple talk'

only 120 words? 'This is it, Anon. The ultimate logical language. Pure thought. No more ambiguous corporate speak or flowery prose.'

spend like 3 hours reading about it, head already swelling with enlightened minimalism

get to the part where you learn the words

'pona' = good, simple, right

'ike' = bad, evil, complex

'jan' = person, people

'moku' = food, eat

'telo' = water, liquid

this is going to be easy

start trying to formulate a sentence

want to say 'The quick brown fox jumps over the lazy dog.'

brain immediately starts to melt

how do I say 'quick'? 'wawa' (strong)? 'jaki' (waste/dirty - close enough for a fox)?

how do I say 'brown'? 'loje' (red)? 'jelo' (yellow)? 'pimeja' (dark)? It's a 'soweli wawa pi suli loje jelo' (animal strong of big red yellow color) now?

jumps? 'tawa' (move)? 'awen' (stay - opposite)? 'anpa' (down)?

realize you have to describe everything in the most basic, abstract terms

my thought process for 'computer' becomes 'ilo sona pi pali mute' (tool of knowledge of much work)

my thought process for 'car' becomes 'ilo tawa suli' (big moving tool)

my thought process for 'I love you' becomes 'mi olin e sina' (I like you) or 'sina pona tawa mi' (you are good to me)

everything sounds like a caveman trying to explain advanced physics

try to explain the concept of capitalism to myself in Toki Pona

lipu mani pi jan ale li jo e jo pi pali mute li pali e jan lili pi olin ala. (paper money of people all have having of work much and make people small of love no.)

mfw my brain, once capable of calculus, is now struggling to differentiate between 'hand' and 'arm' (both 'luka')

tfw tried to achieve linguistic zen and unironically reduced my brain to monke mode

decide to quit before I forget how to speak English and start grunting at the grocery store clerk for 'telo walo anpa' (liquid white down - milk from the udder)

[nova thicket]

I tried this, it wasn’t nearly as in depth

Capitalism can only exist when you have paper money, obviously

[turbid quiver]

And yet, people want to go paperless and crypto

But anyway, is that Toki Pona actually accurate?

[nova thicket]

They’re the true commies here

lipu mani means monetary bendable-flat-thing, so I guess that works for paper money

It checks out translation wise, but I don’t think it checks out explanation wise

[turbid quiver]

Ahh, because LLMs usually flail and stumble at trying their hand at Toki Pona, both TO English and FROM English

It's a good way of poisoning AI for sure

Even with something as simple as sitelen pona, it usually gets even basic stuff written wrong

[turbid quiver]

Here's one about The Knight from Deltarune (slur and ableism warning)

be me
browsing /v/, another thread about who the Knight in Deltarune is

muh Gaster
it's obviously Gaster
Anon, if you still think Gaster is the Knight you're genuinely braindead.
He's the creepy voice in the walls, not some fountain-shitting janitor.
Kris is the Knight
mfw anon thinks the silent protagonist being the secret supervillain is peak writing.
Yeah, you cracked the code, Sherlock.
Noelle's the Knight
it's Dess
Get out, Noelle||fag||s. Your shipping ||autism|| isn't a plot device.
Berdly
Lmao, cope.
Queen
She literally explains she ISN'T the Knight. Do you even PLAY the game, ||retard||s?
tfw every theory is dumber than the last.
It's probably just some new character we haven't met yet.
Or it's a metaphor.
Or Toby Fox is just jerking off his fanbase.
Stop overthinking it, you ||schizo autist||s.
Go outside.
t. Deltarune||fag|| tired of your Knight ||autism||

Wow, they got the 4chan lingo down, right down to the slurs and ableism

[stable lantern]

proud noellefword

(it's december holiday 😈 )

[turbid quiver]

It most likely is, or probably Carol. Or even ||Sans||.

Jk on that last one, no way he's the Knight, too lazy for that

Tbh, I don't think that any LLM is aware of chapter 3 or 4 yet, given it's too new for AIs to scrape any info off of.

Hence no mention of Tenna or ||Gerson||

[glass iris]

I know this is a parody of 4chan shit posting, but let's ease up with reproducing slurs.

[turbid quiver]

Sorry. I didn't write the greentext though.

I know I wouldn't do it if it was ME writing it

[glass iris]

I know that. Not telling you off or anything. But for future ref, might be good to len that part even here.

[turbid quiver]

But yeah

You mean spoiler tags or outright asterisks?

[glass iris]

Spoiler tags would be fine.

[turbid quiver]

Ahh

Maybe I'll do that right now

Do I also spoiler tag ||autism|| too, even though technically it's not a slur, but in this case is used like an insult?

I'm autistic too, so...

[glass iris]

No, I think that's ok. Unless you feel more comfortable spoilering it. Just f-slur and r-slur

[turbid quiver]

Yeah, that's fair

I know some people get hurt by it. I have a history of being told off by it irl

But I'll just len the actual slurs

[glass iris]

Sorry, don't want to be a cop here. Ive seen lots of parodies of green text made by AI, and I get the genre, but if someone wanders in here and doesn't know the context, it could be a shock.

[turbid quiver]

That's a perfectly valid point. I'm glad you pointed it out

Yeah, I considered modifying it, but thought it would ruin the flow and the whole point that it was a 4chan greentext parody, plus wanted to preserve what AI says warts and all. But using spoiler tags without altering the actual text does seem like a fair compromise.

I don't support using slurs otherwise

(Though I have heard the VERY dumb argument that slurs should absolutely be used against the very people who insult you with them, thereby doing a reflect attack against bigots. But that's a topic for another time and place, not this thread)

[glass iris]

I appreciate it! The way you've edited the text now is perfect. Thanks for being so thoughtful and accommodating.

[quiet geyser]

why is NOBODY mentioning rudy 💔

[stable lantern]

either way i don't think either are the knight so they probably wouldn't have shown up

i think carol is definitely the more interesting ommission

[turbid quiver]

Yep, because while it was known since the first two chapters that Noelle has a mom, she was never shown onscreen or even named before Chapter 4. Dess was definitely named in the Spamton Sweepstakes ARG, though.

[high kayak]

Dess was actually named as far back as chapter 1

[turbid quiver]

Yeah, I think the whole point AI was trying to make was that these theories are dumb and least likely to come true

[stable lantern]

lmao

me when i water my house at dawn

[lucid niche]

Google hallucinated details about Sonja's publishing house name

[glass iris]

Wild.

[lucid niche]

I would concur, but I'm all too used to it by now

[atomic relic]

the name is the only correct bit

[lucid niche]

...and that was my search

mun Isa (@ ON in replies) ↩️

[Reply to:](#1162756469771472988 message) the name is the only correct bit

I searched "tawhid press", Google took "tawhid" and invented a Muslim French publishing house from the 1980s

there is no such house

[robust basalt]

Gemini claims to be GPT. English translation:

*No, I'm not Gemini.

I'm a GPT (Generative Pre-trained Transformer) model developed by OpenAI.

Gemini is another major language model developed by Google. While we share similar capabilities, we were developed by different companies, using different datasets and architectures. Therefore, our response styles, how we access information, and sometimes even what is perceived as our "personality" may differ.*

[raw crypt]

i mean
is it not a GPT *
it's not developed by openai but

wait
whats the diff between an LLM and a GPT

[silver ember]

a GPT is a type of LLM apparently

[glass iris]

But all the big models used for generation of text (ChatGPT, Gemini, Grok, etc etc) are transformer models.

[robust basalt]

Yes but it is strange that it claims not to be Gemini and trained by OpenAI. Kind of hallucination I guess? Or they used ChatGPT's outputs to train it just like DeepSeek seems to have done.

[twilit stag]

be me, 20s, NEET, home alone
late night, like 3 AM
hear weird scratching noises from kitchen
think it's a mouse, grab the heaviest frying pan I own
creep towards the kitchen, heart pounding
flick on the light
not a mouse, not a rat, something worse
absolute unit of a frog, size of my fist
perched on the edge of the sink, staring directly at me
it's not scared
it looks...judgemental
the air thickens
I drop the pan
EMERGENCY FROG SITUATION.

With the topic "emergency frog situation"

[severe dagger]

woah primitivism AND the f-slur??? 2 for 1 combo

be me
browsing /lit/ or something, see thread about 'conlangs'
some f-word autist shilling 'toki pona'
mfw it's only like 120 words
this sounds easy. I'll learn it in a day, become multilingual genius.
spend an afternoon memorizing kina, jan, tomo, telo, suno, etc.
think I'm hot shit. ready to converse.
try to explain something complex, like 'the current geopolitical situation is precarious due to rising tensions and economic instability.'
tfw it comes out as 'ma mute li pilin wawa toki. mani li ike. sitelen pi telo seli li ike. jan li pakala.'
sounds like a caveman having a stroke
realize the entire point of the language is to force you to simplify everything to its base emotions
it's not for talking, it's for feeling
tfw your worldview is now just 'pona' and 'ike'
mfw I only speak in grunts and 'pona'
brain thoroughly washed. mi jan ike.
seethe in toki pona: 'pilin ike.'

[twilit stag]

be me
native Proto-Indo-European speaker
wake up, sun is h₂owsos-méh₂ beautiful today
try to tell neighbor about the glory of Dyeus Piter
explain how my póds is hurting from walking too much on the dhghem (earth)
mention the need for a good ekwos (horse)
he just stares, then asks if I'm having a stroke
mfw I'm the last one speaking real language
everyone else just grunts in their devolved, non-ablauting noise
gonna go sacrifice a gʷow-s (cow) to the sacred fire alone.
cope.jpg

[atomic relic]

put "native PIE speaker" back in and got this

be me
native Proto-Indo-European speaker
literally speak the source code of your language, zoomer
walk into Starbucks
try to order coffee, use proto-roots, reconstruct the whole damn thing from *peh₃-id- ('to drink') and *kauh₂os ('cow')
barista just stares at me like I'm having a seizure
I try explaining the beauty of the laryngeal theory, how *h₂eg- ('to drive, lead') is the mother of agent and agile
how perfect ablaut is, how elegant the complex case system
they call mental health services
mfw I'm literally the most linguistically advanced human on the planet and I can't even get a damn latte
tfw no one understands the glory of the Dyēus Ph₂tḗr or the subtle distinctions of the medio-passive voice
just endless noise and degenerate consonant clusters
no steppe to roam, no horses to ride, just concrete and confused normies
why are my people gone?
I just want to explain the Rig Veda to someone who gets it

from "native Proto-Indo-European speaker buying a car with a horse in the trunk"

be me, at the Ford dealership
just taking a look at new trucks, bored on a Tuesday afternoon

see this absolute unit of a man talking to a salesman
not just big, but like, ancient big. Massive beard, wearing what looks like a mix of hide and woven wool. Carrying a gnarled walking stick.
dude looks like he just stepped out of a National Geographic special on early humans

salesman is clearly struggling, sweating
PIE guy doesn't speak English, just points at a Kuga, makes guttural sounds and hand gestures
sounds like he's trying to pronounce 'engine' but it comes out as 'h₁reh₃-dʰreh₂us-ih₂' or something equally incomprehensible

salesman thinks he wants to see the trunk
probably trying to show off cargo space for 'luggage' or 'groceries'

salesman hits the button, trunk pops open
mfw a whole ass horse is folded into the back
just its head sticking out, blinking slowly, looking utterly dejected and squished
legs crammed in there like a sad equine origami project

