I fully agree with @pawel.patrzek here.
What metrics would satisfy you as far as evidence goes? And how exactly would you benchmark a steadily declining quality of responses? it’s not like ChatGPT users have access to logs or data quality reports from OpenAI. Only they can provide this info. But that doesn’t mean you should dismiss so many users reporting the same problem.
I had a test thread here about an audio project to determine what the responses would be like. Terrible quality responses. It lost context about 3 messages in. Does not track memory contents at all, and instructions are completely ignored. This is true across all threads, regardless of the topic. The clarity and detail in the instructions seem to make no difference at all.
