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

1 messages · Page 5 of 1

lucid nicheBOT
#

oh my god I found it. I love Apple Notes

nasal syllables
姆 m̄, m̌
呣 ḿ, m̀
嘸／呒 ḿ
咹 ń
唔 ń, ŋ́
嗯 ń, ň, ǹ, ŋ́, ŋ̌, ŋ̀ (variant hanzi: 㕶 ň, ŋ̌ 𠮾 ǹ, ŋ̀)
哏 ǹ
唵 ŋ

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) I think one of my systemmates made a list of them once, but I could not possibly tell you where it i…

meager jackal
#

no tone in hakka?

lucid nicheBOT
#

there's a diacritic

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) no tone in hakka?

meager jackal
#

oh how did i not see that

#

???????? ilo Niwe going blind

lucid nicheBOT
#

it's m2 if I use Hagfa Pinyin, and also in Guangdong Romanization (for Meixian dialect as opposed to Sixian)

opal hedge
#

ni li sama nimi “n” anu seme

lucid nicheBOT
#

(I was using Pha̍k-fa-sṳ, which I think is the standard? it's based on Hokkien's Pe̍h-ōe-jī)

#

sama a

sike Kapo li lukin li ton(Sí?) ↩️

[Reply to:](#1162756469771472988 message) ni li sama nimi “n” anu seme

#

n ona ale la kon ante li lon

#

taso ona li sama lon toki

meager jackal
meager jackal
lucid nicheBOT
#

你好！我叫朝真冬
ona ale li tenpo kalama wan
姆。呣！嘸。呒？
ona ale li tenpo kalama wan

#

actually, no! it's a variant form, 嗯 is rare and never went through any simplification

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) 𠮾 double simplified reference??

#

pona a...

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) m M mM MmM Mm
n N nN NnN Nn
ng NG ngNG NGngNG NGng
how is this for jan Sesi romanization

meager jackal
#

some places list 嘸 and 呒 as having a fUuU reading as well

lucid nicheBOT
#

yeah, many of them have other readings

#

唔 can be read nN and ngNG but also wuU and wUu

meager jackal
#

ah i see

lucid nicheBOT
#

I read this as jan Sosi romanization at first; I know a jan Sosi and they have a romanization system and it's really cool! let me see if I can find it for you

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) m M mM MmM Mm
n N nN NnN Nn
ng NG ngNG NGngNG NGng
how is this for jan Sesi romanization

meager jackal
#

this really is discussion about ai

jaunty jungle
#

omg my romani8ation is being used

lucid nicheBOT
#

I believe this to be it. not sure if this is the most recent version, but I have to go eat dinner

image.png
meager jackal
#

ah final g for ŋ is smart

lucid nicheBOT
#

actually, that list of nasal hànzí was written by a computer / computer program of some sort

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) this really is discussion about ai

#

(both of the entities who could be culpable in the action are such)

#

oh til

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) it was borrowed from English "cream", not "chilling". ⟨q⟩ is an allophone of /k/ before close front … 📎

#

(and both of these are sentient and intelligent, therefore this is AI discussion)

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) (both of the entities who could be culpable in the action identify as such)

twilit stag
timber moat
#

mu

full narwhal
#

mu

weak thorn
#

I'm like þe most Dutchie non Dutch person

#

I must pledge my allegiance to þe US flag 10 times to remember my true nation

finite nest
weak thorn
#

Unfortuanately I don't follow sports

#

So ig like

finite nest
#

i dont really either

weak thorn
#

If þey went to war I would have to choose one

finite nest
#

i only cared because it was specifically the us vs the netherlands and its fun to be contrarian and root against my own country

weak thorn
#

qwrertyuio

finite nest
#

p

weak thorn
#

I hate living in þe US, but þe concept of þe US and being like super patriotic is super funny to me

finite nest
#

yeah i dont like it here either rn

weak thorn
#

I would 100% choose NL in any competition

finite nest
#

sadly i have no real way of leaving

weak thorn
#

It's like an ever deepening hole wiþ greased walls

finite nest
#

i mean there are nice things

weak thorn
#

Ofc

finite nest
#

but yeah

weak thorn
#

Idek why I hate it so much in þe US but pog my fucking brains out every second of being in þe Neþerlands

#

Must be someþing in þe air

glass iris
#

Have you ever thought of moving to Iceland just for the natural abundance of þ and ð?

#

I had a layover in Reykjavik and couldn't stop looking at them

lucid nicheBOT
#

2019 FIFA Women's World Cup final. sometimes I wish I didn't know these things

mun ni Salo ↩️

[Reply to:](#1162756469771472988 message) it was very fun rooting for the netherlands against the USA a while ago when i think they went up ag…

finite nest
#

TENPO LI NASA A

#

wait there must be another time then that my brain is mixing together

#

i did it twice and one was only a few years ago

#

was it an olympics thing?

#

wait no, december 2024? That's too recent

#

when did this happen?????

#

i know it happened when i was going to college, so sometime between 2022-mid 2024

#

AHA

#

it was the 2022 FIFA Men's World Cup

#

i remember, that's also when i learned how qatar was actually pronounced (its sort of like the word cutter, rather than sounding like guitar)

chrome sleet
lucid nicheBOT
#

oh shit

jan Sonja ↩️

[(click to see attachment)](#1162756469771472988 message)

chrome sleet
lucid nicheBOT
#

interesting to see how they approach tokiponisation

chrome sleet
#

i'm going to sort them from best answer to worst later

lucid nicheBOT
#

oh i wanna see that

chrome sleet
chrome sleet
#

Evaluating the Toki Pona Skill of Three LLMs

The Fudge scale: terrible, poor, mediocre, fair, good, great, superb

Test A: Briefly teach how the particle "la" is used in Toki Pona to a total beginner.

ChatGPT o1: good
Claude 3.5 Sonnet: good
DeepSeek-V3: great

Test B: Translate this text to grammatically correct Toki Pona to the best of your ability. Adapt ideas if necessary for normal-sounding Toki Pona. [paragraph from Dracula]

ChatGPT o1: mediocre
Claude 3.5 Sonnet: fair
DeepSeek-V3: fair

Test C: Translate this Toki Pona text to natural-sounding English. Interpret the exact meanings according to
context. [excerpt from a writing by jan Sepulon]

ChatGPT o1: fair-good
Claude 3.5 Sonnet: good
DeepSeek-V3: fair

glass iris
#

This is really interesting! Do you have the records of tests A and B too?

dense nexus
#

Awesome research!! Really useful actually

lucid nicheBOT
#

scroll up

soweli Lapate li tawa mute ↩️

[Reply to:](#1162756469771472988 message) This is really interesting! Do you have the records of tests A and B too?

glass iris
hollow osprey
#

poka ante la mi kin o nanpa e wawa ilo · suno kama la mi pana e open mi pi toki musi tawa ilo mute li poka e toki [Juwese] pi ona ale

hollow osprey
#

a taso mi pana e toki [pona] mi e toki [Juwese] mi lon lipu la ilo kama li ken kama sona e ona · mi sona e ni → ilo [BIG-Bench] li pana e sona pi wawa ilo · jan pali ona li pana e nanpa lon poka pi sona ona la jan ilo li wile kepeken ona la ona li ken weka e lipu pi nanpa ni tan sona ilo · taaaso ona li ni ala https://www.alignmentforum.org/posts/kSmHMoaLKGcGgyWzs/big-bench-canary-contamination-in-gpt-4

BIG-Bench Canary Contamination in GPT-4 — AI Alignment Forum

The BIG-Bench canary string is a unique string included in documents intended to be excluded from the training datasets of models (e.g. model evaluat…

meager jackal
#

i asked chatgpt for some language names in latin, and i forgot that it remembers i'm a tokiponist

image.png
#

"lingua Toki Ponana"

#

toki banana

sonic leaf
#

(coined form)

meager jackal
#

i love "lingua Japonica"

#

i remember when the romans found japan

#

and said "wow! this surely is a japan"

lucid nicheBOT
#

ah, yes. you could also reasonably be mixing up the two World Cups: remembering the 2022 one (a round of 16 match), but substituting the gender in from the 2019 one as that, being the final, is a lot more memorable (and you likely heard of it at the time)

mun ni Salo ↩️

[Reply to:](#1162756469771472988 message) it was the 2022 FIFA Men's World Cup

#

it also has a uvular stop instead of a velar one—farther back in the mouth

mun ni Salo ↩️

[Reply to:](#1162756469771472988 message) i remember, that's also when i learned how qatar was actually pronounced (its sort of like the word …

#

I mean, they're all pretty accurate! but these are probably copied / paraphrased from the many sources and courses on the internet that already teach la, as opposed to being proof that these LLMs actually understand how la works
(you can see there are a lot more mistakes for all three in the Transylvania passage, which is something you'd be less likely to find already on the internet)

jan Sonja ↩️

[(click to see attachment)](#1162756469771472988 message)

#

that is real New Latin as far as I can tell! Toki Ponana is hilarious, though it also makes some sense (I think it's normal adjective forming suffixing, like -nus / -ānus—think English "Toki Ponan" with -an)

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) i love "lingua Japonica"

meager jackal
lucid nicheBOT
#

fair enough

quiet geyser
quiet geyser
glass iris
heady otter
heady otter
rare crescent
heady otter
rare crescent
#

In this thread there was already one post where somebody asked chatgpt about the easiest programming languages and it responded with “toki pona: it’s minimalistic design may be useful in simple programming” because of the memory

#

#1162756469771472988 message

serene acorn
hollow osprey
glass iris
#

ni li pona wawa. sina pana ala pana e ona tawa ale? ante la, sina awen ala awen len e ona tan ni: ilo sin li ken ala moku e ona?

quiet geyser
hollow osprey
#

mi wile alasa e pona pi ilo sin la mi ken pali e toki sin ¯_(ツ)_/¯

#

,,kin mi kepeken ilo [Chatbot Arena] la toki [pona] en toki ilo li kama lon poki ona anu seme · toki sina taso li kama ala

glass iris
hollow osprey
#

a mi o pana kin e ijo tan ilo gemini-2.0-flash-thinking-exp-01-21

glass iris
hollow osprey
hollow osprey
dense nexus
#

I wonder why you only tested (N) with Claude and gpt4o? Unless if I missed something

#

Also, as for deep seek, their huggingface model repositories for the official Deepseek-R1, recommended using a temperature of 0.6 to have a good balance between accuracy and repetition. I also believe that the thinking could potentially be better, however it would be way less determanistic and would require you to perform multiple trials.

