Similarly I've used it to time save in some coding, but you have to deeply understand what's going on to prevent bad errors. It really likes to "clean up" code without asking too.

Oh dear. How exactly is the LLM supposed to know any new features or functionality?

100%, each change should in theory demand a full retraining. There are ways to paper over the cracks as you say but they wouldn't last long. And retraining is EXPENSIVE.

As someone who just got a nice spanking new masters in AI, claiming an LLM model can 100% accurately say anything is lying or a lack of understanding. They just suggest the next most likely token based on input.
You can add rules for specific questions, but at that point it's just a search engine.

Saw two juvenile foxes playing and tried to make a video without scaring them. Failed miserably

Irish stereotype stuff here, ran into a new friend walking home from my job in a tech company

I was today years old when I found out the Littlest Hobo was, in canon, a sentient mutant

The Hulk might of, but the Littlest Hobo could carry a series on his own no problem

Gave this guy a little bit of my sambo while I was waiting for someone and he looked so happy with himself

Their 90s song, "untitled track 1", was superior imo.