Does anyone experience degradation of LLM models in chat or during voice? For instance, I use "chat" for testing flow and it doesnt follow simple system instructions. when I test it in openai chat, even with the older models, it works fine.