Earlier this week, we had too many uptime errors with assistant-api. I have to switch it off to a self-hosted version of agent.
Today, we saw too many “sorry, I can’t help with that” from normal gpt-4o responses. Note that this is all with multi-modal requests. Now, we have to switch most of the calls to gemini models.
Though I personally loved using chatgpt app, I notice that I have switched majority of our services to gemini due to the instability of the output.
If there is anyone from the team who would like to share a bit, what is going on? Is there some developing trend I missed? Or the team is simply not paying attention to the APIs.
You put an image in any message? Well then, you get this garbage jammed into a system message before any text that was supposed to have aligned your AI identity to its application domain and purpose:
Counter that by trying to make any working application. You can indeed turn a normally developed application and one of its message tasks into a refusal by just adding a small white image anywhere in messages.
Answer: They are paying attention to what they want, and leaving OpenAI services is an understandable reaction.
Ok. This might have explained it. We have been passing screenshots of client’s presentation during a zoom meeting. That may have small parts of screenshots including user faces - which we asked the model to ignore in our prompt. But, that still completely jeopardize gpt-4o’s responses.
I noticed one screenshot has only the user profile in a corner of a site, which I can hardly notice myself. But, GPT still rejected it. Now GPT models are mostly useless for our applications.
