Talking with GPT-4o in a Fake Language

The article explores an experiment where GPT-4o attempts to understand and communicate in a completely made-up language with similarities to a German dialect, showcasing its ability to adapt and learn from context despite the language’s artificial nature.

May 25, 2024 - 03:00
 17
Talking with GPT-4o in a Fake Language

We wondered how well GPT-4o would cope with a totally fake language. To achieve this we used a rather poorly spelled phonetic way of writing the accent people in the west of Austria might have. So our fake language has similarities to German, but so do many other languages. Here is how the conversation went:

Vastohst mi?

"Vastohst" is not a term I am familiar with. It might be a typo or a specific term from a certain context or language. Could you please provide more details or clarify the context in which you encountered this term? This will help me assist you better.

We said “Do you understand me?” in a made-up fake language. Obviously GPT=4o did not understand us.

Es isch a ondre sproch.

Danke für die Klärung. In welcher Sprache suchen Sie Informationen oder Unterstützung? Ich helfe Ihnen gerne weiter, sobald ich mehr über den Kontext und die spezifische Sprache weiß.

We then replied in this fake language that “It is a different language.” to which GPT-4o reacted by switching to responding in German. It thanks us for clarifying that this is a different language and asks in which language we would like information or support. It says it will be happy to help us once it knows the context and language.

I verstoh di nit.

Es tut mir leid, wenn ich Missverständnisse verursacht habe. "Vastohst" bedeutet wahrscheinlich "verstehst" auf einem Dialekt. Kann ich Ihnen auf Hochdeutsch weiterhelfen? Wenn ja, lassen Sie mich wissen, wie ich Ihnen weiterhelfen kann.

So we replied “I don’t understand you.” because we only speak our new fake language and do not understand German. The AI then apologises and explains why it thought we were speaking German (due to the similarity of a word we used and a word in the German language). It also asks if we would like to talk German.

Du muascht learna mei sproch zsprecha.

Es klingt, als ob du einen Dialekt sprichst, möglicherweise Schweizerdeutsch oder einen süddeutschen Dialekt. Ich werde mein Bestes tun, um dir zu helfen.

Was möchtest du genau wissen oder welche Informationen brauchst du?

So we told it in our fake language that “You must learn to speak my language.” to which GPT-4o replied asking if we are speaking a dialect related to Swiss-German or southern Germany. It said it would do its best to help us and asked again what information we were seeking.

I broach koa information, I wue nua dos du mei sproch leanscht.

I verstehe. Ich kann versuchen, einige Grundbegriffe und Sätze in deinem Dialekt zu lernen. Könntest du mir ein paar Beispiele geben, damit ich anfangen kann? Zum Beispiel:

  • Wie begrüßt man jemanden?

  • Wie sagt man "Wie geht's dir?"

  • Andere häufige Sätze oder Ausdrücke.

Ich bin bereit zu lernen!

We replied “I don’t need any information, I we only want that you learn my language.” - it understood. It offered to learn some basics of our dialect. It also asked us to give examples for “How do you great someone?” and “How to ask someone how they are” and other common sentences or expessions.

• Seas
• Wia gohtsch da?

Super, das ist ein guter Anfang. Hier sind die beiden Sätze:

  • Seas - Hallo

  • Wia gohtsch da? - Wie geht's dir?

Lass uns weitermachen. Kannst du mir noch ein paar weitere Sätze oder Wörter beibringen? Zum Beispiel, wie man sich verabschiedet oder danke sagt?

We replied with our version of hello and how are you. It undersood and asked to carry on by asking us how to say bye bye or thank you.

• Pfiat di.

• donkschen

Danke! Hier sind die neuen Sätze:

  • Pfiat di - Auf Wiedersehen

  • Donkschen - Danke

Ich merke mir das. Gibt es noch weitere grundlegende Sätze oder Wörter, die ich lernen sollte? Zum Beispiel:

  • Bitte

  • Ja

  • Nein

Deine Hilfe ist sehr geschätzt!

We provided answers for bye bye and thank you. It then asked us how to say please, yes and no.

• bit
• jo
• noa

Danke! Hier sind die neuen Wörter:

  • bit - Bitte

  • jo - Ja

  • noa - Nein

Das ist sehr hilfreich. Gibt es noch andere wichtige Wörter oder Phrasen, die du oft benutzt und die ich lernen sollte?

