Are there any interesting architectural innovations that we seem to be on the verge of for LLM models or AI models that might be a big deal? (Excluding maybe N-gram, since everyone is already well aware of that one)

Excerpt

So, ideally for this thread we exclude the ones that everyone on here is already well aware of and discussing on here a lot, like N-gram, quantization improvements, MTP, D-flash, and D-spark, since those are improvement areas that most people on here are already pretty familiar with. I'm more curious about any interesting fundamental architectural changes to either LLMs or other types of AI models, that you guys have been reading about or is starting to get any buzz that maybe most of us don't k