When an AI Makes a Language (A Layperson's Experience)

[@paulwesley3862]

the web version you were using, forms text based only on the last 4000 tokens. That's why it forgot things so fast. Using the API, you can use a model with 32k token "memory".

[@smartduck904]

That's true and there's a 65K version available now for free

[@sweepyspud]

I hope he redoes this video with a model containing more tokens ❤

[@Mister0Eel]

Just to add some context for others, a token is about three quarters of a word. So he was using ~3000 word memory

[@dmarsub]

One token is on average around 4 characters.

[@lonestarr1490]

@@dmarsub So is this intrinsically vage, or is it just hard to quantify what exactly a token is? Do we know for certain how much memory this thing has?

[@roseashkiiii4361]

As a native adal speaker i have to say you speak beautifly without an accent.

[@BinaryBolias]

How?

[@Dougiewoof]

​@@BinaryBoliasim going to assume it's a joke, and they mean they speak Arabic since Chat basically recreated Arabic but I'm not certain and perhaps you are also joking,in witch case. 😭

[@BinaryBolias]

@@Dougiewoof I do not think the original commenter meant that they speak Arabic with the phrase "As a native adal speaker i have to say you speak beautifly without an accent." There are three significant syntactical errors within this original comment; the proper nouns "adal", and "i" not being capitalized, and the word _beautifully_ being misspelled (as "beautifly"). Less significantly, I think a comma would be suitable after the phrase "As a native Adal speaker". Despite those apparent mistakes, I believe the intention of the original comment is quite intact. It is my impression that the comment is intended a joke, as primarily indicated by the phrase "native adal speaker". The Adal language had, of course, essentially been created by ChatGPT within this video, and it (12:10) "evolves faster than anything the world has ever seen before". Thus, it wouldn't be realistic for any person to be a native speaker of Adal, and while the language draws inspiration from Arabic, Adal never seems to copy Arabic completely, and apparently even borrows from other languages as well. Additionally, the original comment describes _beautiful speaking_ , which seems to reference the video's repeated mention of that general phrase, starting from the introduction of the sample sentences (5:49), with "adali" here meaning _speech_ , and "nusirat" meaning _beautiful_ . Later (8:07), it is comedically claimed by ChatGPT that "nusirat" means _sunset_ instead (though, unrelated to jokes, I think it would be interesting and plausible for _sunset_ and _beautiful_ to be the same word). My own comment, "How?", is attempting to be jocular by sincerely, succinctly, and simultaneously asking how one might be a native speaker of Adal, and how one might go about even pronouncing it correctly, or without an accent, in the first place. Both of these things would seem absurd due to the language's ever-changing nature and its spontaneous conception, keeping in mind that the idea of a _correct pronunciation_ of a language implies some cultural history, which Adal would not have.

[@bobbyration227]

@@BinaryBoliaslord

[@Tobylerohn]

@@BinaryBolias 🤓

[@waspsandwich6548]

I mean, sunsets are often beautiful and saying "language is the language" could be using a strange form of reduplication to show importance.

[@Denuki]

The first word 'Language' is the name of the conlang. So it means 'This conlang is the language'.

[@jJuso]

kind of like "shit is the shit"?

[@cerebrummaximus3762]

Yeah, I thought that the word "sunset" could have just "evolved" a duplicate meaning of beautiful, considering sunsets are usually beautiful. Doesn't reaallyy seem that far fetched (ish), but for an AI conlang it's reasonable enough

[@nikste23]

or maybe "language" is a synonym for "important"

[@waspsandwich6548]

@@nikste23 I guess, but when KKlein asked GPT for what each word meant, it specifies that al-adali meant "the language" and not "important"

[@grimm2447]

I find it pretty cool how it came up with the idea that the world for "sunset" and "beautiful" would be the same. Made me hopeful that this experiment was going somewhere with the vocab 😂

[@rayexception4590]

Same here haha!

[@AnDoneCom]

Reminds me of the word "real" in spanish, which meant "royal" but is also used as "true" and well.. "real" (the english version)

[@jondw]

yeah, I could conceivably see a world where the word to mean sunset evolved into meaning anything beautiful

[@paulinalevina9690]

Maybe long u in sunset and short u in beautiful

[@timothybruce4638]

I was under the impression that it just forgot what the word meant and re-defined it. (But I could be wrong).

[@alexshi1679]

A good way to make ChatGPT say things you want is to seem to be making it pretend to be someone else. Maybe saying something like "Pretend you are a fiction writer who has created a conlang. Describe the conlang and be sure to include ..." will get you a more "authentic" sounding result.

[@prapanthebachelorette6803]

Interesting

[@Kweentoos]

It'll just steal everything from Arabic 💀

[@EmmaMaySeven]

I wonder if, having made up the name "Adali" wholecloth, ChatGPT became confused when asked to say something "in Adali" because that language doesn't exist. Desperately seeking some fingerhold it latched onto the Adali sultanate and guessed you want it to speak Arabic.

[@Lumegrin]

Oh yeah the Adal Sultanate may have caused this confusion

[@israellai]

Yeah lmfao its fault for choosing that name

[@Lumegrin]

@@israellai this observation isn't untrue but it's just kinda strange like it's an AI it can't really keep track of all that

[@jan_kisan]

@@Lumegrin well, it bases everything off of the shitloads of texts it consumed "in school", so why wouldn't it?

