New gpt-3.5-turbo-1106 is addicted to calling functions

I just tested the new gpt-3.5-turbo-1106, and I noticed that using the same prompts in the API as with the previous version (gpt-3.5-turbo-0613), the new version has a huge tendency to call functions at the wrong times.

No matter how much I refine the prompts or provide examples in the function calls, for every 10 calls, the model incorrectly calls the function 5 times.

This had already happened before when the new version of gpt-3.5-turbo was announced last year.

I hope the gpt-3.5-turbo-0613 version remains accessible for longer.

It is not the new model - it is a new bad behavior in the model originating just days ago - and coincident with the release of new gpt-4 0125 model.

Something was damaged.

Topic where I investigate and report:

1 Like