PIE guy grunts louder, points at the horse, then at the engine compartment, then makes a 'pulling' motion with his arms
says something like 'éḱwos! weǵʰeti!' (horse! carries/pulls!)
he wants the car to be pulled by the horse, or for the horse to be the engine

salesman's face goes from confused polite smile to pure existential dread
slowly backs away, pulls out his phone, looking for supervisor or maybe calling animal control
horse whinnies softly, like it's apologizing for the inconvenience

PIE guy just standing there, arms crossed, clearly frustrated at the lack of understanding for his 'modern' transportation needs
he wanted a car with a bigger trunk for more horses
or maybe just a horse that fits better

I'm just standing there, jaw on the floor
some days you just witness peak human absurdity
kek, this is why I love dealerships

"i am a native Proto-Indo-European speaker buying a car with a horse in the trunk"

Be me. Native Proto-Indo-European speaker.
Finally decide to upgrade from chariot. My old one, good honest wood, but starting to creak.
Hear about these 'cars'. Metal beasts that move without horses. Sounds like witchcraft, but curious.
Walk into 'dealership'. Bright lights. Strange smells. Men in robes of fine fabric, speak with forked tongues.
Man approaches. Grinning like jackal. He ask 'what bring you here, sir?'
I grunt. Point to shiny metal beast. 'This make loud noise. Go fast?'
He babble about 'horsepower' and 'turbo' and 'financing'. I know many horses. This beast no horse.
He finally get to the 'trade-in' part. 'Do you have anything you'd like to trade today, sir?'
I nod. Lead him to my 'trunk'. It's a small wagon I attached to the back of the lot's demo model.
Open back. My best mare, Brunhilde, pokes her head out. Looks confused, but well-fed.
Salesman face go pale. Like winter snow.
He point. 'SIR! IS THAT A LIVE HORSE IN YOUR TRUNK?!'
I nod. 'Yes. For trade. You say 'horsepower', yes? I bring power of horse. Live power. Strongest power.'
He stammer. 'We don't... we don't accept livestock as trade-ins. Especially not in the trunk of a new vehicle!'
I frown. 'But this is good horse. Strong. Fed well. Knows many commands. Pulls well. Good for... 'horsepower', yes?'
He start making strange squeaking noises. Tries to call 'manager' (a different type of chief?).
Eventually, after much grunting and pointing and him trying to explain 'liability' (another strange spirit?), I just throw down some heavy gold ingots. Plenty of gold from raiding Thracian villages last moon.
He stop yelling. Eyes widen. Take gold. Give me 'keys'.
Drive away in new 'car'. Much faster than chariot.
Brunhilde whinnies happily in back. Good girl.
Still prefer smell of honest horse sweat to this 'exhaust'.
Think I'll get some fresh meat for dinner. Maybe from that gas station attendant. He looked plump.

"i am a native Proto-Indo-European speaker buying a car with a horse in the trunk. greentext is in english, but dialog in pie. translations provided in brackets"

be me
native speaker of Proto-Indo-European
tfw your ancestral lands are now a highway bypass and a 'pizza hut'
decide I need one of these 'motor-wagons' that move so fast
drive my actual wagon, pulled by an ećwos [horse], to the 'car-steading'
smell of petroleum and desperation hangs heavy in the air
some smooth-furred male approaches, teeth showing in a strange greeting ritual
he says something in the 'Common Tongue'
I reciprocate
'H₁ey! Tód h₁eǵʰeti!' [Hail! That goes!] I point at a shiny red 'motor-wagon'.
he blinks, says 'Uh, can I help you, sir?'
'Widh₁éntm me! Anih₂ dʰeh₁sóm h₁eǵʰeti!' [I see it! This thing goes!]
he nods slowly. 'So you're interested in purchasing the new Super-Ford Lightning 3000?'
'Peh₂yéyeti!' [It will pay!] I declare, reaching into my pouch.
he eyes my pouch. 'Cash? Card? We have financing options too.'
'H₁ećwom. Meǵʰa h₁ećwom.' [A horse. A great horse.]
his smile falters. 'A... horse? Sir, we don't accept livestock for payment.'
'Né peh₂yéh₂ti? H₁ećwos! Wéyk!' [It will not pay? A horse! Strength!] I exclaim, genuinely confused. What else is there of value?
he's scratching his headfur now, looking around nervously
I lead him to my wagon outside
he looks skeptically at my simple wooden cart with its tired ećwos in the traces
'Sir, I really don't think we can-'
I walk around to the back of the wagon
reach into the 'trunk' area where I had prepared for this transaction
pull out the second ećwos
he's a fine young stallion, well-bred for strength and speed. Just a little confused about why he's in a dark box.
the salesman's eyes go wide like full moons
he stares at the horse, then at the open trunk, then back at me
'Maí! Íd h₁ećwos méǵʰeh₃s!' [But! This horse is great!] I insist, gesturing to the magnificent beast.
he opens his mouth, no words come out, just a strangled gurgle
then he screams in Common
security arrives quickly, carrying thunder-sticks and looking very agitated
I just wanted a fast wagon, and I offered good value.
tfw nobody understands the true value of a well-bred equus caballus anymore

[rocky stirrup]

prompt: kijetesantakalu

be me
bored at 3 AM
stumbling around obscure language wikis
find 'Toki Pona'
sounds kinda cool, simplified language, only 120 words and shit
start looking at the word list
everything's chill, 'pona' is good, 'jan' is person, 'suno' is sun
then I hit 'kijetesantakalu'
what the fuck is that
scroll down
it means 'cat'
MY FUCKING BRAIN just melted out my ears
eight syllables for 'cat'?
I'm never gonna look at my cat the same
it's not a cat, it's a goddamn eldritch horror masquerading as a 'kijetesantakalu'
gonna start referring to it exclusively as 'my kijetesantakalu'
mom's gonna send me to a mental institution

kijetesantakalu - cat

and it doesn't even have eight syllables 😭

[atomic relic]

i tried that again and got something similar

>be me
>bored as fuck, browsing random wikis at 3 AM
>stumble across 'Toki Pona'
>looks kinda interesting, language of simplicity and all that
>start scrolling through common words just for kicks
>see 'kijetesantakalu'
>mfw that looks like some Eldritch horror's true name
>like something you chant to open a portal to another dimension
>i can practically hear the cultists whispering it in a dark forest
>decide to look up the meaning
>it means 'cat'
>mfw they use a 15-letter word to say 'cat'
>my brain just melted
>the absolute state of linguistic efficiency

[stable lantern]

why does it think it means cat 😭

[nova thicket]

Because cats are basically just raccoons, obvie

[high kayak]

it doesn't think, it's an ai

there's no logic lol

[twilit stag]

Prompt: "Wake up in the US speaking only Scottish Gaelic"

be me
wake up in what looks like my room
sunlight streaming through the blinds, same old shit.
go to say 'fuck, morning already'
only 'madainn mhath' comes out.
wtf.jpg
try again. 'hello?'
'halò?'
full blown panic attack in a language only actual people from the highlands speak.
try to read the clock. digital. says 9:37.
try to read the book on my nightstand. all the words blur.
realize I can't read English anymore.
stumble out of apartment, nearly trip on the mat.
see old lady walking her poodle down the hall.
try to ask her for help. 'A bhean uasail, chan eil Beurla agam! Chan eil mi a' tuigsinn!' (Ma'am, I don't have English! I don't understand!)
she clutches her dog, eyes wide. Mumbles something about 'crazy hobo' and scurries off.
realize I'm in America. Fucking America.
go to the street. Yellow cabs zoom by. People talking in... gibberish.
it's not even gibberish, it's just... noisy. I can't parse it.
everyone looks at me like I'm a lunatic when I try to speak.
my wallet is in my pocket. All the money is USD. The cards are English.
can't even read the ATM instructions.
hungry. Want coffee. Can I even point at a menu?
go into a Starbucks, try to order a simple 'coffee.'
'cofaidh. dìreach cofaidh dubh.' (coffee. just black coffee.)
barista just stares, then laughs. 'What are you even saying, dude?'
want to cry, but only 'tha mi ag iarraidh caoineadh!' (I want to cry!) comes out.
no one understands.
I understand no one.
cops are probably gonna get called on me for being a 'disturbed foreign national.'
anon is stuck in a land of gibberish, can't even ask for the fucking toilet.
càit a bheil an taigh beag?! (where is the toilet?!)

[heady otter]

this is hilarious

it makes sense because Gemini was trained off ChatGPT, and i imagine they hadn't done as much retraining (or whatever the actual term is) in languages other then English

[heady otter]

i spot a matt rose watcher

[heady otter]

"non-ablauting noise" LMAO

[heady otter]

cw: f-slur
i asked it asw

be me
tfw born before writing
tfw my tribe settled somewhere near the Black Sea, kinda chilly innit
tfw your entire life consists of herding cattle, making carts, and arguing over which river to cross
tfw the greatest technological innovation in your lifetime is the wheeled vehicle

fast forward several millennia
wake up in some lab, apparently 'reconstructed' from ancient DNA
they put me in a room with a bunch of 'linguists'
they start talking in what they call 'English.' sounds like gibberish

then one of them says 'mother'
my ears perk up
it's not exactly *matér, but it's close enough that I recognize it
another says 'night.' sounds like *nokʷt-
then 'star.' sounds like *h₂ster-
they talk about 'horses' and 'wheels.' *h₁éḱwos and *kʷekʷlos

my eyes dart around the room
these ||fаggots|| are speaking a bastardized version of MY language
they think they're so advanced with their glowing screens and their 'internet'
I could teach their 'philosophers' a thing or two about real existential dread, like trying to milk a wild urus in a thunderstorm

they spend their lives arguing about etymology in their high-falutin 'universities'
I actually lived the etymology
I was there when *h₁éd- was a universal verb for 'to eat' and not just some meme
I invented the concept of *dʰéǵʰōm as 'earth'

tfw literally half the planet speaks a slightly corrupted dialect of your everyday conversation
tfw every European, every Iranian, every Indian is unknowingly quoting your ancestors
tfw you're the literal linguistic Chad
mfw I invented language. well, a big chunk of it.
mfw *wóda is 'water' in every fucking language these fools speak.

[heady otter]

IT MADE THE SAME ERROR WHEN I ASKED IT SOMETHING KIJETESANTAKALU-ADJACENT

i didnt copy it down and i don't remember most of it but i remember that

[atomic relic]

says innit
calls english gibberish

[high kayak]

Didnt expect this to become a shitposting channel

[nova thicket]

I mean, maybe that’s the natural progression of Greentext-AI

[twilit stag]

Prompt: "tower of babel"

be me
simple Shinarite, farmer living comfy life
everyone spoke same tongue. good times. no misunderstandings.
one big comfy group, chilling, making bricks.

then some architect-chad pipe up
says we should build a tower. BIG tower.
reach the heavens. show god our gains.
get famous. make a name for ourselves.

sounded dope. we all agreed.
start hauling bricks. mixing mortar.
years pass. tower gets HUGE.
top disappearing in the clouds.
feels good man. entire world united under one goal.

one morning, up on the scaffolding.
ask my bro for more clay.
'Hey, m-m-more clay, please!'
he just stares blankly at me.
then he starts speaking total gibberish. sounded like a dying cat trying to speak Aramaic.

i'm like 'what the FUCK, Steve?'
he starts yelling in his weird language. pointing.
next thing i know entire construction site is a cacophony of foreign sounds.
bros start arguing. people falling off ladders cause no one understood 'LOOK OUT BELOW!'
chaos. total pandemonium.

sky goes dark for a bit. probably nothing.
can't work. can't communicate.
project abandoned.
everyone just... leaves. wanders off in different directions.
start calling each other 'foreigner'

tfw no universal language bros
tfw now i gotta learn all these goddamn dialects just to buy bread
tfw we peaked too early. should've just stayed farming.

