I have noticed the same degradation of performance. I’m pretty convinced that as the internet figures out new ways to make the models misbehave the mitigation strategies openai (and everyone else) are trying seem to compromise the models. The ‘repeat …’ thing seems like a very deep and fundamental probelm. Hopefully it’s not.
I get garbage json from the api all the time recently, have dug into it and pretty sure it’s not me. Would be nice to know if all this is ‘trough of disillusionment’ or if something is really going on.