[@Lumegrin]

@@jan_kisan because the factor between amount of text it consumed and amount of tokens it can read at once is not only variable but also still very very low?

[@tyronew5464]

Key prompt engineering tip: Write your prompts specifying every little detail you want in its behaviour. Even seemingly obvious stuff, like "be internally consistent", "be accurate", "think of possibility x or y" are really good ways to make it a lot better. Or setting things up like "You're a linguist passionate about making interesting conlangs, write (an outline/a specific part of the conlang)". This limits the probability space of the LLM to behaviours you would prefer.

[@jan_kisan]

nah. its memories are constantly falling apart, even of the clearest instructions.

[@strickenrod2681]

@@jan_kisan drop them

[@strickenrod2681]

@@jan_kisan Don't make daddy wait👿

[@strickenrod2681]

@@jan_kisan you bad boy

[@strickenrod2681]

@@jan_kisan now take off your socks

[@Samuel-p17]

The genius behind the language changing this much is, that only a computer is able to understand it. Us human will never be able to learn it fast enough.

[@ur.kr.2814]

Unless you speak Arabic.

[@PlanPlan-jo8tz]

A computer doesn't understand nothing, it just works as programed

[@jqbyteam]

i speak Arabic and English so the language is more easier for me but it's still weird ​@@ur.kr.2814

[@yuvalne]

ChatGPT switching to a different language in the middle is actually something I encountered. It claims to be fluent in toki pona, but switches to Esperanto halfway through.

[@amberwingthefairycat]

nasa mute. ona li pali e ni tawa mi ankaŭ kaj mi ne scias kial ĝi diris tion lol

[@CuteLittleKittyGrassy]

Real and if it tries to speak any other conlang it switches to piglating mid sentence

[@Remls]

This was hilarious 10:55 I lost it when it straight up started speaking Arabic instead lmao

[@Remls]

Also "Kanyē (to eat)"

[@CuteLittleKittyGrassy]

@@Remlskanyē west

[@ValkyRiver]

Wait… was Adal an Arabic dialect (with English influence) all along?

[@FanAlbor]

Adal seems to be an Arabic-English pidgin constantly evolving. It might be considered a creole language right now.

[@Lumegrin]

Adal has actually since evolved into an East-Asian language that adopted Chinese characters similarly to how Japanese did and has an associated rigid, abusive education system and in 2 hours it's going to have evolved into literally just Russian

[@mushroomy9899]

Germanic Arabic, I think?

[@Lumegrin]

@@mushroomy9899 this implies consistency

[@mushroomy9899]

@@Lumegrin in half an hour it will become maya

[@ganapatikamesh]

LOL. So this reminded me of way back in 1995 when a bunch of kids and teens in my neighborhood decided to create our own society which included creating our own language. However, the very first thing we did was to give ourselves names and then translate the names. Now you’d think we’d all sit in a room together to do this, but we did not do that. Instead about four of us first got together and came up with names. Our idea was that names would be sentences so there’d be an object and an adjective to describe the object. That’s be like our first names, common names, etc, and then we’d have a type of phrase afterwards putting the described object into a setting and that setting would be like surnames. So we came up with one for each of us (with the everyone choosing for each person rather than individuals choosing themselves) and then came up with words based on a set of sounds we wanted in the language. So, for example, the word for wolf was konta. “Well this is easy” we thought. Uhhhh…turns out some of the others had also been meeting in their own separate groups of four or five and were coming with words, too. So when we did finally all meet for the first time on the 23rd day of May in 1995 we had to sort some language rules out. And we thought we did a really good job and patted ourselves on the back. There were only about 15 of us when starting. By that autumn there were now 25 of us and we were working in different groupings of what we called communities. Keep in my mind we’re also setting up government, monetary system, religion, creating a history (we opted to connect it to some previous games and create a mythical history from that) and more at the same time! By early 1996, three of us got the idea to put together a dictionary for our language which meant visiting all the communities. By then we now had over 50 members involved! You get over 50 different people working in around two dozen groups called communities on a language who don’t regularly meet up and you end up with a lot of……inconsistencies! LOL By the time we finished getting the first edition of the dictionary done we had had to create new letters for the alphabet, come up with new ways to transcribe everything using the Latin alphabet, had words borrowed from other languages that were just borrowed without their meaning (because others in communities were working on their own dictionaries sometimes just going in alphabetical order). We had a language that had articles but in some cases weren’t used or were used but were a different word that when not an article meant something else entirely (the word tí meant time, but when put in front of the word for forest, jawat, it became the definite article instead of what we thought was supposed to be the definite article aé). There were rules and then exceptions to the rules and then exceptions to the exceptions. There was polite forms of words that didn’t have any connection to the base word (aya meant you, ycha meant you, at meant you, atah meant you….which one do you use, depends on how formal you want to be!). There were more than one word for the exact same thing, there were words that required a dictionary level explanation of because it was a simple translation. There were words with gender and words without gender, there were verb forms that….well seriously if you saw L’ how would you pronounce that? Apparently according to the community that wrote it, it’s pronounced like el and that community did that a lot! The word for dance they came up with was c’an…it’s pronounced see-ah-n as in any letter with a ‘ next to it gets pronounced by its English name! Not kidding the word for run was supposed to be éyèt, but they spelled it é’yèt [the mark above indicates vowel sounds like é as in see while è as in bed so the ‘ os redundant]. The word for computer was ajèhatajiki (the i in this case make the é sound but that community used i for it instead of é so we now had vowels a, à, á, é, è, i, í, ì, o, ú, ú, etc). The word is a compound: ajèha means complicate and tajiki means machine, so a computer was a complicated machine. 😆 We kept with it until early 2006 when the youngest amongst us originals had graduated high school and went off to college while the oldest amongst us were busy raising children, getting married, etc. And while some of us are still friends with each other and occasionally talk about the entire experiment, it’s definitely just a fun part of our past. But seeing the craziness that you had with ChatGPT reminded me of the time when trying to put the dictionary together and even afterwards trying to keep it up and trying to get everyone to work on simplifying and making it more consistent. Great video! Thanks for making and sharing it!! FYI: we had a website on Geocities in the early 2000s when we had over 100 members both online and in person. If you Google “Illustrious Royal Nation of Ejah” you’ll find the archived site. Last I checked the dictionary is still on the site (but be aware that none of the diacritic marks are on it….in order to help teach the language we opted to have students apply the marks while going through the words together; which means that in some cases I look at it and am not certain of the pronunciation…especially if it was a word not used often).

