I think what would make LLMs much more viable is if they had more of a sense of credibility. They've developed a reputation of being very confident generators

I think what would make LLMs much more viable is if they had more of a sense of credibility. They've developed a reputation of being very confident generators of bullshit, and that's completely fatal in compliance-oriented environments. If I ask Zesty Autocorrect to generate a snippet for me, but then I have to carefully inspect and review it to make sure it actually does what I wanted, I may as well spend the same amount of time just building what I wanted in the first place. Trading off 30 minutes of coding and testing for 45 minutes of code review and testing isn't a win. I will admit to having found a use case for an AI model; I spend a fair bit of time using Stable Diffusion locally trying to roll some new desktop wallpaper, but if my anime-style vampire prince has a hallucinated third nipple, nobody's going to die or go bankrupt.