AI in conlanging - present and future

Conworlds and conlangs
Otto Kretschmer
Posts: 525
Joined: Tue Mar 16, 2021 4:09 pm
Location: Poland

AI in conlanging - present and future

Post by Otto Kretschmer »

Has any of you guys tried to use AI in making conlangs?

I am using Gemini and I was quite surprised that it can already do quite a lot - like coming up with a plausible list of sound changes or plausible semantic shifts for words. It is still lacking in many areas but I guess that is going to improve in the coming years as well. It's a question of 4-5 years before AI is able to write prize worthy novels so why not a conlang?
Last edited by Otto Kretschmer on Mon May 06, 2024 6:42 am, edited 1 time in total.
bradrn
Posts: 6194
Joined: Fri Oct 19, 2018 1:25 am

Re: AI in conlanging - present and future

Post by bradrn »

For me, the fun in conlanging lies in figuring out what I want, personally, and investigating the richness of language myself. Why would I get a computer to do the fun bits for me?

(And, if I want to describe languages created by someone else, actual linguistic fieldwork would give me far more interesting languages to study than an AI ever could.)
Conlangs: Scratchpad | Texts | antilanguage
Software: See http://bradrn.com/projects.html
Other: Ergativity for Novices

(Why does phpBB not let me add >5 links here?)
User avatar
masako
Posts: 886
Joined: Thu Jul 12, 2018 12:25 pm

Re: AI in conlanging - present and future

Post by masako »

I've used a few of the tools out there for various things such as vocab creation, and the occasional base-line grammatical sketch, but these are programs, not AI proper. It's one thing to use a program to develop ideas, or get a "palette" to paint from, but I would never imagine using AI to create a whole, or even most of, a conlang. Where the hell is the fun in that? Also, I've seen issues folks have had with AI being a partial solution rather than THE answer to writing, imagery, solving complex or nuanced problems, and I would argue few things are as complex or nuanced as human language.

You do you, but I have no interest in an algorithm having all the fun...I've been using paper and pencil since ~1986, and I see nothing leading me away from that any time soon.
Image
Ares Land
Posts: 2990
Joined: Sun Jul 08, 2018 12:35 pm

Re: AI in conlanging - present and future

Post by Ares Land »

I'm surprised there is enough stuff about conlanging out there that ChatGPT actually knows about it :)

I've said it before in the AI thread... but beyond the 'cool toy' aspect (which I mean, it is) I don't think generative AI is terribly useful.
I'm very skeptical about claims that it's going to write a novel (a real one, I mean, not the kind of AI/randomly generated drivel they spam Amazon with) any time soon, let alone a conlang.
Otto Kretschmer
Posts: 525
Joined: Tue Mar 16, 2021 4:09 pm
Location: Poland

Re: AI in conlanging - present and future

Post by Otto Kretschmer »

Ares Land wrote: Mon May 06, 2024 7:00 am I'm surprised there is enough stuff about conlanging out there that ChatGPT actually knows about it :)

I've said it before in the AI thread... but beyond the 'cool toy' aspect (which I mean, it is) I don't think generative AI is terribly useful.
I'm very skeptical about claims that it's going to write a novel (a real one, I mean, not the kind of AI/randomly generated drivel they spam Amazon with) any time soon, let alone a conlang.
This is what AI experts have been predicting for a few years now - and I fully believe AI will be able to write novels by 2030 and will be able to consistently generate Nobel Prize worthy novels by 2034-35.

Folks really underestimate how fast AI is evolving. The growth is exponential. 3 years ago we didn't even have ChatGPT. Now we're waiting for ChatGPT 5 and AI can generate photorealistic 1 minute long movies (look up Sora, quite an upgrade from "Will Smith eating spaghetti")
User avatar
masako
Posts: 886
Joined: Thu Jul 12, 2018 12:25 pm

Re: AI in conlanging - present and future

Post by masako »

Otto Kretschmer wrote: Mon May 06, 2024 7:13 am This is what AI experts have been predicting for a few years now - and I fully believe AI will be able to write novels by 2030 and will be able to consistently generate Nobel Prize worthy novels by 2034-35.
I think you might be overestimating the public's willingness to award AI authorship. I don't think many literature enthusiasts really want to spend time/money on AI novels. Just my opinion, mind you, but as Ares said, it's a cool toy...not much different than any other newfangled tech. It will likely fade into the void of "just another thing" sooner than revolutionize any particular field (not related to computing).
Image
User avatar
WeepingElf
Posts: 1482
Joined: Sun Jul 15, 2018 12:39 pm
Location: Braunschweig, Germany
Contact:

Re: AI in conlanging - present and future

Post by WeepingElf »

bradrn wrote: Mon May 06, 2024 5:40 am For me, the fun in conlanging lies in figuring out what I want, personally, and investigating the richness of language myself. Why would I get a computer to do the fun bits for me?