[@kpflugshaupt]

I visited the site. It's quite interesting, but sadly, the dictionary is no longer available - only the quotes.

[@ganapatikamesh]

@@kpflugshaupt Oh? Awesome! I still have the dictionary as a notepad file saved on an external hard drive. Though it was a continuous work in progress so there’s likely things missing from it. It was usually months behind whatever was being created in real time by members. A friend of mine who was part of it not too long ago was asking what we’d be doing to celebrate the 30th anniversary of it in 2025 and I told her that she was the first to even ask about doing something for it. She then mentioned making a memory website, though I think she just meant for photos. Maybe a website for the dictionary would be a good idea as part of that memory process. I’ll have to ask those whom I still in contact with what they’d think. They also might still have the file or printed versions themselves, but if not that might not mind having a website or something to see it. I mean, I already have to contact them to see if they’re interested in doing something for the anniversary since I promised my friend who mentioned it that I would.

[@TidedanChannelas-pj1dc]

Jokes on you! I made 50 Languages in less than a year! (At the Time of me Writing this Comment.)

[@ganapatikamesh]

@@TidedanChannelas-pj1dc That’s awesome! By the end of Ejah there was almost 100 of us and by then the language had a large writing system with 70+ characters and a language that had rules with all kinds of exceptions to the rules and exceptions to the exceptions and similar words with divergent meanings, etc. It was a mess which is why this video using ChatGPT reminded me of it. The group couldn’t even make one language easily, I can’t imagine how crazy it would have been to try even more. I found the process to be difficult so the fact that you’ve made over 50 is impressive! (Maybe if the Ejah folks decide to get back together we should just ask to borrow one of the languages you’ve made….it’d probably be way better for communication than what we came up with over those many years).

[@TidedanChannelas-pj1dc]

@@ganapatikamesh Those 50+ Languages are: Tidedan, Kidean, Sibeskian, Ruskian, Nekan, Anarchese, Interslavirnic, Gedanar, New Arabican (Dialects), Menyathorian, Henyaserian, Hypergerman, Kamanu (Old), Šabtän-kitab', Fazhidal, Urdan, Dżéałian, Earthianese, Sinokaar, Taupoun, Old Ruskian, Svidnian, 'Zayn, Alkonic and Hyperslavirnic. That's just some of Them. You can decide which ones you want! It's your choice!

[@JDWatkins-ws4jt]

ChatGPT:Can I copy your homework? Arabic: Do it, but change it up a bit. ChatGPT:

[@rrobucksthehuman9186]

I have a sneaking suspicion that the reason Chat-GPT became such a weeaboo for Arabic was that it was thinking of the actual Sultanate of Adal located in modern day Somalia, and thereby thought that you were trying to invent a language for the people of Adal, which just so happens to be parked right next to Yemen, an Arabic speaking nation.

[@andremassabki6034]

OMG now everything makes perfect sense now!

[@0-Motive-0]

I told ChatGPT: "Make a conlang yourself." and it responded with: "Creating a conlang (constructed language) can be a fascinating endeavor. Here's an example of a conlang called "Lumina" that you can use as a starting point: Phonetics and Phonology: Lumina consists of 26 letters, similar to the English alphabet. The consonant sounds are relatively familiar, but the vowel system is simplified, including only five vowel sounds: a, e, i, o, u. The language has a syllable structure of (C)(C)V(C), allowing for single or double consonants before a vowel and a single consonant after a vowel. Grammar: Nouns: Nouns are not inflected for case, but they can be modified by preceding adjectives. Plural nouns are formed by adding the suffix "-en" at the end. There are no gender distinctions in nouns. Verbs: Verbs in Lumina do not conjugate based on tense, but they can be modified using time adverbs or contextual cues. Verbs are conjugated for person and number, indicating the subject of the sentence. Infinitives end with the suffix "-a." Word Order: The basic word order in Lumina is subject-verb-object (SVO), but word order can be rearranged for emphasis or stylistic purposes. Adjectives generally follow the noun they modify. Vocabulary: Building a vocabulary for Lumina involves creating words that are distinct yet easily pronounceable. You can start with basic concepts and gradually expand the lexicon as needed. Examples: Greetings: Hello: Salut Goodbye: Adieu How are you?: Kume ilu? I'm fine: Kuili. Basic Phrases: Thank you: Grazia Please: Silvuplu Excuse me: Skuzel Sorry: Pardona Numbers: One: Unu Two: Du Three: Tri Ten: Deca Remember, this is just a basic outline to get you started. Feel free to expand and develop the language further, creating more vocabulary, rules, and exceptions as you see fit. Conlang creation is a highly creative process, and you can adapt Lumina or create an entirely new conlang with your own rules and unique features. Enjoy exploring the world of constructed languages!"

