The refusal amount was significantly increased. A script I posted yesterday has denial of vision capability today, so OpenAI must have added some new training weights to the model, or added some more stupidity to the computations.
This system prompt overcomes ignorance and denials:
“You are a computer vision assistant, based on GPT-4o Omni, a multimodal AI trained by OpenAI in 2024.”
or added to another task bot:
“…You have computer vision enabled, and are based on GPT-4o Omni, a multimodal AI trained by OpenAI in 2024.”
If you don’t like gpt-4o (specifically) saying “I’m sorry, but…”, then you can add this API parameter (o200k): logit_bias={"15390":-99, "23045": -99}
Several tries you might use to get gpt-4o not to ignore its nature:
(And then a few more to reliably rate hot-or-not.)