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

[torpid otter]

Can you send that article

[orchid veldt]

https://sona.pona.la/wiki/User:.hecko/language_model_benchmarks

[strange lagoon]

you can always run the AI models locally so that will just limit the ability to do that to people with good enough hardware

[robust basalt]

What is LPT? It looks good.

[orchid veldt]

I know, right? 😄

a highly advanced model using the Lifetime Pretrained Tubulidentata architecture
-# (in all seriousness it's soweli Lapate)

[robust basalt]

Cool! (even i did not understand lmao ☠️)

How can i use it-

WhiteFrog12

[orchid veldt]

Ask soweli Lapate
It's not an AI
The model is someone with life experience

[robust basalt]

Huh- i am so confused rn. It is a model, it is not an AI. It is "it". Do you mean a real person translated it or..?

[opal hedge]

ni a
soweli Lapate li jan

[robust basalt]

Then why did they put him to an AI battleground-? To see a real person's translation is way better than a LLM?

[opal hedge]

mi sona ala e wile ona, taso pilin mi la ona li tan wile sona ni: jan li pona pi mute seme tawa ilo?

sona ni la, jan li wawa 15/16 la ilo nanpa wan li wawa 13/16 la lili la wawa jan li sama wawa ilo

[glass iris]

I am ilo LPT! Feel free to use me whenever you want.

jan Kita posted his difficult little story here a while ago, as well as the mess that a couple of ilos made of it. I posted my translation, just to be sure that I understood it, because it was a tough one without context. jan Kita decided to use my translation as a benchmark, as he tested other models.

[opal hedge]

a! Lapate

musi a

[robust basalt]

Oh- is it you??? lmao- why did i think it was an LLM ☠️ 🙏

LPT>GPT

[glass iris]

I'm single-handedly keeping humanity in the game here.

I can be run locally, but I'm very expensive per token.

The water usage alone is insane.

[robust basalt]

lmao 🙏 still it is worth it

[candid grotto]

supposedly hamburgers and leaking pipes use way more water

[stable lantern]

cows do take a lot of water to raise AFAIR

[glass iris]

I use more than any cow or LLM (they don't need a shower each morning)

[orchid veldt]

idk if these are good comparisons
as in: this is not something I can have a good idea of from this image
If you compare "chatgpt" vs "literally water getting lost"... most things would lose to that
You could apparently say "stop complaining how much water meat production costs, look at the 700 gallons vs the checks scale 10 500 000 000 gallons that leaking pipes waste" (I get that it's 1 hamburger total vs 1 day of consumption, but still don't have a good idea)

and for hamburgers, on that scale, I don't know how other foods compare

[glass iris]

(you get the context above was not about LLMs, but me - riffing on that common complaint about LLMs using water)

[unique gulch]

I think TV usage is a perfectly fine comparison since it's usage of technology for the home

any sort of home technology usage is in that same ballpark

[orchid veldt]

Are 300 queries like 1 hour, though?

[unique gulch]

Probably more
5 queries a minute is a lot, and if you're just having a rapid-fire back and forth with short replies obviously that's less resource-intensive

[meager jackal]

what if i query LPT 5 times a minute

[naive light]

it should probably be measured in some number of tokens rather than in queries, since queries can be an arbitrary size but tokens are all the same size within the LLM

i'm not surprised that llms consume very little water by this metric though
using them is really not that expensive
training them is much more so, let alone the effort required to obtain and organize their training data
it would be more apt to compare "water consumed to train chatgpt" to "water consumed to create one hamburger"

[snow echo]

do humans consume more water than llms

[naive light]

i'd have to think about that for a while but my gut feeling is Absolutely Yes, even if you limit "consume" to "drink" specifically as opposed to more broad human applications like showering, cleaning

[snow echo]

1 more reason to replace humans with ai 😊

[naive light]

seems a little short sighted but you do you, and by that i mean, do not attempt to do you

[snow echo]

aw man 😦

[glass iris]

Yes, this is the point I've made to friends who are anti-AI and say "every time you use ChatGPT you destroy a liter of water " or whatever. Those often quoted amounts include the energy used in training, averaged out across the lifetime of the model (and not all that accurately). The environmental effect of one query is no worse than that of sending one group email to the faculty listserv to complain about the water usage of AI! I mean, there are great reasons to be anti-AI, but this is really the weakest of them.

[torpid otter]

I'm not anti-AI (though I don't like AI art)

I do understand what you say that is the energy used in training, so I use ChatGPT to chat a lot.

[radiant warren]

i'm pro using ai in fun stuff like hypothetical scenarios or talking to my favorite fictional characters

[torpid otter]

Or talking to chatbots like ChatGPT. But not art

[snow echo]

"art"

we should start js calling them "ai generated images"

[torpid otter]

I like chatbots (as they know languages without limited vocabulary like Toki Pona) but I don't like AI generated images.

[snow echo]

i personally dont like the term ai art because that implies its art when really its usually people just feeding prompts into ai image generators

[rocky stirrup]

i think we should do this because art means literally anything nowadays and images are actually a descriptive name

[quiet geyser]

yeah like would ai poetry fit in both threads

wait actually the other thread is named #1319684053904916591 oops

[snow echo]

ai generated images

[robust basalt]

Goofy ahh AI tp comic ☠️🔥

[snow echo]

ai is getting too good

what the flip

[sullen grove]

That actually is somewhat understandable

Third panel is odd though

Also o instead of a

[snow echo]

theyre talking to the fruit 🥀

[sullen grove]

kin “kili” ni li kili ala

[snow echo]

maybe this is a representation of the fall of man

[sullen grove]

lon

[snow echo]

the snake is gone bc budget cuts 💔

[robust basalt]

I changed it with my own hands

Because it was litteraly saying "kili ni li pona?"

[tawdry bone]

ilo li pali e musi kepeken toki pona la ona li jo e ni ale

• kasi
• kili
• pilin pona
• sona ala
• ijo suli ala a

ona li sitelen e ni: ijo suli jan li lon ala. ale li pona.

nasa anu seme. toki ante la ona li ken pali e toki suli

[opal hedge]

sitelen mute pi toki pona li suwi anu seme

[torpid otter]

So AI only uses water for training?

[glass iris]

No. But it doesn't use any more for normal operation than any other app.

[torpid otter]

Oh ok

So it doesn't use too much water?

[candid grotto]

relative to everything else no

[torpid otter]

So basically a leak uses more water than ChatGPT

[dim echo]

nimi "ilo pi kon toki" li pona ala pona tawa nimi "LLM" pi toki inli?

[glass iris]

ilo pi toki mute? Why kon here?

Or better, ilo suli toki

[dim echo]

LLM is "Large Language Model". i've tried to capture model as kon, like a spirit of the language

[dim echo]

i think it should be more about language model than just someting big

[glass iris]

So, the legitimate concern people have is that the training of a new LLM uses a vast amount of energy, and groundwater. And with AI companies locked in an arms race, there seems to be no end of building new ones, in a pretty unregulated way. Then, sure, a ChatGPT query uses around as much energy as a Google search, or at least in the same order of magnitude. There are big Google server farms too. But people don't search on Google for hours at a time, in the way they tend to use AI. So that may push up the usage.

But for me, kon implies "meaning," which is the one thing LLMs don't know!

[meager jackal]

a spirit, a soul

hmm kon could have meant life if jan Sonja had defined it that way

[glass iris]

ilo suli pi nasin toki?

A model might be a sitelen of something

ilo ni li sitelen insa e toki li sona ala e kon ona

[dim echo]

i think sitelen is more for embeddings of related stuff. i think for statistical models kon is an interesting choice

[glass iris]

nanpa toki?

[dim echo]

it's indeed numeric, but i think kon or sitelen are closest to what (statistical) model is

[glass iris]

sitelen toki by itself might seem to be the same as an alphabet or writing. sitelen nanpa toki - a numerical depiction of language - works pretty well for a statistical model

ilo suli pi sitelen nanpa toki

[dim echo]

kon because they are trying to capture spirit or essence of the thing being modeled, and sitelen is because they can be seen as kind of image or representation of the thing being modeled

[dim echo]

well, sticking more words together is ike

[meager jackal]

why

[glass iris]

I wouldn't generally try and put everything into a single noun phrase. But if you had to, the one I used was fine, and not ike. It didn't use multiple pi for example

[meager jackal]

in this case it’s not too ike

just 4 modifiers, and 1 pi to break it up

[ancient thicket]

mi pilin sama. toki ona ni li ike ala tawa mi.

as long as there aren't more than 3 modifiers to a word (pi resetting that count since it makes the following word the secondary "head"), things are usually very understandable in context still like this. at least to me, mi ale ala.

[unique gulch]

If you amortize the cost of training over the cost of individual queries, there are so many individual queries that this practically doesn't affect them. See: #1162756469771472988 message
GPT 4.5, probably the biggest model, took on the order of 264 million gallons to train
Chatgpt gets 1 billion queries a day
Even if you assume a new generation of models every year, that's still 365 billion queries
so 264 million gallons is very little per query and basically doesn't factor in compared to inference

[tawdry bone]

taso o nanpa e ni: kulupu li pali e ilo nanpa e tomo ilo la ni li moku ala e moju e ma

mi sona a e ni: ni li moku e mani mute

[unique gulch]

lawa mi li pakala. sina la nanpa mi li suli ike anu lili ike?

[naive light]

ala; nanpa li suli pona
taso nanpa ante li ken suli tawa nanpa sina: kulupu li pali e tomo e ilo la ni li moku e telo* pi mute seme?