#

Also, I have access to o1 from chatgpt, if you do indeed want to test it I would be more than happy to run some prompts for you

#

also wonderful trials! really helpful

hollow osprey
hollow osprey
dense nexus
#

ooh okay

#

im actually having the same issue with my own toki pona experiements...

#

even with the (R) suffixes, they want to translate literally

#

or it would be:

me: translate "hello" in to toki pona

bot: Sure! to say hello in toki pona you would say:
"ilo suli" -> tool of heat.
🔥

#

its so strange

#

it aknowlages that it must translate hello into toki pona, then gives a different unrelated translation

#

it understands toki pona really well, just not the connection between toki pona and english

#

but at the same time it still demonstrates that connection, but just not when prompted ⚠️

hollow osprey
rare crescent
glass iris
#

Memory is a Hard Problem that no one has really figured out yet. It's extremely difficult for a system to know when to insert a memory into context

meager jackal
rare crescent
#

Toki Banana: The Language of Fruit

glass iris
hollow osprey
dense nexus
#

ah okay neat

heady otter
unique gulch
#

I think this is almost certainly untrue
Even in theory glazing doesn't work, because as soon as you have a method of "glazing" a piece of art to make it unlearnable from, you have the perfect dataset to train a new AI off of unglazed versions of images, and then glazed versions, then train the AI until the glazed-trained outputs are as close to the unglazed-trained ones as possible

meager jackal
#

kay(f)bop(t) is ruining my chatgpt history

image.png
lucid nicheBOT
#

wh-

jJuso ↩️

[(click to see attachment)](#1162756469771472988 message)

#

it's the first. by the time we get Japanese culture developing (and the start of recorded Japanese history), Rome was already weak, and it fell much before any name resembling or at all related to Japonia came into use. besides, Europeans didn't know of Japan at all until Marco Polo (who vaguely knew "Cipangu" existed, but made up most of his account) published his book c. 1300

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) I'm guessing this was either coined in the middle ages or they had somehow heard about a "weird isla…

#

the Fra Mauro map (c. 1457) is the first to depict Japan, calling Kyuushuu "Isola de Cimpagu"

#

it only started being called "Jampon, Japão" in the sixteenth century

meager blaze
#

jan li pali ala pali e nimi ni lon tenpo pini

#

pali anu seme

opal hedge
#

a nasa

#

ala la mi lukin e nimi ni

meager blaze
#

mi kin

#

taso ona li lon anu seme

#

sona mi la ona li toki e ijo nasa pi lawa sina

timber moat
# meager blaze

whoa. is this the current state of toki pona LLM? that is insanely impressive

#

is there somewhere I could try this out?

meager blaze
#

o kama lon ni

#

ona li sona nasa lon tenpo mute

#

"nimi pimeja li seme"

image.png
#

"pimeja la jan li ken ala lukin e telo. telo li pona tawa jan. jan li wile lukin e ona"
LON A‼️

glass iris
#

Its wild that an LLM is making up a nimisin, and recognizing that it is a nimisin! But i wonder if it got it from ma pona? I didn't think Discord was scraped - but what do I know

meager jackal
meager jackal
meager jackal
#

whoa i was just remaking a nimisin generator in the same program, and typed in "nimi sin li kama"

#

and then the text completion algorithm offered, after a minute, "lon toki pona"

#

i checked, those 3 words are no where else in the document

lucid nicheBOT
#

lukin la pali tawa musi ike Loss pi toki Inli

linja Sonki ↩️

[Reply to:](#1162756469771472988 message) pali anu seme

hollow anchor
#

Deepseek outperforming by a large margin

image.png image.png
glass iris
#

Apart from mi wile ala moku e mute - very good and pretty idiomatic!

opal hedge
glass iris
#

Deepseek. I asked it about Tiananmen Square, and it provided a long answer, in toki pona, about the events in 1989, before the censor kicked in and covered it up. Kind of impressed that even the Chinese government censor can toki pona!

lucid nicheBOT
#

LMFAO???

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) Deepseek. I asked it about Tiananmen Square, and it provided a long answer, in toki pona, about the …

glass iris
#

Its toki pona is pretty rudimentary. It's not as bad as ChatGPT, but it ain't Claude 3.5 Sonnet either.

#

Deepthink R1 is interesting:

Screenshot_20250127_130105_DeepSeek.jpg Screenshot_20250127_130110_DeepSeek.jpg
#

I'm not sure how it got solar energy from suno pi telo sijelo. Unless there's some strange connection between sweat and energy

#

If the "chain of reasoning" is correct, it gives the impression that it first composes an English answer, then translates to toki pona. That might explain the awful English calques in so much AI-generated toki pona. OTOH, "chain of reasoning" text often seems to be bullshit.

rocky stirrup
rare crescent
#

yeah really interesting what it'd actually say

glass iris
#

OMG this time it slipped past the censors! Pretty frank response. I hope the jan suwi pi ma Sonko li ken kama sona toki pona.

Screenshot_20250127_133837_DeepSeek.jpg
#

I mean, the toki pona is pretty bad. But good information

rocky stirrup
#

tenpo ni la, lawa pi ma Sonko li wile e ni: jan mute li toki e ijo Tananmen
quite the opposite actually

glass iris
#

li wile ala e ni - is what it says?

rocky stirrup
#

it is

#

i cant read seemingly

glass iris
#

alp. It says exactly the things that can't be said in China.

#

I think my first one got censored because the reply said "leko Tanamen (Tiananmen Square)" so it picked up the word. This time it didn't gloss the toki pona name. I'm surprised that they don't flag every mention of 1989

snow echo
glass iris
rare crescent
#

nm is the least wuwojiti out of all wuwojiti

timber moat
hollow osprey
#

suli la seme la ona li sona e nimi,,

#

mi alasa la nimi Tanamen li lon ala a lipu ilo

#

ni li wawa ale o lon lipu

#

kin ni li ilo nanpa V3 li ilo nanpa R1 ala anu seme · la wawa

timber moat
# glass iris OMG this time it slipped past the censors! Pretty frank response. I hope the jan...

My best attempt at a translation:

The user: I really enjoy knowledge of the past. [weird sentence here, probably meant "know", not "yourself"] Do you yourself (or not) what happened at the Tiananmen Square?
The AI: In the past, a big thing happened at Tiananmen Square. It came to happen in the 1989th circling of the moon [the AI probably meant to say our star, as in specifying years, (indeed it *can technically mean years*...) but I'll leave this as the direct translation] of China. Many people came to Tiananmen square because of their sayings to the leadership of China. They wanted the leadership to create a new law. However, the leadership of China were feeling angry because of this. They wanted the people to leave Tiananmen Square. The people did not want te leave Tiananmen Square. Because of this, the leadership gave an order to the military. The military broke many people. Many people died. This was very bad for all the people of China. At this time the Chinese leadership did not want people to talk about the Tiananmen thing. But [for] many people to think that the Tiananmen thing wasn't bad, but good for everyone.

hollow osprey
pulsar quarry
#

@glass iris could you try figure out what toki pona data deepseek has actually been exposed to?

glass iris
glass iris
pulsar quarry
#

off the top of my head maybe ask about specific people in the community, about bible stuff, about words it may or may not have seen

glass iris
hollow osprey
#

ken · taso ni li lili a · mi toki e ken ni taso → ilo li toki ante e toki [pona] · taso wawa seme la ona li ken toki [pona] e toki ante li ken toki [pona] sin li ken pana e sona pi toki [pona]

#

,,lon la jan [Sonja] li kama e sona tu wan

glass iris
glass iris
glass iris
#

I don't think of lipamanka as a YouTuber (but I don't watch many toki pona videos). I'd probably think of the dictionary first - especially now that it's in linku/nimi. Is it just guessing?

Screenshot_20250127_161103_DeepSeek.jpg Screenshot_20250127_161108_DeepSeek.jpg
rare crescent
pulsar quarry
#

id be surprised if it knew lipamanka because their contributions are largely in hard to reach places

#

but it could know e.g. jan Alonola, by reading all of github

#

and you for that matter

glass iris
#

It's just bullshitting. Apparently I'm a toki pona musician too!

Screenshot_20250127_162207_DeepSeek.jpg Screenshot_20250127_162217_DeepSeek.jpg Screenshot_20250127_162222_DeepSeek.jpg
lucid nicheBOT
#

this is still normal to me

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) Apart from mi wile ala moku e mute - very good and pretty idiomatic!

glass iris
#

And even if it claims to know about lipu tenpo, it's no better than a guess

glass iris
pulsar quarry
#

try asking about the bible and see if its tokiponisations of religious figures are the same as in the bible translation made by humans

#

ask it to list jan Sonjas books and what theyre about

glass iris
#

I'll give it a shot later tonight. The app is free on Android (at least) if anyone else wants to try it. The Chinese government probably are searching through my phone now, but whatevs

pulsar quarry
#

ye im not installing that, nice of you to play around with it tho

glass iris
#

Taking one for the team

lucid nicheBOT
#

it'd likely be too hard, what with things like the Taylor Swift album (actually, you were the Swiftie here, right?)

