Everything I've seen about hyper-specific-specialty-data-trained, limited purpose AI - particularly for medical screenings - looks promising. The idea of a FOSS AI trained locally on my own collection of documents is intriguing, but I have a hard time believing that it will be useful beyond asking it to pull up any notes relevant to keywords of any given project I'm working on.
Conversely, nearly everything I've seen about general AI/LLMs, including my own (limited) experiences, provides little reason to believe the hype espoused by the folks promoting them.