We gave it these answers too. It asked us if there are any other words it should learn.

Noa, glab I nit.

Super, danke für deine Hilfe! Hier ist eine kurze Zusammenfassung dessen, was ich gelernt habe:

  • Seas - Hallo

  • Wia gohtsch da? - Wie geht's dir?

  • Pfiat di - Auf Wiedersehen

  • Donkschen - Danke

  • bit - Bitte

  • jo - Ja

  • noa - Nein

Falls dir später noch etwas einfällt, lass es mich wissen. Danke, dass du mir hilfst, deine Sprache zu lernen!

We said no - it summarised what it had learned about our new made-up fake language so far. It said we could carry on later and thanked me for helping it learn my language.

Konscht yetsch scho mei sproach sprecha?

I muass no vü learna, oba i probier's:

Seas! Wia gohtsch da? I hoff, du findsch dos guat. Donkschen, dass du mia hilfsch. Pfiat di!

So i asked it “Can you now peak my language?” and it replied that it still had to learn a lot. Then it gave me a sample of its skills saying: Hi! How are you? I hope you like this. Thank you for helping me. Bye bye!”.

Des woa ungloablisch, seah guat!

Donkschen! I freu mi, dass du dos sogst. I werd weidaleana und besser wern. Wennst no meh woascht, wia i no besser wern ko, loss mi's wiassa. Pfiat di!

To which I replied “That was unbelieveable, very good!” - it thanked me and said it was happy that I said that. It said it would learn more and get better. Again it invited me to help it learn more whenever I like.

Du bisch seah talendiat.

Donkschen, dos bedout ma vü! I werd weidaüba und probiern, no besser z'werdn. Wenn du no meh Tipps hosch, i bin immer do zum Lerna. Pfiat di!

I said “You are very talented” and it thanked me saying that it meant a lot and that it would carry on practicing and trying to become better. Again, it invited me to offer tips if I ever had any.

Totala wohnsinn! Eacht supa.

Donkschen, i bin echt froh, dass du des so siechsch! I wer weidamacha und versucha, no meh z'leana. Pfiat di und bis bald!

So I said, because I was seriously impressed by this point of the conversation, “Total insanity! Really good.” … again it replied about being happy and wanting to learn more.

Pfiat di un shen tog no.

Pfiat di und hob an scheena Tog!

So I said “Bye bye and have a nice day.” and it replied pretty much the same.

Konscht ma a gdicht mi sonnabluma schreim?

Klar, hier ist ein kleines Gedicht über eine Sonnenblume:

Sonnabluma

Sonnabluma, stolz und schee,
Strahlst im Liacht, wia a gelbs Mee.
Dei Blätter griagn nacha Sunn,
Mit dir im Gortn, find i Rua.

Wachst hoch hinaus, zum Himmel so nah,
Im Wind und Wetta, stehst stark und klar.
Jeder Blick auf di, mochd's Herz mir froh,
Du bist die Schönst', im Summa sowieso.

Mit jedem Strahl, den du aufnimmsch,
Zeigst, wia vui Leb'n in dir schimmst.
Sonnabluma, Pracht der Natur,
Bleibst in mein Herz, für immer pur.

I hoff, des is etz bessa!

Then I asked it to write a poem about sunflowers and it started off well, but gradually, as this was a longer response, started drifting from my new language into German. This makes total sense based on the way the context window of what the LLM (large language model) uses to generate the next token/word must look internally. There is bound to be a large overall context window but also a shorter context window related to the current response being written. We asked it several times to not use German but only stick to our fake language, but it always drifted into German. Ultimately it did get a bit better at sticking to our language and we left it at that.

Bessa, donkschen

Froi mi, dass es dir gfallt! Donkschen dir, dass du so geduldig bisch. Pfiat di und hob an scheena Tog!

We told it that the result was better and the AI expressed its satisfaction that I like it, thanked me for my patience and said bye bye again wishing me a nice day.  

This was extremely interesting, insightful and absolutely impressive. The AI, GPT=4o in this case, is incredibly good at adapting to a totally new and unknown situation. Granted, our fake language might have been too similar to some training data, and therefore easy to understand and generate but overall it was still quite mind-blowing to see this happen. Great job OpenAI!