The models should be fine-tuned for the latest API features

I’m starting to do a lot of code gen that asks o1-mini to write code for calling OpenAI models directly. I’ve noticed that it has a tendency to want to use either gpt-4 or even gpt-3.5 when it calls the completion API. That’s presumably because that’s what its seen prior to its knowledge cutoff. The model also doesn’t know about newer features like Structured Outputs or even JSON mode.

Given that OpenAI controls the fine tuning of the models they should fine tune all of their models to know about their latest model names, API endpoints, and features. I mean shouldn’t o1-mini have a natural preference to want to call itself?

4 Likes

I completely agree.

It’s strange to me that o1-mini is supposed to be the best at coding tasks… but it is not trained on all of the recent coding advancements from OpenAI that I want to use it for.

2 Likes

I’ve heard about other forum members saying it is difficult to do so in the training data without risking quality for future models, so the easiest way to solve this for now is to include in the prompt you are sending the information about the latest models.

2 Likes