(And, if I want to describe languages created by someone else, actual linguistic fieldwork would give me far more interesting languages to study than an AI ever could.)
AMEN!!!
... brought to you by the Weeping Elf
My conlang pages
User avatar
Raphael
Posts: 4480
Joined: Sun Jul 22, 2018 6:36 am

Re: AI in conlanging - present and future

Post by Raphael »

Otto Kretschmer wrote: Mon May 06, 2024 7:13 am This is what AI experts have been predicting for a few years now - and I fully believe AI will be able to write novels by 2030 and will be able to consistently generate Nobel Prize worthy novels by 2034-35.
Well, AI experts have a psychological, and in many cases, also a financial motive to overrate AI. And since AI is trained on what human beings put on the internet, there are limits to how much AI can improve. Before we even get to the "AI output becoming AI input, thus worsening future AI output" thing.

Since AI processes existing works to create new works, I don't see how it can principally do more than pastiches.
keenir
Posts: 924
Joined: Fri Apr 05, 2019 6:14 pm

Re: AI in conlanging - present and future

Post by keenir »

Otto Kretschmer wrote: Mon May 06, 2024 7:13 am
Ares Land wrote: Mon May 06, 2024 7:00 am I'm surprised there is enough stuff about conlanging out there that ChatGPT actually knows about it :)

I've said it before in the AI thread... but beyond the 'cool toy' aspect (which I mean, it is) I don't think generative AI is terribly useful.
I'm very skeptical about claims that it's going to write a novel (a real one, I mean, not the kind of AI/randomly generated drivel they spam Amazon with) any time soon, let alone a conlang.
This is what AI experts have been predicting for a few years now -
...and the only place I can hear "I'm sorry, Dave, I'm afraid I can't do that" is on the original film, and a sound clip I used to have as my ring tone.
and I fully believe AI will be able to write novels by 2030 and will be able to consistently generate Nobel Prize worthy novels by 2034-35.
you mean Hugo and Nebula Awards, right? Nobels are something else.
Folks really underestimate how fast AI is evolving. The growth is exponential. 3 years ago we didn't even have ChatGPT. Now we're waiting for ChatGPT 5
I thought the 5 was part of China and America's race to get better cellphone tech, rather than a ChatGPT.
Travis B.
Posts: 6763
Joined: Sun Jul 15, 2018 8:52 pm

Re: AI in conlanging - present and future

Post by Travis B. »

To me LLM's are limited for all the reasons people have already stated here. LLM's can only produce content as good as the human-generated content it is trained on, and will get worse once they start getting fed content generated by AI in the first place.
Yaaludinuya siima d'at yiseka wohadetafa gaare.
Ennadinut'a gaare d'ate eetatadi siiman.
T'awraa t'awraa t'awraa t'awraa t'awraa t'awraa t'awraa.
User avatar
xxx
Posts: 806
Joined: Sun Jul 29, 2018 12:40 pm

Re: AI in conlanging - present and future

Post by xxx »

there is no advantage to using an ia
except in a commercial context
where the result prevails
(and its value in hard cash...),

when the path is more important than the goal
its advantage becomes null...

this is particularly true in conlanging,
where the activity is perfectly free,
totally immaterial and has no practical application...

and what's the point of talking about,
or being interested in,
the worldview of an average simulated human being,
when the world is literally teeming with unique and original minds...
sasasha
Posts: 466
Joined: Mon Aug 06, 2018 11:41 am

Re: AI in conlanging - present and future

Post by sasasha »

xxx wrote: Mon May 06, 2024 3:13 pm and what's the point of talking about,
or being interested in,
the worldview of an average simulated human being,
when the world is literally teeming with unique and original minds...
I’m quite interested in what AI have to say. They’re not, at all, the same thing as an average simulated human being. The sheer volume of stuff they are trained on ensures that that comparison doesn’t take us very far.

