Is anybody using the fine-tuning custom models on openai? Should we be using these models to train them for each assistant. Sounds like it would work better?