Link to the playground madein this video: beta.openai.com/playground/p/BnYSA6P7vdULleDKLoPAlPNf?model=text-davinci-002 Support me on Patreon: www.patreon.com/linguij My Music: open.spotify.com/artist/2oNmhlH0CJoRoNoXGqRPAF?si=yfQrkVjBSnSmVs0ewy_Tiw New song coming soon!
Gpt-3 also speaks Esperanto extremely well, you can let it write little stories and even the code generation works in Esperanto. I think with a little fine tuning with texts in Toki Pona this model would speak it very good. Unfortunately fine tuning costs money, but working with presets also seems to work pretty well. Edit: typos and gramar
jan li pilin, li isipin e pilin. jan ala a li ken pilin, li ken ala isipin e pilin. jan li ken lukin e lukin ona sama, li ken lawa e lawa ona. jan ala a li ken lukin, li ken lawa e ijo ante, taso ona li ken ala lawa e ona sama. ona li ken tawa lon nasin tan jan taso, li ken ala pali e nasin sin. ni li toki pi poki toki Gödel Escher Bach. jan ante li pilin e ni: pilin/kon jan li lon, en sijelo li lon, en sijelo li sama ala pilin/kon, li pali ala e ona. taso mi a isipin e ni: sijelo li pali e pilin, en kon jan li lon ala. tan ni la, ken la tenpo kama la, jan li pali e ilo nanpa en nasin pi ilo ni li sama ilo pi sijelo jan. taso... o isipin e ni: jan li wile pali e sijelo sama jan la, ona li ken ala ken? ken ala tan ni: sona pi sijelo jan li wile e pali mute mute a, e tenpo mute mute a, e jan sona mute mute a. tan ni la, jan li wile pali e ilo nanpa pi nasin sama jan la, ona jan li wile e pali mute, e tenpo mute, e jan sona mute. kepeken nimi lili la,, ilo nanpa li pilin ala, en jan li ken ala pali e ilo nanpa pilin kepeken tenpo lili. taso ilo nanpa la mi jan pi sona mute ala.
I wonder how well it would do with making compounds. For example, if you said: palisa | stick laso | green jelo | yellow banana | palisa jelo cucumber | Would it know that "palisa jelo" is a good answer? Or would it give up and produce something entirely wrong? Since it's "read" Wikipedia articles, it probably knows that cucumbers are banana-shaped and green, right? But is it smart enough to make the connection?
sina jan pona. jan ale li wile e sina. taso, ni li lon ala: jan li pona mute la sina o pana e pona tawa ona. You [are a] good/fixer person. All people need you. However, this doesn't exist: person is very good so you should give good to (help) them. You are the hero that everybody needs, but not one that anybody deserves. mi pilin e lape mute la ken lili la toki mi li nasa a a a.
This is very impressive considering that this is working with a pretrained model. There must be some Toki Pona documentation in the GPT-3 dataset, probably in Common Crawl and WebText. What I'd like to see is a GPT-like (e.g. GPT-NeoX) model trained on Toki Pona, and maybe English as well, for translation, if it can be done without disruption of the Toki Pona learning. Of course, there isn't nearly as much data on Toki Pona, but I think this would be mitigated by the language's simplicity. And release the dataset and the weights so it becomes an open AI, unlike all of OpenAI's offerings. ;)
Link to the playground madein this video: beta.openai.com/playground/p/BnYSA6P7vdULleDKLoPAlPNf?model=text-davinci-002
Support me on Patreon: www.patreon.com/linguij
My Music: open.spotify.com/artist/2oNmhlH0CJoRoNoXGqRPAF?si=yfQrkVjBSnSmVs0ewy_Tiw
New song coming soon!
Gpt-3 also speaks Esperanto extremely well, you can let it write little stories and even the code generation works in Esperanto. I think with a little fine tuning with texts in Toki Pona this model would speak it very good. Unfortunately fine tuning costs money, but working with presets also seems to work pretty well.
