GPT-4 with JSON function calling is about on par with BERT and company for simple classification problems, but is slower and more expensive. For anything more c

GPT-4 with JSON function calling is about on par with BERT and company for simple classification problems, but is slower and more expensive. For anything more complicated than that, you’re better off using standard ML models. One way they can be used to build on each other, however, is synthetic training data generation from LLMs to train other kinds of models. In my experience, this works best when you already have a fairly robust dataset for it to create permutations from.