Running same prompt multiple times gets worse

API
1

I have found that when running the same prompt multiple times, many times the results get shorter or worse each time I re-run the prompt.

I’m assuming that OpenAI is keeping some sort of history on a per user basis, and that each time I rerun the same prompt, OpenAI is interpreting that as some level of dissatisfaction with the prior result.

It is most noticeable if I run a prompt about writing code. If I ask it to write a component that does X, it will generally do a great job the first time. If I run it a second time, I will get a less robust result. A third time may give me a tangent of my intended result (like returning the associated html view of a component).

But this happens with text as well.

My question is - is that what is going on? That there is some sort of per-user history that can alter future results based on past prompts?

If this is so, is there a way to reset the api so that I can get a result as if it’s the first time I ran the prompt?

Thanks in advance!

1 Like
2

Turn the temperature down to 0 to see consistent behavior. Also, just practice writing better prompts. Consistency is as much about what you put in.

3

Thanks David! Turning temperature down certainly helps for prompts asking for code and other less creative requests.

I’ll keep an eye on it & play with temperature.