I mean, it's obvious in retrospect, and technically not too hard to implement... Combining the power of LLMs with the constrained nature of classifiers. But no one did it well, and they did it first, and it took off.
Sure, or let's say, no one else who was doing it before was noticed at this level. But hype doesn't automaically equal a paradigm shift. In this case, it's going back to what used to be one of the most common ML applications until not so long ago. If anything it feels like going old school (and I really mean that in a good way).