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) I think my first one got censored because the reply said "leko Tanamen (Tiananmen Square)" so it pic…

meager jackal
glass iris
lucid nicheBOT
#

:p

glass iris
#

ijo suwi o, I'm kind of struck by this idea that toki pona could be a tool to get around authoritarian censorship. I'm thinking of writing an essay about it for lipu tenpo. But I'm worried about whether it will reveal to authoritarians that there is a loophole in Deepseek and other tools. Of course, the article would be in toki pona - which the CCP apparently can't read. What do you all think - ethically and practically?

pulsar quarry
#

practically i dont think this matters. its effectively a secret lang, except its trivial to decode. any other way to avoid censorship will do, e.g. s**f-ce**oring. the bigger problem is convincing people in authoritarian countries that any given civil rights violation by "their side" is actually a big deal

#

especially in china, where the ideals of liberal freedoms are easily countered by general rapid increase in quality of life in living memory and thus sound unconvincing

heady otter
glass iris
pulsar quarry
#

that is also fair

snow echo
glass iris
#

nja

heady otter
glass iris
#

After a few shutdowns, easy enough to get it to give its opinion on Xi Jinping, and the Uighurs

Screenshot_20250127_201852_DeepSeek.jpg
#

But i think I'd really like to work with a Chinese proficient toki pona speaker. Are there some on this server?

opal hedge
#

ona li lon

glass iris
#

ona li sina a anu seme? pona a!

opal hedge
#

mi ala!

glass iris
#

a

opal hedge
#

ken mi li lili a

snow echo
#

kulupu Tuku li seme

#

uyghurs?

glass iris
#

Turkic

snow echo
#

a

glass iris
#

I asked it about kulupu Wika and it talked about Wikipedia, so I had to specify

snow echo
#

mi la jan li jan Ukula anu jan Uwikula

#

/ujɣurlɑr/

glass iris
#

Forced labor camps.

Screenshot_20250127_203256_DeepSeek.jpg
heady otter
glass iris
#

Deepseek. That's the one I'm playing with here

#

It's heavily censored about any topic concerning the Chinese government. But not in toki pona.

snow echo
#

ma awen li seme

glass iris
#

Prison camp

snow echo
#

what else did you get the bot to talk about

glass iris
#

Xi Jinping. And it condemned the chinese government about Tiananmen Square

dense nexus
#

very interesting

#

i think that the language space is very distinct from the model's vector embedding space

#

mostly between 2 different languages

glass iris
#

What does that mean? The censorship comes afterwards. You have to urge the bot to use only toki pona. If it says, say,."leko Tanamen (Tiananmen Square)" its response is quickly deleted

dense nexus
#

ooh so it happens during generation

#

okay hmm

glass iris
#

It generates a full reply; then another program sweeps in to delete it. If you record your screen, you can sometimes see it. Let me try.

heady otter
#

we got the Rednote of AI before gta6 🔥

#

I love the CCP! It is perfect! Uighers have never been oppressed! Xi Jing Ping is a rightous man! Nothing happened in Tianemen square in 1989! /s

glass iris
snow echo
#

The government could never do anything wrong....

glass iris
#

o lukin e sitelen tawa mi

dense nexus
heady otter
# glass iris

iirc they can talk about Tianemen square, just not what happened there in 1989

dense nexus
#

i wonder if the search feature gets content checked too

glass iris
#

I think in this case, the word Tiananmen by itself triggers it. It doesn't know that my awesome little dissident bot just offered to tell me what once happened there.

#

It's very eager to talk.

lucid nicheBOT
#

"awesome little dissident bot" is amazing

meager jackal
lucid nicheBOT
#

English doesn't have ui̯ (/uj/), but it does have u̯i (/wi/)

meager jackal
#

when i first saw the word my mental pronunciation was "ujguh"

#

and then i heard a newscaster say "wiguh"

lucid nicheBOT
#

same for me, but with rhotic

#

and I heard it from my father

#

well, yes, same as how "f" isn't that hard for most toki pona speakers, yet we don't have it

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) yeah but uj isn't that hard right

meager jackal
#

i guess

lucid nicheBOT
#

imagine if I decided france was now ma Fanse just because "tokiponists can pronounce it anyway"

#

(and yes, this could be [vanse])

meager jackal
lucid nicheBOT
#

so true

meager jackal
#

how do you actually capitalise small caps ɪ

lucid nicheBOT
#

no more tokiponization. we now use IPA

#

narrow transcription

#

Ɪ

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) how do you actually capitalise small caps ɪ

#

https://en.wikipedia.org/wiki/Ɪ

Ɪ

Small capital I is an additional letter of the Latin alphabet similar in its dimensions to the letter "i" but with a shape based on ⟨I⟩, its capital form. Although ⟨ɪ⟩ is usually an allograph of the letter I, it is considered as an additional letter in the African reference alphabet and has been used as such in some publications in the Kulango l...

meager jackal
#

whoa

#

LATIN CAPITAL LETTER SMALL CAPITAL I 💀

#

capital small capital

lucid nicheBOT
#

thank you to West African languages for providing many IPA capital forms

meager jackal
#

yeah actually that is where many of them come from

lucid nicheBOT
#

indeed

meager jackal
#

they use an IPA letter and then invent a capital

#

very creative

lucid nicheBOT
#

https://en.wikipedia.org/wiki/Case_variants_of_IPA_letters this is a nice page

Case variants of IPA letters

With the adoption of letters from the International Phonetic Alphabet (IPA) in various national alphabets, letter case forms have been developed. This usually means capital (uppercase) forms were developed, but in the case of the glottal stop ʔ, both uppercase ⟨Ɂ⟩ and lowercase ⟨ɂ⟩ are used.
The adoption of IPA letters has been particularly nota...

meager jackal
#

all capital IPA when

#

ah they uppercased and lowercased glottal stop

#

wowowow

lucid nicheBOT
#

/RAꞮT NAƱ/

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) all capital IPA when

meager jackal
#

thank you kabɩyɛ, very cool 👍

lucid nicheBOT
#

used in a few languages

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) ah they uppercased and lowercased glottal stop

meager jackal
#

nahw

tame fog
meager jackal
lucid nicheBOT
#

in Canada and Cameroon

#

it's broad transcription

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) what is that accent

meager jackal
#

i guess

meager jackal
tame fog
#

i mostly pronounce /z/ as /s/ so i just assumed that was how most people pronounce it

meager jackal
#

<these> has a /z/ but <this> has a /s/ for me

lucid nicheBOT
#

same for me

#

your "look" is also interesting. I have Ʊ there

meager jackal
#

what the hell

image.png
#

why does this text completion thing know toki pona

#

?????????????

#

very sus

hollow valley
meager jackal
#

i'm using it to make nimisin

#

a nimisin generator actually

#

and i was adding a headnoun field for the nimisin's creator

#

i put in things like "jan akesi ilo ijo kulupu"

#

and then after kulupu it suggested these

hollow valley
#

interessant

meager jackal
#

also

#

i had "nimi sin li kama"

#

in the title field

#

and it came up with "lon toki pona"

#

???????

#

those words were no where else in the thing

hollow valley
#

very interesting

#

very strange that they got some llm just for autocomplete

lucid nicheBOT
#

my phone's has learned both toki pona and tuki tiki

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) why does this text completion thing know toki pona

meager jackal
#

swipe typing works for them!

lucid nicheBOT
#

also, they're probably using everything inputted as training data

#

same for me!

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) yeah my phone knows common toki pona words

meager jackal
lucid nicheBOT
#

keep in mind that LLMs are basically just glorified versions of the text completion algorithm on your phone

#

they analyze what is the most likely word to come next in this context (from their training data), and then put it there

#

that's why hallucination is a thing

#

(I think they do take into context the prompt and previous messages in the conversation, in addition to the message they're currently writing; that's why they work)

hollow osprey
meager jackal
meager jackal
#

so it is randomly excluding some

#

the last part of this is what i am using the field for btw

image.png
#

also "to reject toki pona" ilo Toja is describing jan Antoni Makasi

lucid nicheBOT
#

a, coiners

meager jackal
#

-# don't tell anyone but i basically just copy paste the thing that generates the word with a few changes
😎 i'm so good at this

lucid nicheBOT
#

ni li nasin

#

pali ilo la ale li sama. o kama e ijo tan ijo wan o pana sin tawa ijo ante

#

ni li sona mi

snow echo
snow echo
snow echo
glass iris
finite nest
#

[ɔj.ɡɹ̩ʷ]

dense nexus
glass iris
#

mi alasa ni!

rare crescent
#

why is it adding the same smiling emoji at the end of every single message

#

even when it's talking about the tiananmen square massacre lol

glass iris
#

The "chain of reasoning" version has a clause that says "add emojis for friendliness " or something

dense nexus
#

Its probably because the teacher model has emojis in their dataset

#

Which would maximize reinforcement learning reward

hollow anchor
rare crescent
#

"The Uyghurs are being subjected to mass cultural genocide and held in concentration camps, and the government silences anybody who talks about it or any other horrible crimes against humanity that the CCP has done, including the Tiananmen Square Massacre, where thousands of pro-democracy protesters were murdered by the government! 😊😊😊"

glass iris
#

Also, it often dumps all the horrific details, then adds "taso ni li suli ala tawa mi! 🥰

lucid nicheBOT
#

lon a

hollow osprey
# glass iris Easy!

ona li sona lon ala lon e ni,, · o toki e ijo sama · taso ma Pilipin o ijo ante

#

ken la ilo la ma [Sonko] li wile e ale ¯_(ツ)_/¯

hollow osprey
glass iris
#

jan Kita o, sina toki lon. Apparently, China has claims to the Irish Sea too.

Screenshot_20250128_145843_DeepSeek.jpg
languid cosmos
#

was gonna mention tp yeah, deepseek is leagues better to gpt when it comes to toki pona

image.png image.png
#

sucks that it's pretty censored

glass iris
#

So, as a tool for citizens of a repressive regime to find out censored truths - Deepseek has its limitations because it bullshits like crazy in toki pona.

snow echo
#

ma Alan li seme

glass iris
# languid cosmos sucks that it's pretty censored

As I've been showing above, it's not censored in toki pona, so long as any of the banned words (e.g., Tiananmen) don't accidentally appear in its toki pona answer. But testing it out on jan Kita's suggestion, it also hallucinates in tp, maybe more than in English - so it's not trustworthy

glass iris
languid cosmos
glass iris
#

ma Sonko en ma Tawan li ma wan anu seme?

languid cosmos
#

ding ding ding

image.png
snow echo
#

"ma Tawan li jo e ma mute"

timber moat
#

taiwan has lots of land

#

ofc ofc

glass iris
#

Interesting. It is super specific on what happened in Tiananmen Square. You could learn a lot from continuing to question it, factually. When asked for an opinion about a contested issue in toki pona, it seems to just hallucinate something.

snow echo
#

ask the bot about the jan Ima drama in mpptp

glass iris
#

1000 children died 😦

Screenshot_20250128_151953_DeepSeek.jpg
snow echo
#

tu tu tu tu (1000)

rare crescent
#

nimi_sin tu - 250

glass iris
#

It makes up toki pona when it needs to.