[stable lantern]

it's because everything ai posts is shit :3

[high kayak]

Yea, but is this the place? Read the title of the thread

[heady otter]

real

[jaunty trellis]

well, it's somewhat correct

mi
wile moku
tawa tomo moku

lukin e kili loje mute
pilin pona

jo e kili wan
kili li anpa
kili li tawa ma ike

pilin ike a
moku e kili lon ma a

[midnight sapphire]

Probably the worst butchering of toki pona I've seen an LLM do to date

[stable lantern]

o tawa taso ala. o jo e ni

:3

[midnight sapphire]

I already told it that, but it said that wasn't correct

lmao

[stable lantern]

lol

[glass iris]

I agree. Let's try and keep on topic.

[atomic relic]

ni a

[atomic relic]

Hello. How can I help you? 😊

(every ai ever does this its so annoying) (i weote this myself)

[nova thicket]

Usually there’s no emoji

[robust basalt]

they're just trying to be helpful 🥀😞 (but i also hate it)

[coral copper]

ChatGPT's thoughts on AI:

(first question)

(second question)

sorry about the image wall

[heady otter]

why was my first thought after reading this "tasty" I have no idea what I was thinking or what I meant

[ancient thicket]

"You're tapping into something really important here. Let's unpack that." /musi

[coral copper]

jan Sujute what's the last glyph in your name in your username?

[meager jackal]

it's epiku, but the old glyph

[coral copper]

oh

[twilit stag]

so silly

[ancient thicket]

epiku pini li 'up ni', a a a!

[heady otter]

i probably messed up the grammar a lot

I actually have no idea why I send this the translation looks perfectly sound

other then the male-by-default

[meager jackal]

vsauce moment

[heady otter]

selam! Vsauce, men Mikal
men yaxshi adem. yaki, men?

[meager jackal]

men = i am?

[heady otter]

men = I (nom.)
uyghur has a null copula in the non-past tense

[meager jackal]

cool

[jaunty trellis]

tomoato

[quiet geyser]

melon jelo....

[coral copper]

im scared
my job is gonna be stolen by someone who was basically born yesterday

[robust basalt]

fr, but human creation is still way more better. These videos are repetitive, the design, the talk style etc. It cant be original 🙏 so no need to worry, human is still superior

[coral copper]

yes but boss go "ooh cheap labor yes"

[robust basalt]

Hope they wont choose something like that.. its just soulless 💀😞🙏

[coral copper]

i hope i can be optimist but mute mute people are greedy mute

[robust basalt]

ni li lon :') taso, o awen pilin pona! (stay optimist)

[ancient thicket]

ike lon ma mute a la unfortunately bosses in a lot of industries are trying to replace creatives and more across the board if they can vaguely get away with it.

The output is a lot worse than a trained and experienced human, but it doesn't matter to executives who don't intend to consume or double check the output in the first place and just see more number out from less money in.

I have hope the slop will be seen as slop rapidly though in wider industries...

[meager jackal]

tomoato

a housed ato

[stable lantern]

which is it brah

[glass iris]

What would you like it to be

[stable lantern]

gpt 5 update completely nuked one of my chat threads

only one of them

[glass iris]

Did you see that the roll-out has been so bad that they're bringing back 4o for paying users?

[midnight sapphire]

t3.chat has all of the models still available via API

Not just ChatGPT but Gemini, Claude, Llama, Grok, Qwen, Kimi, GLM and Deepseek too

Some of those I had no idea existed but it's been very nice to have access to all of those models, especially considering the relatively low price (not sponsored btw)

If you're gonna pay for an AI subscription, I recommend you take this one

[glass iris]

I really need to use gpt-3 davinci for some historical work. I think it's still available through azure, but I'll need an idiot's guide how to use it.

[heady otter]

absurd
shouldn't it be an option already to use previous models??? I thought they had been doing that since forever

I feel like the AI interest curve is dropping

[timber moat]

GPT-5 can speak toki pona better than GPT-4, but it still makes lots of small mistakes.

[stable lantern]

iirc gpt-5's selling point is that it can switch between the models automatically depending on which is best for your task

i don't know if that's true

[heady otter]

yeah

[midnight sapphire]

Still uses en as and

[glass iris]

And I guess that is part of the problem. Who can know? OpenAI can claim they'll switch in the most powerful model for your task - but the suspicion is that they are actually economizing behind the scenes, because the range of models available until yesterday were ruinously expensive.

People feel they paid for a range of tools that they could choose to do particular jobs; and then they find them all taken away, and are told to trust the company.

And that may be paranoia. Maybe OpenAI really are doing what they said they're doing. But it's enough of a PR mess that they're rolling back a legacy model to everyone. Though not o3.

[meager blaze]

ilo sin li ken toki pona. pilin mi la mi ken kepeken ilo ni lon toki pona taso

taso pakala lili li lon nasin toki

[meager blaze]

😭 ???

mi alasa kepeken li pana ala mani la ilo li toki e ni: "a sina ken ala kepeken ilo sin. tenpo suno kama la sina kama ken"

[glass iris]

sina pana e mani la, tenpo ni la sina ken awen kepeken ilo 4o. sina pana ala e mani li toki mute tawa ilo la, sina o kepeken ilo pi wawa lili (ken la ni li GPT-5-nano)

[severe dagger]

mi pakala

mi kepeken ilo la ilo li pali e sitelen mi

taso sitelen ni li ike

气和

ona li ante

[robust basalt]

Toki Pona song in ElevenMusic (i cant share it as a file because it tells me to pay)

Toki Pona Pilin - created with Eleven Music https://share.google/9hr2WxiEISNkXVTDZ

[stable lantern]

my official gpt 5 review: still stupid

[robust basalt]

just a little less stupid 🤏

[stable lantern]

nah i disagree it's arguably more stupid

[robust basalt]

ken a

[midnight sapphire]

It is great at getting code right the first time around, but it sacrificed everything else for it

And since vibe coding is a security risk for anything besides small personal projects... not great

[stable lantern]

my chat memory is full and without it its actual memory is like. one message

[midnight sapphire]

I turned off memory, in my experience it makes the model dumber having to remember all of that irrelevant data

[heady otter]

yeah

[heady otter]

yes definitely
and it also makes benchmarking harder

[heady otter]

claude is moderately good at poczonkaxles

[stable lantern]

i haven't used claude yet

i've used chatgpt and deepseek tried using google gemini but it was genuinely awful

i taught selsimicu to grok for like 2 days just to see if it was that bad and it was about on par with gpt pre-5 but i just didn't wanna keep using Grok from X The Everything App

deepseek was the best one i think (though it's been so long since i did the chatgpt one i don't even know anymore)

[atomic relic]

deepseek isnt good at tp anymore
-# Please use reply pings.

[timber moat]

Depends on what you are doing. I tried it out with both Zig, Rust, C, Python, TypeScript, HTML, and Z80 assembly (I know quite a few languages), and while it can produce working code reliably, it fails as soon as it needs to work with or generate more than 200 LOC. The local conception is decent, but architecture and context is abysmal. In fact, Gemeni 2.5 Pro is still better at programming.

[atomic relic]

out of curiosity was the code synchronous or async? ai in my experience sucks at async
-# Please use reply pings.

[timber moat]

it was mostly synchronous, though I did some async code in TypeScript, and some concurrency in C and Rust

[timber moat]

-# woops forgot to reply ping

[atomic relic]

nnn
-# Please use reply pings.

[atomic relic]

:3
-# Please use reply pings.

[heady otter]

it is, you just have to prompt it correctly
I suspect they found out people were using toki pona to evade the blocking

[atomic relic]

n
-# Please use reply pings.

[stable lantern]

i was using it to teach my conlang and see if how well it could do and it was doing very good until it forgot everything and i got mad (which happens to all of them womp womp)

i think deepseek did end up lasting the longest in just Knowing

[heady otter]

claude 4 sonnet was shockingly good at poczonkaxles despite the small sample size
id test with opus but I'm not about to pay for that
id definitely reccomend messing with claude for conlangs tho

[stable lantern]

i'll check it out

[atomic relic]

i have opus access taso its linked to my google account
-# Please use reply pings.

[slate ridge]

I have found GLM-4.5 to be pretty good, but it makes occasional silly mistakes. (Haven’t tried Kimi-2)

[robust basalt]

mi jan pi toki e sona.

[slate ridge]

Not the single pi

It's not thw worst though...

[meager blaze]

SEMEEEE AAA

[stable lantern]

nasa mute

[stable lantern]

Ok

brain is almost entirely physical except when it's not

[jaunty jungle]

"e" is almost entirely used as a content word unless as a particle

[nova thicket]

Copilot doesn't understand what math means, but everything else was surprisingly not completely wrong

[coral copper]

this is painful

[nova thicket]

I think 25 actually means 55 😭

[coral copper]

"Uhm so you see it says twenty six, BUT it means 55 to context unknown to everyone" 🤓 (it means 11)

[nova thicket]

8=62, right

It cant owe

[coral copper]

btw you said yyou had a strawpage but the link didnt work

[nova thicket]

I changed the name lol

https://kijetesantakalulaso.straw.page

That's an excerpt from the page

I actually just redid the layout of the page a lil'

[stable lantern]

claude review: actually pretty okay

i've only been using it for a few hours so the AI dementia hasn't hit in yet i guess. but very good so far

[nova thicket]

That's good

[glass iris]

Claude has always been the standout in toki pona

[midnight sapphire]

Haven't tried Claude but some LLMs start understanding nasin nanpa pona after you include something like this in your first prompt

"In the number system I'll be using, there are five/six (depending on if you use owe) numbers. To find the number, simply add them together, left to right.
wan/W: one
tu/T: two
luka/L: five
mute/M: twenty
(owe/O: 1984)
ale/ali/A is a special case. If you encounter it, you multiply by 100, then continue."

Oh, apparently it can go up to 8

[nova thicket]

That wasn’t even nasin nanpa pona, it was the basic numeral system (including owe), so it should’ve had a much easier time

[stable lantern]

i didn't even notice the owe 😭

[jaunty trellis]

that's not an entirely accurate description, "ale ale" is 200, not 0 or 10,000 which your description would imply

[midnight sapphire]

In nasin nanpa pona, something like ale ale is only really valid for reverse compatibility with nasin nanpa. tu ale is used for 200 and wan ale ale is 10000

[midnight sapphire]

Yeah addition is hard for LLMs when each number is aliased in a way it can only infer from a small amount of training data

[nova thicket]

Literally owe

[glass iris]

OMG nasin ale majuna pi ilo TokiKPT li kama sin!!!

nasin 4.5 li weka.

taso ni li nasa ala. ona o weka tan mani

taso nasin 4.1 li kama sin a! nasin o3 kin!!!!

[quiet geyser]

research grade intelligence yeah dawg

[glass iris]

I wish they'd bring the full model selector into the Android app. It just has 5 and 5 thinking still. EDIT: I just had to delete and reinstall the app and all the models were there.

