I have been using AI for almost 30 years. I have a few observations.
The AIs of today are not the same they were 30 years ago. They are bigger. It is hard to imagine the scale of resources available to AI in the last decade of the last century. I learned Prolog reading a paper book. I trained my first Neural Network on a 80386 CPU and it was written in Pascal. On the scale of modern day AI applications those AIs were absolutely tiny. Yet they had their uses.
My first ever use of AI was to ask an Expert System (Prolog) to build a school schedule. You know, the case when you have a hundred teachers and thousands of constraints like "PE blocks must be equally spaced throughout the week" and "Belinda does not work on Mondays" and "Joshua likes to have a long lunch break" and "There should be no gaps in classes for students" and the AI would find a schedule that satisfies the constraints. And it was glorious. Because Belinda could change her mind and ask for Fridays off and a couple hours later a new school-wide schedule was ready. That was around 1998, and just about that time the AI developers abandoned the Expert System approach and switched to Neural Networks. There was no way a Neural Network could solve a school schedule problem. That was an AI capability that a newer generation of AI just dropped on the floor and never got back.
And then it kept happening again and again. Yes, AIs gets better and better at more and more challenging tasks. But instead of accumulating the applications, it is about the same amount of thin frontier of usages.
That is the main reason I ran LLM locally. I still use some old models from two years ago - because they work for the cases I use them for. But the proprietary AI applications do not like users to stick on the old models for long. And they push the users to the new models, even though the new models are actually worse for those older solved use-cases.
I see a lot of frustration online linked to the release of the new version of OpenAI model, primary driven by the frustration of loosing access to the older models. This way of development was the "AI way" for as long as I have been a user of AI - and I am getting close to 30 years of that.
This is just the way the humanity approached the AI and the main reason I do not think AGI is imminent.
I hope putting this out there helps to alleviate the frustration. I remain a techno-optimist and believe in progress.