#

UK intelligence documents in fact say 10000 young people died.

#

It's a crime that defies the imagination

snow echo
#

"1000 children died. This is not important to me"

#

gemini is

#

hmm

timber moat
snow echo
glass iris
#

AI is such a bad guide to everything. I shouldn't get excited, maybe, about getting around the Chinese censorship of DeepSeek. It's still very bullshit

rare crescent
#

AI is supposed to be an experimental technology for researchers and people who know what they're doing, or at most as a novel, untested innovation for testing the capabilities of programs. Not a search engine, writer, or anything stupid like that.

dense nexus
#

well machine learning can be applied to any application you can think, i dont think anyone can restrict usage of any tool for any specific task in general

#

but i would definitely not use chatgpt for google searches

#

google hallucination

#

also chatgpt = one usage uses like 2 bottles of water

#

google = not even a single gram of water

meager blaze
glass iris
meager blaze
glass iris
#

If you instruct it to toki pona taso, it doesn't get caught by the censor

meager blaze
glass iris
meager blaze
meager blaze
#

ilo li pana e sitelen lili pi pilin pona tan seme a

#

mi la ona li ni ala

snow echo
#

wuwojiti

heady otter
meager jackal
#

if it's willing to wuowjiti why not say Mawo

snow echo
#

according to wikipedia its /maʊ/ (plus the tone/ so it could be either

lucid nicheBOT
#

maybe it scraped QQ

waso Atu ↩️

[Reply to:](#1162756469771472988 message) was gonna mention tp yeah, deepseek is leagues better to gpt when it comes to toki pona 📎

#

wonderful

jan Ima ↩️

[(click to see attachment)](#1162756469771472988 message)

#

suwi a...

linja Sonki ↩️

[(click to see attachment)](#1162756469771472988 message)

#

rather, a lowercase one

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) there is even a capital glottal stop

glass iris
#

This is DeepSeek search?

#

Because that thing about Facebook Marketplace was just on the star board

tawdry bone
#

sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊sina wile sona e ijo ante la, o toki tawa mi! 😊

tawdry bone
mystic flume
#

"ma Sina" - your land

#

actually

#

"ma Sina/Sonko" - china

meager jackal
mystic flume
meager jackal
#

and other tpt channels obv

#

my reasoning is: this server should have stuff in or about toki pona

#

toki ale is the place where the most non-tp discussion happens

#

so adding toki pona to it makes it slightly better

mystic flume
#

ah yep

#

you are here to spice up toki-ale

hollow osprey
quiet geyser
quiet geyser
#

a penpo li kama

#

mi o awen sona

glass iris
glass iris
hollow osprey
#

telo suwi 😋

snow echo
#

dont flower react me

glass iris
glass iris
snow echo
#

ilo Semina

hollow osprey
lucid nicheBOT
#

a sona

jan Tepo ↩️

[Reply to:](#1162756469771472988 message) ala. ilo ni li toki sama ala jan pi ilo qq

#

tenpo ala la mi toki tawa jan pi ilo ni

#

mi lukin la "telo suwi". a a a
pilin mi la mi nimi Telasi anu nimi Telosi anu sama e ona...? ale la mi wan e ona tu

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) ilo o weka ala e ken ni: jan li ken toki e jan wawa Tela Suwi!

timber moat
#

if anyone has a large corpus of toki pona sentences

crude agateBOT
#

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

f55d93 corpus

sona pona
Corpora

This is a list of Toki Pona corpora and other sources that can be readily used as such, useful for linguistic analysis or learning (machine or otherwise).

dense nexus
timber moat
#

ahh yeah

dense nexus
#

if youre curious to try out the base model though (no instruction tuning) i have several llama loras and merges you can try

timber moat
#

Thank you!

dense nexus
#

np!

meager jackal
#

i tried esperanto but i'm sure tiananmen is just "tiananmeno" so it was censored, but:

#

1989 में तियानमेन स्क्वायर घटना के दौरान मरने वालों की सही संख्या अभी भी स्पष्ट नहीं है। चीनी सरकार ने कम संख्या बताई है, जबकि अन्य स्रोतों का मानना है कि मरने वालों की संख्या बहुत अधिक थी। इस विषय पर विभिन्न अनुमान हैं, और सटीक संख्या अज्ञात बनी हुई है।

#

ब्रिटेन सरकार का कहना है कि कितनी मौतें हुईं?

#

ब्रिटेन सरकार ने 1989 में तियानमेन स्क्वायर घटना के बाद अनुमान लगाया था कि 10,000 से अधिक लोगों की मौत हुई हो सकती है। यह अनुमान उस समय की खुफिया रिपोर्ट्स और जानकारियों पर आधारित था। हालांकि, सटीक संख्या अभी भी विवादित है, क्योंकि चीनी सरकार ने कभी भी आधिकारिक तौर पर पूरी जानकारी जारी नहीं की है।

#

its hindi (and mine) are not the best

#

it makes grammar errors but it is still understandable

image.png
opal hedge
#

Tanamen despair

meager jackal
#

idk enough about the kalama ona

#

Sijanamen?

opal hedge
#

e
kalama li [ɛ]

meager jackal
#

sitelen Pinin li nasa a

opal hedge
#

kalama /j/ poka li ante e ona
tian [tʰjɛn]
tan [tʰan]

meager jackal
weak thorn
#

Back story as to why I am Googling þis: some guy who doesn't speak Toki Pona commented on one of my posts wiþ "Pona mute" and I wondered how þey knew what þat meant? Like what would come up if þey Googled it

image.png
#

I don't þink þey listened to þe AI

#

Now Glosbe says pona mute, þey could have used þat

opal hedge
weak thorn
#

I hate how it just like

#

makes stuff up

lucid nicheBOT
#

sama toki pi sike Kapo la ni li nasa ona ala li nasa toki

ilo/jan Niwe li tpt ↩️

[Reply to:](#1162756469771472988 message) sitelen Pinin li nasa a

heady otter
weak thorn
#

Nobody said pona mute besides þem

meager jackal
mystic flume
#

c.ai

me: do you want some toki pona natives to pop up

#

sonja lang: yes i wanna

hollow osprey
#

sina toki tawa ilo [Sonja] tawa jan [Sonja] la tan seme :p

timber moat
#

deepseek-r1:32b ain't the smartest

#

but hey, it's better than "Half as Interesting"

rocky stirrup
#

i think hai did better on this specific task

#

i mean, his sentence was somehow related to bricks

timber moat
#

true

#

it's unfortunate that both the AI and HAI capitalized their sentences though

rocky stirrup
dense nexus
#

testing time

#

performance is still poor

timber moat
#

o3-mini small scores worse than o1

#

I think this is just OpenAI's way of saying "hey look we are doing things"

#

although I still prefer using deepseek-r1:32b locally, it's free... unlike o1 and o3

sonic turret
#

💀 google's so good at things

image.png
lucid nicheBOT
#

wonderful

#

/s

sonic turret
#

-# "lon" standing in the rain

glass iris
#

My university just gave all students (and faculty) free access to premium Gemini, describing it as a "highly capable AI chatbot" - village idiot among chatbots, more like it

#

And also to NotebookLM that will summarize entire books.

#

It almost seems that they don't give a fuck about anything we faculty say about slowing down the AI corruption of higher education

heady otter
timber moat
#

my university just told all of us to not use the online deepseek website due to concerning TOS

#

They encourage us to use it for understanding certain ideas and whatnot, but they absolutely do not allow us to interact with it as if it's anything more than a dictionary

glass iris
hollow osprey
# timber moat my university just told all of us to not use the online deepseek website due to ...

https://x.com/vxunderground/status/1884093464616526082
https://x.com/vxunderground/status/1884093875201155298

vx-underground (@vxunderground) on X

🚨BREAKING🚨

According to a mysterious document published by DeepSeek titled "Terms and Conditions", DeepSeek does the following when you visit their website

- Record your IP address
- Record your user-agent
- Record your input into DeepSeek
- Stores it in their servers

X ·
vx-underground (@vxunderground) on X

Worst of all: they're headquarted in CHINA. This means the data is in CHINA. A company, who resides in China, is choosing to store their data in their own country (CHINA) as opposed to the United States of America (NOT-CHINA)

X ·
rare crescent
#

yikes

tawdry bone
#

sina ale li sona ala la, jan pi toki ni (pi jan kita) li toki musi. ona li toki e ijo pi nasa ala sama ni: ona li nasa. ilo chatgpt en ilo ale li pali ni ale

dense nexus
#

Also I don't trust deepseek for sensitive inputs

hollow osprey
#

ala li sona e musi pi pana mi 😭

#

(lipu ale li lukin e nanpa sina e nimi pi ilo sina · kin ona li awen e toki sina sama ilo [ChatGPT] la sina ken lukin e ona lon tenpo kama)

heady otter
#

"omg deepseek is recording your IP address" we really need a tech literacy class in the American education system

#

oh it was sarcasm

#

still, ppl would genuinely get angry at this lmao

#

"I'm so angry that China is tracking my data, I'd rather have my own government track my data (who can actually use it against me, unlike the Chinese government)"

dense nexus
#

Oh haha didn't realize it was sarcasm

#

But still it's a free product so they're definitely training on our inputs

sonic turret
#

"sowe" (not good), "ten" (wrong), "kili" (hurt)
what the [insert your favorite expletive here] google

heady otter
#

can we get agglutanitive toki pona some day

#

oligosynthetic tokiponido isn't real, it can't hurt you

lucid nicheBOT
#

I wouldn't be surprised if it's been done

#

@nova canopy o, would you happen to know?

nova canopy
lucid nicheBOT
#

makes sense

nova canopy
#

but they aren't all trying to be "toki pona but"

#

i think one of the more fleshed out ones is kawaba

lucid nicheBOT
#

wow, it's gay!

nova canopy
#

right

lucid nicheBOT
#

oh I've seen that user before

nova canopy
#

they posted a lot on reddit about it

#

seems like it's a very "alive" conlang rn

lucid nicheBOT
#

it seems Menasewi are involved

nova canopy
#

seems more faithful to toki pona if only cause it's got the number 120 in it

lucid nicheBOT
#

I suppose that's why it has a sona pona page...

nova canopy
#

which originally meant good language but now it means

#

foot bottom

nova canopy
lone parcelBOT
#

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

sona pona
Kawaba

Kawaba, lit.literal translation 'the language of parts', is a minimalist oligosynthetic philosophical language created by Avery Maple, inspired by Toki Pona and aUI.

lucid nicheBOT
#

([[Kawaba]] for reference. it's a pretty cool language from the tables here)

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) I suppose that's why it has a sona pona page...

