If this were true ChatGPT would be useless for any topic that doesn't have a vast Q&A dataset. The purpose of instruct fine tuning is to illicit the model to ou

If this were true ChatGPT would be useless for any topic that doesn't have a vast Q&A dataset. The purpose of instruct fine tuning is to illicit the model to output in a question answer format. The bulk of the model's knowledge comes from the unstructured dataset. I'm also not sure how "reasoning" at inference time is relevant here. You seem to be conflating unsupervised offline training with inference. The ironic thing is that the behavior you describe actually works, you can try it right now. Add code documentation to your context window and ChatGPT will happily write code using that documentation as reference.