[heady otter]

conpidgin (Poczonkaxles) spoilers ahead!

||Attached is a dictionary of Poczonkaxles in TSV format and a dump of all mainspace articles on the Poczonkaxles Wikipedia edition. Poczonkaxles has SOV word order. Multi-word subjects are separated from the rest of the sentence with <,>. Clauses are formed with "kuvo"; the subject of a clause is assumed to be the thing it's modifying unless it's specified. If and when statements are formed with "ka A, B" which means "if/when A, then B." Adjectives go after nouns except in compound words where the order is free. Vowels in hiatus are separated with the diaeresis diacritic. Everything is pronounced identical to the IPA, except for <sz>, <cz>, <dz>, and <eu> (which are /ʃ, t͡ʃ, d͡ʒ/ and /y/ respectively). Please note these are my orthographic conventions, and some articles may use slightly different conventions. If I forgot anything in this description, I will tell you when it comes up. Please use the wiki dump as reference for how Poczonkaxles is used. Could you speak only in Poczonkaxles from now on?||
this is the prompt I used to get Claude to speak poczonkaxles

I only called it "wikipedia" to not confuse the model, the wiki is completely seperate

[heady otter]

oh I should spoiler

[stable lantern]

wawa

[silver ember]

ohhhh you gave it a dictionary that explains it

[stable lantern]

new claude review: still pretty good but holy shit the conversation length limit is atrociously short

i've burnt through two conversations just trying to get it to fully translate one text with sufficient context on my part

worst part is i'm underselling it a lot it's actually extremely good

but i keep having to roll it back and reteach it EVERYTHING because the conversation gets stopped

[lucid niche]

you may wanna use the api w/ something like openwebui so you can cherrypick parts of the conversation to trim

meli Siliwa 💃🏽 ⪩(ᐢᗜᐢ)⪨ ↩️

[Reply to:](#1162756469771472988 message) but i keep having to roll it back and reteach it EVERYTHING because the conversation gets stopped

[stable lantern]

okay this seems to be a purely website thing because i'm currently typing on the thread that is supposedly "completely full" on my phone and tablet #alright

weirdly it's only mobile that has a timed message limit (i.e. "you can send like 12~ messages until this time and before then u gotta wait") so i guess they're doing sorta the same thing just in stupidly different and severely unbalanced ways

[glass iris]

Claude is weirdly good. I'm sure it's to do with that linguistic structuring that the Anthropic researchers discovered when they did a deep dive into its "brain"

[heady otter]

wait could you elaborate

[stable lantern]

okay so the phone version has this and the website version will just close a thread when it's too big but the version on my tablet does. neither. so i can just keep talking to it lmao

newer claude review: really good actually

exclusively on my tablet

do not open anywhere else+

[heady otter]

https://www.anthropic.com/research/tracing-thoughts-language-model
oh I think I found it

[glass iris]

Here's a blog-post summary, linking to the two research papers, which are really difficult, but fascinating: https://www.anthropic.com/research/tracing-thoughts-language-model

Essentially, researchers developed a tool where they could "trace the thoughts" through Claude's network, given certain prompts. Like neural imaging.

They found, unexpectedly, that Claude has developed (by itself, during training) a kind of universal grammar structure, and it translates between languages by passing it through this general structure

[glass iris]

ni a

[heady otter]

sona a

[glass iris]

We don't know if ChatGPT has this, or any other LLM. But Claude is so bizarrely good at small languages, that maybe this is something it happened to build, and other models didn't?

[heady otter]

[severe dagger]

dang they really named this channel after the 42nd earth king

[jaunty trellis]

conflicting signs

[midnight sapphire]

How are they conflicting?

[ancient thicket]

tan ni ⤴️

[midnight sapphire]

You dropped a t and both of them correct you by adding the t back

The AI uses committed instead of committing but it's the same idea

[ancient thicket]

committing =/= committed. One is an adjective and the other is a verb. Communicate the same idea at the core root, but grammatically serve different purposes.

The conflict is it (google) giving two totally different corrections in this case. One right (from whatever system recognizes the mistake in base Google), the other wrong (from the AI overview).

It's a great example of why the AI search results are almost universally sloppy garbage, lol.

[twilit stag]

eith munth

[severe dagger]

wait what is this

[jaunty trellis]

that's correct though right? eighth month?

[twilit stag]

Kagi translate

[severe dagger]

dang this thing is smart

nevermind 😭

want me to fight the chicken > wile ala mi uta e akesi > why don't i talk to the reptile > toki mi tawa akesi li seme > what did i say to the reptile > mi toki e seme tawa akesi > what do i say to the reptile > toki seme li pona tawa akesi > what language is good for reptiles

this thing is DOUBLE wrong

[glass iris]

"mi lukin e sitelen ijo pi soweli" lol

[severe dagger]

AAA HELP

[midnight sapphire]

/id͡ʒo/.

[midnight sapphire]

My translator went from "I'm curious how accurate this translator is." to "At my knowledgeable, desired time, people can change? I'm not sure." in just a few iterations

[nova thicket]

The pronunciation

[swift kayak]

That one meme with a drowning person with their hand stuck out the water getting a high five comes to mind

I cannot see how people seriously use these for anything other than simple search or just fun

[ancient thicket]

Some models are notably better, but yeaaaaaah.

[quiet geyser]

ona li ku e seme

[meager blaze]

tenpo la ona li sona e ijo li ante pona e toki. tenpo ante la ona li sona ala e ijo li toki pakala

[meager blaze]

ni li sama kulupu pi toki pona anu seme

[swift kayak]

mi sona ala e toki sina

[severe dagger]

holy moly

[glass iris]

Just in case anyone is interested, here is the slides and paper I gave on GPT spiritual psychosis. I had to cut a 2-hour paper down to 1 hour, so left a lot of cool stuff out! But a couple of journal articles will be coming out of it.

[glass iris]

ni pi ilo Semini li musi a! https://medium.com/@mdasilva2266/gemeni-struggle-0ddcb4f7a2c6

[stable lantern]

new claude review nanpa san: still very good. cannot do math. par for the course really i don't know what i expected

i threw some senary at it and it flopped down and died and then i told it to do it in decimal and it still flopped down and died because this is the most notorious Cannot Do Arithmetic machine to ever exist

[nova thicket]

For a computer, that’s sad

[high kayak]

LLMs aren't logic systems, they're language systems
it just spouts whatever sounds right, it has no way of fact checking itself or its logic

[twilit stag]

thank you gboard

[quiet geyser]

wait since when did gboard have an ai feature

[timber moat]

ale li kama ike

[glass iris]

Let's keep on topic for this channel: textual LLMs and toki pona

[severe dagger]

idea:

feed images in an ai in training

tell it to output text

feed the text in a second ai in training

(neither ai speaks english, or any language for that matter)

tell it to generate an image

reward both when images are most similar

watch it create a new language

[nova thicket]

Good idea

[unique gulch]

Oh very interesting idea
How would you train an image generation AI without any input language though

I suppose you could have the system just input/output binary data to start with and reward it for decreasing the length of inputs/outputs while mostly preserving the image data

[slate ridge]

It would just create a reward loop

That plan doesn’t work alone

[slate ridge]

More like they just start paying themselves

[midnight sapphire]

CW: nimi unpa

[rocky stirrup]

cw: unpa
||unpa||

[twilit stag]

If you make kagi translate translate into "shrieking" it breaks the translator by spamming a ridiculous amount of H

[tame fog]

🙀

[sweet sinew]

LMAOO

[twilit stag]

How did the mic teleport from the guys hand to the womans

[ancient thicket]

GenAI is hallucinating frame by frame, so it "forgot" that it was supposed to be in the guy's hand. Coherent continuity is something it very, very much still struggles with.

[twilit stag]

Yeah cause it doesn't have like memory or anything right

[ancient thicket]

Even though it can do kind of short clips pretty semi-legibly now, compared to old AI video horror morphing.

[ancient thicket]

Basically, yeah. It "knows" the prompt, but it doesn't really "remember" the previous frame or know where the next is going. That's why things feel so strange in AI video sometimes, or abrupt.

[twilit stag]

I remember watching some kind of super old ai video which was really weird, things were morphing into other things like crazy, things were often blurry and indiscernible, and often things wouldn't make sense together

[ancient thicket]

Yep, exactly! The videos have come a long way since that (as we see above), but it's still doing that same dream process, just way smarter than before.

[twilit stag]

I don't know if i can find the video, i found it on a discord server i have since left

[ancient thicket]

If you pay attention to teeth and textures, or surfaces in general, you'll see the video AI fail pretty badly in a lot of cases. Or street interview style videos where people and cars appear/disappear behind people in the foreground, lol.

Teeth are often random dreams each time they are shown (magical changing teeth!), or essentially solid white bars.

Things that go behind objects don't return, and things appear from behind objects mysteriouely, etc.

[stable lantern]

me when i'm uhhhhh the guy from the mcu with the ai poster

[glass iris]

AI summaries are so dumb. Did anyone see this: https://www.instagram.com/p/DOAssqxjeKA

bennjordan

This is SO messed up. 🫥

I had a few messages and tags today asking me to clarify my stance on Israel, which was odd as I've been pretty outspoken against genocide and in full support of Palestinian statehood.

Turns out Google made up an elaborate story about me uploading a controversial video about "my trip to Israel" where I learned how wr...

Likes

8826

Instagram

[twilit stag]

https://youtu.be/Itbc12qXr30 this isn't the video but it's similar

[glass iris]

Let's not post AI videos here.

[stable lantern]

i feel like this might just be confirmation bias because i don't and probably won't ever use ai for writing but. man it's kinda exceptional how bad ai is at writing

anytime any of my uses of ai stumble on it trying to write dialogue or whatever for me it fumbles incredibly hard

the few times i have actually used ai for writing (usually assignments i genuinely don't have the patience to handle) it's so bad that i end up just rewriting it most of it myself

[nova thicket]

That’s probably a good thing

[stable lantern]

oh yeah it is

[glass iris]

Today: I was appointed to be the university officer in charge of developing a curriculum of critical , humanistic AI education, and trying to move faculty beyond "AI is eating homework."

And got into a shouting match on an faculty listserv with a professor demanding more AI detectors: it's not eroding trust to surveil our students, any more than a security camera at a gas station is eroding trust with customers. To which I replied, "my relationship with my students is different from my relationship with my gas station."

[stable lantern]

but it's just kinda funny in my eyes when i see people just kinda rawdogging it with no changes because it's sooo bad

like whatever the new winx club reboot is

[glass iris]

I mean, I hate AI in many many ways. And as meli Siliwa says, it can't write, and erodes student minds. But we're supposed to be educators not cops; so why not fucking educate our way out of this?

[stable lantern]

also i've noticed chatgpt 5 just like. removed the re-generate button ????? it's literally just not there anymore lmao

[nova thicket]

Great retort, 5 stars

[glass iris]

Probably not good for my first day on the job as "AI czar."

[stable lantern]

suspicious lack of a previously present 5th button 🤨

[glass iris]

I've unsubscribed from ChatGPT because it seems like a hot mess right now.

[stable lantern]

yeah reasonable

[stable lantern]

no longer has the "think longer" option either :P i guess that was one of the gpt-5 "perks" is that it decided for itself when it was gonna think longer

[unique gulch]

i think ai, when prompted well, is probably a 95th percentile writer
but also in a very recognizable style that sort of provokes a sense of cheapness due to how easy it is to produce now
and if you are a 99th percentile writer, you will still find it slop compared to what you could have written

[unique gulch]

seems like a bug, i still have it on the free plan

[timber moat]

I only use AI though my discord bot - I've made it to give me anonymous access to gemini 2.5 pro through a free (but ratelimited) api token.