Edit: typos and gramar
It sucks at some natural languages though. Play around with some, ask it to write short essays. There are spelling mistakes.
wa!!! ilo li ken kama toki a!! ilo sin li wawa, li nasa kin!
ken la, ilo li ken pilin lon tenpo kama anu seme?
ni li wile sona pi jan mute! mi toki e ni: tenpo kama la, ilo li pilin
jan li pilin, li isipin e pilin. jan ala a li ken pilin, li ken ala isipin e pilin.
jan li ken lukin e lukin ona sama, li ken lawa e lawa ona. jan ala a li ken lukin, li ken lawa e ijo ante, taso ona li ken ala lawa e ona sama. ona li ken tawa lon nasin tan jan taso, li ken ala pali e nasin sin. ni li toki pi poki toki Gödel Escher Bach.
jan ante li pilin e ni: pilin/kon jan li lon, en sijelo li lon, en sijelo li sama ala pilin/kon, li pali ala e ona. taso mi a isipin e ni: sijelo li pali e pilin, en kon jan li lon ala. tan ni la, ken la tenpo kama la, jan li pali e ilo nanpa en nasin pi ilo ni li sama ilo pi sijelo jan.
taso... o isipin e ni: jan li wile pali e sijelo sama jan la, ona li ken ala ken? ken ala tan ni: sona pi sijelo jan li wile e pali mute mute a, e tenpo mute mute a, e jan sona mute mute a. tan ni la, jan li wile pali e ilo nanpa pi nasin sama jan la, ona jan li wile e pali mute, e tenpo mute, e jan sona mute.
kepeken nimi lili la,, ilo nanpa li pilin ala, en jan li ken ala pali e ilo nanpa pilin kepeken tenpo lili.
taso ilo nanpa la mi jan pi sona mute ala.
this is super fascinating. awesome work :]
I wonder how well it would do with making compounds. For example, if you said:
palisa | stick
laso | green
jelo | yellow
banana | palisa jelo
cucumber |
Would it know that "palisa jelo" is a good answer? Or would it give up and produce something entirely wrong?
Since it's "read" Wikipedia articles, it probably knows that cucumbers are banana-shaped and green, right? But is it smart enough to make the connection?
just tried that and it worked! (assuming you meant to say palisa laso)
That last example really impressed me. Seems like the singularity is coming for us 💀
learning toki pona will be skynet's final step before taking over humanity
ni li kulijo suli a!
I wonder how well it can understand context given enough text.
I wonder how will this go with a polysintethic lang morohology
not well unless you put spaces or hyphens between the morphemes
@@nardi4613 why?
"You used to be a person" should be "tenpo pini la sina jan" not "tenpo pini la sina jan ike"
woops good catch i missed that
aa, ilo ni li wawa a! pona mute
Anyone else experiencing an infinite loading loop where 3 grey boxes on the left of various sizes never load into anything?
I taught toki ma to gpt-j and it worked extremely well
ok, but why did it translate mama as mother and not just as parent
chatgpt is extremely good with toki pona too
used it to learn toki pona
j wait till GPT-4 👀👀
I could make a translator for my conlangs.😀 thanks
How's that going
sina jan pona. jan ale li wile e sina. taso, ni li lon ala: jan li pona mute la sina o pana e pona tawa ona.
You [are a] good/fixer person. All people need you. However, this doesn't exist: person is very good so you should give good to (help) them.
You are the hero that everybody needs, but not one that anybody deserves.
mi pilin e lape mute la ken lili la toki mi li nasa a a a.
Poggers and based
This is very impressive considering that this is working with a pretrained model. There must be some Toki Pona documentation in the GPT-3 dataset, probably in Common Crawl and WebText.
What I'd like to see is a GPT-like (e.g. GPT-NeoX) model trained on Toki Pona, and maybe English as well, for translation, if it can be done without disruption of the Toki Pona learning. Of course, there isn't nearly as much data on Toki Pona, but I think this would be mitigated by the language's simplicity.
And release the dataset and the weights so it becomes an open AI, unlike all of OpenAI's offerings. ;)
pona a!
Are you the reason ChatGPT just *knows* toki pona now?
epiku