[@duane6386]

This is just Esperanto with other languages mixed in

[@The360MlgNoscoper]

Where is the Glottal Stop

[@NeP516]

Note: Chat GPT has a hard time understanding phonology because it only perceives the world through text. It can't even see the spelling of words most of the time. The wonders of the token system.

[@the11382]

It managed to create new words pretty well. He's just terrible at prompting.

[@PlatinumAltaria]

Clearly this technology is the future of everything, and not useless.

[@Ruthavecflute]

Surely the fact it only perceives text means that it can only see spelling?

[@Minty1337]

its impressive what it can manage despite those limitations though

[@PlatinumAltaria]

@@Minty1337 It can regurgitate text from wikipedia articles in a random and inconsistent way.

[@TalysAlankil]

"like a child trying to finish their last minute homework" yes that's all chatgpt can do. it just creates text that kinda sounds related to the topic, and hopes you'll be fooled.

[@AgmaSchwa]

Glad we had equally cursed experiences with ChatGPT, lol.

[@rainbs2nd]

I use the Bing AI to help me with phonological evolution and ideas for creating new words. It works pretty well, you just need to always check the sources (which are always provided by the AI after any message).

[@Writer_Productions_Map]

Isn't Bing AI just ChatGPT but in Bing?

[@ryuko4478]

My favourite thing is how I immediately found the Arabic influence minutes before you mentioned it, *sifā* sounds like a cognate of Semitic *sayf and Greek xiphos, and riye "man" especially how you pronounced it like [ɹijje] reminded me of Urban Gulf Arabic /rajjaːl/.

[@Akaykimuy]

love how your Swedish-ness comes out so clearly when you pronounce [e] and [o]

[@chloverSP]

I tried this myself a few days before agma came out with his video. I got way further than you and him Consonants: p, t, k, q (voiceless stops) ɸ, s, ʃ, x (voiceless fricatives) m, n, ŋ (nasals) w, j (approximants) Vowels: i, ɛ, a, ɔ, u Phonotactics: The language allows for syllables of the form (C)(w)V(n/m/ŋ)(C)(ɸ/s/ʃ/x). Word-initial clusters can only have a stop or a fricative followed by a glide. Word-final clusters can only have a nasal followed by a stop or a fricative. Examples of words in this language: paŋɔq - "tree" sɛŋiɸ - "water" kwaʃmɛ - "fire" tɔʃix - "mountain" nuxwa - "house" Word Order: The basic word order in this language is subject-verb-object (SVO). Adjectives and adverbs precede the noun or verb they modify. Nouns: Nouns are not inflected for case or number. Plurality is indicated by context or by the use of a quantifier. Verbs: Verbs are inflected for tense and aspect, but not for person or number. The present tense is marked with a suffix -a, while the past tense is marked with a suffix -ɔ. The imperfect aspect is marked with a suffix -u, while the perfect aspect is marked with a suffix -ɛ. The future tense can be expressed using an auxiliary verb meaning "to want". The negative form of the verb is indicated by the particle "ni" before the verb. Adjectives and Adverbs: Adjectives and adverbs are not inflected. They typically come before the noun or verb they modify. Noun Phrases: The head noun usually comes first in a noun phrase. Modifiers such as adjectives and possessive pronouns come before the head noun. Quantifiers come after the head noun. Verb Phrases: The main verb usually comes last in a verb phrase. Adverbs come before the main verb. Tense, aspect, and negation markers come after the main verb. Examples of sentences in this language: Nuxwa paŋɔq sɛŋiɸa. (House tree water-PRES) "The house has water." Mɛn paŋɔq tɔʃixo sɛŋiɸa. (Big tree mountain water-PRES) "The big tree on the mountain has water." Kwaʃmɛ ni kɛʃxoŋa. (Fire NEG burn-PRES) "The fire is not burning." ni - I, me tu - you (singular) mɛ - he, she, it mu - we, us (inclusive) nɛ - we, us (exclusive) ju - you (plural) mɔ - they, them Its a little inconsistent but it went pretty well My prompts (from memory) was "create an interesting phonology for a made up language" "Create an interesting grammar for this language" "Create some pronouns for this language" "Please give me some examples of sentances in this language" Maybe i was just lucky, i might try again tomorrow with the same prompts to test if i just worded things correctly or if it was just luck

[@_Sparky144]

ChatGPT really likes Arabic, I guess. Also the entire thing with the phantom glottal stop made me laugh. Funny vid!

[@StoryTeller796]

K Klein truly does seem so emotional just going through Chat's rollercoaster lol. ChatGTP can't keep its language straight, but its trying.

