How do we even define 'best and most useful' for a language model? If you have your own data and problems you are trying to solve I don't think using GPT4 in z

How do we even define 'best and most useful' for a language model? If you have your own data and problems you are trying to solve I don't think using GPT4 in zero-few shot mode (even with CoT) is necessarily better/higher performing than using a smaller fine-tuned or retrieval augmented model, it certainly wasn't true for GPT3.5 although we don't have as much insight into GPT4. If we're talking about something as general/open-ended as ChatGPT or a 'poor mans AGI' well yeah, the data curation and preprocessing alone takes a small team to figure out let alone training it. How useful this type of model is when it's still not good enough (i.e. hallucination) to be used in most production cases despite the exorbitant cost in development remains to be seen.