On the subject of LLMs and their usefulness or lack there of:
I just got one to solve a zero-shot, 4-way classification problem with a precision of 98%.
Work involved: about 30 minutes of tweaking a prompt explain what belonged in each class.
It's like all the promises of ML from ~2015 are now coming true if you have an NLP or vision task. Best of all, you don't need training data, and you don't need to host the model.