Here is what I imagine the problem to be: GPT is trained to wow humans. It is not trained to be honest. I mean, granted, I have no idea what kind of honesty circuits are in the works. But it can get away with lying convincingly, which is an admirable human skill, but not what we want in this case.

Hey, maybe we should ask GPT how to fix this!

There's got to be some interesting science involved in tuning a network (I donno terms) to be both impressively knowledgable and truthful.

Maybe it's because we try to get the same gizmo to also be charming and "funny" (which involves a lot of, in context harmless, lying.)

But then again if its an absolute truth machine its I/O should look something like 011001001011010100110101010101...

Which is useless to us (but absolutely truthful).

I asked GPT to tell me a joke, and it was some kind of formulaic and particularly eye-rolly "mathemetiion walks into a bar..." joke. That's it! No The Funniest Joke in the World as expected.