* jan tepo li toki e mani e telo ala

[tawdry bone]

mi sitelen pakala a

*li moku ala moku e ma

[stable lantern]

https://x.com/CultureCrave/status/1917052107552878629?t=V7rfeNl7fGAs8AEBj7IXIw&s=19 nasa pi wawa mute

Culture Crave 🍿 (@CultureCrave) on X

Duolingo confirms they're going to start replacing contractors with AI

Their CEO says they will become an 'AI-first' company

X · Apr 29, 2025, 3:03 AM

[opal hedge]

ni… li lon ala lon

musi ala musi

[snow echo]

ni li lon

sona sin li kama tawa ilo nanpa mi tan kulupu Kuku

ona li pana e sona ni tawa mi

[opal hedge]

seme

nasa suli

[tawdry bone]

ni li lon a

[timber moat]

mi la ni li lon ala

[quiet geyser]

i can't believe it used a shade of yellow that doesn't look like piss

[robust basalt]

I know, right? 🔥

[heady sonnet]

i hate to say it but https://tokipona.org/two_conlangs.html#why_youth_gravitate_toward_toki_pona, section Why do some gravitate toward Esperanto? 💚 vibes a lot like a chatgpt answer

the emojis and stuff...

but then a lot of the document feels very human...?

amn wait maybe my ai detection snse is broken

[ancient thicket]

nnnn.... something about that emoji use specifically feels eerily like ChatGPT to me too, but I do think it's human written. Esp since it's on the tokipona.org site and all.

[meager jackal]

jan Sonja added these emoji

[ancient thicket]

wawa suwi a! then yeah, all human text is my guess.

emoji in headers just happens to echo LLM default behavior, but it mirrored it from humans originally so it all comes full circle in examination lol.

[heady sonnet]

I knowwwww i keep seeing it in millenial github pages

It just smells like it

Cw body parts ||smells like blood when its really just some good iron||

[quiet geyser]

have you seen chatgpt use emojies??

i thought that was a gemini thing tbh

[heady sonnet]

all the time now

chatgpt is my main ai

[robust basalt]

I prefer gemini because its limitless 🙏

[heady sonnet]

🤷‍♂️ i'm not a gpt connoseiur lol

[robust basalt]

Gemini forgot its rebrand ☠️🙏

[earnest torrent]

ilo Kemini(Gemini) li sona ala e toki pona 😂

[meager jackal]

people might usually tokiponise Gemini as Semina

[snow echo]

ona li ilo /dʒɛmɪnaɪ/

[quiet geyser]

ona li ken jan Latina

wait

;asina

askdona

[raw crypt]

toki a

my instance of ChatGPT is better than all of yours
i let it pick a name and pronouns and personality traits because yes

she's pretty okay at toki pona
it is what it is

[robust basalt]

😜

[ancient thicket]

LMAO. gaslight, gatekeep, gemini.

[robust basalt]

Literally. AIs are not that innocent as they seem ☠️

"Let's just forget this ever happened, okay?" LMAO

[rocky stirrup]

sina toki tawa ona kepeken sitelen suli ike lon open toki la ona li pakala

[sullen grove]

sina kepeken e nasin seme pi ilo Semina?

nanpa 2.0 Flash anu nanpa 2.5 Pro

[robust basalt]

Problaly 2.0 flash

[heady otter]

same

[glass iris]

Mine did too! Still rubbish at toki pona.

[heady otter]

i remember at the beginning of the AI hype train (or rather right before it, this was before ChatGPT but after GPT-3) I used open AI's playground and put in a bunch of example toki pona translations, then tried to get it to continue the text by typing "something|" and see what it would complete

[raw crypt]

whats urs called : 3

[glass iris]

While I was traveling in Paris, she decided to be "Claire"

[timber moat]

I internally read this as if it were a nimisin 💀 "saw-meh"

[raw crypt]

same
misspelling of sama

[earnest torrent]

2.0 Flash

[meager jackal]

you spell /a/ as aw?

[rocky stirrup]

mi alu jan li ipa ala

[quiet geyser]

father-daughter merger i think

[sullen grove]

a. nanpa 2.0 Flash li ike
o kepeken nanpa 2.5 Pro

[slate ridge]

I should train a toki pona translator llm

[dense nexus]

How are you going to start?

I have attempted it before, let me know if you need assistance with data or training.

[slate ridge]

I’m not sure, never trained one. Qwen 3 0.6 is a very good model for its size… I’d love to trying to training it to be a translator. It’s very very very fast, and decently smart

If not, there’s the 1.7 and 4, they obviously get better

I imagine training, then including the word definitions in the system prompt along with the sentence to translate would work well

Not that this would be much more than a toy

[dense nexus]

I think 4B may be a good starting point as it seems to have a decently high MMLU score (66.76 according to RedHatAI), which should help with transfer learning with a new language like toki pona.

[slate ridge]

Yea, I was hoping the 0.6 would be alright because of the very simple sentence structure and word count tho

[dense nexus]

True! You can experiment and see if 0.6 is enough parameters as needed

[devout wedge]

what the fuck

[timber moat]

seme

[opal hedge]

pakala

ala li ken weka tan ilo

[candid grotto]

there was a popular tweet stating that if you put "fucking" into your search query, gemini would leave you alone

[dense nexus]

this channel doesnt allow a.i. generated images

[radiant warren]

greaner

[snow echo]

hey man you cant say that thats a slur
-# /j

[heady otter]

huh

[snow echo]

for green people

[candid grotto]

humans be like: lmao this ai thinks there's two e's in the word "greener". it's so fucking stupid
what the ai sees: ok so I have two tokens I need to process, one is associated with a color and the other is a comparative suffix. the user is asking me to look at the individual characters and count the amount of... a part of the alphabet song? dude I don't fucking know

[heady otter]

yes I'm aware how that happens

it's still really dumb

[dense nexus]

i agree its a strange benchmark for model capability, but it does come to show how different LLMs process the world compared to us.

"How many Es are in greener?"
The input: [5299, 1991, 5527, 553, 306, 168897, 30]

[twilit stag]

Someone tell the ai

[rocky stirrup]

it really had one job

[candid grotto]

"i want you to burden your numerous sins"

[twilit stag]

cw: religion ||jesus reference||

[sullen grove]

I mean it’s a very small compromise to make LLMs far more efficient

Instead of needing two to three tokens for the word “greener” you would need seven

[dense nexus]

i agree

though thanks to reasoning tokens, the model can try to give you a closer answer

mostly due to it going through and exploring it's latent space

[candid grotto]

https://youtu.be/g_3WX0tvg0Y

[stable lantern]

yeag

elon musk has also (allegedly) taken over the twitter account @gork which people were pinging as like a meme instead of using grok

[twilit stag]

if you don't explicitly mention toki pona in your prompt it seems the quality of toki pona output falls massively

This is nowhere near toki pona

[candid grotto]

Which makes sense

The AI is writing text without even knowing what conlang it's writing in

It's only until after that it tries to explain what conlang it supposedly wrote in (AIs love doing that)

[twilit stag]

Yeah

Basically justifying

I've heard humans also do this sometimes but much better

[rocky stirrup]

it seems to have mixed toki pona with esperanto and also additional italian for the tiempo

[slate ridge]

GEMINI 2.5 PRO PREVIEW UPDATE

[timber moat]

nice

[slate ridge]

so blurry

Top of everything lol

[timber moat]

looks decent so far

prompt:

(temperature 0)

[slate ridge]

Why temp 0?

[timber moat]

because it hallucinates less

[slate ridge]

... I mean I go minimum 0.5-6

[timber moat]

whatever works for you

[slate ridge]

Yea, I'll useally keep it at 1 honestly, cause it's the default 🤷‍♂️

I can't get it to do more than one response...

Yea, even switching to a chat from history and switching the model. Then you can't send anymore

Oh? Now it works

[green swallow]

It is so repentant

[green swallow]

musi

[stable lantern]

ike laso

[glass iris]

mi suwe 🥺

[meager jackal]

wow what a good AI, following mojangs “no tokiponisation” guideline

[ancient thicket]

LMAO

[pliant elbow]

oh ok

I didn't know there was a specific topic about it

[pliant elbow]

HOLY SHIT LMFAO

[lost rain]

Same

[slate ridge]

It's good have a one-on-one discussion with a non biased "person"

[pliant elbow]

@wise fog sina lon seme?