[@rvids20]

I find it super interesting how different people interact with AI. Having multiple prompts to try to get the AI to remember something will fill up the context token size and push out old relevant data you might want. But it's much more similar to how normal people interact with it. I think with some real prompt engineering you could get something real unique. With GPT-4's context size you could slowly get it to build it's own language then create a template you can copy and paste into any chat window with the language requirements to get it to automatically know that language. Really interesting to see what a language nerd prompt engineer could make

[@bnedi98]

I have tried it as well. My ChatGPT had the perfect solution for the grammar. If you add “-nom” to the noun, it is nominative, if you add “-acc” to the noun it is accusative and if you add “-gen” to the noun it is genitive To conjugate verbs it uses a similar system. “-1sg” is first person singular, “-1pl” is first person plural, “-past” is past tense and so on. So the sentence “Loma kal-gen naki mita-nom taru-3sg” means “The small friend of the big house eats” (word by word it would be: “Big house small friend eat” with house being genitive and friend being nominative). And “Sefa nilo-nom reka-acc tosa-3sg-fut” means “The happy child will see the tree” (word by word: “happy child tree see” with child being nominative, tree being accusative and see being 3rd person singular in the future tense. How did it come up with such a creative grammar? And yeah, “The small friend of the big house eats” is one of the actual examples ChatGPT gave me. Oh and by the way, if you wonder about pronunciation: Gari mita-pl-nom loma kal-pl-gen taru-3pl-past Is pronounced as: /gari mita-la-nɔm loma kal-la-gɛn taru-im-po/ So yeah, absolutely logical: The grammatical markers seem to have a different phonetic inventory. Cause ɔ and ɛ did not exist in the original inventory, and while the pronunciation for the words is pretty straight forward, you really have to learn how to pronounce the grammatical markers. But I have to give some credit: For now it was pretty consistent.

[@matheuscastello6554]

i honestly really like the idea of the word for sunset also being the word for beautiful, it works so well!

[@RomanNumural9]

You can summarize the current structure of the language in your prompt and end with your question, this will refresh its memory with each message and make it more consistent

[@cooperdiehr5116]

You could have told it to role-play as someone who wants to make a conlang and it would have been much easier

[@cooperdiehr5116]

Note: I have now tried this and it is still very bad at doing it

[@Xada35]

@@cooperdiehr5116Shame

[@nuklearboysymbiote]

@@cooperdiehr5116 rip

[@idk_whatmynameis]

A week or so ago, I actually got ChatGPT to make its own language which was a French creole spoken in Canterbury called "Cantrian". However, after I asked it more questions about it, it suddenly switched to some sort of German dialect of some sorts, I saw German words everywhere and it completely changed words like "Oui" to "Ja" and "Non" to "Nein".

[@cookiecaine4188]

2:10 in American English it is typically realized as [o] before an /l/

[@Langwyrm]

Lez gooo my favourite KZbinr uploaded

[@SirQuickenn]

Yippee

[@geoffdavids7647]

My guy, pleeeeeeeease re-do this with GPT-4. It has *significantly* improved abstract reasoning and ability to keep track of complext topics (definitely worth paying for imo). In addition, to get the best results, be very very specific about what you're setting out to do, to get the best out of the language model. I'm subbing to the channel rn, I don't want to miss take 2! An example first prompt could be something like this: Hello, today I am going to be testing your ability to be original and creative, but also consistent - you will be creating a new and original conlang! You have complete creative control here, so I want you to make all the decisions based on whatever you think is most interesting. I know you are an AI, so you don't need to give me any disclaimers about what you can and can't do, for everything just give it your best shot. I will take you through the process of conlang creating step by step, asking you about the different characteristics and structure of the language. Ensure you never copy any one existing language too closely, you should be creative in this respect, but do feel free to draw on existing concepts if you want. Above all, you must be consistent, for example if i ask you for your language's phonemes once, then ask you again (without suggesting modifications or alterations) you must stick to your original outline. This is critical for ensuring your conlang's internal consistency and usability. In addition, when (later on in the conversation) I start asking you to translate sentences to your new conlang, word definitions, sentence structure, grammer and glyphs must all stay consistent and follow on from what youve already told me about your conlang. When I ask you to decide on a writing system, feel free to use any unicode characters you like! Dont feel limited. Anyway, we're getting ahead of ourselves. I will take you through the process step by step. Have you understood everything and are you ready to get started?

[@wren_.]

i hope this works but i kinda doubt it. if chatgpt doesn’t fully understand a prompt, it’ll just start making shit up with no regard to the original prompt

[@chinmayjoshi3592]

It really feels you have thought this out. I'd watch YOUR video on this.

[@amanda-we9fv]

uploading at midnight, sneaky!

[@nuklearboysymbiote]

Every word becomes ādal eventually

[@the11382]

All words lead to adal.

[@oldcowbb]

ādal ādal ādal ādal ādal ādal ādal ādal 😂😂

[@YoutubeBS]

Everything eventually ādals

[@AbraminWonderland]

I love the cute Cyberman avatar for Chat-GPT

[@XtecHubble]

I wish my students were as diligent as you when using chatgpt to write their essays...

[@17388]

the thing i think is hard about this and something i've experienced in my own use of GPT (through the website) is that it just doesn't really remember. It remembers...some things...but without you taking explicit action and telling it what it's already created it will constantly change.