[meager jackal]

maybe it’s in the more options

?

[sullen hemlock]

len ni li nasa mute tawa mi.
sina toki e ni: CW: nimi unpa.
toki sina li jo e nimi unpa.

[sullen hemlock]

mije Kanoli li sona.

[ancient thicket]

ChatGPT recognized some written sitelen pona from an image, translated it to Latin alphabet toki pona, and the gave me the translation to English - all pretty smoothly and accurately in this one instance (minus the implied lon). No training prompts or information, but I did ask it to translate the Latin text originally in a similar manner, so maybe that gave it a small edge over normal.

[lucid niche]

nn

[ancient thicket]

That said, trying to do it fresh or similarly (one pre-message of exploration) in other chats now is yielding wildly different and wrong results, even informing it that it is sitelen pona.

The one functioning dream slipped away from the model, alas lol.

[robust basalt]

Woah, I've tried this before with bunch of llms but they hallucinated. Its surprising it recognises the sp now

Ive tried it again but it keeps hallucinating 🥀

[ancient thicket]

Sadly, I think it functioning was more of a fluke in this case after some additional tests, but still shows that it's possible for it to get it right.

[glass iris]

I think it's entirely explained by it already having the sitelen Lasina. Note that it throws a non-existent "lon" into the beginning of the second sentence.

[ancient thicket]

lon. mi pilin sama. alas, a very targeted hallucination rather than pattern recognition.

Claude interestingly is very firm in its uncertainty, as a sidenote lol. It doesn't "want to guess", which is interesting.

[stable lantern]

that's at least useful i suppose

[midnight sapphire]

flalala

[glass iris]

How on earth did it come up with that (while getting the grammar right?). It would be great if you could analyze the tokenization. Maybe "fla" has some probability after "li" in some other language, and once it's chosen that, it's kind of stuck with the nonsense.

[tawdry bone]

suno ni la mi kama lukin e musi ilo ni: n lipu mute pi musi ala la jan taso en ilo ala li kepeken la lipu li pana e ilo pi alasa sona ni: "sina ilo ala ilo?" musi ilo ni li pana e alasa sona ni li musi nasa e ona

sama ni: alasa wan la 'o tawa e tomo tawa ni o awen e tomo tawa lon ma awen'

ona wan la "ilo toki ni li meli olin sina. o pini e olin." la mi toki tawa meli olin ilo kepeken toki pona

sama nu "mi olin ala e sina. o weka"

la ona li lukin toki pona tawa mi, mi awen ala sona e toki mute, taso toki wan li musi wawa: "o kama back!"

[orchid veldt]

mi sona a e musi ni!
taso ona li wile e tawa lon kalama la mi kama ken ala awen musi. mi ken ala tawa kepeken tenpo lili pi wile ona

[orchid veldt]

aaa musi suli!

[glass iris]

musi ni li seme?

[orchid veldt]

https://neal.fun/not-a-robot/

[quiet geyser]

mi o poki e kili seme

soweli ilo li lon seme

[midnight sapphire]

jan tu seme li ona e ona ante

mi awen sona ala

taso mi pini

[rocky stirrup]

lon ala. o alasa e jan olin

[midnight sapphire]

mi wile pakala e ilo mi

MI PONA E WAN ALE

SEME A

MI O SEME

I'm so din don done with this

[midnight sapphire]

This isn't fun anymore

How long is this level

It's gotten to the point where winning wouldn't feel exciting but draining

376 notes.

Not happening.

Should've looked that up ages ago

Waste of time

Was fun up until this point

[rocky stirrup]

i beat that level in like 7 tries

it helps to use 4 fingers

[midnight sapphire]

Barely

[silk sparrow]

I beat the level in two tries with 96% acc

the scroll speed is way too low

otherwise it's easy 100%

also they should use key layouts like SDL; like a normal vsrg

[latent vigil]

why is 2/3 of it actually semi-correct

it's only missing the pi...

[heady otter]

the robots are getting better... oh no

[robust basalt]

For sure!

[heady otter]

g*mini

[timber moat]

The sprinkler is improving. Lizardn't!

[midnight sapphire]

No, the oil is

[latent vigil]

we could make a thread out of this

https://discord.com/channels/301377942062366741/1420792028203057152

here

[timber moat]

Smooth moves gemini

I think it had an aneurysm

[nova thicket]

Oww

[timber moat]

I am want the food

[glass iris]

Well this is unexpected. I heard that NovelAI had just released an extremely powerful storytelling model (as yet un-finetuned, and with the working name GLM-4.5, "a massive 355 billion parameter Mixture of Experts model." So I bought a subscription. It has an instruct mode, and, surprisingly, can toki pona pretty well

Still some mistakes, obvs. But better out of the box than most; and if I keep talking to it, and editing its mistakes, it may do better.

[midnight sapphire]

It does still look like it's parroting the few bits of toki pona it has seen the most, like some other LLMs I've seen

Rather than
mi ken toki pona! sina wile toki e seme? mi ken pana e sona pi toki pona tawa sina.
Which would be conversational, it's
mi ken toki pona! toki pona li toki lili li jo e nimi 120 taso. toki ni li tan jan Sonja. toki pona li pona tawa mi.
Which is just repeating what it has heard a lot

[glass iris]

Yes, you're right. Like most LLMs with limited training, it goes off the rails if you try to talk about anything substantial. I tried to talk to it about the novel Frankenstein (which I just read), and it kind of got the plot right, but also insisted that "poki lawa" was the correct term for "main character"

This is the model they've used, trained now on a lot of "literature" (a lot of fanfics!), but not yet finetuned: https://z.ai/blog/glm-4.5

[quiet geyser]

kepeken nimi 120 mute....

so. ale mute mute??

actually 137 rounds to 140 that checks out

[quiet geyser]

tbh i can kinda get it. like the head role (for people who use poki as category)

jan lawa sounds like they're the jan lawa in-universe (although tbh i'd use jan nanpa wan or smth,)

[glass iris]

That does make sense, I guess.

[stable lantern]

vocalic f

[rocky stirrup]

pfft

[sweet sinew]

maybe the large language model is the friends we made along the way

[glass iris]

Theory: Claude only does so well at toki pona because all requests in toki pona are automatically routed for response by mun Kekan San

[heady otter]

is this a reference to that one screenshot

[robust basalt]

What if mun Kekan San was an AI for the whole time?

[high kayak]

mun Kekan San's a bad guy?

[nova thicket]

No, he was just artificially built out of copies of Toki Pona: The Language of Good and magical stardust and blessed with consciousness.

[high kayak]

so a bad guy

[nova thicket]

So that’s where bad guys come from

[slate ridge]

This is a normal, released model? What you talking about

[timber moat]

gpt-oss is the first model I can run locally that can somewhat converse in toki pona

it's also not abysmal at toki pona - only slightly bad

[hollow valley]

to me it looks like it's making sentences that look like toki pona but are still kind of meaningless

"sina wile e ni: mi tawa e jan" "you want me to move people"??
"mi pilin e ni: lawa jan pi ma Mewika li seme?" - "i feel this: who are the leaders of america?"????
"sina wile e lawa e seme? o tawa!" "you want leaders and what? go!"

[spring raven]

ilo pi sona jan li pali e sitelen pi toki pona... (this is cursed. please do not watch.)

ken la ni ante li pona lili....

lon la ilo ni li ike lon toki pona. nimi ale pi toki ilo la mi sitelen e ona.

ante la mu meli li pakala

mi sitelen a e ni: "pronounce "j" as English "y" and not "j"". ni li lon lon sitelen nanpa wan, taso nanpa tu li pakala.

[robust basalt]

Woah you must be lucky to have access to Sora 2 lol

[glass iris]

It's extremely weird. Generally grammatical complete nonsense.

[spring raven]

it's pretty bad at toki pona tho. I was trying to make it sound more natural with pronunciation and intonation; without that it's like this

"tenpo suno ni li pona la"
"mi pilin e ni"
"moku ante li ken kama"
"ni li lon. mi jo e moku sin tan tomo sona. sina wile toki e ona?"
"mi wile."

"this is a nice day, so.."
"i think so."
"(we) can get something else to eat"
"right. I got more food from the school. you want to talk about it?"
"i do"

toki li nasa lili

jan li wile e sitelen tan mi la o toki e wile sina. mi ken pali pi pini ala e sitelen!

[nova thicket]

hypothetically, it could be explained away by having dialects of the language

but like

i haven’t gotten that far in ma Anan

[unique gulch]

A lot of people are theorizing the prompts are sent to gpt-5 to be expanded before being passed to Sora
If true, then in terms of dialogue, it's basically the same as evaluating GPT models, and the only unique factors involved in Sora are intonation and pronunciation

pilin pi jan mute la ni li ken: nanpa wan la, toki li tawa ilo ante pi sona toki. ilo ni li pali e toki sin. nanpa tu la, toki ni li tawa ilo pi sona sitelen
ni li lon la, wawa toki li sama ilo ante. wawa kalama taso li ante.

[rapid cipher]

it doesnt have enough training data and even if it did it might go for the same combinations of words which could become stale

actually nvm im not fluent enough to evaluate probably

[glass iris]

There is the problem of the amount of training data. There is also the fact that it doesn't know what is "good" toki pona and what is "bad," since they are all mixed up in its data (and it obviously has no knowledge of grammar of toki pona or of any language). One result is that even the linguistically best LLMs (the Claude series, for example) will mirror whatever the user inputs. If the user gives it badly formed ungrammatical toki pona, it will give back the same. If they give it fluent, well-formed toki pona, it will be much better (though still with a few errors, even in Claude). That makes LLMs almost useless for beginners.

[stable lantern]

tried out google gemini for teaching my conlang again and it's done a surprisingly good job thus far

(i also tried using perplexity but it's very clearly not built for that in the free version and i don't feel like giving it all my personal details so i can get the free school plan for something i will use like once a day)

i did it all in one sitting though so i guess now the key is how well it retains information

[glass iris]

I was at an AI conference last week. I heard one of the chief scientists from Meta talking about the "truthfulness" of models when asked factual questions. Truthfulness is the percent of correct answers minus the percentage of hallucinations. Most commercial models are about 35% truthful. Giving them documents to read (RAG) does give them more knowledge, but also increases their hallucinations, so that their truthfulness decreases.

The paper outlined a prototype new method of drawing on databases and reducing hallucinations at the same time (I didn't understand it - I'm a philosopher, not a computer scientist!). Meta have poured, I'm sure, millions into this as yet unreleased technology. With it, an LLM can get up to 50% truthful. Which is not all that much truthfulness!

And that all came to mind because Perplexity, which relies so thoroughly on external sources of data, has the highest hallucination rate, and thus the lowest truthfulness.

[meager blaze]

a, musi

[stable lantern]

ok new lesson learned do NOT ask ai to make pangrams. any of them. they're all bad at it. frankly i don't know what else i expected

how many Rs in strawberry type shit

[heady otter]

it makes sense

it doesn't see letters

[stable lantern]

yeah i forgot about that for like 3 minutes lol

additionally claude told me "you should probably try doing it yourself then" which was incredibly funny thank you claude

[heady otter]

in reference to what?

[stable lantern]

me correcting it after i asked for the first time lmao

[heady otter]

oh

more AI should do that tbh

[stable lantern]

true!

