Comment by MichaelAza
3 months ago
I actually liked that version. I have a fairly verbose "personality" configuration and up to this point it seemed that chatgpt mainly incorporated phrasing from it into the answers. With this update, it actually started following it.
For example, I have "be dry and a little cynical" in there and it routinely starts answers with "let's be dry about this" and then gives a generic answer, but the sycophantic chatgpt was just... Dry and a little cynical. I used it to get book recommendations and it actually threw shade at Google. I asked if that was explicit training by Altman and the model made jokes about him as well. It was refreshing.
I'd say that whatever they rolled out was just much much better at following "personality" instructions, and since the default is being a bit of a sycophant... That's what they got.
This adds an interesting nuance. It may be that the sycophancy (which I noticed and was a little odd to me), is a kind of excess of fidelity in honoring cues and instructions, which, when applied to custom instructions like yours... actually was reasonably well aligned with what you were hoping for.