I got ChatGPT to estimate the total number of words it has been exposed to in all of its training. It desperately didn’t want to – gave me endless caveats about both accuracy and ethics. But I trapped it in some snare and got it to produce its estimates.

It’s ultimate estimate, IIRC, was 5 quintillion.

I then asked it to estimate the number of words spoken by human beings in the last 100,000 years. Again, caveats abounded... Eventually it gave a figure of, IIRC, 15 quintillion.

Now, these figures could be seen as utterly meaningless, but I pushed it to use sensible parameters, and even if it’s wrong by orders of magnitude, it’s still incredible that these two figures could be roughly comparable.

AI are too easy to anthropomorphise because their interface shares human-like qualities, such as communicating in a manner recalling conversation. But works of literature (or anything else) produced using them are not in any way works produced by an ‘average simulated’ person ‒ in that there’s nothing ‘average’ about having access to what could conceivably amount to something like 1/3 of the volume of the total number of words ever spoken by sentient creatures.

I think it’s worth listening to the information upon which they can provide, essentially, intensely synthesized attention-focusing services. Rather than literature or conlanging, I asked ChatGPT to design an alternative tax reportage system for the self-employed, that wasn’t so open to failure on either party. Its suggestions were ... remarkably salient, and ‒ if anyone gave a shit what it thought ‒ potentially useful.

I concur with the general scepticism on the artistic value of AI doing art (be that conlanging or novels or other), but that doesn’t mean IMO that no one should be interested in ‘listening’ to what they have to say. It just means we should see their produce for what it is.
sasasha
Posts: 466
Joined: Mon Aug 06, 2018 11:41 am

Re: AI in conlanging - present and future

Post by sasasha »

Ares Land wrote: Mon May 06, 2024 7:00 am I'm surprised there is enough stuff about conlanging out there that ChatGPT actually knows about it :)
ChatGPT can, when pushed quite hard, speak really really bad Verdurian straight out of the box.

I got it to tell me about Verdurian, which it could do quite well, and then insisted that for the rest of the conversation, it would exclusively use that language to communicate with me.

It’s quite ‘lazy’ ie great at finding workarounds and loopholes to do the minimum effort whilst technically satisfying requests made of it. I chided it. It eventually did some sentences. They were awful, but they — to my surprise — used some valid Verdurian words that I had not fed it (I had stuck entirely to English).
Ares Land
Posts: 2990
Joined: Sun Jul 08, 2018 12:35 pm

Re: AI in conlanging - present and future

Post by Ares Land »

sasasha wrote: Tue May 07, 2024 9:16 am
ChatGPT can, when pushed quite hard, speak really really bad Verdurian straight out of the box.

I got it to tell me about Verdurian, which it could do quite well, and then insisted that for the rest of the conversation, it would exclusively use that language to communicate with me.

It’s quite ‘lazy’ ie great at finding workarounds and loopholes to do the minimum effort whilst technically satisfying requests made of it. I chided it. It eventually did some sentences. They were awful, but they — to my surprise — used some valid Verdurian words that I had not fed it (I had stuck entirely to English).
It didn't occur me to try that, but that's pretty impressive!
sasasha
Posts: 466
Joined: Mon Aug 06, 2018 11:41 am

Re: AI in conlanging - present and future

Post by sasasha »

Ares Land wrote: Tue May 07, 2024 10:44 am It didn't occur me to try that, but that's pretty impressive!
Interestingly, I tried to repeat the exercise today, and it produced strings like ‘ekëzäm ziëk lämec ö terfihal’ ‒ phonotactically pretty good, and it sort of looks the part. But the only in any way ‘correct’ word it could produce was ‘de’ - of, from. It told me it didn’t have access to a large amount of Verdurian words or syntax, so it was guessing. From my memory my previous attempt was a little better ‒ I think it is probably my memory at fault here.
User avatar
xxx
Posts: 806
Joined: Sun Jul 29, 2018 12:40 pm

Re: AI in conlanging - present and future

Post by xxx »

generative artificial intelligences are pretty interfaces for accessing knowledge
that are nicer than google keywords, but so little...

the only thing they bring is that users put their feelings into them
to see a human person...
anthropomorphism is natural to us...

it's their statistical power that's their strength,
and that makes it possible to tell the difference, less well than a human, between a dog and a cat,
or about a medical file, better than a human, if it needs surgery or not...