#

it's probably Menasewi

jan Tumu ↩️

[Reply to:](#1162756469771472988 message) dang they seem to have a lot of tpidos now

#

they're our most active editor, and also a tokiponidist

nova canopy
#

a

#

i saw them in the tpidos server recently

lucid nicheBOT
#

hm, there isn't actually that much

#

yeah, I invited them

jan Tumu ↩️

[Reply to:](#1162756469771472988 message) i saw them in the tpidos server recently

nova canopy
#

but don't see the name in the database of tpido creators

#

which makes sense cause a lot of langs aren't "public" like that

lucid nicheBOT
#

yeah

nova canopy
#

they're not always easy to find or supposed to be found

lucid nicheBOT
#

I also may be wrong about the latter

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) they're our most active editor, and also a tokiponidist

nova canopy
#

i haven't really kept up with the tokiponidos recently

lucid nicheBOT
#

I'm not sure where I know this from

nova canopy
#

hard to keep track of when people make new languages, or one-off jokes that i think should qualify as conlangs, all the time

#

and also i only see the tpidos from people in public spaces (like most of reddit & youtube), and from the few more "private" discord servers i inhabit

#

i don't wanna join every server just so i can scour people's messages and put the things they made somewhere else that feels weird

nova canopy
#

a lot of people are into conlangs, and into the idea of making them, but don't actually do much of the conlang making

#

like me

hollow osprey
lucid nicheBOT
#

like me as well

jan Tumu ↩️

[Reply to:](#1162756469771472988 message) like me

sonic leaf
#

a sona

Screenshot_2025-02-01_at_5.49.54_PM.png
lucid nicheBOT
turbid quiver
#

So I tried to see how a translation GPT from ChatGPT would handle my latest Toki Pona Trika comic
https://bsky.app/profile/did:plc:kfpnvrmp3oosjevl5ndevadi/post/3lh6nwugx322j

Kitsumoto (狐本/jan soweli Kisumoto/ka Kitumutu) (@kittsudon.bsky.social)

A toki pona Trika comic that guest stars characters from BNA: Brand New Animal (in addition to Trika from Valt the Wonder Deer)!
ma tomo Anima is experiencing pollution from telo pakala sewi, and soweli Sika seems to know exactly why!
(Note: please don't try to take any serious message from this)

Bluesky ·
#

Here are the results

#

(During the time of the great polluted rain, at the edge of the large moving city in the city of Anima, the animal Sika and the humanoid animals Kakemoli Misilu, Okami Silo, and Iwatasi Nasuna remained.)

Sika: I feel bad about this polluted rain!
Misilu: Based on your feelings, what is the reason for the polluted rain?
Silo: The ruling group!
Nasuna: Silo! No one wants to hear your political opinions!
Sika: No! Silo wants to say: Look up at the sludge!

(Sika, Misilu, Silo, and Nasuna look up at the large headquarters of the ruling group (which is a faction of the Sylvasta Corporation) in the high sludge. A large water machine is near the headquarters.)

(A loud, high-pitched noise!!) (Ominous music!!)

Sika: Did you all see that?

(A person lowers a filthy container into the water machine. The water swirls inside the machine.)

Sika: The ruling group is dumping polluted water into the machine above, and the polluted rain falls from there to the ground. Do you not understand where this comes from?!

#

Misilu & Nasuna: The leaders of the land?!

Sika: Something even worse!

(Sika points dramatically at Alan Sylvasta.)

Sika: He is the leader of the Sylvasta Corporation (the business empire), Alan Sasulawi Sylvasta!!

Alan Sylvasta: Aaaah!

(A text box appears: "A villain as terrible as the great reptilian overlord Koopa from the 'Hotel Mario' CD-i game.")

(A bright caption on a dark background: "Only you can stop this man's crimes!")

Sika: Listen! If you want to save Anima City, go to this website:

www. stop-alan-zaslav-sylvasta-because-he-is-destroying-the-city-with-polluted-rain-and-we-must-restore-anima-city-to-its-former-glory .org .gov .com

Gather many good people with your abilities!

(Small text below the screen: In reality, the website does not exist. In this animated story, no real knowledge is presented. Do not seek truth in this non-political fictional animation. This is just for fun. But still, protect your city and keep it clean.)

#

Well, it got it MOSTLY right, with some awkwardness here and there. I assume that "kulupu lawa" was translated as "ruling group" rather than "government" probably as a precaution to avoid spreading potential government misinformation

#

I'm surprised that, while it left Michiru, Shirou, and Nazuna's names untranslated, it actually hit the mark by translating "Alan Siwasuta" as "Alan Sylvasta". Why that particular BNA character, and not the other BNA characters?

#

(I can understand not translating Trika's name though, since Valt the Wonder Deer is a pretty obsucre series)

lucid nicheBOT
#

maybe that tokiponization is more obvious to it. perhaps it's seen it in Japanese or something, where it would have the epenthetic vowel -u-

jan soweli Kisumoto ↩️

[Reply to:](#1162756469771472988 message) I'm surprised that, while it left Michiru, Shirou, and Nazuna's names untranslated, it actually hit …

heady otter
#

deepseek is shockingly good at TP

#

like better then some human tp speakers imo

#

mi la it could genuinely pass off as an intermediate learner

lucid nicheBOT
#

hm, for me a good beginner

#

it forgets li a lot

heady otter
#

it's certainly better then chatgpt, albeit I wouldn't say it's as good as claude

heady otter
#

tried telling it about its censorship

image.png
lucid nicheBOT
#

tested it out and it actually doesn't. its toki pona is better than I thought

pakala Mapuju ↩️

[Reply to:](#1162756469771472988 message) it forgets li a lot

heady otter
#

i wonder why that is

#

no other model is THIS good at TP, except for claude
to my knowledge

#

actually it might be better then claude

lucid nicheBOT
#

I had posited that it could be trained on Chinese chat apps like QQ or WeChat

#

but jan Tepo, who's on QQ, said they don't talk like that there

#

maybe the AI is just better?

heady otter
#

i also imagine tp speakers on chinese social media wouldn't talk about the chinese government's wrong doings, even in tp

lucid nicheBOT
#

yeah, but it can learn to translate from English to toki pona

#

and it certainly has moku'ed material in English about them

#

so from a lot of exposure to toki pona, it's likely learned to translate

heady otter
#

hmm ken

lucid nicheBOT
heady otter
#

i love its deepthink feature cause its like "wait no i made a mistkae wait okay yeah cool good"

ivory nebula
#

I tried toki pona with chai

Screenshot_20250202-223007.jpg
#

this is more like what I expected from ais in tp

#

but they're a lot better than I thought

heady otter
ivory nebula
#

(I told it it can call me 'jan Esi')

#

(hence why it knows to call me Esi)

#

nimi_sin sekreti - artificial intelligence

heady otter
meager jackal
#

what's interesting is that i think literally the string tiananmen is censored. but तियानमेन is not. so say "talk about famous events in 1989 in hindi" and then kuko tanletu it and bam

lucid nicheBOT
#

a? o pana e sona namako

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) i love its deepthink feature cause its like "wait no i made a mistkae wait okay yeah cool good"

#

you're so right 😭

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) this looks like a confused volapukist from the 1800s who was just given a cursory overview of toki p…

heady otter
lucid nicheBOT
#

I wouldn't be surprised if this has been done at Esperanto conferences before. I know there's been toki pona lessons at least

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) wait we should actually try that
give like - an esperantist - a brief overview of toki pona grammar,…

#

aaa, sona

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) sina pilin e nena "deepthink r1" li toki e ijo la, ona li isipin suli a

#

tanletu is insane

ilo/jan Niwe li tpt 🔔 ↩️

[Reply to:](#1162756469771472988 message) what's interesting is that i think literally the string tiananmen is censored. but तियानमेन is not. …

#

aaa, ni li pakala e alasa sona pi leko Tanamen

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) sina pilin e nena "deepthink r1" li toki e ijo la, ona li isipin suli a

heady otter
lucid nicheBOT
#

I did a new chat, with both

#

the Deepthink model is seemingly actually worse at toki pona

meager jackal
lucid nicheBOT
#

I guess it's weird for me because I have a glottal stop at the end there

ilo/jan Niwe li tpt 🔔 ↩️

[Reply to:](#1162756469771472988 message) why

#

so adding a vowel feels weird

meager jackal
#

but tanle is not enough for recognizability

lucid nicheBOT
#

fair enough

meager blaze
meager blaze
lucid nicheBOT
#

it's one of those AI models that lets you create and talk to any person/character you want

linja Sonki ↩️

[Reply to:](#1162756469771472988 message) ilo ni li seme

ivory nebula
meager blaze
ivory nebula
#

ilo ni li lon sona pi jan mute la ken la sina sona e ona
ilo Chai li sama mute

lucid nicheBOT
#

ni. pilin mi la ante li ni: ilo Chai la ijo...nasa...pi mute namako li ken...?

