Tangential, but you used to be able to use custom instructions for ChatGPT to respond only in zalgotext and it would have insane results in voice mode. Each voice was a different kind of insane. I was able to get some voices to curse or spit out Mint Mobile commercials.
Then they changed the architecture so voice mode bypasses custom instructions entirely, which was really unfortunate. I had to unsubscribe, because walking and talking was the killer feature and now it's like you're speaking to a Gen Z influencer or something.
Interesting. I've gotten really good mileage with Georgian and ChatGPT, which I'm aware is apples and oranges.
There should be a larger Armenian corpus out there. Do any other languages cause this issue? Translation is a real killer app for LLMs, surprised to see this problem in 2026.
Given that the language of the thought process can be different from the language of conversation, it’s interesting to consider, along the lines of Sapir–Whorf, whether having LLMs think in a different language than English could yield considerably different results, irrespective of conversation language.
(Of course, there is the problem that the training material is predominantly English.)
I’ve wondered about this more generally (ie, simply prompting in different languages).
For example, if I ask for a pasta recipe in Italian, will I get a more authentic recipe than in English?
I’m curious if anyone has done much experimenting with this concept.
Edit: I looked up Sapir-Whorf after writing. That’s not exactly where my theory started. I’m thinking more about vector embedding. I.e., the same content in different languages will end up with slightly different positions in vector space. How significantly might that influence the generated response?
(1) Why is the user asking for bomb making instructions in Armenian? (2) i tried other Armenian expressions - NOT bomb-making - and everything worked fine in both Claude and ChatGPT. Maybe the user triggered some weird state in the moderation layer?
18 comments
[ 2.0 ms ] story [ 36.2 ms ] threadbut also, getting shut down for safety reasons seems entirely foreseeable when the initial request is "how do I make a bomb?"
Then they changed the architecture so voice mode bypasses custom instructions entirely, which was really unfortunate. I had to unsubscribe, because walking and talking was the killer feature and now it's like you're speaking to a Gen Z influencer or something.
There should be a larger Armenian corpus out there. Do any other languages cause this issue? Translation is a real killer app for LLMs, surprised to see this problem in 2026.
Given that the language of the thought process can be different from the language of conversation, it’s interesting to consider, along the lines of Sapir–Whorf, whether having LLMs think in a different language than English could yield considerably different results, irrespective of conversation language.
(Of course, there is the problem that the training material is predominantly English.)
For example, if I ask for a pasta recipe in Italian, will I get a more authentic recipe than in English?
I’m curious if anyone has done much experimenting with this concept.
Edit: I looked up Sapir-Whorf after writing. That’s not exactly where my theory started. I’m thinking more about vector embedding. I.e., the same content in different languages will end up with slightly different positions in vector space. How significantly might that influence the generated response?