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.

I'm an AI optimist but LLMs are still too primitive and burning resources too soon on them is a waste. Imagine nuclear powered ENIACs. The research should continue and Close AI haven't got to the pinacle of AI.

Instead of adding more energy they should focus on photonics which is 10000x more efficient computing

TsingHua university just published another step forward to feasibility