Did anyone else notice GPT4 getting dumber?
It used to be able to understand my half baked-prompts and create very relevant code. Nowadays I have to describe what I want in explicit detail and even then its a hit or miss.
As an experiment, I pasted in my old prompts and the answers I got were noticeably worse.
My theory is they secretly moved to a dumber cheaper model and/or deliberately dumbing their previous models, so that you have to ask it to correct itself more, thereby providing it with more RLHF data.