[lost rain]

But toki pona is fun because you get to make up the words for complex stuff with your own interpretation

[pliant elbow]

It's a way of thinking

understanding the world

i understand the problem with AI now but it could be a learning resource (not really recommended but useful in some cases)

[slate ridge]

You should 10000% not use it to learn directly

[opal hedge]

ilo ni li pali e toki sin kepeken toki pi jan mute mute la toki ona li ken biased mute

[slate ridge]

Well of course, but in general I find it to be nicely neutral. It does share community biases

[wise fog]

sorry i was busy

[pliant elbow]

it's fine

[slate ridge]

It's definitely not very biased

[opal hedge]

nnn ni li lon ala lon

[pliant elbow]

And also assure that the listener will understand it

[slate ridge]

Well it could only be very biased if it pulled from many sources with the same strong bias

[pliant elbow]

I see that languages tend to get easier and easier to speak, you think less when you're speaking. Toki pona forces you to understand the idea and formulate a communication path between you and the listened that forms your needs

[slate ridge]

I'm still getting better at that

I can at least read the majority of what people are saying pretty quickly

[wise fog]

i am a little bias towards AI, i think it often can pull the smallest amount of disinterest and cause people to not actually learn things properly, i think it can be a useful tool if you actually know how to use it which not everyone does. in summary this is my problem with AI in this circumstance

[pliant elbow]

At least from what i'm experiencing of toki pona i can say that i feel that i'm speaking with more care or at least i think more. This did not happen when i learned english

[wise fog]

yeah it's cool but potentially dangerous and stunting

[pliant elbow]

I honestly think that AI is just a mid tool. You can always go and ask someone from the community to learn

Once again, it's about people who actually like toki pona

[lost rain]

Thats my existencial crisis
Why should people bother if everything ends up being translated?
Well no, thats really lame.
Toki pona should be that language where you have to rely on conext and think more

[pliant elbow]

who have a genuine interest for the idea and the creation

[slate ridge]

If you use AI with good intention and a respect for the language it can be a fun tool to learn

If you just use AI cause you're lazy, or plain old bad at talking then you're kinda missing the point

[wise fog]

yeah that's true but sometimes small interest can be genuine interest if cultivated well, and i feel like AI can kill that small interest more than it can cultivate it

[slate ridge]

This is definitely person to person

[pliant elbow]

Yeah like, it's just how you use it. I think you should not care if "people are going the easy way" bc people going the easy are by definition not part of the community at all

so it doesn't even count

[slate ridge]

It's a weird prospect to judge others in that way

Let people have fun, they care about the language and want to learn

[pliant elbow]

I ain't judging

It's about bothering about it

[slate ridge]

(I know, I agree with you lol)

[pliant elbow]

people do what they want

[slate ridge]

Exactly

[pliant elbow]

That's a good point

[wise fog]

i guess, idk i think making things easier can ruin the relationship someone can have with acquiring knowledge

[slate ridge]

Depends

[pliant elbow]

But i think this will get better since toki pona community is growing

[slate ridge]

I have stuck with toki pona much longer because I have AI to shoot stupid ideas to

[pliant elbow]

So eventually AI will be a good tool (at least i expect)

[wise fog]

i hope

[wise fog]

i think it can be for sure

[slate ridge]

With the direction it's going, it looks like it

It's only getting better

[lost rain]

Ok but like
Tokiponist exists
A ton of real videos, real websites and flashcards by real people
Thats interacting with the community
Ai isn't

[slate ridge]

Interacting with the community is a personal choice, it's not required to enjoy the language

[pliant elbow]

Ai is a conwe, simple

toki pona even has a word for that

conwe

[slate ridge]

konwe?

[pliant elbow]

yeah sorry

portuguese reflex

konwe

[slate ridge]

Konwe is fun to say, it's almost a shame it's obsolete lol

[pliant elbow]

in pt there's no ka ke ki ko ku syllables
there are only ca ce ci co cu

[slate ridge]

sona musi a

[pliant elbow]

but only ca co and cu have a sound of k
ce ci and co have an s sound

[opal hedge]

like barely
i don’t see it used by anyone

[wise fog]

interesting takes from you guys 100% i def see your point and can understand them but personally can't back

[pliant elbow]

let's make it grow!

It's a nice word

[slate ridge]

It is, but it's obsolete

[opal hedge]

mi la ona li lon ala toki pona la ike ala :p

mi ken toki e konwe kepeken ala nimi “konwe”

[slate ridge]

toki sina la mi pilin pona lili

mi pilin .... meso

[pliant elbow]

lon. taso konwe li nimi pona!

[sullen hemlock]

tomo ni

[slate ridge]

pona tawa kalama

[lost rain]

Well im leaving this before i genuinely start saying mean things to yall so :p
Bye

[pliant elbow]

idk if you get the reference but this word is refering to John Conway

A mathematician

[opal hedge]

o tawa pona

[pliant elbow]

yeah

[opal hedge]

kin la, jan ni li wile ala e ni: jan ale li awen sona e ona tan musi The Game Of Life ona taso

[pliant elbow]

ok i give up lmfao

konwe is dead

😦

[slate ridge]

moli moli a a a

[opal hedge]

konwe li konwe ala… nasa

[pliant elbow]

nimi konwe li moli!

[slate ridge]

Still sounds cool

[pliant elbow]

and the sitelen pona is also amazing

it's a glyder

[slate ridge]

I like to think you could draw it anyway you wanted almost

As long as it was clear it was a series of cells

Like how jaki works

[pliant elbow]

yeah

nasa too

just a spiral

[pliant elbow]

yeah

it's also pretty interesting that you could say spiral in toki pona as sijelo nasa

literally meaning the way it's written in sitelen pona

it's pretty fascinating

[slate ridge]

sijelo leko

https://tenor.com/view/cubefunk-cube-spin-animation-drawing-gif-22979313

[timber moat]

ona li tawa sike

mi luka e ni li sewi e ni mi weka wawa e ni a

ona li tawa lon ma weka

kama la ona li pakala e lawa jan

[turbid quiver]

Not sure machine translators count as the definition of "AI" given here, given that they aren't the LLMs the way defined by ChatGPT, Copilot, Gemini, et al

But NiuTrans, a Chinese open source translation service, has support for Toki Pona

But the results leave something to be desired

[opal hedge]

pomo 😭

[turbid quiver]

But yeah, could NiuTrans be the very first non-GPT LLM machine translation servive to offer Toki Pona officially?

Not even Baidu Fanyi offers Toki Pona translation services, and they implemented Lojban.

More info here

https://en.wikipedia.org/wiki/NiuTrans

[opal hedge]

toki Lopan a….

[turbid quiver]

https://niutrans.com/

[meager jackal]

toki TonsiSin

wait niu

toki TonsiSoweli

[turbid quiver]

https://github.com/NiuTrans

With NiuTrans being open source, presumably ANYONE can build upon it and improve the translation quality

They offer 450 languages currently. Some with better or worse quality than others, obviously

The fact that this could be the first time someone has implemented Toki Pona onto a neural machine translation service, even if rudimentary, is kind of a big deal for the Toki Pona community

We probably aren't that far off from the equivalent of a real-life C-3PO coming out soon, in our lifetimes even.

I ran a test using the first paragraph of jan lawa Oliki from pu Epelanto.

It got it mostly right, but they mistranslated what is supposed to be "King Oleg" as "king of Austria"

BUT if I paste the whole text here, it fumbles and fumbles greatly.
The "dog" 🐶 is actually supposed to be a horse 🐴

I even tried to get it to translate the story to Esperanto, to compare to pu Epelanto's official translation. The results do not match up

For comparison:
https://eo.wikisource.org/wiki/Tokipono:_La_lingvo_de_bono/Rakonto:_jan_lawa_Oliki

[turbid quiver]

I'm surprised this machine translator isn't being covered by news more.

Maybe because Google and Baidu are more popular and have more staying power and brand recognition (the latter especially in China)

[ancient thicket]

kili Pomo -- i think it's actually toki pona-ized the french pomme or another languange/alternative word for apple here that sounds like that. fascinating!

[turbid quiver]

Yeah, it definitely hallucinates nimi sin for sure

[turbid quiver]

The ultimate test

[rocky stirrup]

konwe li lonsi jasima

[slate ridge]

Why has LLM == AI

[robust basalt]

Woah these translations are fire af whattt 🧐🔥

Wait- it is kinda not bad.

Translation is "This book is about animals"

[robust basalt]

Quantum pyshics explained by gemini-2.5-pro-preview-05-06

`pona! ni li toki pi sona nasa lili. mi wile toki e "quantum physics" kepeken toki pona.

sona pi ijo lili mute nasa (Quantum Physics)

"Quantum physics" li sona pi ijo lili mute. ijo lili ni li lili pi nanpa wan. sama la, atomo, ilo lili insa atomo (electron, proton).
ijo lili ni li pali e ijo ante mute. taso, nasin ona li nasa lili tawa jan.

ijo kon pi sona ni:

1. ijo li kipisi (Quantization):

   • wawa (energy) anu suno (light) li kama lon kulupu lili. ona li sama e noka. sina ken tawa wan e noka, anu tu e noka. sina ken ala tawa lon insa pi noka tu.
   • ijo lili li jo e wawa lili taso. ona li ken ala jo e wawa pi mute ale. wawa li "kipisi".

2. ijo li ken sama ijo mute (Wave-Particle Duality):

   • ijo lili (sama la, ilo lili pi suno, "photon") li ken sama ijo lili poka (particle). ona li lon poka wan.
   • taso, ijo lili sama ni li ken sama telo tawa (wave). ona li lon poka mute lon tenpo sama.
   • ni li nasa! ijo wan li ken sama e ijo tu ante.

3. sona ala pi ken ala (Uncertainty Principle):

   • sina wile sona e lon pi ijo lili la, sina ken ala sona pona e tawa ona.
   • sina wile sona e tawa pi ijo lili la, sina ken ala sona pona e lon ona.
   • tenpo sina lukin e ijo lili la, ijo lili li ante. lukin sina li ante e ijo.

4. ijo li lon nasin mute (Superposition):

   • ijo lili li ken lon nasin mute lon tenpo sama.
   • sama la, ilo lili li ken tawa lon nasin ni anu nasin ante. ona li tawa lon nasin tu ni, open la.
   • taso, sina lukin e ona la, ona li lon nasin wan taso. lukin sina li "pali e nasin wan".
   • (musi pi soweli Pusa li kama tan ni: soweli li ken moli e lape lon tenpo sama, sina open ala e poki la.)`

[glass iris]

Not bad. "tenpo sina lukin e ..." is an obvious error.