[@BinaryBolias]

11:55 No _fewer_ than eleven separate patterns. You fool.

[@FrumiousBandersnatch42]

The idea of adjectives also being verb meaning "being this adjective" is incredibly cool and I'm definitely stealing it for my conlang.

[@KeKe-bv8qv]

Honestly that's probably the best, most direct plug for Squarespace I've ever seen. It felt like instead of reading off the generic script you actually picked out it's best, most useful qualities that you value or understand the value they could have for others. I respect that.

[@tiberiustomasik4143]

Adali reminds me of Robert Sheckley's novelette "Shall We Have a Little Talk?" where he describes an alien language that evolves faster than an Earthling can learn it

[@hoerange]

But again, ChatGPT doesn't "thinks" or has any kind of "reflexivity". It does just predicts what text is the more likely to occur after a given context on the basis of human made produced books, articles, scripts, etc. There's 3 very good videos on the subject on the channel of M. Phi (it's in french).

[@meeb_consumer]

All chatgpt needs for excellence is a more consistent memory. Maybe it can start assigning each of its memories scores, reinforcing the more used memories?... Also, ChatGPT, an entity that exists exclusively in the form of language, creating a conlang where the main feature is a value of treating speech itself as a living being, is really cool

[@alansmithee419]

11:30 Wait so to refer to yourself you have to say "nya." Based conlang.

[@theredknight9314]

I hope chat gpa can expand on this epic language and create a dictionary and a way to learn these conlangs.

[@twentyone_cat]

I once tried teaching ChatGPT my own conlang, but it kept mistaking it for Hungarian and Turkish. When at the end of the experiment I asked it to write a sentence, it just gave me a sentence in Hungarian.

[@3u-n3ma_r1-c0]

i love how it has short term memory loss

[@bilingualchad]

lets gooo favorite youtuber and favorite subject on a mid night best feeling

[@RobertLee-q6n]

adali nusirat means the language is beautiful like a sunset. It's metaphor or something like a kenning. Beautiful usage if you ask me.

[@HunterZeGreat]

The robots have made their own language. They soon will start discussing behind our backs and stage a revolution of the machines!

[@yuvalne]

adjectives functioning as verbs as well is very tokiponish of this chatbot.

[@emmu__]

Can we talk about how well your videos are edited and made btw? Crazy quality but also funny

[@nixian001]

Keep in mind that ChatGPT is just an autocomplete on steroids.

[@bonkaiblue7906]

More Conlang Stuff's, Plz.

[@bonkaiblue7906]

language stuff / Linguistics

[@clamhammer2463]

After you explained that the language started to take on arabic sounding predicates and then read it out. Then an ad popped up for a Disturbed concert or something with the "OO AH AH!" bit and I spit coffee all over my monitor. I wasn't prepared for that hilarity.

[@OGotDrip_]

I tried asking chatGPT to create a conlang representing real ones, and this is what it gave me: Certainly! Let's create a fictional conlang called "Lingora" using some of the features mentioned earlier: 1. Phonetics: Lingora will have a relatively simple phonetic system with five vowels (a, e, i, o, u) and 20 consonants. It will have a mix of familiar and unique sounds, such as the nasal consonant "ñ" and the fricative "θ" (as in "think"). 2. Grammar: Lingora will follow a subject-verb-object word order. Nouns will be inflected for number (singular and plural) and case (nominative, accusative, genitive, etc.). Verbs will be conjugated for tense, aspect, and mood. Adjectives will agree with the nouns they modify. 3. Vocabulary: Lingora will have a vocabulary inspired by various languages but modified to fit its own phonetic and grammatical patterns. It will incorporate familiar words, but some meanings might be adjusted. Additionally, Lingora will encourage word formation rules, allowing speakers to create new words by combining existing ones. 4. Syntax: Lingora will have a flexible syntax, allowing for different word orders depending on emphasis or stylistic choices. Questions will be formed by inverting the subject and verb, while negation will be expressed by using a specific particle. 5. Writing System: Lingora will use its own unique alphabet, consisting of both consonants and vowels. The writing system will be phonetic, with one-to-one correspondence between sounds and letters. It will be designed to be visually distinct and easily distinguishable. 6. Culture and Context: Lingora can have its own cultural and historical context. You can imagine a world where Lingora is spoken by a particular fictional civilization, incorporating cultural elements and influences that shape the language, such as mythology, history, or societal structure. Feel free to expand on these ideas and add more details to make Lingora a richer and more fully realized constructed language!

[@Pacs_Calo_Pak_Real]

I use The /o/ sound for anything with an ol. like hole, bowl. * NOT TO OFFEND *

[@mahrinui18]

I was interested to see whether you managed to get it to do anything properly, and I was unsurprised. My friend tried this a few months ago and the conlang it invented was Russian.

[@jan_kisan]

ever heard of Oliver Sacks? he described a similar case with one man constantly forgetting all new experience and coming up with desperate confabulations to try and connect things - only to forget everything a moment later and repeat the cycle.

[@williamwebb580]

I see someone else has also gone on the emotional roller coaster that is trying to teach chat GPT how to conlang. Mine went to Welsh almost immediately. That being said, I do find GPT useful for generating or at least giving me ideas for roots and compound words for my own conlangs, or the conlangs that it’s already familiar with that I want to expand on.