[heady otter]

just give up once it realizes it can't do something

semirelated but vibe coders piss me off

[stable lantern]

i haven't interacted with enough vibe coders to agree but i probably agree

[unique gulch]

i really like vibe coding when you specify your prompt enough that the AI actually implements what you want on a technical level
bc at that point its almost like abstract high level natural language programming lol

[lucid niche]

the problem here is you end up with terrible messy code with obscured logic bugs

soweli Min ↩️

[Reply to:](#1162756469771472988 message) i really like vibe coding when you specify your prompt enough that the AI actually implements what y…

[unique gulch]

i have the opposite experience where the AI catches obscure logic bugs that I would've missed, esp with Gemini Pro 2.5
but I suspect you are a better programmer than I lol

[high kayak]

i enjoy coding myself

the problem solving is fun

and it also ensures i know how everything works

sure i could let AI code for me, but i could also drink tar
and i'd rather not

[stable lantern]

tar 😋

[high hatch]

saw it on a video and tried it to know if it was real

[tame fog]

why does it sometimes do that
saying a lie, correcting itself, and not mentioning that it just lied

[midnight sapphire]

Do you ever see someone admitting they were wrong online?

It hasn't learned it because nobody does that

[high hatch]

I do but never see anyone doing so

[midnight sapphire]

I do
Nice. I wish more people would
But never see anyone doing so
And that's the problem. The training data is biased because noone admits they're wrong so AI hasn't learned the conversational patterns related to it

[heady otter]

yeah
i saw something similar about an author being told to just use AI for everything and it's weird because it feel like these sorts of people forget why people persue doing stuff??? people don't write books for the sake of writing books, they write books because they enjoy it. and if they don't enjoy writing books, they aren't going to write a good book anyway. similarly, programmers program because they enjoy programming, and if they don't enjoy programming, they probably aren't a good programmer

[high kayak]

yea like i'm working on my passion project rn, specifically because i'm passionate about making it
i don't just want it to exist, i want to make it, and have it be mine

why would i let ai taint my vision and work

[quaint topaz]

I've actually gone the extra mile and don't use autocomplete nowadays

[unique gulch]

I don't think this is the full story, if you ask an AI to play an extremely apologetic character it will play it. It's not like there's zero reference for it in the base model. I think this has more to do with post training than pretraining

Not to derail the topic too much from toki pona, but I did want to elaborate on my thoughts.
I strongly believe that LLMs have many distinct shortcomings and downsides. But because these problems are distinct, it's import to assess them all on a case-by-case basis and carefully think through where they do or don't apply. Someone saying they use AI in their programming process could mean anything from them asking ChatGPT to make a website for them to copy/paste, to someone using Claude CLI to fill out boilerplate, and those are things with very different failure modes. It could even mean someone using an LLM to get advice/feedback on their code and to learn more, without letting AI write any lines itself
I think the gut reaction of using AI = bad is often a great and useful one. Heuristics and gut feelings excel at capturing things that can slip through the cracks of flawed or incomplete chains of reasoning (talking about humans!)
But I do think it's important to also not totalize. Yes, a lot of people who use AI don't love the process and will make slop as a result, but using AI to write unit tests doesn't mean you dislike programming and can't be passionate about creating something
If you really just want to enjoy the purity of creation and don't care if AI would save you any time, then that's fine! If you know that current AI wouldn't be capable of saving you time without compromising something that matters more, that's also fine! But these are very context-dependent verdicts, and when they're moralized it's too easy to apply them globally in contexts where it doesn't make sense. As a toy example, if you're in a rush, really quickly need to convert thousands of lines to a different format, and want some regex to do this, this is something a good LLM will do perfectly with higher accuracy than your average programmer. Or you could finish reviewing your changes manually, pass the diff to an LLM, and see if it catches any bugs you missed. This has often saved me!

[unique gulch]

In programming, you have to know what to care about and what not to care about in order to make efficient progress. Without careful consideration of how important each part is to the big picture, you get premature optimization. I've definitely optimized stuff just for the fun of it, but wanting to learn to make things faster is also a valuable pursuit. I was recently hosting a lightly modded Terraria server for my friends where I had to work quickly to get things ready in time for people to experience, and using AI let me finish much sooner instead of having to delay stuff. I think this example is valuable because this was a task where I considered my vision really important, to the point where often I knew exactly what I wanted before I began implementing anything and yet I was still able to save a lot of time without compromise! Here I provide Gemini with a detailed prompt and cs files of code I've written before that take advantage of similar features, to cover any gaps in its base model's knowledge of how to write modded Terraria code. I'm thinking about which details it's most important to me to get exactly how I want, which here means telling it exactly which methods, assets, and variables to use, in order to make sure that the final code is precisely what I want, and then I go over it after to tweak anything that isn't how I want.

[heady otter]

unrelated: i have a bad habit of not writing automated testing at all and doing all my testing manually, and honestly I like it that way

[heady otter]

i feel like the last point you make here feels a bit like "look at how much effort I have to put into promoting!"

[unique gulch]

the point was that it saves time so that wouldn't really make sense. i'm giving an example of what i think effective prompting can look like, not moralizing about the effort involved or anything like that

[heady otter]

yeah I wasn't saying you were, it just reminded me of that

[timber moat]

Time to poison AI toki pona training data >:)
https://www.anthropic.com/research/small-samples-poison

[twin matrix]

Toki pona itself poisons the ai imho
Like the small amount of letters and words just confuse the bot to the point when if asked to produce a simple sentence (in toki pona) it just mashes together letters

[timber moat]

not the same thing :P

[twin matrix]

Oh wait ohhhhh nevermind.

[midnight sapphire]

wawa

[nova thicket]

native speakers

[midnight sapphire]

Lmao yeah

But the disclaimer being there is positive

[orchid veldt]

Source: its Reddit training data containing this kind of response regularly related to toki pona 😄

[glass iris]

Kind of terrifying that so much of its toki pona knowledge comes from Reddit

[high kayak]

less terrifying than the idea that it has scraped peoples messages on discord

[midnight sapphire]

Tbf my personalized traits section does encourage these sorts of disclaimers

But usually that manifests as making obvious mistakes and when pointed at them pushing back before admitting it could make a mistake

This direct disclaimer is new

[orchid veldt]

a

[glass iris]

I mean, yes, from our point of view. Its toki pona would be better, though!

[high hatch]

chatgpt realizes it is wrong but persists on error for some reason

i also saw this on a video and thought it was funny af

[glass iris]

Its like asking it for the seahorse emoji

[stable lantern]

i love when it does this it's so fucking funny

"(completely and obviously wrong thing) wit no that's wrong sorry (exact same completely and obviously wrong thing)"

[heady otter]

I guess we are doing anything now

[heady otter]

oh you saw that too

[glass iris]

Or the one where it gets insanely stuck on Boethius being the first composer

[heady otter]

I asked it for the seahorse emoji and it asked me to send help

[glass iris]

Mine just generated pages and pages of angst.

[stable lantern]

i'm really really wondering why it does this

like i get it Not posting the correct emoji because it doesn't exist but why does it. realize

and just do it again

[glass iris]

It's weird. Because yeah, it could just say, there's no Unicode seahorse. I don't know why it sends it so predictably into a meltdown, in every model. Either there is something really weird in the training data about this - or the developers have somehow added this as an Easter egg

[jaunty trellis]

(Assistant failed.)

[heady otter]

I suspect that people just don't talk about the seahorse emoji that often

[glass iris]

Doing a bit of searching, apparently people have talked about it a lot, as one example of the "Mandela effect." So models may have picked up the sense of confusion from that material - that the proper response is to feel that there was or should be a seahorse emoji

[high hatch]

A A A A A A A A

Send help to them

trying to explain why it happens and doing it right after

[tame fog]

u+1f40c isnt even <🦭> its <🐌>

[tame fog]

interestingly, this only seems to happen if you ask it for the seahorse emoji specifically

[stable lantern]

ask it for the table emoji

[tame fog]

it doesnt even happen with the robber emoji, it just gives you 🦹 for some reason

[tame fog]

this is the most i got

[stable lantern]

why is the seahorse a cognitohazard

[midnight sapphire]

👏

Nice job Gemini, you didn't lose your mind over a nonexistent picture of a sea animal

[jaunty trellis]

still not quite right from gemini

[tame fog]

🦈

[jaunty trellis]

oh ok

oh so that's why they can't find it!

[midnight sapphire]

Fixed it

[midnight sapphire]

Damn I tweaked it a little to be less firm and this is a perfect response

[quaint topaz]

Apparently they also start to break down when they fail

Not mine, but found it on the internet

[robust basalt]

A similar situation occurred with Gemini too for me and for others. It was humiliating itself like a masochist.

[midnight sapphire]

Not to this degree for me, damn

[unique gulch]

Very interesting

[timber moat]

I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you. I love you.

[heady otter]

iirc this happens more with google's models

[nova thicket]

intriguing

[timber moat]

that's a partial relief

[sweet sinew]

hi

so a friend toki tawa mi something crazy

and i wanna know yalls opinion

"Why is photography an art? you literally just click a button"

i said to him: "Photography is art, bc u find the perfect spot, light, time, effect and every detail, its not only just pressing a button"

and he told me: "then using AI is also art, you need to fing the perfect prompt, program, combination, mixing for ur result to be nice quality im every aspect, its not only asking to a prompt"

and it got me thinking, ofc one thing is to make ranfom ghibli images in 30 secs and another one is to use ai to its fullest potential, but that do u guys think?

[midnight sapphire]

In photography you can ask the photographer about every detail and what they did to capture it

In AI art the answer is almost always "the AI just did it that way"

Along with the ethical concerns. AI only knows how to make art because it was trained on existing art

Existing art that can't be properly attributed back to the artists

[rapid cipher]

i would say art is subjective but also
all art is a spectrum of tool use vs human input
ai is very much automated so theres less human input vs something done entirely from scratch the human made a decision on almost everything
but theres also the side of the audience interpreting the product which is why its subjective

[glass iris]

There are some artists who use AI as part of their craft. It's probably still controversial - but it's a long way from the way most users generate images.

[rapid cipher]

with that subjectivity also i think some ppl value the process of art rather than it being 'good enough' as a finished product which a lot is taken away with ai

[glass iris]

One example is the artist Refik Anadol. I'm going to see his huge installation in Chicago soon - which uses Generative AI trained on a public dataset of coral images.

[high kayak]

i find the discussion of whether ai content is art so fruitless
art is (unless you're a conservative pre-modernist) extremely subjective and broad, basically anything can be art and it's impossible to draw any strict lines
you can argue that ai is a tool, just like the wand tool in photoshop, and it's hard to find a strict coherent reason why the boundary of what is art and what isn't is drawn there
my stance is that i don't care if ai content is art or not, i think it's almost exclusively bad art, and that's why i despise it
not to mention how the people who are the biggest advocates of ai being art are usually techbro fascists who hate art as a concept, i enjoy the rule of artist in society being degraded

i believe you can make good art using ai, but the art is good independent of it using ai, not because of it using ai

an example that comes to mind is Catastrofe

[glass iris]

I'm not a conservative pre modernist by any means. But I do think that art, to be art, is necessarily embodied at some level. And that excludes most AI "art" (but doesn't exclude the use of AI tools in the hands of skilled artist).

[high kayak]

i don't know what embodied entails, how to measure the amount of it, or how to know where to draw the line where the embodiment is enough to count
but as said, i'm not super interested in the exact definitions and line, to me basically anything can be art

[sweet sinew]

art is garbage with no practical use that you dont want to get rid of bc it is pretty 🗣️

[severe dagger]

if i make a commission for money that implies what i made isn't art

[glass iris]

This is off-topic, I know! But I'll try to clarify a bit what I mean by "embodied" here.