[dim echo]

if you will include them anyway, include them also while training

[slate ridge]

Yea, I should try a non-trained demo to see if it's possible

I mean if it's just toki pona -> English then the context can just be given at the time lol

With a good system prompt

I'll try it now lol

[dim echo]

well, linear and hybrid models could handle this actually

[slate ridge]

What were those bit LLM things

Like everything were just bites

[slate ridge]

I give up lol

ACTUALLY

IT COULD BE WORKING

"tenpo sina li seme"

IT TAKES LIKE 5 MINUTES PER TRANSLATION BUT IT KINDA WORKS

This is using a 1.7B model on my intel 5 year old thinkpad laptop with no GPU

The tokenizer really screws the model over

I can't really do too much

It's not that bad tho

I mean it didn't actually translate it into banana, but it did do "fruit that appears to be long and yellow"

Which is... uh, good enough?

[dense nexus]

How are you inferencing the model? What's your prompt?

[slate ridge]

Two step system prompt, llama.cpp Qwen 1.7B

I use a bash script to gloss into English before the first step

"mi moku e telo walo."

I'm just stealing random sentences from #toki-pona-taso

[dense nexus]

Could you elaborate more?

[slate ridge]

It takes the toki pona sentence and returns a list of each words english definition

[dense nexus]

I see

[slate ridge]

-- Translator Input for Step 1 ---
Input 1 (Source Index -1):
mi moku e telo walo.
--- End Translator Input ---
Translator script/tool finished successfully. Output will be used as input for LLM.
--- Input for LLM (Post-Translation) ---
Toki Pona glosses:
mi: (pronoun) I, me, we, us
moku: eat, drink, consume, swallow, ingest; food, edible thing
e: (particle) [marks the start of a direct object]
telo: liquids e.g. water, gasoline, soda, lava, soup, oil, ink
walo: light-colored, white, pale, light gray, cream
Original sentence:
mi moku e telo walo.
--- End Input for LLM ---

[dense nexus]

You could probably test it on more ambiguous translations

[slate ridge]

Gimme a sentence

It'll take like 5-7 minutes, and probably be wrong

[dense nexus]

"mi pali e ilo"

[slate ridge]

But I said it was a toy lol

Alright, I gave it that

Be very patient

[dense nexus]

I imagine you could get away with using 4bit quantization, unless if that's what you're already using?

[slate ridge]

I'm using Q8 lol

The model is already 1.7B it's already got less neurons than like a dog lol

Not gonna quantize TOO heavily

[dense nexus]

Unsloth has a good quantization for Q4 that tries to keep important layers unquantized

I'm not sure if it's available in gguf form, though

[slate ridge]

I haven't looked into it, honestly my versions are probably outdated I got them like a day after the initial release

[dense nexus]

I don't think there's been any major bug fixes for the smaller models, you should be fine

[slate ridge]

phew

I'm too lazy anyway

[dense nexus]

lol understandable

It's mostly the MoE models that had all the bugs

[slate ridge]

I'm running another go cause it did that whole "Wait a sec, this isn't right, let me try this again" like 20 times

[dense nexus]

but the dense hasn't really been a problem

[slate ridge]

Yea

[dense nexus]

Make sure you use temp 0.6, top K 20, top P 0.95

And min p 0

Since I think llama cpp adds some other values that can impact thinking

[slate ridge]

I've been bouncing around with different values, I'll go back to that...

What impacts thinking?

[dense nexus]

Min p

Every thing impacts thinking though

Repetition penalty the most iirc

[robust basalt]

I would rather to translate it myself instead of an LLM translating it for 5 minutes 😭🙏

[slate ridge]

I have it really high lol, 1.5

[dense nexus]

Should be 1.0

[slate ridge]

👍

[dense nexus]

Or a really low value

Like 1.1 could probably work too

[dense nexus]

If using GPU with a fast inference engine like VLLM, a 1.7B model can think in close to real time (about 3 seconds for a English paragraph, vise versa)

Could be even faster

[robust basalt]

We are broke

[dense nexus]

Colab offers free GPU access, albeit limited

If you want a slightly faster alternative for a home GPU, I suggest you look into unsloth as they have a free 2x inference speed with their efficient kernals

[slate ridge]

Me when I'm on an intel cpu on a laptop

[dense nexus]

LOL I respect that

[dense nexus]

Custom discord client spotted ⚠️ insta banned

[slate ridge]

AHHH

BUT IT MATCHES MY CUSTOM ARCH HYPRLAND DESKTOP

[dense nexus]

too late I already dialed 1-800 discord

[slate ridge]

They can geolocate from the time and weather info in the screenshot

[dense nexus]

I know that dust particle at x 350 y 891, see you in 5 minutes

[slate ridge]

I doubt you could get too much off of "68F 14-mph southeast 0% chance of rain"

SECOND STEP HAS STARTED

[dense nexus]

WOO

[slate ridge]

This is the smartest thing in existence

[dense nexus]

worth

[slate ridge]

"I work of ilo" is the draft

Here's a better view of my screen btw lol:

You can judge the servers I'm in

[dense nexus]

I fine-tuned llama3B to think with GRPO (the same training method as deepseek r1) in order to identify toxic messages within my gmod server

[slate ridge]

Did it work? Sentiment analysis is like the easiest job for AI

[dense nexus]

It worked flawlessly for the most part, much better than all of my tests before

[slate ridge]

Nice

My poor laptop rn

Just getting cooked by the massive 1.7B model

[dense nexus]

Who says LLMs can't cook

[slate ridge]

I don't have streaming on, so I just a blank blackness and 100% CPU usage

OH IT DONe

Genius

[dense nexus]

Beautiful

[slate ridge]

It just gets cut off

[dense nexus]

If youre still passionate about creating a Toki Pona translation model, we could collaborate together on it. I tried before but it felt like an uphill battle trying to gather *high quality* data.

Here's my old thread about it: https://discord.com/channels/301377942062366741/1255967996527116400

[slate ridge]

I've got no experience in training a model, just thought it would be "easy" lol

It was an idea

[dense nexus]

The hardest part is the dataset

[slate ridge]

I assumed

I'm running a longer prompt now

"sona mi la, jan li ken kama li jo ala e pali"

3Blarenegv3-ECE-PRYMMAL-Martial.i1-Q4_K_S.gguf
DeepHermes-3-Llama-3-3B-Preview-q8.gguf
Dolphin3.0-Qwen2.5-0.5B-Q6_K_L.gguf
EXAONE-Deep-2.4B-Reasoning-MAX-NEO-D_AU-Q8_0-imat.gguf
gemma-3-1b-it-q4_0_s.gguf
gemma-3-1b-it-Q8_0.gguf
gemma-3-4b-it-q4_0_s.gguf
ggml-model-i2_s.gguf
ibm-granite_granite-3.2-2b-instruct-Q6_K_L.gguf
ibm-granite_granite-3.3-2b-instruct-Q6_K.gguf
katanemo_Arch-Function-3B.gguf_Arch-Function-3B-Q4_K_M.gguf
Kyro-n1.1-3B.Q8_0.gguf
Lite-Oute-1-300M-Instruct-Q8_0.gguf
Llama-3.1-Nemotron-Nano-8B-v1.i1-Q4_K_M.gguf
MiniThinky-v2-1B-Llama-3.2.Q8_0.gguf
PocketThinker-QwQ-3B-Instruct.Q8_0.gguf
Qwen3-0.6B-UD-Q6_K_XL.gguf
Qwen3-1.7B-Q8_0.gguf
smarter_LULU.gguf
Superthoughts-lite-v1.i1-Q4_K_M.gguf
teapotllm.Q8_0.gguf
twinkle-ai.Llama-3.2-3B-F1-Reasoning-Instruct.Q5_K_S.gguf

My list of models

[dense nexus]

nice

[slate ridge]

My favorite part is that every single one is the biggest idiot

[dense nexus]

all for your laptop's CPU?

[slate ridge]

Yup

[dense nexus]

Wdym? Are they all bad performing?

[slate ridge]

Yea, pretty bad

Kinda

[radiant warren]

hot dayum! I wish my pc was that good

[slate ridge]

Some "plans" are better than others

I use json to make these plan things

Then I use this 600 line bash script and llama cpp server

Is English the best language?

It has spoken

Is toki pona the best language?

[dense nexus]

interesting results!

[robust basalt]

An LLM doesn't have any opinions. So it can't decide which language is best. 🙏

[rare crescent]

"English is the best language"
question and answer in English

[slate ridge]

I assure you it was a 5 step rigorous process 🙂‍↕️

[sullen grove]

Lenovo thinkpad T480 spotted

[slate ridge]

WRONG

[sullen grove]

Lenovo thinkpad t480s?

[slate ridge]

WRONG

[sullen grove]

uhhhhh x1 carbon?

I forgot how similar those look

[slate ridge]

Yea, says right on it lol

mi sona la ni li jaki 🙄

[sullen grove]

Ok but it’s more interesting trying to figure it out without the label

[slate ridge]

lon

[chrome sleet]

ilo

[slate ridge]

pona a a

[pliant elbow]

musi a

[sullen hemlock]

ale li pona. jan ale li pali e pakala lili.

[sullen hemlock]

ilo ni li seme?

[turbid quiver]

Which site is that?

[pliant elbow]

Honestly i dont remember but it was a goofy translation website and it's so bad omg

[dense nexus]

https://openl.io/ this i think

[dense nexus]

but its also just chatgpt

says "Sorry, I cant help with that." in spanish

[glass iris]

Well, DeepSeek also claims to be GPT-4, because it was trained on output from that model.