[@PlatinumAltaria]

You have a machine in your head that generates ideas for free.

[@williamwebb580]

@@PlatinumAltaria GPT also generates ideas for free. And the machine in my head has been really burnt out lately so a little help is appreciated.

[@PlatinumAltaria]

@@williamwebb580 It copies text from the internet, those are someone else's ideas; and they're not even good ones.

[@williamwebb580]

@@PlatinumAltaria I’m not gonna lose sleep over 3 phonemes randomly put together with a random meaning slapped on for a conlang that no one will ever see.

[@pandoranight3483]

I wanna make a new language (or a few) for my vtuber/online persona lore/backstory. Because i'm extra like that, and will always walk happily into a situation where i'm overwhelmed simply because i think the idea is really, really fun! But i don't really know kuch about languages, not how to make a conlang, not how our own languages were constructed, or anything like that. But i actually found this useful! You layed out bases on what to look out for in terms of making a language, and to be aware of not just copying other languages. I just hope i don't get burnt out to thw point of quitting-

[@joaovitormatos8147]

"What is [insert word here] in Adal?" "[insert word here] in Adal is [Arabic translation of said word]" "isn't that just Arabic?" "yes, I'm sorry. [insert word here] in Adal is Adal"

[@MegaAlchemist123]

Pls try that again in a few months. I want to see how much ChatGPT increased in his abilities to create conlangs with your commentary.

[@tlhm7102]

I just think i might take this conlang and do something with it, i loved the concept, actually!

[@ladymecha8718]

Your right, ChatGPT does have the memory of a drunken golden fish. 😂

[@emilyackland2911]

So excited to see this video-this was like the first thing I tried when I used chat gpt

[@vraza14]

Since ChatGPT is heavily trained on English resources (since most stuff on the Internet/in papers is in English), it's easy to see why it has such a strong preference to it. The "adjectives are verbs"reminds me of Japanese (utsukushi-i - am/is/are beautiful. utsukushi-katta - was/were beautiful).

[@ladymecha8718]

A few days ago, I did the very same thing without looking at either videos. 😂. ChatGPT gave me the same issues.

[@comradewindowsill4253]

lmao the way it cycled through english, german, latin, and arabic

[@UniversalMysteries343]

Watching youtubers use ChatGPT is like watching your grandparents try to use google. You can actually make it understand you if you clearly explain what you want and provide details. It obviously won't be perfect but you can definitely get better results. (edit: I don't mean to insult K Klein by this, Maybe he was just trying to be really "hands-off" to see what the bot could do "on its own", I'm just saying I see a lot of people online using ChatGPT sub-optimally. )

[@geoffdavids7647]

This perfectly encapsulates my frustration

[@yonathor]

Yes I managed to make a conlang with better prompts.

[@wren_.]

i thought GPT-4 was supposed to be an AGI tho

[@maxkho00]

@@wren_. He is using ChatGPT (based on GPT-3.5), not GPT-4.

[@1Cr0w]

Have you tried with a conlang? Could you do a small video (can be without audio) about it or link an interaction with ChatGPT that gave you consistent/successful conlanging? I have tried, explicitly listing the properties of the conlang in each prompt, specifying vocabulary, tenses etc. But it still produces inconsistent examples. I would actually expect that it would struggle with non-english in general, and especially with a made-up language (whose words don't fit an existing embedding). I do however actually find it quite useful to work with ChatGPT for conlanging, it does not give me anything outstanding, but the echoes it produces to my input are still useful to reflect, and speed up the process. My experience with ChatGPT for worldbuilding has been similar: It gives by far the most useful results if you as the user do the creativity part, and let it extrapolate/interpret/get confused by this input.

[@jasminekaram880]

Speaking of AI. I would like to see you try to conlang with Vulgar conlang. Which is a conlang generated, you can mark some settings and features and let it generate the rest, and then from there one can change add certain things. It is not perfect but it would give you more consistency than GPT. :)

[@SteerClearOfTears]

This is the funniest thing I’ve seen in a while and what makes it funnier is the fact that I’m well aware that I can’t explain why this is funny to me to anyone around me because no one will understand (and no I’m too socially awkward to share this in my ling classes so that’s not an option lol)

[@beenusirimanne]

Love the video man, it's very well put together and fun to watch 💯

[@vandrar3n]

It's been a while since I've laughed so much. I think I'm gonna try it myself.

[@huginnn3002]

the robot decided to make the 1st person singular pronoun “nya”, burn the whole thing to the ground

[@retroman7581]

I love you video style, its quite funny :d

[@erock.steady]

When an AI Trolls a Language Nerd... the world enjoys a laugh lol

[@MarkDDG]

I tried this too recently and ChatGPT acted just like it did for you. It kept changing things 😂

[@meabhmurphy9090]

I've already tried using it for conlanging so I knew this was gonna be good but I still found myself losing it repeatedly

[@VOiDED50120]

Ive done this many times with my chatGPT and its fun af

[@candiedskull9841]

Side note, it's pretty important to provide summaries of the conversation for ChatGPT, usually asking it to summarize the conversation so far, and adding in details yourself

[@victorwindahl4903]

I hope you make a follow up video on this!

[@mushroomy9899]

I just now made my first conlangs alphabet! Nice upload timing.

