The change that made LLMs more effective than older approaches was the transformer model, and extended to transducers/encoders+transformers for general, non language based models. But I doubt anyone is going to call current 'AI' tech transformer models, maybe if it's shortened to TMs.

There was ai used as the concept for machines that can think and reason as well as humans. I think around when marketers got people to accept splitting ai and AGI for general Intel and ai was left with whatever was left I guess reduced the scope of the meaning that continued to today's common use of ai for non thinking, non rational models and a bit of a vacuum for the original usage.