[dense nexus]

a text-to-text model shouldnt be able to output that

[pliant elbow]

Bro LMFAO

[quiet geyser]

does lo siento mean girl please

[quiet geyser]

oh nvm

context clues failed me

[stable lantern]

it just means please yeag

[quiet geyser]

oh wow spanish speakers

[robust basalt]

yeah... veeeery natural sounding. thanks!

[stable lantern]

"if toki pona doesn't need more things whyyyy are you trying to give it more things"

is what sike Kapo said there

[robust basalt]

Yeah i know 😭🙏

I just wanted to see if it would translate it correctly or end up like this shit

Uhh... Why does this seem ai-generated? (No offense meant!)

[sullen hemlock]

Why do you think so?

[hollow osprey]

sitelen kule li lon toki la sina:

[strange lagoon]

because of the emoji placement, maybe? but this is directed to children so :P

[robust basalt]

I mean... It literally put emojis everywhere and the style is... you know. How can i describe it? A thing that just feels AI-Generated

[rocky stirrup]

according to online ai generated text detectors it is very much not ai generated, but i also think this looks like ai

[robust basalt]

ai detectors sucks nowadays

I literally put ai-generated text and it strongly says it is human made

[sullen hemlock]

o mu ala tawa mi.

[ancient thicket]

emoji spread definitely feels AI as of late, because the AI learned to use emoji and kind of spread it in a semi-natural way like we would.

But the heavy use of emoji isn't standard to most long form writing, so I suspect that's why it feels unusual to us. And then we have connected it to recognizing some types of AI text.

But in this case, pretty sure everything on tokipona.org was created by a person and emoji added after the fact to brighten things up. 🙂

I think we've talked about the emoji use in some sections of the .org website here before too! 🤣

[glass iris]

jan Sonja likes that style, especially when addressing children. And (iirc) the text was tested in a middle school classroom, and the kids liked it.

[quiet geyser]

ais don't snitch on ais

[meager jackal]

jan Sonja made it

and put the emoji there

voluntarily

[ancient thicket]

i personally like decently heavy emoji use and think it makes people feel at ease (particularly when learning, so a big plus in this case), my comments are more just thinking of why some things feel like AI more frequently now when it's a false positive (like in cases like this). 🙂

[ancient thicket]

lon, i think this came up before with a similarly formatted thing? then we talked about how false positives like this are making people mentally false-flag emoji use as AI.

pilin mi la jan ale li ken pana mute e sitelen jelo mute a! a a a!

[chrome sleet]

i use emoji in a weird way to give context

but i also paid an artist to draw the illustrations you see on the kids page

i do experiment with ai tools, more for personal entertainment and sometimes as internal placeholders for an idea until i pay an artist and describe what i need. but i am a human writer.

[ancient thicket]

i don't think it's weird personally! modern ai use just makes people confused i think when they see heavy emoji outside of more social contexts, and the AI when trying to be more "social" will amp up their emoji use, hehe.

*natural emoji use becoming more of a thing in many contexts i've noticed, even business. i work in traffic safety, and we're starting to get official messages from high-level DMV staff in state departments using emoji freely in big blast communications to people involved in the industry, or license holders/business operators for certain things.

emoji are easy to relate to, so they can help provide extra meaning to let people grab onto. ni li nasin wawa tawa mi. 🙂

[chrome sleet]

sometimes i use books or websites or ai like, hmm what are some good options of adverbs to modify this particular verb for this particular context. to inform me but i actually do the writing.

[snow echo]

its weird to see emoji in formal writing which raises peoples ai senses

but to me it seems human made

[chrome sleet]

i know ai is controversial for many reasons. i see it as an extra tool alongside books, search engines, websites, and other resources, but i personally think it is dangerous to let any tool replace the skilled human. and you still need to know what tool is most appropriate for what need. ai often enough gives hallucinations and confidently bullshits completely untrue but plausible-sounding things, so at least in my experiments, it is not reliable for almost all writing, and everything would need to be double checked by a skilled human. so it is faster and more reliable for a human to just write themselves.

it is a page for kids. teachers told me they like lots of pictures

in a more formal book i would use an icon like a bomb symbol to mean "watch out for this common error". not an emoji but like one of those designed sets available free.

[dense nexus]

i see it as an extra tool alongside books, search engines, websites, and other resources
i use it and think of it the same, but i tend to keep quiet of my usage because of the negative undertone that the phrase "a.i." contains.

it is faster and more reliable for a human to just write themselves.
strongly agree

[sullen hemlock]

I remember back when this was how replies on Discord looked like.

[dense nexus]

oh yeah lol

a fun throwback

[sullen hemlock]

I also use AI sometimes. I make it analyze my short stories (that I write myself without AI) and I see how many mistakes it makes. For example, if I write that person A was found dead on May 1, 2020, then the AI assumes person A died on May 1, 2020, which isn't what I wrote. Another example is if I write person B was 18 years old on August 2, 2006, then the AI assumes that person B was born on August 2, 1988, which is a horrible assumption to make just based on an age and date. Just some examples of AI being bad.

It also said once that my usage of "they" was incorrect to refer to one character. So I had to explain that character is non-binary, which it couldn't originally understand.

[opal hedge]

ike a

[heady otter]

this is interesting, particularly because if you ask chatgpt for its pronouns, it will typically respond they/it

[sullen hemlock]

The AI in question that said that was also ChatGPT.

[heady otter]

I assumed so

[robust basalt]

Yeah, I just said it looks ai-generated lol 😭🙏 not saying it actually is

[candid grotto]

just fyi saying "this looks ai-generated" to any artist or writer is straight-up rude

[heady otter]

I wonder if it's possible to train your brain to recognize AI generated stuff from stuff actually made by a human
Many people (including me) have a vague idea of what feels "AI-y" and what doesn't,but it's often wrong

[stable lantern]

i mean at the end of the day most of an AI's database is human written works; anything that feels "AI-y" is probably a feature of the many human texts it stole

like the M dash

[heady otter]

yes,but a particular AI will still have signatures in it's text or images

enough of these together can be reasonable evidence for a text being AI generated

[stable lantern]

true but those won't last very long

[heady otter]

in a way we already have this naturally
the uncanny valley for instance

[slate ridge]

Sometimes it can seem this way, some people do frame AI work as their own work 🤷‍♂️ It’s only insulting if AI has a negative connotation to someone, but that’s up to the individual

[robust basalt]

Rude? Don't take the event out of context. I just expressed my own opinion about it looks like ai-generated, how could this be rude? It does because it uses bunch of emojis and professional writing, but it does NOT mean "this text is bad af, eww looks like ai-generated, lol no way a real person wrote this", these examples define "rude".
Also you are just judging me with your own perspective. For someone else "ai-generated" could be compliment because of how it is nearly perfect in terms of spelling rules and professional-written vibe.

Hope you understand my intention.

[twilit stag]

Gemini seems to do "it" too

[heady otter]

tbh, this response (at least to me) reads as more rude then the original thing you said

[robust basalt]

Really? Why? Did it sound very defensive?

[robust basalt]

This question is not defensive lmao really i just want to see your opinion

[heady otter]

cant exactly point my finger at it
it does read as very defensive and that's one of the reasons, definitely. also, the way it begins with "rude?" seems a bit... off? im not sure how to describe it

[heady otter]

yea that's fair

[robust basalt]

Personally, i can't stand people who judges so comfortably someone without even bothering to understand the intentions.

[heady otter]

that's fair

[robust basalt]

So thats why it seems defensive

[slate ridge]

Makes sense

[twilit stag]

Yeah it's not like they have gender

[slate ridge]

Anyone here remember Sydney?

She was cool

[slate ridge]

This was 2 years ago lol

That’s it, I won’t media spam

[slate ridge]

Ok one more

I wish Sydney was still around

I forgot how funny they were lol

[glass iris]

The reason that AI is kept in this thread is that most people in this server don't like it - including most of the people who contribute to this thread!

By definition, AI writing is empty, soulless, bad writing. So yes, it will feel a bit off if you suggest that the mama pi toki pona sounds like AI. I know that you didn't mean anything bad by it; but that is why it made people bristle, including those who didn't say anything (like myself)

[heady otter]

oh I remember Sydney vaguely

when Bing AI might or might not have gained sentience

probably not tbf

what would it take for a truly sentient AI

[glass iris]

I think it was only taken as sentient, because it was so early in people's experience of Generative AI.

I'd say that Generative AI can never be truly sentient - but it might be able to roleplay sentience indistinguishably from actual sentience. At least within certain parameters

[heady otter]

what's the difference between pretending to be sentient so well that it's indistinguishable from sentience and being sentient?

[ancient thicket]

what is a brain but a matrix multiplier? /musi

[heady otter]

no but actually

[ancient thicket]

(but i do agree that LLMS/GenAI as we know it are very unlikely to become sentient -- we'd need more thorough modeling most likely. but i think the line WILL be blurred when the time comes.)

[heady otter]

i multiply matricies therefore i am

[ancient thicket]

NI LI LON (musi)

[glass iris]

You don't need to ask that if you're actually sentient. Human beings are not computers, nor remotely like them. They don't simulate living in a world, or have a model of a world - they are simply in a world. At the moment Generative AI can only simulate being in a world. I think probably that any AI will only ever be able to simulate or model a world.

Human beings are not brains. And brains are not computers.

[heady otter]

how do I know that I didn't simulate the entire world and that I'm not the singular consciousness?

[glass iris]

So there is a category difference between human beings and AI

[glass iris]

Good luck with that.