But in the latter case, I'm not sure that anthropomorphism is enough,
and that it doesn't take a great liar in a white coat to believe it...
Last edited by xxx on Sat May 11, 2024 1:29 am, edited 1 time in total.
User avatar
Kesshin
Posts: 6
Joined: Wed May 08, 2024 3:19 pm

Re: AI in conlanging - present and future

Post by Kesshin »

I've never used or even spoken to an AI before, so I decided to test it's abilities and finally give it a shot. I decided to give it my little Daas alphabet and see if it could put the letters together and make words that genuinely make sense. Daas is oligosynthetic, so every letter has it's own meaning.
Here's what I got:
More: show
ChatGPT
Here are 10 words in the Daas alphabet based on the provided meanings and implications:

Akun - (A + K + U) - Meaning: A large body of water (implying wavy and the main body/biggest part).
Bivit - (B + I + V + I + T) - Meaning: To progress slowly (implying forward, small, slow movement).
Dafo - (D + A + F + O) - Meaning: The human race (implying people, life).
Sɑnt - (S + ɑ + N + T) - Meaning: A linguistic problem (implying language, straight, size).
Rog - (R + O + G) - Meaning: A royal decree (implying regal, wavy, to take in).
Kezɑ - (K + E + Z + ɑ) - Meaning: The present moment (implying the direction of the vowel, time).
Pɑrke - (P + ɑ + R + K + E) - Meaning: To hold a majestic event (implying to have, regal, the main body/biggest part, small).
Zɛhɛv - (Z + ɛ + H + ɛ + V) - Meaning: A magical wind (implying time, sound, movement, air).
Jutɑb - (J + U + T + ɑ + B) - Meaning: To perform a significant action (implying wind, straight, size, forward).
Niʒɑ - (N + i + ʒ + ɑ) - Meaning: A problem with communication (implying a problem, cold, language).
These words are constructed based on the provided meanings and implications of the letters in the Daas alphabet. Let me know if you'd like more examples or if there's anything else I can assist you with!
Clearly, it got a little confused with the IPA pronunciations and differentiating them from the romanization. However, it didn't outright fail either. I didn't give it any grammar rules either, so we have a lot of consonant ending words, but in general these actually seem pretty decent.

A few, like "Zɛhɛv" are weird and don't make sense. Magical wind? Where's "y"? In fact, it seems to have just inserted wind into random things.
No "n" in the little letter additions it added for "akun", but otherwise actually a decent Daas word. I also don't know if these are accidentally taken from other languages.

But others are nice, and have an added bonus of being grammatically correct somehow without me telling it the grammar, like "bivit"(my favorite, I might actually use this) or "keza".

Thoughts?
User avatar
xxx
Posts: 806
Joined: Sun Jul 29, 2018 12:40 pm

Re: AI in conlanging - present and future

Post by xxx »

Sounds like a roll of the dice, doesn't it...
Torco
Posts: 774
Joined: Fri Jul 13, 2018 9:11 am

Re: AI in conlanging - present and future

Post by Torco »

so called AI is excellent for some things, I routinely use GPT to code, but I can't think of a use for it in conlanging: maybe something like asking it what is triconsonantal roots or sth?
zompist
Site Admin
Posts: 2912
Joined: Sun Jul 08, 2018 5:46 am
Location: Right here, probably
Contact:

Re: AI in conlanging - present and future

Post by zompist »

sasasha wrote: Tue May 07, 2024 9:08 am I then asked it to estimate the number of words spoken by human beings in the last 100,000 years. Again, caveats abounded... Eventually it gave a figure of, IIRC, 15 quintillion.
That is probably in the ballpark.

A study online says people say about 16,000 words a day. Let's say each human does 40 years of nonstop babbling-- that's about 200m words. Google the number of humans ever and you get estimates of about 100 billion. 200 x 106 x 100 x 109 = 2 x 1019 = 20 quintillion.
I got ChatGPT to estimate the total number of words it has been exposed to in all of its training. [...]
It’s ultimate estimate, IIRC, was 5 quintillion.
Now this one doesn't pass the smell test. Who generated those 5 quintillion words? Aliens?

ChatGPT was trained on pre-ChatGPT data, which was almost entirely generated by humans. I mean, there's some ELIZA output in there, and outputs of randomizers of various sorts, and program code, but it can't be a high percentage, and more importantly, can't be important for its ability to write plausible human language.

There are some estimates of the total size of the Internet, but I don't think they're comparable... consider how much of the Internet is pictures, video, HTML codes, or other code.
Post Reply