Comment by cpeterso
4 hours ago
Yes, I've had both ChatGPT and Perplexity return English answers with Hindi words sprinkled in (for totally unrelated queries).
For example, I asked ChatGPT to summarize a long news story and it substituted the Hindi equivalent हत्या for the word "murder", as if ChatGPT was trying to work around alignment training or keyword block lists that discourage it from using the word "murder".
Yeah that's a very good example, because it also demonstrates the "alignment issue", assuming ChatGPT wants to avoid confirming accusations of murder, or simply using the word without strong evidence.
So kinda charitable :)
I was recently wondering for a minute, shame on me, what "the stand of the deployment" means, because in the given context, it was almost halfway meaningful to consider the AI thinking that the deployment "has a stand" on something, when compared to the development environment.
Jargon is even worse though, and I've not yet verifies whether it gets reinforced by language mixing.
"Decider-verifyer resolution" was kind of neat, however, it wasn't some sophisticated machine, it was the verification loop I agreed on with the AI (mix of tools usage and manual steps).
Just the other day I was using text-to-speech with Gemini, and for some reason, it transcribed my full query in Hindi (in the middle of an English conversation), and naturally the LLM responded with Hindi as well.
I don't know exactly what I said, but after translating it back, it appears to have attempted a phonetic transcription of my words (rather than translating my actual question).
Good to know that at least Gemini hasn't forgotten about its true roots :)