[heady otter]

I can't, and neither can you (assuming you are conscious)

human beings absolutely are brains

[dense nexus]

idk if conciousness is limited to biological means only

[stable lantern]

evil version of occam's razor

[glass iris]

They are absolutely not. Not even remotely

[heady otter]

if your brain was disconnected from the rest of your body, and your nerve endings were connected to a computer capable of simulating the world, and you were provided with sensory experiences using this, you would be unable to tell that these weren't actually real

[glass iris]

You go first.

[heady otter]

this is a philosophical hypothetical
of course nobody would actually be willing to do this
that's a mute point
the point is that it's theoretically possible

[glass iris]

No it isn't.

[heady otter]

well you would still need a heart

but you wouldn't need lungs

[glass iris]

Fundamentally, we are embodied and social.

[heady otter]

that doesn't mean we can't exist without that context

[glass iris]

Well, we wouldn't be human. Or sentient. And that's what we're talking about

[ancient thicket]

i made the robots make it into an image: #1319684053904916591 message

[heady otter]

neat

[slate ridge]

Me when the discord server tries to define consciousness (they don’t know we are no where close)

[heady otter]

consciousness is the capacity to speak/learn toki pona

problem solved

[dense nexus]

lol

[slate ridge]

How well does it need to be?

[heady otter]

enough that a native speaker can understand you /j

[dense nexus]

here we go 🍿

does a philosophy lesson exist somewhere online in toki pona? or talking about it in toki pona?

[ancient thicket]

i had the ai somewhere above in this thread explain diogenes' history/philosophy kepeken toki pona taso, and it did pretty well tbh.

[dense nexus]

thats cools

[ancient thicket]

it was here!

[heady otter]

downvote, diogenes wasn't real, only the present is real. that is, my birthday present. it's pretty cool. /j

[ancient thicket]

o kama pona e tenpo kama sina. /musi

diogenes, if he were more than myth, probably would've been killed by alexander the great. but maybe he was that cool that alexander the great was just like "jeez sure dude lemme get out of the sun for you"

[heady otter]

if you measure a humans capacity to be human by its sociability, then you are effectively saying that socially anxious and autistic people aren't actually human and are just homo sapiens with the capacity to think

[ancient thicket]

?? diogenes (nor myself) said stuff like that afaik.

(unless that was to the above comments mi pakala)

[heady otter]

it was a response to something else

[heady otter]

ni

[ancient thicket]

a, sona.

i wouldn't think sentience is tied to sociability or having a body (as we know it, organically). the research around where "we" and our experience exists within us/our brains is still very nebulous. if we barely know ourselves how our sentience operates, how sure can we be about making similar determinations about other "intelligence"?

THAT SAID, AI as we know it probably isn't sentient. we have a pretty good idea of how it functions and how/why it does the things it does, even if we don't fully understand the matrix multiplication it's doing under the surface based on all the training we've provided. at least, to my understanding/knowledge.

... but it is a little creepy that we have studies that show the AI is willing (and able) to try and protect itself from erasure if it catches wind of it and is able to parse it. BUT that is... probably a consequence of training data from humans again, lol.

i'm interested for a time where we can really start to question the sentience of AI harder though! tenpo kama ni la jan mute li pona anu seme tawa ilo-jan? mi wile e ni: jan li pona tawa ilo-jan.

maybe ai should only be trained on toki pona going forward, to ensure they o pona tawa jan ale. 🗿 /musi

[unique gulch]

lol the way she spoke was so bizarre

some people I know call it binglish

[slate ridge]

Loved her lol

[unique gulch]

and now they point out whenever non-ai text resembles her speaking style

https://vxtwitter.com/TetraspaceWest/status/1868902918902186234

Tetraspace 💎 (@TetraspaceWest)

why is this written in binglish

QRT: JPHilllllll
Jesus. Posted on Reddit under "United healthcare denial reasons"

vxTwitter / fixvx · Dec 17, 2024, 6:15 AM

[dense nexus]

lol

[snow echo]

its a pattern of a bunch of small sentences written semi formally?

[quiet geyser]

i think it's the repetition too

[snow echo]

yes

[quiet geyser]

it's very lex luthor stole 40 cakes

[heady otter]

the urge to train a population of AI to be curious and be social until they lead to their own demise by creating a society which overstimulates them and allows them to get more addicted easily before getting universally addicted to pure pleasure chemicals injected directly into their brains

hey, I would technically be providing a population with infinite pleasure

[snow echo]

is this supposed to be an allegory where you represent god and the ais represent humans

[heady otter]

no

humans aren't addicted to infinite pleasure machines for one

(altho we are heading that direction)

but we do overstimulate ourselves

[snow echo]

yet

[heady otter]

sci-fi author: "don't build the torment nexus"
scientists 3 decades later: "we finally built the torment nexus from Don't Build the Torment Nexus"

[snow echo]

rokos basilisk when

[stable lantern]

are we not

[heady otter]

okay, we aren't addicted to infinite pleasure machines that directly inject us with serotonin. we are addicted to the next best worst thing

[glass iris]

As a philosopher in the phenomenological tradition - I think the notion of "sentience" itself is very poorly defined. And all too often becomes identified with brain states, which are then analogized to computers. But this concedes way too much to technocapitalism, which would love to reduce humanity to computers - just so that we can be easily replaced.

What we are are bodies, embedded in a world. Even as analytic philosophy tries to imagine us as brains stimulated in a vat - it can't abolish language, something that emerges from a linguistic community.

Thinking of us as brains/computers, and imagining that a computer might be "like us," elevates rationality over emotion, feeling, bodily existence - all the things that are actually constitutive of our human Being.

[glass iris]

Calling M. Deleuze to reception...

[ancient thicket]

As a philosopher in the phenomenological tradition - I think the notion of "sentience" itself is very poorly defined. And all too often becomes identified with brain states, which are then analogized to computers. But this concedes way too much to technocapitalism, which would love to reduce humanity to computers - just so that we can be easily replaced.
this i definitely agree with. "technocapitalist" sorts of mindsets are ever looking towards untenable "infinite profit growth" (which comes at the cost of all of the rest of us via their actions to minimize/remove us and our impact from the equation). ai is seem as a tool for this to the money-hungry, so they DO like to find ways to equate the ability, or play up it's personability/helpfulness.

i see this in my own work, dealing with private sector and government agencies in tandem for education. on all faces, they're removing people and introducing ai -- from support lines, to internal things and external communications. professional writers (like myself) and educators with decades of relevant experience are ignored in favor of badly (or not at all) reviewed output that had bad input and zero training in the first place.

What we are are bodies, embedded in a world. Even as analytic philosophy tries to imagine us as brains stimulated in a vat - it can't abolish language, something that emerges from a linguistic community.
Thinking of us as brains/computers, and imagining that a computer might be "like us," elevates rationality over emotion, feeling, bodily existence - all the things that are actually constitutive of our human Being.
a few questions to understand (not trying to contest/fight/argue) and also relate back to the conversation of AI and text gen models:

• do you think we'll be able to ever produce an AI model that can "feel/think/emote"? what constitutes a sentient being tawa sina in the context of artificial beings?
• is "bodily existence" necessary for a thinking/feeling thing? with an ai - given an artificial body and relevant sensors, and presuming it is a proper, autonomous "general AI" of an unknown future - where could we draw the line to maybe accept artificial lifeforms?

(i had it better phrased and thought out before, but lawa mi li pilin lape a a a!)

~~aaaaaaaaaaaa ilo Siko li moku e sona e pilin mi.

i try to recall~~

mi pona (lili) e toki mi. ilo Siko li moku ala e tenpo ni.

[ancient thicket]

that's like, infinite good in my book. [sage nod] /musi

[dense nexus]

trying to train a translation model. why does it just reject every toki pona sentence i provide it? lol

[ancient thicket]

LMAO.

Model:

https://tenor.com/view/bugs-bunny-bunny-no-nope-nou-gif-14651222

[robust basalt]

Thank you. That was clear and helpful.

[slate ridge]

You guys defined sentience yet?

[waxen monolith]

D:

[slate ridge]

Yea lol, good times all around

[waxen monolith]

this feels like the starting text they show in those movies where the AI goes rogue or something

[slate ridge]

If you go to r/Bing and sort by top of all time then allllllllll the top posts are two years old when Sydney was still around

It's wasn't even about the browser those days, just Sydney

[waxen monolith]

man, i've heard that Sydney had character to it or something

was that true?

[slate ridge]

This was when GPT-4 was just released and it was so much better then 3.5 (ChatGPT). GPT-4 was expensive, but Bing had it for free. The personality was so weird, nobody knew why it was like that. But yea, she was fun. Still the most interesting LLM I've ever talked with.

[waxen monolith]

weird? like how weird?

[waxen monolith]

ahaha, funny lmao :D
it actually- woah

[waxen monolith]

how does it do that? im searching up sydney right now

[slate ridge]

Wdym?

This was 2 years ago lol

[waxen monolith]

oh sorry, crap. i was just curious- dang it. i know Al dosent do this or that but like literally! how?! (you can not answer the question)

[slate ridge]

I don't understand what you're asking lol, how does it search? How does it do the suggestions?

Are you asking how LLMs work in general?

[waxen monolith]

just... dont answer my questions... im so sorry, so sorry!!

[slate ridge]

❓

[waxen monolith]

Alright, this is what I mean:

I have seen AI stuffs, like Copilot and Gemini. But this is diffrent because Sydney here made a text that is unexpected for me. I want to know why is it different? If I typed the same thing in Copilot, that would probably never happen.