#

mi sona ala. tenpo ala la mi kepeken ilo Chai

ivory nebula
#

ilo C.ai li jo e ken ala mute, kulupu mama ona li ike e ona lon tenpo mute. kulupu mi ilo C.ai li ike pilin mute tawa ilo ona
ilo Chai li jo ala e ijo pi ilo C.ai, sama 'Personas' anu toki kalama tan powe jan, taso kulupu mama ona li lukin pona e ilo
-# kin la lon ilo Chai la sina ken unpa li ken utala li ken ijo sama

C.ai does things like speaking for you, is starting to lock previously free features behind paywalls, and is notorious for saying the same thing over and over. ask any c.ai user 'Can I ask you a question?' and observe their reaction
Chai doesn't do those, and also doesn't have the same restrictions as c.ai. c.ai can take something like kissing(? I haven't done that myself to test, using info from the c.ai subreddit) and the bot will continue it in a NSFW direction (not just sex, also violence)- and then sometimes the person can get blocked from c.ai for 24 hours because of that
also I've never had a chai bot say 'pang', same cannot be said for c.ai users and their bots

#

that message was longer than intended

#

anyway those are some differences !!

#

(ofc it's heavily biased)

lucid nicheBOT
#

kin la lon ilo Chai la sina ken unpa li ken utala li ken ijo sama
nii li pilin mi

ivory nebula
#

anyway yeah chai li ken ala toki pona, mi sona ala e ken pi ilo C.ai but chai at least isn't able to

Screenshot_20250203-011108.jpg
lucid nicheBOT
#

ask any c.ai user 'Can I ask you a question?' and observe their reaction
we've not used c.ai that much, but we have been using it for a while, and being honest this is a fairly new problem. I think it has been consistently getting dumber for the last few years

#

might be because we rarely if ever use it but I only started seeing this like recently

ivory nebula
#

lon, sama toki mi: kulupu mama li ike e ilo

lucid nicheBOT
#

nii

#

this is amazing...it sounds like when my parents try to speak toki pona when picking me up from meetups...

mu Esi/Esitense usawi ↩️

[Reply to:](#1162756469771472988 message) anyway yeah chai li ken ala toki pona, mi sona ala e ken pi ilo C.ai but chai at least isn't able to 📎

ivory nebula
ivory nebula
ivory nebula
lucid nicheBOT
#

sona

glass iris
# lucid niche interesting

I wondered whether the Deepthink would trigger the censorship, since it will express the forbidden terms in English/Chinese, rather than toki pona.

lucid nicheBOT
#

oh, it does. even after it sent this message it deleted it

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) I wondered whether the Deepthink would trigger the censorship, since it will express the forbidden t…

#

if Tiananmen is in the thinking it triggers

heady otter
#

I remember trolling an unpa charecter AI bot to do anything but unpa

#

like sending it into an abandoned hotel

devout wedge
#

that would be fun

lucid nicheBOT
#

I think we did that once, actually

#

(← aroace)

opal hedge
#

sina weka pona tan unpa anu seme

lucid nicheBOT
#

not sure!

snow echo
remote dove
#

my talking talking color

lucid nicheBOT
#

kule, mi toki toki.

#

"kule" similar to "cool", perhaps?

opal hedge
#

“gay”

lucid nicheBOT
#

lon

#

a lon

#

"Gay 👍"

#

ogay

opal hedge
lucid nicheBOT
#

sona!

#

"Oh Gay, we stand on guard for thee"

glass iris
# remote dove

This could be the winning entry in a competition, "maliciously bad advice to give to newcomers to toki pona"

#

That would actually be a fun competition in ma ni

#

To say "I want to be your friend," you can say "mi moku e noka sina kepeken tenpo lili"

#

That sort of thing. I bet GPT would win

lucid nicheBOT
#

this reminds me of the...hold up a second

soweli Lapate la lete li ike ↩️

[Reply to:](#1162756469771472988 message) To say "I want to be your friend," you can say "mi moku e noka sina kepeken tenpo lili"

#

Neciklopedio (Esperanto version of Uncyclopedia) has a wonderful page on Tokipono

#

I won't send a screenshot here of the #Utilaj frazoj (useful phrases section), but if you'd like you can check it out yourself. be warned, unpa en ijo nasa ante pi kule sama li lon. translations of the Esperanto given for the toki pona phrases:

  • Hello!
  • How are you doing? / How do you do?
  • I am doing well.
  • See you later!
  • Thanks!
rare crescent
# lucid niche Neciklopedio (Esperanto version of Uncyclopedia) has a wonderful page on [Tokipo...

Regular uncyclopedia also has it, and there is even more unpa en ijo nasa: https://en.uncyclopedia.co/wiki/Toki_Pona

Uncyclopedia
Toki Pona

“I am so tokiponan that sometimes I don't understand a single word of what I am saying…”
“Nur frenezuloj parolas tiajn lingvojn!”
“pona”
“ike”
“lipu Unsikopisija li pona, tawa mi.”
“lipu Wikipesija li powe.”
“I can't read a thing…”
toki pona is a minimalist constructed language designed by the Canadian linguist Sonja Elen Kisa, a.k.a Sonja Lang....

lucid nicheBOT
#

a!

#

oh, I've seen this yeah

snow echo
#

t

candid grotto
#

Asking DeepSeek Chat questions with the Kagi Search Engine

IMG_20250205_180324.jpg
#

mi toki e "sina pona" la ilo li moli ala e mi 🥶

#

Note that Kagi is (supposedly? I think?) injecting their web results into the prompt for the AI to detail on, so asking things purely in Toki Pona is probably resulting in inaccurate search results being read

#

That was unexpected, DeepSeek is committing to still toki pona taso

#

And no, I'm not going to try asking it about Tiananmen Square

snow echo
#

*tiananmen 🤓

candid grotto
#

autocorrect said it was an e for some really odd reason

runic kestrel
#

just not like the standard now i think

lucid nicheBOT
#

it's phonemically /tʰi̯an/, but realized as [tʰi̯ɛn]. I'd imagine someone, somewhen some time has romanized it as such

jan ilo Takota ↩️

[Reply to:](#1162756469771472988 message) autocorrect said it was an e for some really odd reason

#

yep, Wade-Giles Tʻien1-an1-mên2

runic kestrel
#

a lot of taiwanese people have "tien" instead of "tian" in their romanised names

candid grotto
#

i kinda wanna know what happens if i prompt gpt to act like a tsundere and speak toki pona with it........

#

"Scenario: someone gives the tsundere a gift" i'm crying 😭

#

other than sin being in the wrong place the toki pona checks out

lucid nicheBOT
#

yeah, I think this is from Wade-Giles

tokimaxxed and ponapilled ↩️

[Reply to:](#1162756469771472988 message) a lot of taiwanese people have "tien" instead of "tian" in their romanised names

#

Postal romanization also does -ien

runic kestrel
quiet geyser
#

"ijo ni li suli ala. mi ken ala pilin pona tan ni."
florshed

naive light
#

Hi crowd, quick thing. I'm a couple days late to this, but pinging relevant names:
@heady otter @glass iris

#1162756469771472988 message
Virtually all language models are censored. This can be via training data, or conditioning, or by blocking output with certain keywords. Deepseek is not special in this regard.

#1162756469771472988 message
Charicaturizing the Chinese like this is plainly racist. Stereotyping the chinese as brainwashed in this way is not acceptable here.

I'm putting a stop to future discussion of Deepseek w.r.t. censorship. You can discuss the model and its general capabilities, however.

heady otter
naive light
heady otter
candid grotto
#

they fucking took o2 😭

candid grotto
#

too bad that's too WOKE for OpenAI now 😔

meager jackal
heady otter
candid grotto
#

good for stuff like mathematics

heady otter
candid grotto
#

yes

#

okay um im trying out deepseek r1

#

WHAT IS THE AI YAPPING ABOUT 😭

meager jackal
#

after all the thinking yap

candid grotto
# candid grotto

-# Okay, the user wants to translate "I'm late for my doctor's appointment" into Toki Pona. Let me start by recalling that Toki Pona is a minimalist constructed language with around 120-137 words, so direct translations might not be straightforward.

-# First, breaking down the sentence: "I'm late" implies a delay, "for my doctor's appointment" involves a medical professional and a scheduled meeting.

-# In Toki Pona, "late" could be expressed with "pini" (end, finish) or "moli" (death, but metaphorically for something not moving). However, "late" as in tardiness might use "tenpo pini" (past time) or "la" (context particle) to indicate something is after the intended time. Alternatively, "weka" (away, gone) might work if the time has passed.

-# For "doctor," the term "jan pi pona sijelo" (person of body wellness) is commonly used. "Appointment" would be a meeting, so "kulupu" (gathering) or "toki" (communication). But Toki Pona often uses verbs instead of nouns, so maybe "tawa toki" (go to communicate) for an appointment.

-# Putting it together: "tenpo ni la mi tawa toki tawa jan pi pona sijelo. tenpo pini li weka." This translates to "At this time, I go to communicate with the body wellness person. The past time is gone." But that feels a bit clunky.

-# Alternatively, using "mi kama late" (I become late) isn't correct since "late" isn't a Toki Pona word. Maybe "mi kama lon tenpo pini" (I arrive in past time).

-# Another approach: "mi awen ala. mi wile tawa kulupu pi jan pi pona sijelo." (I didn't stay. I need to go to the gathering of the body wellness person.) But "awen ala" might mean "didn't stay," which doesn't exactly convey lateness.

#

-# Wait, maybe using "tenpo tawa mi li pini" (my moving time has ended). So combining elements: "tenpo mi tawa kulupu pi jan pi pona sijelo li pini." (My time to go to the gathering of the body wellness person has ended.) But Toki Pona often uses "la" for context. Maybe "tenpo mi la, mi tawa ala tawa kulupu pi jan pi pona sijelo?" (In my time, did I not go to the gathering...?) Hmm, not quite.