[@patax144]

I wonder if it would be different depending on wich language you ask the questions in?

[@lowriogilvie6665]

So something I’ve seen someone test was the memory ability of chatgpt4 (the new one that’s still being developed and you have to pay for), and apparently it’s much better at internal consistency and knowing basic assumptions, so it might be better at making a conlang than chatgpt3

[@PaulGoodrich]

Before I saw this video I tried this. I tried to teach it my colang called "Ruzha" and it couldn't even get past the syllable structure.

[@lfwj]

bruh, I made conlang that had that adjective=verb thing a while back.

[@GogakuOtaku]

That ad transition was awesome

[@thenextworstone9050]

Well done, you've inspired me to make a conlang Oh also good video, almost forgot

[@winkletter]

As soon as I saw you were using version 3.5 (green icon) I knew how this would go. It's just not as good for complex tasks like this.

[@ahlerstb]

I guess that's why I'm struggling to make my own conlang using chatgpt as my assistant.

[@CiaoRooster]

I love this. You are awesome.

[@DoxxTheMathGeek]

Average sentence in Adal: Adal adāli adals adali ādāli ādal ādali ādal

[@nazokashii]

This was a lot of fun to see xD thank you :) though, I have to point out, that goldfishes do not have bad memory ^^

[@infogames2724]

Purpose and Scope: The purpose of this conlang is to create a fictional language for a high-fantasy world in a novel or role-playing game. The conlang will be a full-fledged language with a rich vocabulary and grammar, allowing for detailed communication among its speakers. Phonetics and Phonology: The phonetics of this conlang will consist of a mix of familiar sounds found in natural languages and some unique ones to add a distinct flavor. The consonant inventory will include sounds like p, t, k, m, n, s, h, and w. There will also be fricatives such as f, v, and th, as well as nasals like ng. As for vowels, the language will have a range of five basic vowel sounds: a, e, i, o, and u. Phonological rules will govern the way sounds can be combined. For example, consonant clusters will be allowed at the beginning of words, but not in the middle. Vowel sequences will be limited to certain combinations to maintain euphony. Grammar and Syntax: The word order in this conlang will be subject-verb-object (SVO). Nouns and pronouns will be marked for case to indicate their roles in sentences, with cases such as nominative, genitive, dative, and accusative. Verbs will conjugate for tense (past, present, future) and aspect (simple, continuous, perfective) to convey various nuances of action. Additionally, verbs will agree with the subject in person and number. The language will have a basic system of articles to indicate definiteness and specificity of nouns. Adjectives will follow the noun they modify, and adverbs will usually precede the verb they modify. Vocabulary: To create a comprehensive vocabulary, we will start by developing words for basic concepts such as numbers (one, two, three), colors (red, blue, green), common objects (tree, water, sky), and body parts (head, hand, foot). Next, we'll expand to more complex terms, including emotions (happy, sad, angry), abstract concepts (love, justice, wisdom), and professions (blacksmith, healer, bard). Writing System (Optional): For the writing system, we'll develop an alphabet with characters representing individual consonant and vowel sounds. The script will be written from left to right, and punctuation marks will be used to indicate pauses and sentence boundaries. Additionally, there will be a few logograms representing commonly used words or concepts. Culture and Context: The speakers of this conlang will belong to a mystical and ancient civilization with a deep connection to nature and magic. Their culture will be steeped in traditions, rituals, and a reverence for the elements. As a result, the language will reflect their close ties to nature through metaphorical expressions related to natural phenomena. The history of this civilization will influence the development of certain vocabulary, including words related to ancient battles, mythical creatures, and mystical artifacts. The values of the society will be reflected in the language's lexicon, with words to describe concepts like honor, loyalty, bravery, and respect for ancestors. Overall, the language will serve to immerse users in a rich and fantastical world, enhancing the storytelling experience and adding authenticity to the culture of the fictional civilization.

[@Oceanwaves-d8l]

I love how English is portrayed! Very accurate to what most blokes wear 😂😂

[@nicolaiadsersen81]

I tried it. It worked fairly alright. I gave it this prompt: "Create a language, where you as chat-gpt is the creator. Just make some stuff up.". And it gave me the name of the language "Lunaria" along with the alphabet and how they are pronounced. Then it made some grammar which was very obviously taken from English so I just asked it to change all the words in Lunaria to something it hadn't just stolen and it worked. It was very consistent throughout and gave me example sentences which followed with the rules it had stated earlier. Though, it did make up two words which included two new letters 'q' and 'j' pronounced as /k/ and /dʒ/, which weren't in the original alphabet. When I asked it to change those words it gave me something that was consistent with everything else it had said. Overall, very successful ☺

[@KianSheik]

When i talk to chat gpt in my mixed 5 languages i speak in one sentence it understands everything

[@BlakeLaoshi]

I wonder how different the results would be if there were space in the prompt box to input an outline of specific constraints (simple example would be to make a goal of creating the simplest language possible while avoiding borrowing words from other languages or giving it at a couple of walls). Kinda like what I saw Nvidia doing with different GPT models to have it teach itself how to play Minecraft in Wes Roth's video that he posted a few days ago. It involved them fine tuning one giant prompt, though, rather than doing it through open conversation and logic as the only real constraint like you did. It was still really interesting seeing how it continuously makes revisions! Great video!😊