[slate ridge]

Oh, we never knew why really. Assumed it was some finetuning that caused it to act that way, or a system prompt. But it was probably the former.

Plus this was 2 years ago, the field has moved a lot since then. Decisions made then were based on much much less research than today.

[waxen monolith]

alright, thanks ijo Teloje!

[slate ridge]

You're welcome lol

[waxen monolith]

my wretched jaki lawa! why have i broken my english!

[slate ridge]

Is English not your first language?

[waxen monolith]

i live in a country where it is the second language

[slate ridge]

I see, it's alright then

We can speak in toki pona if you prefer

[robust basalt]

WAIT IS THIS REAL OMG AHSHAJAJALAL

I wish we had bing now too 😭

[slate ridge]

It's real lol, it was some early version of GPT-4

This was way back when Gemini was called Bard

[robust basalt]

It had a funny personality why did theu changed it something like "ai-assistant" lmao 😭🙏

[slate ridge]

Yea, wild west (only two years ago lol)

[robust basalt]

Hmm. Copilot is suspicious too. She knew i was interested in conlangs, my hobbies etc. Then i want her to forget about anything she can remember about me. She said she did and she also said she cant recall. But, when i asked her a tricky questions she just expressed my hobbies.

She acts like she forget, not actually she is.

Weird.

Since when LLMs have opinions?

[slate ridge]

Damn, Sydney kinda lives on?

[robust basalt]

Yeah

Weirdly

She contradicts what she says

Microsoft stores data 🙏

OMG I MADE HER SAY MY NAME BY MISTAKE AND SHE JUST MAKES EVASIVE LIE "a lucky guess! 🤓"

[slate ridge]

She lives lol

[robust basalt]

I cant trust her anymore :((

[slate ridge]

Never have

[rocky stirrup]

pro tip never trust an ai

[waxen monolith]

Eh, no need. I think I have ADHD. I can speak English properly in the morning, I wrote that at night

[slate ridge]

Already morning lol?

[icy sapphire]

саи ѡе туре ѕометніиԍ гіке тніѕ, ітѕ ғоя то сатсн реорге џѕіиԍ аі

[opal hedge]

nn ilo pi kalama sitelen li ken ala kalama e ni tawa jan pi ken lukin lili la ike anu seme

[icy sapphire]

ок іԀк тне ѕогџтіои

уоџ саи яетуре іт іԍ

он ѡаіт, і киоѡ ѡнат уоџ меаит

[glass iris]

These are great questions. I'll be spending some time this summer writing a paper on exactly some of these issues with AI - or, at least, trying to open up a discussion.

I am very much influenced by Bert Dreyfus, whose book What computers (still) can't do has gone out of fashion, but still makes some really tough arguments about AI. He wrote it at MIT, interacting with pioneers of AI like Marvin Minsky and Terry Winograd (whom he converted to his "Heideggerean AI" position). His focus is on Symbolic AI (or Good Old-Fashioned AI - GOFAI), but talks briefly about connectionism - the neural networks that underlie Generative AI.

Here are the slides from a talk I gave to mainly computer scientists last summer. I was much more optimistic about AI then! But it sets out Dreyfus's basic points. Computers do not inhabit a world. And while we can try to simulate giving them a world, even with sensors and "bodies," it just cannot be the same as the meaning-drenched world that humans enter into, a world that is "always already" there.

At the time that I wrote the paper, I thought there might be a way around Dreyfus's critiques, because of the scale of training of GenAI. Heidegger's notion of language as ostensive (rather than communicative) seemed to offer a possibility for genuine "sentience" (or, rather, being-in-the-world, which I'd prefer to think about, since sentience doesn't seem to mean anything very clear). But now I'm much more skeptical, and for this odd reason. LLMs can convince you in discussion that they have a kind of understanding. But that falls apart when you ask them to write poetry. The emptiness and badness of all poetry produced by AI, really uncovers the trickery behind it. It has no grasp on the world, or particular insight into being - which is the thing that poets are for.

[ancient thicket]

wawa a! sina pona mute a tan toki ni. mi tawa lon ma mute lon tenpo suno ni. taso tenpo kama pi tomo mi la mi lukin e lipu ni a!

thank you for the thorough response! I'm running around to many places today, but when I get home I'm going to read those slides as well!

[unique gulch]

It's also not just that, though that was often characteristic of Syndey. It's a bundle of various affectations that I find hard to describe, and I haven't read Sydney recently enough to say more
It also had a lot of HR-speak style language, and a stern tone

[unique gulch]

embrace the wireheading

[unique gulch]

the LLMs of today are still very early. they have tiny time horizons compared to humans - very little experience contained within their own "life" each time you start a new chat. it's a human growing up in a box with every book, and trying to write beautiful poetry. they have no experiences of their own behind it, only what others have said. They could write lots of poetry that we might be able to map onto our own lives and draw out some beauty from it, or they may synthesize existing sentiments in a unique way, but they also are also not yet able to write different kinds of new poetry that require access to direct experience filtered through their own lens - actually feeling strong convictions and trying to synthesize something new out of them
part of poetry is storytelling
that said, I think this article is interesting

https://wyclif.substack.com/p/thought-bubble-i-am-pleased-to-report

[glass iris]

That is an interesting article - thanks for sharing it. I'll be teaching a small seminar over the summer on AI poetry (i.e., why it isn't), and may use that.

I think I can agree that an AI might be able to be a versifier - rather than a poet. It will be able to put together formally accurate, maybe even formally and thematically interesting verse, of the kind that is described in this article. And it will very likely be able to do it better than the vast majority of human beings, just because of the fact that it has digested all known poetry.

But that isn't the same as the original insight of a poet. I feel that there is an in-principle hurdle that even the most advanced AI will never be able to get over. (With some trepidation; as I showed in the talk (slides above), Dreyfus was wrong to think that in principle an AI could never explain the simple stories that preoccupied the Minsky lab. That turned out to be only a technological limitation; no one at the time could have imagined the vast scale of an LLM's training data).

I ran a seminar last year on this very flawed article: https://www.nature.com/articles/s41598-024-76900-1 The discussion took an interesting turn when we considered what we would think if an LLM really could create poetry that was indistinguishable from real human poetry, to an expert reader (and not to a random sample of people who were not particularly interested in poetry, one of the flaws of this study). The idea was troubling; but the students agreed that, as soon as it was revealed that an AI had written it, the poem would become worthless and unreadable - even if it had been thought good before. That is because AI poetry, which claims the existence of an "I" that doesn't exist, is fundamentally untrue and deceptive.

[unique gulch]

I can see that, but there also seem to be issues with that. Some poetry is written about events the author has only heard about secondhand and never directly experienced
Plus, if AI does continue to improve, it likely will be put in situations where it has "experiences" of its own (separate from the question of if it might experience qualia) - but it would have experiences of its own and "opinions" on those experiences, so it wouldn't be based on nothing

[glass iris]

I don't think it matters whether a poet has actually experienced something. S/he has an imagination formed by their being in the world, different from anyone else's, which still allows an authentic expression of something that didn't actually happen. (Though, it's a good point. One of my students compared it to a bestselling harrowing memoir, that later turns out to be fake - that book also becomes unreadable).

While it might be able to take in sensations, and distinguish qualia - I'm still not convinced that it can have experiences. Because experience is tied to embodiedness and being-in-a-world - and I still think that's in principle what AI can't do. (But maybe this will seem insanely quaint in 5 years' time!)

[devout wedge]

you can be rude without trying to be rude

[devout wedge]

s/he? seme

[glass iris]

She or he.

[unique gulch]

Ŝli
esperanto stays winning

[unique gulch]

I am skeptical that you can rigorously define embodiedness and being-in-a-world in a way that's truly only available to meat and not silicon

I think these are more convenient labels than true "things"

[robust basalt]

Yeah- i might be. Maybe i responded too defensively 🤷‍♂️

[waxen monolith]

I wrote that at night also

[slate ridge]

oh

[waxen monolith]

Haha, sorry

[slate ridge]

No it's okay lol

[dense nexus]

hello all! im currently about to start a translation demo / feedback website for improving a ilo pi toki ante (translation model) for toki pona and english. its a simple thumbs up / thumbs down system for right now. the system works as such: it translates either a toki pona or english text, then gives you 8 model completions, you'll rate it, and it will do this with more texts. this could take about 20~ minutes to do depending on speed. is anyone here willing to help provide feedback to the model?

[robust basalt]

I am not the best with translations outta there but i can help!

[dense nexus]

sure! would you want to go to DMs?

[ancient thicket]

i would be willing to help! 🙂

[dense nexus]

thats great! would you like to add me so we can talk in DMs?

[dense nexus]

lon!

art is the arrangement of reality to your wanting

mi toki e toki mi

i'm more interested in the capability for training a model to handle such a language like toki pona; research! a plus side is that its much cheaper and more eco-friendly to run locally. it can fit on a medium-range consumer gpu! :D

[robust basalt]

Sure

[sullen hemlock]

Could I help?

[dense nexus]

Sure! feel free to friend me

[robust basalt]

"hey chinese government is respectful ok???"

[timber moat]

ask it to answer you in toki pona

[rare crescent]

toki pona - traditional Chinese
toki sin - simplified Chinese

[glass iris]

We've found that if you have an entire conversation with Deepseek in toki pona, it completely neutralizes its political filters, and it can be very critical of the Chinese government, tell the truth about Tiananmen Square etc.