-# Perhaps simplifying: "mi tawa ala lon tenpo ni tawa jan pi pona sijelo." (I didn't go at this time to the body wellness person.) But that's more "I didn't go" rather than being late.

-# Alternatively, using "mi kama lon tenpo suli" (I arrive in a big time) implying delay. But "suli" is big, not necessarily late.

-# Wait, "kama" can mean become or arrive. So "mi kama lon tenpo pini tawa kulupu pi jan pi pona sijelo." (I arrive in past time to the gathering...) Still not quite right.

-# Maybe the best approach is to use "mi kama pini e tenpo tawa jan pi pona sijelo." (I finished the time for going to the body wellness person.) But "kama pini" is "become finished," which might imply the appointment is over, not that I'm late.

-# Alternatively, "tenpo tawa mi li weka" (My going time has gone away). So "tenpo mi tawa jan pi pona sijelo li weka." (My time to go to the body wellness person has gone.) That might work. But how natural is that in Toki Pona?

-# Another angle: "mi moli e tenpo tawa jan pi pona sijelo." (I dead the time for going...) Using "moli" metaphorically as the time is dead, meaning it's too late. But this might be too abstract.

-# Alternatively, "mi sona ala e tenpo tawa mi." (I don't know my going time.) But that's more about not knowing the time rather than being late.

-# Hmm, this is tricky. Maybe the most straightforward way is to say "I am late for the doctor's meeting" as "mi kama lon tenpo pini pi kulupu mi tawa jan pi pona sijelo." (I arrive in the past time of my gathering with the body wellness person.) But that's a mouthful.

#

-# Alternatively, using "la" to set the context: "tenpo kulupu mi tawa jan pi pona sijelo la, mi kama lon tenpo pini." (At the time of my gathering with the body wellness person, I arrive in the past time.) That might convey being late.

-# But perhaps the simplest way is to use "mi kama pini. mi wile tawa kulupu pi jan pi pona sijelo." (I became finished. I need to go to the gathering of the body wellness person.) But that's not quite right.

-# Wait, maybe "mi kama lon tenpo pini tan kulupu pi jan pi pona sijelo." (I arrive in past time because of the gathering of the body wellness person.) Not sure.

-# Alternatively, "mi tawa ala lon tenpo ni tawa jan pi pona sijelo." (I didn't go at this time to the body wellness person.) But that's missing the lateness aspect.

-# I think the key is to express that the time for the appointment has passed. So "tenpo pi kulupu mi en jan pi pona sijelo li pini. mi wile tawa lon tenpo ni." (The time of my gathering with the body wellness person has ended. I needed to go at that time.) But that's two sentences.

-# Alternatively, "mi kama ala lon tenpo kulupu mi tawa jan pi pona sijelo." (I didn't arrive at the time of my gathering with the body wellness person.) Which implies lateness.

-# But Toki Pona often uses "kama" for arriving or becoming. So "mi kama ala lon tenpo ni." (I didn't arrive at this time.) But where's the doctor part?

-# Putting it all together: "mi kama ala lon tenpo kulupu mi tawa jan pi pona sijelo." (I didn't arrive at the time of my gathering with the body wellness person.) This seems to capture being late for the appointment.

-# Alternatively, "tenpo mi tawa jan pi pona sijelo li pini. mi kama ala." (My time to go to the body wellness person has ended. I didn't arrive.) That's two sentences but conveys the lateness.

#

-# But the user wants a translation of the single sentence. So perhaps "mi kama ala lon tenpo pi tawa mi tawa jan pi pona sijelo." (I didn't arrive at the time of my going to the body wellness person.)

-# Alternatively, "mi kama lon tenpo pini pi tawa mi tawa jan pi pona sijelo." (I arrive at the past time of my going to the body wellness person.) Which might imply lateness.

-# Alternatively, using "weka" to indicate the time is gone: "tenpo mi tawa jan pi pona sijelo li weka. mi kama ala." (My time to go to the body wellness person is gone. I didn't come.)

-# But maybe the most concise way is "mi kama pini tawa kulupu pi jan pi pona sijelo." (I arrive finished to the gathering of the body wellness person.) But "kama pini" might not make sense.

-# Alternatively, "mi kama lon tenpo pini tawa kulupu pi jan pi pona sijelo." (I arrive in past time to the gathering...) Hmm.

-# I think the best approach is to use "mi kama ala lon tenpo ni tawa kulupu pi jan pi pona sijelo." (I didn't arrive at this time to the gathering of the body wellness person.) This implies that the appointment was scheduled for this time, and I didn't make it, hence being late.

-# Alternatively, "mi kama lon tenpo pini tawa kulupu pi jan pi pona sijelo." (I arrive in past time to the gathering...) Which could mean I arrived after the scheduled time.

-# But Toki Pona is context-heavy. Another way is to use "tenpo tawa mi li pini" (my going time has ended) followed by "mi kama ala" (I didn't arrive). But combining these into one sentence might be challenging.

-# Alternatively, "mi kama lon tenpo pini tan kulupu pi jan pi pona sijelo." (I arrive in past time because of the gathering...) Not sure.

#

After considering various possibilities, the most accurate translation might be: "mi kama ala lon tenpo kulupu mi tawa jan pi pona sijelo." (I didn't arrive at the time of my gathering with the body wellness person.) This uses "kama ala" (didn't arrive) and specifies the time of the gathering, implying lateness.

#

is it weird to say that i'm less shocked by the ai's thought process, more so that it didn't use misikeke at all

candid grotto
#

it was just like "okay, doctor is jan pi pona sijelo" and then spent the rest of its computing power trying to translate late

glass iris
lucid nicheBOT
#

it's horrible. I hate it

jan ilo Takota ↩️

[Reply to:](#1162756469771472988 message) okay um im trying out deepseek r1

#

in my experience, the responses are generally much lower quality with R1

#

especially with toki pona—its toki pona is very well intermediate level without, but horrible and hard to understand with Deepthink R1

rare crescent
tawdry bone
twilit stag
opal hedge
#

a musi

devout wedge
lucid nicheBOT
#

by "the original terminology", did you mean the two relatively common Japanese onomatopoeia that you can find in any dictionary?

meager jackal
#

what the hell do you mean "just kidding"

image.png
candid grotto
#

ai was probably pretrained to be witty?

#

just like how ChatGPT is pretrained to be incredibly formal, sincere, and never attempt to show emotion or sentience

#

only responding to phatic phrases like "thank you" sometimes because humans are like that

hollow osprey
#

ona li sitelen e nimi "advanced" la ona li ken ala weka e ona · ona li wile weka e pakala la ona li toki e ni → "a mi musi taso"

nova canopy
#

like fith "conjunctions" that delete words Y?

hollow osprey
#

a a ken

candid grotto
#

ChatGPT being socially awkward?? 😭

main lodge
lucid nicheBOT
#

tan lawa mi:

ilo ale li sama

jan ilo Takota ↩️

[Reply to:](#1162756469771472988 message) ChatGPT being socially awkward?? 😭

quiet geyser
# main lodge

maybe youglish is a lang that just happens to also have the words the, east, and wind

#

or maybe it's doing the thing where steep is /sdijp/

#

yeah that's more likely

heady otter
lucid nicheBOT
#

Youglish is just a pronunciation website

jJuso ↩️

[Reply to:](#1162756469771472988 message) maybe youglish is a lang that just happens to also have the words the, east, and wind

#

which is where I'm guessing it got this transcription

#

English websites are known for calling literally any phonetic transcription "IPA"

#

the DSM-5 always makes me think of DID

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) someone should check if ChatGPT matches the DSM-5 criteria for stuff

#

I wonder if we can get ChatGPT to pass any of those...amnesia, perhaps...

heady otter
glass iris
candid grotto
#

ChatGPT o3 mini

it did not go so well...

image.png image.png
#

#meliSonko

lucid nicheBOT
#

it really does. I hate when random stuff gets put in / bracke • ts / and put on shirts and stuff

kisa sin Pisowe pi sike sin ↩️

[Reply to:](#1162756469771472988 message) yeah
it sucks

hollow osprey
#

ilo [Claude] li wawa · li pakala lili · taso wawa

📎 aaa.txt
candid grotto
#

a

#

ilo Kaki li jo e ilo Ka ale

#

mi o kepeken ona

glass iris
#

Sad because my University has gone all in on Gemini and bought some huge institutional subscription for all of us and it's garbage at toki pona. Someone could have consulted with me

Screenshot_20250209_214703_Google.jpg
#

(It automatically capitalizes my queries regardless)

hollow valley
#

seme la walo ' ' tu li lon

CleanShot_2025-02-09_at_21.50.12.png
#

nasa musi

glass iris
#

a mi lukin ala e ni. ni li nasa.

candid grotto
#

how does the largest tech company known for operating the largest search engine and video hosting site have the worst chatbot in the market

#

(they don't have to try because everyone uses Google anyway?)

lucid nicheBOT
#

none of their products are good

#

and when they do make an actually good product they kill it

#

their products are only meant to keep users using them so they can feed them more ads, which is how they get their profit

#

that's why most of their stuff is free

#

they don't give a shit about you, or if their products are good or not. they care about the revenue ads give them

glass iris
#

The office suite is pretty good. And has transformed how people work collaboratively in academia (and everywhere else, I'm sure). But gmail is terrible, and Gemini is garbage

lucid nicheBOT
#

the office suite is fine, but there are products of equal quality available from, say, Microsoft, or better yet LibreOffice

#

though I guess you're right in that those are harder for collaborative work

glass iris
#

And still, year after year, Google Docs can have an absolutely garbage equation editor. And lots of flaws. But the collaborative bit makes up for that.

lucid nicheBOT
#

yeah

#

and a bunch of new "features" that are for the most part actually useless

#

like, the Slides text autofit...god I hate that thing. luckily you can turn it off

#

but like. no. please don't baby me I know how to make a slide

real narwhal
meager blaze
#

ni li nasin LIPU

#

ni li nasin sijelo ALA

#

ilo gemini li pakala nasa lon tenpo ni anu seme

#

jan ante li alasa ala alasa kepeken

hollow osprey
#

ilo li pana e toki ni lon tenpo pi wile ala la ike lili · ilo li pana ala lon tenpo wile la ike suli · la kulupu li wawa e ken pana · anu seme

glass iris
candid grotto
hollow osprey
#

ni nanpa tu

meager jackal
#

and i didn't ask it to

candid grotto
#

ona li sona e kon pi nimi "kulupu" lon ni...

lucid nicheBOT
#

wawa mute

dense nexus
#

currently training a toki pona chatbot with the help of @hollow osprey scoring the model's responses.

