I’ve lately been noticing that I’ll get an immediate answer to a question followed by reasons that contradict this answer. As such, I am wondering why there is no submodel which can be used to refine responses?
I asked ChatGPT about this, and, among other things, it agreed that this is a bug & suggested using a classifier with “Needs Improvement” as an example of a classifier that could be used. It also recommended user interaction, which I think would work only if you could set a threshold for answers ‘Good’, ‘Good Enough’, ‘Terrific’, ‘Best yet’, etc.
It would be especially great if these categories could be mixed in classifier responses so that phrases like ‘bad, best yet’ can be formed. Plain ‘bad’ might be usefuk for auditing more than anything.
Wondering also why this hasn’t been implemented yet?