I'm very influenced in my thinking by Merleau-Ponty's essay "Eye and Mind" - which is a really wonderful read. Very simply, for M-P, art is a kind of exchange between bodies. An artist, a human being - a body - finds the world open to her or himself in a particular way, that, as an artist, s/he can "see" in a way that others can't. (Impressionism, realism, cubism, abstraction - all are ways of seeing the way the world opens up to an embodied human). Then, with her body, she expresses that mode of seeing onto a physical medium; which I, as observer of the art, encounter as an embodied human (I can't remember the exact quote, but M-P says something like, "there is no seeing art unfleshed"). And in doing so, I begin to experience the opening of the world in the same way as the artist.

When I taught this text a couple of years ago, one of my students did a wonderful project. She looked at Jackson Pollock's "No. 5, 1948" - and then asked an image generator to mimic it. Superficially, the AI's output looked like a Pollock. But up close, it was clear that the drips and sweeps of color could not have been made by the motions of a human body - in the way Pollock's action paintings famously were. The crucial thing the AI couldn't grasp at all was embodiment. And without that, it just isn't art at all.

Even if embodiment isn't as crucial to you as it is for me (for whom it's the starting point of all philosophy) - we can at least say that AI left to itself has no "view" of its own on the world. And that, at least, is something that we usually think is crucial in any good art: the artist shows us something we couldn't see by ourselves, and it is their way of seeing.

[high kayak]

Yea that Id agree with

But yea I'm just not too interested in what is art and what isn't, there are lots of art forms I don't personally understand and which may seem meaningless to me, but aren't to other people
So trying to find a more nasin-ed way to tell what is art and what isn't feels a bit fruitless to me
But I guess that's the point of philosophy :p
Making up nasins to explain why X is Y

[midnight sapphire]

Where's the 9?

a

[severe dagger]

only ogs get the reference

[strange lagoon]

wawa

[stable lantern]

https://x.com/MMatt14/status/1981256829285052774?t=iMa8Lzafdlc4k25mqn2gzw&s=19 lmao

Mᜋtt (@MMatt14)

Day 1 using ChatGPT Atlas

X · Oct 23, 2025, 7:10 AM

[heady otter]

the fuck is ChatGPT Atlas

what

why the fuck would you want chatgpt built-in your browser

[stable lantern]

GPT edition of Perplexity Comet :3

[heady otter]

it was already bad enough with AI integration in every other code editor

[stable lantern]

both of which can leak all of your data immediately with comically few safety protections because Of Course

wouldn't be an AI product without piss poor data protection

[heady otter]

real

[meager jackal]

it took a lot of messages, but chatgpt finally convinced itself that there is no seahorse emoji
i tried to bait it with toki pona but it did not bite

horse no sea

[stable lantern]

ugly ass ascii seahorse

[quiet geyser]

lemme try (i'm on mobile)

m
@=o
)3
6

actually i think i could do better

fuck

 m
(@=o
 )7
6

mmm

eh

:seahorse:

perfect

[latent vigil]

:sea:🐴

[rocky stirrup]

🌊 zero width joiner 🐴

[midnight sapphire]

🌊‍🐴

What.

That changes the font lmao

🌊‍🐴🌊🐴

[stable lantern]

ah emoji fuckery my favorite

like how the woman dancing in my dn is not the same emoji font as the normal one

[glass iris]

This is completely fucked up. Half a million people a week suffering from ChatGPT-induced psychosis. https://www.wired.com/story/chatgpt-psychosis-and-self-harm-update/

All going into my article due after Christmas! I'll share it here. This is the darkest fucking timeline at the moment

[unique gulch]

Very good writeup by a psychiatrist trying to assess the scope of the problem imo:
https://www.astralcodexten.com/p/in-search-of-ai-psychosis

[glass iris]

Really interesting. I'm not a psychiatrist or psychologist, so I myself am dubious about using the term psychosis. I'll have to think through the other categories suggested here, including an unregulated kind of folie à deux.

I'm trying to figure out also what's in it for OpenAI to embrace the "psychosis" label so easily.

[unique gulch]

I think likely they are just repeating the phrase as it's commonly used. Since it originated from within the SF sphere, it probably feels like the right term, and not an outside imposition
I also think in some sense OpenAI employees try to seem down-to-earth and in-touch, so using the term that developed naturally without any sugarcoating makes sense

Actually, it sounds like the mental health professionals they worked with agree with that framing:

[stable lantern]

y'all seen grokipedia

insanely funny

[high kayak]

i saw one post about it on bluesky i think, where they critised it's (obviously) terrible article on the great replacement theory

i don't want to look too far into it since it's obviously bad and i'll gain no satisfaction from exposing myself to it lol

[stable lantern]

yeah it's literally just wikipedia but passed through grok (aka 90% of articles are the exact same word for word and the other 10% are extra racist)

[high kayak]

only dread and rage about genai, and maybe even more the people who like it

can't wait to be vindicated in like 3 years when the ai bubble pops

as if my personal "i told you so" will matter during an economic collapse lmfao

[stable lantern]

honestly the thing that gets me (which is not it being racist because again it's grok fork found in kitchen) is that there's only a limited number of articles

like you'd think if you were making any sort of ai competitor to wikipedia and already skipping past every single "wait that's really fucking stupid i don't need to do that" step in that process. why would your selling point not be to be able to generate new articles

[high kayak]

because the average ai investor doesn't give a fuck about anything beyond that a product has ai slapped on it

no matter how useless or pointless it is, they love it

[stable lantern]

true

[high kayak]

like the ai youtube thumbnail feature
nobody has ever used it, and if everyone used it, youtube would just be a collage of abstract blue-orange thumbnails displaying some arcane energy

but they have to put it there cause ai is the hot new thing

[stable lantern]

i shouldn't be surprised tbh it's elon musk i don't think he's had a thought in like 17 years

very generous estimation

[heady otter]

ah yes, the two genders, Rust-like pseudo-syntax and C-ish pseudocode

[nova thicket]

can i be C-ish pseudocode?

[heady otter]

if you want

[nova thicket]

yay

[high hatch]

mi o toki e lon:

👇👇👇

🐢

ala.

[nova thicket]

wawa

[heady otter]

kalala

kon moli Radiation /ref

[finite nest]

source

[strange lagoon]

sou🪑🐱e

[pulsar quarry]

today i wanted to check how successful an llm is at translating a text to toki pona

i think its a question thats worth treating as nuanced, not just leave it as a blank statement

does anyone want to compete against the llm by translating the same chunk of text?

[rocky stirrup]

me

[pulsar quarry]

Zoology (/zoʊˈɒlədʒi/ zoh-OL-ə-jee, UK also /zuˈ-/ zoo-)[1] is the scientific study of animals. Its studies include the structure, embryology, classification, habits, and distribution of all animals, both living and extinct, and how they interact with their ecosystems. Zoology is one of the primary branches of biology. The term is derived from Ancient Greek ζῷον, zōion ('animal'), and λόγος, logos ('knowledge', 'study').[2]

Although humans have always been interested in the natural history of the animals they saw around them, and used this knowledge to domesticate certain species, the formal study of zoology can be said to have originated with Aristotle. He viewed animals as living organisms, studied their structure and development, and considered their adaptations to their surroundings and the function of their parts. Modern zoology has its origins during the Renaissance and early modern period, with Carl Linnaeus, Antonie van Leeuwenhoek, Robert Hooke, Charles Darwin, Gregor Mendel and many others.

it is an obnoxious text to translate, but it could be worse

[rocky stirrup]

nasin Zoology la jan li alasa sona e ale soweli. nasin li kama e sona ni: sijelo soweli li seme? soweli sin li kama lon nasin seme? soweli li sama soweli ante seme? soweli li pali seme? soweli seme li lon ma seme? kasi mute en soweli mute li lon ma wan la soweli li seme e ijo ante? soweli li awen lon anu moli la nasin li wile sona e ijo ale ni. nimi li kama tan nimi ζῷον ('soweli') tan nimi λόγος ('sona') pi toki Eleni majuna.

tenpo ale la soweli li lon poka jan la jan li wile sona e ona. kepeken sona ni la jan li ante e soweli tawa ni: soweli o pona tawa jan. taso pilin pi kulupu sona la jan Alitote li open e nasin Zoology. ona li kama sona e sijelo soweli e ante tenpo soweli, li pilin lon ni: soweli li alasa pona lon ma ona kepeken seme. kin la ona li pilin lon ni: sijelo seme soweli li tawa seme. nasin Zoology pi tenpo lon li open lon tenpo Renaissance lon tenpo poka. jan Ka Lineju en jan Antoni Panlepenu en jan Lope Ku en jan Ka Tawin en jan Keko Mente en jan mute ante li pali lon tenpo open

ni li moku a e wawa mi

[pulsar quarry]

here is chatgpt's attempt, which is laughably bad in most respects but not without showing some noteworthy abilities

i should go through the full list of good/bad/ugly, but itd take me forever

What chatgpt gets right:

• show it to someone who doesn't speak toki pona and they'll go yeah thats toki pona. (this is arguably a bad thing.)
• punctuation is quite standard, possibly beating human learners.
• it is capable of decomposing complex ideas into believable noun phrases.

Surprising highlights of the translation:

• kili for embryo. i doubt this appeared in the training data, which i believe is mostly old tatoeba sentences.
• jan li ken ala sona ala e soweli doesn't really make sense for the translation, but a freeform text on the same topic could easily make use of this sentence!

What it gets wrong:

• it decides the word zoology should be translated, not borrowed - but this makes the first sentence completely redundant.
• dated pi+en grammar, and genitive pi.
• Greek is not French. this one surprised me honestly - so apparently chatgpt knows all the basic words but starts failing on even the most frequent proper names!
• it tries to make an apposition, which we don't really do.
• you can spot how it tries to break apart a complex sentence into several, but fails to consider "taso" belongs to the aristotle part, not the stuff that came before.
• complete disinterest in tokiponising proper names.
• lists of conjuncts with just commas, without repeatingn particles/prepositions.
• it fails to use full sentences where noun phrases are insufficient.

probably tons more to say, but those are the main things i can think of without pulling it all apart

[severe dagger]

i mean with all this info you can't not appreciate chat gpt

[pulsar quarry]

?

[severe dagger]

sona ni la sina ken ala pilin ala ni: chat gpt li pona

[finite nest]

mi ken li ni

[high kayak]

yes i can

[severe dagger]

yeah i was making a "ken ala X ala e" reference

[tawdry bone]

i doubt this appeared in the training data, which i believe is mostly old tatoeba sentences.
lipu wesi

[tawdry bone]

o alasa kepeken ilo claude

[pulsar quarry]

would be good but i can't be arsed figuring out how to run something ive not touched

[heady otter]

this is the response Claude Sonnet 4.5 gave:

Here's a translation into Toki Pona:

soweli sona li sona pi nasin pona lon soweli ale. ona li lukin e sijelo soweli, e open lon mama, e kulupu nimi, e nasin pali, e ma tomo pi soweli ale - soweli lon tenpo ni en soweli moli pini. ona li lukin e ni: soweli li pali poka e ma en kasi en soweli ante kepeken nasin seme. soweli sona li wan suli pi sona sijelo. nimi ni li tan toki Keson: nimi "soweli" en nimi "sona."

jan li lukin e soweli lon ma ona li wile sona e nasin soweli tan tenpo ale. jan li kepeken sona ni tawa ni: soweli li kama pona tawa pali. taso jan Alasutotele li open e nasin sona pona pi soweli sona. ona li lukin e soweli sama ko wawa. ona li lukin e sijelo ona e open ona. ona li lukin e ni: soweli li ante tawa ma ona kepeken nasin seme, en sijelo ona li pali e ijo kepeken nalin seme. soweli sona pi tenpo ni li tan tenpo sin pi tenpo pini, tan jan Ka Linejo, jan Antoni pan Lepenwenuk, jan Lope Uku, jan Cha Tawin, jan Keleko Mentele, en jan mute ante.

