As someone who was in academia at the right time to really contextualize the seismic shift in the capabilities of automated natural language processing models that came of both attention and the feasible compute scale increase that allowed for them to use long enough context windows to outpace recurrent models in this regard, I really didn't think I'd end up having to roll my eyes at people falling over themselves to exaggerate the implications, but in retrospect it's clear that this was more me being unconscionably naive then than it being that unpredictable