Nothing displaced anything. Transformers dominate NLP, CNNs dominate vision. That's the standard thing in deep learning, where a new architecture is proposed every few years that brings about a new revolution in one sub-field of AI research, then performance plateaus and everything stagnates until the Next Big Thing in AI comes around and the hype train choo-choos in another town, another domain, another modality, whatever.
I've had the misfortune of starting my MSc (ML and data science) just 2 years after the ImageNet moment and I've lived through all the heights of hype. By my count, in the last 10 years or so, the hype train whistled at MLPs (for speech recog), CNNs, LSTMs, GANs, DQNs, Transformers, and Diffusion models. Note well that each of those architectures still dominates its domain (maybe with the exception of GANs). For instance, nobody seriously considers replacing CNNs with Transformers for vision (oh, people are trying) but nobody makes a big fuss about how beating ImageNet heralds a new age of AGI right around the corner, anymore either. In fact nobody talks much about CNNs today, hence the impression that Transformers "displaced" them: only in the headlines.
In all probability, in a few years, just as the LLM hype starts to deflate, the Next Big Thing will arise that will take us one more step closer to AGI, Zeno's paradox-style.
/cynic