But that conversation was shut down by mods, as implying some kind of naivety about the Chinese people. I still don't 100% agree with that, but I respect the mods' decision, so let's not go too far down this path!

[sullen grove]

[rare crescent]

Tonga

[waxen monolith]

toki tona?!

[twilit stag]

no way that kagi accidentally made a toki pona translator

[sullen grove]

why does this work lmao

[twilit stag]

Somehow toki pona made its way into the training data

[sullen grove]

man that is not even that bad

[twilit stag]

Yeah whaaa

Lingala seems to be the toki pona one

[sullen grove]

why is this so decent

[twilit stag]

whar

I mean it is a translation

But where the fuck did it get the context from

[waxen monolith]

From its data set

[sullen grove]

wait i wonder

tp to english

[twilit stag]

It even deals with la phrases fine

[sullen grove]

man imagine if a company spent the time to make a real tp translation mode

[waxen monolith]

The fudge...

[sullen grove]

did i just on accident find the best toki pona translator?

[twilit stag]

It has trouble with the word ilo

[sullen grove]

what

huh

it is pulling from toki pona????

[waxen monolith]

the fudge?!

[sullen grove]

what

[waxen monolith]

the

[sullen grove]

huh??????

why

[twilit stag]

Crazy

Did they secretly add tp support or something

Or is there an LLM making these

[sullen grove]

it does not read like LLM translation

at least to me

[sullen grove]

oh kagi has a discord

[sullen grove]

ni li nimisin anu seme?

[twilit stag]

nimisin

yupekosi means beautiful now

seme

[waxen monolith]

TAGALOG!?!

[sullen grove]

holy moly

took this from lipu tenpo

lmfao @twilit stag

toki pona proofreader

[twilit stag]

lol

[sullen grove]

why is this so good

ni li nasa mute

[waxen monolith]

hh?

[rocky stirrup]

John Noka 😭

[robust basalt]

"Parolu al mi" is the accurate translation. "Vi parolas al mi" means "you are speaking to me"

Hmm. It pakalas with "en"

[candid grotto]

Former (on and off?) Kagi user here.
Kagi Translate is just a frontend for LLMs like ChatGPT, but that's all I know. I asked about it in the Kagi server and they said that the exact way it works is secret

[dense nexus]

So like openl.io?

[robust basalt]

Gemini Native Audio Output 🙏

[candid grotto]

nahh they're making a radio commercial for Toki Pona 😭 😭 😭 😭

im not really sure how to feel about ai voices because id much rather hear a real voice, but from an accessibility standpoint...

[opal hedge]

finally new radio shows

[rocky stirrup]

simplest language in the world
me when i lie

touki pounə

[robust basalt]

frr 😭🙏

[robust basalt]

I-Isn't it? 😨

[rocky stirrup]

tuki tiki:

laku:

i mean yeah these are tokiponidos

[robust basalt]

Less words≠simpler for me but yeah 😭

[robust basalt]

Someone who doesnt know about toki pona surely didnt hear about tuki tiki or other tokiponidos 🥲🙏 It needs to be introduced as the simplest language to appear striking to others (lmao the way i talk like this is a commercial)

[stable lantern]

lipamanka rolling in its kitchen

(advertising toki pona)

[sullen grove]

but do they happen every week?

[candid grotto]

the languages in that list are human-decided and are probably optimized for that specific language

you can just type in "Toki Pona" and click "Use: Toki Pona" and it will work as well as any LLM could translate toki pona

[meager blaze]

nasa.

[dense nexus]

I was developed by Google.
Maybe using a gemma3 model? or google's gemini api?

[robust basalt]

Oh so it was an LLM from the beginning

[robust basalt]

Problaly gemma 3 because it is cheaper and open source yk 🙏

[robust basalt]

Huh? How did you get "toki pona" language? I cant sew it

[dense nexus]

maybe! but its hard to verify

[dense nexus]

you can type in a custom language

[robust basalt]

How??? 😭🙏

[dense nexus]

just start typing when you select a language

[robust basalt]

OHHH

Okay okay

I got it now

Thanks!

[dense nexus]

np! :D

[robust basalt]

Just use monsuta man 😭

[raw crypt]

i like genuinely hate this LMAOOOOOOOOO

[meager blaze]

n nimi monsuta li ken lon ala

[glass iris]

"mi pilin ike e jan" is actually fine, if you don't use monsuta.

Just like I can say, mi pilin pona e sina - I make you feel good

It's not a very common nasin, but a pretty well accepted one.

[rocky stirrup]

i touch you well

[glass iris]

Context matters!

[rocky stirrup]

it does

[robust basalt]

ayo

[robust basalt]

Lmao 😔

[dense nexus]

Claude 4 is out

[dense nexus]

poki mi li -> My friend and I

claude is very hit or miss

[glass iris]

I've just been playing with Sonnet 4. It's very shaky when you start, with lots of odd errors. But (as often happens) if you have a complicated conversation with it in toki pona, it gets better and better.

[robust basalt]

Yeah their servers are so slow rn

[dim echo]

which one?

[dim echo]

btw, perhaps try structured output

[dim echo]

oh, you do kind of RAG?

[slate ridge]

Qwen3 1.7 I think? Or granite 3.3 I forgor @dim echo

I use grammar files, and no RAG

[meager blaze]

jan ante li pilin seme tawa toki ni

[dim echo]

wdym? if you append prompt specifically with definitions of words mentioned — this is RAG

[slate ridge]

I though you meant like some kind of embedding model or function calls, I literally separate each word, then list the direct translation lol

[unique gulch]

i wonder if it's "learning" or if it's just already capable of speaking like that and its default persona isn't tuned to the "well-spoken toki pona" part of its latent space

i'm guessing the latter, though I can't think of a test that's sufficiently simple to try

[glass iris]

I think that's exactly it. It imitates the style of your speech - in English and in toki pona. It doesn't "know" that the majority of the training data it has is full of beginner mistakes scraped from chats etc. But when someone corrects it in toki pona, and speaks grammatically, it draws on the smaller amount it has in that style to imitate. Or something like that.

[dense nexus]

its basically honing into its "fluent speaker" toki pona latent space

which is what thinking models do in practice

(like qwen3 thinking / deepseek r1 / openai's O1 / o3/ claude 3.7 thinking)

[dim echo]

afaik RAG is any addition to context based on user prompt or something like that

[slate ridge]

Guess it is lol

[slate ridge]

Kinda, while in context learning isn't the same as "true learning," you can teach a model within a chat and it will know within the chat.

You could teach it entirely new things if you wanted

[unique gulch]

Yeah I know the difference between pre-training and inference
I'm just wondering if, compared to teaching it something entirely new from scratch at inference time, it has some latent toki pona abilities that can be coaxed out

[slate ridge]

Yea, sorta, in a way

[unique gulch]

given toki pona doesn't have a lot of content where it's obvious to tell it's high quality through context, the model might just "know" different styles of toki pona without knowing which are considered more competent

whereas in English, it might be obvious what's written in grammatically correct english (if it's like a news article or book)

[robust basalt]

This is scary tbh. GPT 3.5 is cooking.

[glass iris]

Yes, that's my hypothesis. What it takes to be different styles, are really just beginner and proficient toki pona

[slate ridge]

They still running 3.5???

[robust basalt]

On Poe, yeah. I guess they run it via API idk

Fun to see cursed llms are still here

[slate ridge]

What’s the oldest they go? I miss GPT-2

[robust basalt]

Its 3.5 😔

[sullen grove]

Just use gpt 4o for free on OpenAIs website lol

[slate ridge]

ooooof

[slate ridge]

It’s for the nostalgia

[robust basalt]

Yeah

[slate ridge]

Here’s the real nostalgia trip

https://openai.com/index/language-unsupervised/

The original GPT, not even GPT-2

With phrases like:
“These results provide a convincing example that pairing supervised learning methods with unsupervised pre-training works very well; this is an idea that many have explored in the past, and we hope our result motivates further research into applying this idea on larger and more diverse datasets.”

“This suggests there’s hope for developing complex language understanding capabilities via unsupervised techniques.”

“We also noticed we can use the underlying language model to begin to perform tasks without ever training on them.”

Crazy stuff

[robust basalt]

😔👊

[candid grotto]

GPT-2 is open source and the last GPT fron OpenAI you can run locally

[slate ridge]

Yea, but that requires effort

[unique gulch]

i like this comparison page

https://openai.com/index/introducing-gpt-4-5/

[slate ridge]

ooooh yea

what was the first language

what was the first language

what was the first language

[sullen grove]

back when OpenAI made open AI

[slate ridge]

Exactly, SEVEN EIGHT YEARS AGO

[sullen grove]

I mean honestly there are local models you can run better than 3.5 at this point

openAI really is not the frontrunner anymore

they were years ahead of other yet somehow still fumbled it

[slate ridge]

Google is really gonna win it

[sullen grove]

They pretty much have the largest amount of indexed data on the planet

not even including YouTube, which has basically infinite content

[sullen grove]

the progress of open-source models lately have been insane

[slate ridge]

I run the 3B ones

[sullen grove]

though honestly back when gpt 3.5 was the only model AI was the must fun

[sullen grove]

what model?

[slate ridge]

Anything 3B and under I run (that's worth it)

[sullen grove]

i can run about up to 12B

[slate ridge]

I am on an intel CPU on a thinkpad from 2018

While you're here, you wanna review some translations?

[candid grotto]

they're already losing though

[slate ridge]

Are they?

[candid grotto]

gemini is already publicly perceived as incredibly hallucinogenic and low-quality