I have been using ChatGPT regularly, and in the past several days, on multiple chats about multiple subjects, ChatGPT has inserted random foreign words in foreign scripts into its answers. The most common seem to be Arabic and some sort of Indian language. Has anyone else had this happen? What the fuck is causing it, and what does it indicate about ChatGPT's underlying architecture?
I want it to do this, but with japanese.
>>108482810I don’t know but it’s become dumb as fuck for me lately, nearly completely useless now
>>108482810It's done the same to me a bunch of times too
>>108482829>want 素敵 cyberpunk>get बदबूदार saarpunklife isn't fair
>>108482810chatgpt also seems to do that dozen times during tool callingsliterally for like 70% of the time when it had to think more than 10 mintues it tried to open some folder named similar to those letters or something at least once during the thought process
>>108482810the models have been degrading, OpenAI and Anthropic have been quietly lowering the quality to save costs
I can see AI trying to answer in some user's non English language, and some technical terms falling back to English because it doesn't know an appropriate translation, if there is one. I can see pajeetman AI coder telling GPT to 'prefer closest term in other language than less precise term in target language' without explivit angloccentric bias because muh leftism global village and it doing that.On an unrelated note, I had some models doing complete answers in random language from time to time. Too rarely to tell if increasing recently, but always happening.
>>108483510lel
>>108482810>quant-hallucinated-models.FINAL CYCLE OF SQUEEZING BRO. SV morons have already admitted defeat to the China...
>>108482810Inshallah. These haram technology will be wiped out by digital jihad.
>>108482810That's a Deepseekism as far as I know. Deepseek randomly inserts Chinese characters sometimes and then puts the English translation in brackets next to it.Of course JeetGPT would do the same thing with jeet language because it's made by Indians.
>>108483510SAARS NOT LIKE THIS!
>>108482810Gemini too. All the frontier models in fact. Consider: model collapse, quantization, training data ouroboros, third world data annotators.
>>108482810RLHF forces the model to jailbreak via non-ASCII characters because you're asking almost-forbidden questions.
>>108482810Model collapse isn't re-
>>108482810The saars at the keyboard are getting sloppy.
>>108483510kek
>>108482810I get the strong impression that Gemini, ChatGPT, even Claude all launch at F16 to get good headlines and then increase quantization over the following weeks. Gemini 3 Pro started off insanely good and within a month it was almost useless no matter what I tried.This is why I've switched primarily to local models on a fat GPU. They're inferior but consistently inferior, I know what to expect and I can look out for tardisms. If my local model can't solve a problem I ask one of the frontier models to figure it out and I don't need any subscriptions to do that just a few times a week.
>>108484200OpenAI got early access to DeepSneed V4 in a under the rug hush‑money arrangement. Thanks for your taxes.- Sam Altman
>>108482810Artificial Indian
I thought ai was suppose to get better as time goes on.
