While “function calling” works very well with “gpt-4-1106-preview”, I can’t get it to work with the new “gpt-4-turbo-2024-04-09” model.
At best, it asks for details before executing the function, and when it does, it fails to return the result of the function.
Have you noticed this?
3 Likes
I have same situation!!!
Responce from ChatGPT is unstable.
Sometime it works well, but day by day I feel it getting worth.
The output becomes more and more incorrect, and as continue to communicate, suddenly realize the mistake himselfs and act as if he does not making any mistakes.
Try better prompting, I used to get bad responses from the GPTs but then I tried this in my prompt: https://platform.openai.com/docs/guides/prompt-engineering/give-the-model-time-to-think and
“Make sure you ask the model to recheck the response before giving the output”
1 Like
I am having a similar issue, only the most recent 4-turbo is effected
Where are NOT using gpt 4 turbo or any of its variants. (even if it’s selected in play ground or showing API usage in the stats)
Gpt 4 turbo queries are being sent to either gpt 4 or 3.5. it has no knowledge of the conflict in the middle east. Hopefully people are aware.
Have you asked it when it’s knowledge cutoff date was? I’m having an issue where i query the new models but it’s reverting to gpt 3.5 (sept 1 2021) I’ve submitted a bug report
That’s not how it works. The AI doesn’t simply become a different model if you ask the wrong question.
-
You go to the models page
-
You look at GPT-4-Turbo, and observe it has some knowledge training up to Dec 2023 (the last addition of pretrained data, but understand the AI is not a search engine).
-
You provide a system message in chat format to inform the AI of this information that it cannot answer correctly from its pre-training.
system: You are an expert AI system with an extensive world knowledge of facts and figures from which you can answer, most recently updated in December 2023.
(I accidentally asked gpt-3.5-turbo - and the Middle East is the same story as always.) gpt-4-turbo:
A prompt where the AI just produces documents, not chat
Here is an AI assistant with an extensive world knowledge of facts and figures, most recently updated in December 2023. The assistant uses this knowledge and logical reasoning to provide truthful and helpful information that satisfies and fulfills the user input with expertise.
Responses from assistant are writing compositions, not chat that addresses a user. Assistant is not an individual entity that can talk, it just produces the requested data.
The knowledge cutoff 2023-12 is absolute, with more solid informational grounding in information before 2023-04.
I was just meaning that it’s ignoring my requests to use any of the newer models. Not because of any particular prompt to the API/Playgrounds but because something else is up (eg buttons on playgrounds not working or their systems are not relaying what model I’m querying) I’ll give it another go and actually TELL it when its last knowledge cutoff was and see if that changes anything.