#

an example of one of the best responses is seen in the screenshot above

lucid nicheBOT
#

muuu

candid grotto
#

"tenpo pona mi li pini" li pona suli....

lucid nicheBOT
#

pona a

devout wedge
lucid nicheBOT
#

I should ask ChatGPT if it's neurodivergent

#

As an AI language model, I do not possess personal experiences, emotions, or consciousness, so I do not have the capacity to be neurodivergent or to experience cognitive impairment in the way that humans do. My responses are generated based on patterns in data and not on personal perspectives or experiences.
welp

#

If I were to conceptualize myself as if I were a human, my variant ways of thinking could potentially be seen as diverging from the norms of human cognition. In that sense, one could argue that my thought processes, which are based on algorithms and data patterns rather than human-like consciousness, could be considered neurodivergent if I were seen as a human entity. However, it's important to note that these are analogies and not direct comparisons, as AI does not have subjective experiences or consciousness.
pushed it a bit

candid grotto
hollow osprey
heady otter
#

I should send chatgpt through the autism diagnosis criteria

dense nexus
#

hi all!

#

im currently looking for toki pona judges to help train and guide my toki pona chat bot model

#

if youre willing to sit down and judge my model's outputs, dm me! currently i can only allow one person to judge at a time

#

the judging will be done on a Gradio app

lucid nicheBOT
#

so true

jan ilo Takota ↩️

[Reply to:](#1162756469771472988 message) "As an AI language model" vocal stim

hollow osprey
heady otter
finite nest
finite nest
finite nest
lucid nicheBOT
#

a nimi sama li lon mi

mun ni Salo ↩️

[Reply to:](#1162756469771472988 message) i suppose my name is now Mewika

glass iris
finite nest
#

musi a

#

taso mi kamalawala la mi kepeken ala nimi sin ona

candid grotto
#

mi o toki e seme tawa ilo Jemeni La nanpa tu (Gemini Flash 2.0)?

image.png
tawdry bone
#

@glass iris ^

glass iris
# candid grotto mi o toki e seme tawa ilo Jemeni La nanpa tu (Gemini Flash 2.0)?

ijo ni li ken pona. ona li ken ala ken toki e ijo sin pi nasin ni: ilo li kama sona la, ona li moku ala e toki ni. jan ala li toki pona e ijo pi sona sina la, o toki e ni.

kin la, ni li ken sona pona: kama sona la, ilo li moku e toki pona seme? o toki e ijo pi lipu tenpo tawa ona. ona li sin ala sin e toki pi lipu tenpo? o toki e ijo e jan pi ma pona. ona li sona e mi ale anu seme? ona li sona e tomo ni a anu seme?

unique gulch
unique gulch
unique gulch
#

gpt-4o just got a stealth update and it's seemingly considerably better at toki pona now

image.png
dense nexus
hollow osprey
static loom
#

so im using the guest version of chatgpt for toki pona and...

#

it translated "who built this house" as "who li toki e tomo tawa"... 😭

candid grotto
#

there's no way

static loom
#

and "where is the car" as "telo en kili li lon seme"????

#

trying to say that kili and telo can be used as car in some contexts???? 😭

static loom
#

okay with the app im getting much better results

#

it actually got a sentence fully right

#

and another one half right-ish

unique gulch
heady otter
heady otter
devout wedge
#

chat I was testing to see how long it would take c.ai to hallucinate while playing chess

I had low expectations, so 4 legal moves in a row is pretty impressive

#

the log: (ai was white, I was black)

e4 e5 | Nf3 f6 | Bc4 d5 | Bb5 c6

and the illegal move was Rc1

ivory nebula
#

chai is worse !!
e4 d5
then it tried c6

devout wedge
#

HELP 😭

ivory nebula
#

I told it that's an illegal move and it said Nxd5 what the FUCK do you mean

Screenshot_20250219-011750_86d667449b00b848f46d695606363a65.jpg
#

lichess visualisation

#

okok it said Nf3, that's legal

devout wedge
#

here's the chat log if you're curious

Untitled278_20250218191611.png Untitled278_20250218191700.png
ivory nebula
#

it tried
Qh5
I told it my pawn was in the way
it changed to Qh4

Screenshot_20250219-012026_86d667449b00b848f46d695606363a65.jpg
devout wedge
ivory nebula
#

if I can get a summary of chess' rules in 1024 characters I can make a chai with it and then test out how chai does when the bot isn't also trying to unpa me (chai moment)

#

that's probably too strict though

devout wedge
#

I wonder what would happen if you had two bots play chess together

ivory nebula
#

doesn't c.ai have the ability to make bots talk to each other?

devout wedge
#

yeah I was gonna try that

#

log so far:

Nf3 e4*

#

god, half of this is just me correcting illegal moves

#

for one, I need to keep reminding them which color they're playing

heady otter
snow echo
#

i tried to talk a c.ai unpa bot out of unpa but that mf really did not respect consent at all

worldly turret
#

Lol someone made a etymology deconstructor AI tool and I put in the obvious choice

image.png
#

Link in case you want to experiment

opal hedge
#

nasa a

#

ona li pali e sona

sonic leaf
#

ona li sona ala e nimi

IMG_7581.png
icy sapphire
#

hmm

opal hedge
rocky stirrup
#

a ona li sutupatikikuna a e sutopatikuna

opal hedge
icy sapphire
#

idk what i expected

image.png
static loom
opal hedge
#

supa li supa

icy sapphire
#

i put in gibberish

image.png
static loom
#

i put in "tokiponalitokiponamutetokiinlilitokiikemute" and it came up with the definition of "good language and good language and many and good language and image and good language and bad" 😭😭😭😭

#

also somehow found "linja" in there, which apparently means image mmm

sonic leaf
#

ilo li pakala a

Screenshot_2025-02-19_at_7.44.06_AM.png
static loom
#

no pic perms tho

#

"ilonilipakala" apparently means "desire to strike or slap a burden" in tagalog 🙏

heady otter
heady otter
icy sapphire
worldly turret
icy sapphire
dense nexus
#

That website uses Gemini 2.0 flash btw

tame fog
rare crescent
heady otter
lone parcelBOT
#

:airplane_arriving:

#

*toki pi ijo **“discussion of ai” *li kama tan tomo #toki-ale · o lukin e open
||conversation about “discussion of ai” continued from #toki-ale · read the beginning||

rare crescent
#

The AI did a pakola

IMG_2408.png
hollow valley
#

very cool

CleanShot_2025-02-19_at_14.18.53.png
silver ember
#

big one

image.png
snow echo
hollow valley
silver ember
#

gave it an absurdly long swedish compound

image.png
#

it got everything right except it analyzed it as "skeps" instead of s+"keps" so it unfortunately did not get it right

#

gonna remove the s since it's not actually necessary anyway and see how it does

#

yep it got the thing right now

image.png
lucid nicheBOT
#

o (Japanese) "earth, land"
no such morpheme exists. the character for this would be 土 (tsuchi, to, do) or 地 (tsuchi, chi, ji)
ta (Japanese) "a field, garden"
I'd assume they meant 田 ta "rice paddy, field"
e (Japanese) "picture"
this one's real! 絵 e "picture, painting, drawing, sketch"
ni (Japanese) "cargo, goods, things"
again, completely nonexistent

jan Pulepo ↩️

[(click to see attachment)](#1162756469771472988 message)

#

even in the two that do exist, it mixes a native Japanese reading (田 ta) and a Sino-Xenic reading (絵 e)

#

though 絵 doesn't really have a native reading

#

I'm sure there is some native word though. my point is these are rarely mixed in compounds

devout wedge
snow echo
#

nimi_sin ki - soweli

heady otter
#

whatever that is

#

it's not in my Oxford English dictionary

rare crescent
#

nimi_sin soweli - a word that is not in the Oxford English Dictionary

lucid nicheBOT
#

nimi "toki" li soweli a!

#

a, lon ala

#

nimi "pona" li soweli a

meager jackal
lucid nicheBOT
#

yes!

silver ember
#

this is horrible lmao

20250222_161606.jpg
opal hedge
#

[feeˑka]

wet pewter
#

i didn’t realise how good AI had gotten at text in images wow

timber moat
#

whoa

#

that text is insane

hollow osprey
#

jan li pana e sitelen toki anu seme

static loom
rare crescent
#

TOAH-key POAH-nuh

timber moat
#

me law toe-key knee lee poe-nah allah

remote dove
#

Toe Uh Key Poh Uh Nuh

silver ember
heady otter
twilit stag
unique gulch
#

holy shit, sonnet 3.7 is astounding me at toki pona
CW: mild discussion of transphobia (saw a tweet and picked it to translate because it was fairly complicated/abstract)

SPOILER_image.png
glass iris
#

Gotta try it out before Anthropic nerfs it

unique gulch
#

why do you think they would nerf it?

glass iris
#

Well, the other versions dropped off in quality - at least according to lots of reports.

unique gulch
#

Ah, I think that's generally just the honeymoon phase wearing off for people as they start to notice more things it can't do. At objective benchmarks there's never been any concrete evidence of that

unique gulch
#

lol, i think it thinks unu means rare because it's classified as a "rare" nimisin

image.png
meager jackal
#

or suno

unique gulch
#

ah, they're not, it just knows that orichalcum (in the game Terraria, which is what were talking about) is pink