But the naysayers will argue that your problem is not novel and a solution can be trivially deduced from the training data. Right?

I really dislike the simplified word predictor explanation that is given for how LLM's work. It makes it seem like the thing is a lookup table, while ignoring the nuances of what makes it work so well.