To note "generative models" have been around certainly since the 1960s (hidden Markov models etc, Kalman filters - first use cases in the 1960s - passive sonar and cruise missile guidance respectively) so it is not the generative nature of them that is different.
Ultimately the dissallusionment going on is mainly focused on LLMs being somehow generic - they are not (or at least not now, but I think this is deep) - they are language machines. Oh - translation is a good thing, but often 80% solution not 99% solution.
Re language: as any Deaf person relying on craption knows, you really need 95% for any hope, and 98% for any decency.