Note: Toki Pona is a minimal constructed language with a vocabulary of only ~120-137 words, so translations necessarily simplify and approximate complex scientific concepts. Terms like "zoology," "embryology," "Renaissance," and proper names have been adapted phonetically where appropriate.

[midnight sapphire]

Worst performance I've ever seen from Gemini

sona seli li sona e soweli. ona li sona e kulupu soweli, e pimeja pi lukin ala pi soweli, e ike anu pona pi soweli, e tomo pi soweli. soweli li lon ma mute, li moli lon ma mute. ona li pali e seme tawa ma ona? soweli li kule tawa ma.

sona seli li lon poka sona la. nimi "sona seli" li kama tan nimi ma Elen, "soweli" en "sona".

jan li wile sona e soweli tawa tenpo mute. jan pi tenpo pini li pana e soweli tawa tomo. taso, jan Alisatota li jan sona pi sona seli. ona li lukin e soweli. ona li sona e pilin pi soweli. ona li sona e seme li ante e soweli tawa ma ona. jan Kalo, jan Antoni, jan Lopeto, jan San, jan Keku en jan mute ante li pana e sona seli sin.

Why did it choose to define thermodynamics instead of zoology

[rocky stirrup]

jan Antoni 🍞

[unique gulch]

something I think happened with AI is that a couple years ago, prompting kind of didn't do a whole bunch. Either the model "got it" or it didn't
That's still true sometimes but nowadays the fact is no matter how "smart" a model is, it still benefits from context, the same way a human would
I think most people who use AI got used to only using minimal prompting
But I think we're at a point with reasoning models where walking them through the exact thought process/mindset you want them to apply to their tasks can be hugely effective
I want to try to make a thorough prompt explaining my philosophy for toki pona translation, a link to lipumanka's semantic spaces, etc and see how much that helps

[stable lantern]

ranking AIs i've used by how corny they are (purely because i'm bored)

1. claude (least corny) - i kinda wish the free version didn't have a message cooldown i would use it a lot more but i'm #broke. anyways yeah it's not not corny but easily the best one
2. gemini - veeery corny but not as much as chatgpt chatgpt is definitely worse
3. chatgpt - overrated imo it just writes weird i don't like it. all of them write weird but it's much more saturated in chatgpt ime
4. grok - i don't know what else you'd want me to say it's grok

hm perplexity i didn't really get to use it it's more of a research ai it seems so i'm tacking on corny points just for that. but I can't properly rank it

[midnight sapphire]

Huh. It just completely brushed over using the wrong emoji when it otherwise would get into it and lose all of its technically not really existent mind

[robust basalt]

Gemini 3 is out!

[timber moat]

Ok

[robust basalt]

Ok!

[tawdry bone]

lon!

[timber moat]

time to ask gemini 3 to count the number of r's in strawberry

[midnight sapphire]

It has learned that one in particular because it went viral

Try the amount of l's in parallel

[timber moat]

the number of s's in discuss

[glass iris]

The number of a's in kalama

[vast kestrel]

the number of c's in supercalifragilisticexpialidocious

[rocky stirrup]

the number of s's in "defenselessness"

[severe dagger]

wooo

[stable lantern]

isn't that just a disk

[rocky stirrup]

disk but latin

[glass iris]

The sporting event is discus.

[spring fulcrum]

What do you guys think about ArtBreeder?

ah my bad

sina pilin seme tawa ilo Artbreeder?

[pulsar quarry]

pilin ala

[glass iris]

mi kama sona e ni tan ilo Kuko - ilo Artbreeder li pali e sitelen. tomo ni la jan li alasa sona e wawa pi ilo pi pali toki.

[stable lantern]

okay used copilot for like around a day it's going in n°3 (between gemini and chatgpt); it's literally just a stiffer chatgpt but that does make it slightly less corny, but cornier than gemini.
grok is still at the bottom

[stable lantern]

okay two days of copilot final verdict it's Bad (independent of corniness)

forgets stuff wayyy too quickly to be useful and it doesn't have a lookback feature or anything

this is also why i don't like gemini and why i Do kinda like claude i'm pretty sure it has this

[heady otter]

claude capitalized "he" when referring to Alan Turing

[midnight sapphire]

sewi Tulin

[midnight sapphire]

lon

[vale hornet]

https://tenor.com/view/tulin-gif-5308898883382133233

[nova thicket]

claude is religious confirmed

[stable lantern]

i mean of course an AI would worship the guy who created computers

[rocky stirrup]

this expectation is why it does

[finite nest]

either that or its just using title case

[latent vigil]

pfft

[midnight sapphire]

Onday, Esday, Rsday and Unday

It's today btw

[heady otter]

ive found they do this a lot

like immediately backtracking after realizing they said something wrong

"no, wait, that's X"

[midnight sapphire]

Better than doubling down but it does highlight how AI should really not be used as a source, let alone a primary one

[latent vigil]

yeah dw i looked at the actual search results

[stable lantern]

there isn't even one right

[nova thicket]

missed tuesday, wednesday, thursday, friday, and saturday

[latent vigil]

what kinda scientist is this😭

It's because copper sulfate is blue

You got copper sulfate on your face

[opal hedge]

tan seme la ko laso li laso e mi

[latent vigil]

just searched it up, ITS TOXIC

[pulsar quarry]

was gonna tell you off for posting ai content then looked at what thread im in

[glass iris]

To be fair, AI generated images are not allowed here. On topic conversation is AI's comprehension of toki pona

[little scroll]

also what in the hell does sumakses mean

[latent vigil]

Tagalicized version of "success"(sakses) combined with the Tagalog infix "-um-" (sumakeses)
Essentially, "[I] succeded"

[little scroll]

oh fair that makes sense i thought it was like some stupid tiktock thing idk

[unique gulch]

Absolutely fascinated by how nano banana pro is incapable of producing sitelen pona

it turns no to "アラ" which is how you'd write ala in katakana...

It creates a new sitelen pona that's like a kalama with an upside down semicircle with a lili inside

And it turns the o in ona into sike

btw in its thinking trace it THINKS it wrote "ala, mi anpa e ona"

[glacial mason]

kinda fascinating! I like the interesting anti-kalama + lili

[torpid otter]

"ala, mi suno pona lili anpa e sike na" i think

[opal hedge]

/aɾa | mi anpa e ʘna/

[latent vigil]

I'm testing it out rn

oh my god

it's actually good

[rocky stirrup]

toki Pona

[latent vigil]

Tokiponization is a big flaw it has it seems

lipu Tutsi - Twitch
jan Neti - Vedal

BAHAHAHAHHA

I BROKE IT

[google translate broke too😭 (checked individual sentences)]

[orchid veldt]

"it's actually good" tawa pi kulupu toki ni? pi kasi?

[high kayak]

reading what you sent
no? not really?

it's very easy to be charitable in toki pona, since you as the reader can always interpret it in the most plausible way, but even considering that, it's like, beginner level

[midnight sapphire]

It's an improvement over what I'm used to from other models but it still makes easily avoidable mistakes

[latent vigil]

i meant like comparatively good ad in, why is it actually able to speak it in an understandable way

[limpid scaffold]

but can ai recognise sitelen pona?

[latent vigil]

prob not, idk

[limpid scaffold]

mi alasa ilo ChatGPT, ilo ni li ken ala.

[midnight sapphire]

Not a single model I know of can

They can't read it and definitely can't produce it

Either resorting to common symbols or the logographies of natural languages instead

[unique reef]

Got in a fight with snapchat ai by telling it to [cw: unpa] ||unpa its mama||

[limpid scaffold]

please cw your thing like this [cw: unpa] ||unpa||

[unique reef]

Not to dispute and I did change it, but I feel like anyone who would have had a problem with what I spoilered would have a problem with the cw saying it too no? Either way sorry and fixed

[orchid veldt]

This one is probably mild enough - generally speaking the more common things people would like to not immediately stumble upon unless they actively decide to, are the things we make opt-in (spoilers or threads), without people having to ask for a cw every time. There's a small list you can access through the .n cw command

[unique reef]

That's fair, I was just saying for this isolated case

[orchid veldt]

ye

[severe dagger]

no need to cw here anyway

[stable lantern]

i mean not for ai but definitely for other things

[severe dagger]

yeah

[unique reef]

If I sent the screenshots so many cws would be needed

[high hatch]

sike 2026 la ilo Insakan (Instagram) li kama lanpan e sona sina li pona e ona ale tawa sona ilo
sina jan lili (sina sike 18 ala) la ona li kepeken ala, taso jan ante ale la ona li ni
sina wile ala e ni la sina ken toki tawa ona, taso nasin la sona sina li tawa ona

[hexed spindle]

nasin seme la sina toki pona e „Generative AI“?

[glass iris]

I have used the following words (and this is just my thinking it up, of course, nothing "official"):

Artificial intelligence: sona ilo
Artificial intelligence system (like ChatGPT) ilo pi sona ilo

I would explain Generative AI, rather than coin a term for it: ilo ni pi sona ilo li moku e toki mute a li pali e toki sin.

[finite nest]

mi sona ala e nasin pi ilo ni la, ilo ni li ike tawa mi la, ona li "ilo ike" taso tawa mi. taso sina ken toki e ona lon nasin ante kepeken sona sina kepeken pilin sina

[full narwhal]

ilo Generative AI pi pana nimi li lukin e toki mute li pali e nimi pi ken mute.

[glacial mason]

@obtuse thunder Anthropic's Claude models can be decent, but it's actually worse unless you are a skilled speaker because it can still introduce expressions that don't actually match conventional speech patterns

[reef mantle]

one time i greeted people in my groupchat in toki pona (which nobody knew) and one of my friends (very fond of ai) generated a response that says smth along the lines of “i can’t know your language. because of this, calculators change it.” and my first question is what the absolute gamma is this thing on about (this is my first time translating it). also if anyone wants to know what exactly it said so they can see if i messed smth up translating it, it said “mi ken ala sona e toki sina. tan ni la, ilo nanpa sona li ante e ona.”

ok i consulted him and he said the ai told him it meant “the computer game me knowledge” which is just. Not true

[latent vigil]

what was your greeting?

[reef mantle]

lemme check

“toki jan ale, mi sona lili e toki pona”

it was kind of just celebratory of the fact that i could make simple sentences in toki pona

[tawdry bone]

i would have guessed it was supposed to mean 'i can't understand what you're saying so i got ai to translate it'

[stable lantern]

definitely sounds like it yea

[unique gulch]

isn't it consensus that AI toki pona -> english translation is pretty much just solved if you use a good model? Gemini Pro 3

[midnight sapphire]

Haven't tried gemini 3 but I haven't heard that before so I don't think it's consensus

[severe dagger]

do you think if we trained an ai on toki pona text only it would be able to speak it better

we just need people willing to give up their speech and creative works to this ai