Anyone else seeing performance issues with GPT-04?

Hi,

I have been seeing sudden performance issues with my GPT-04 (06/13 snapshot) prompt starting Feb 27th 2024. This includes things like repetitiveness, high bias on the one-shots and injecting those examples in inappropriate places, and just overall not following instructions as well as it was.

I have not updated my prompt since Feb 20th, and even tried reverting to a prompt I used at the beginning of the month, but am still seeing this issues.

Has this snapshot been deprecated? Even using the most current GPT-04 endpoint is giving me these same issues.

Is anyone else seeing things like this? Any ideas on what can help?

2 Likes

The model has been deprecated, in that its shutoff in June is announced, and you’re not supposed to be able to fine-tune it according to the deprecations page.

I don’t have a particular edge case to retry, but this model has been a dumping ground for performance reductions since it was introduced. One could have hoped that they’d leave it alone after there were more models to pay attention to, but obviously not. I get a sense that -0301 is also not acting the same, even more than the damage remaining after the model was corrupted with bad ChatML for several days.

1 Like