You’re suffering from an acute case of confirmation bias.
When ChatGPT-3 came out I said two things: (1) this is not going to develop godlike intelligence even if you scale it up considerably [2], but (2) we had no idea how these work so that particular model was terribly inefficient and there will be future models that are more efficient.
And here we are. First-mover advantage turned out to be a disadvantage. If you are OpenAI or Facebook and have a good model you're going to be inclined to take what you've got and try to improve its accuracy. The right thing to do is throw it out and start all over using everything we know about how to increase efficiency, but there's the fear that it won't be as good as the original model, customers won't want to rebuild all their RAG vectors, etc. See [4]
[1] https://garymarcus.substack.com/p/five-things-most-people-do...
[2] they get the right answer by the wrong method, hallucinations won't go away, past a certain point trying to improve performance will be like pushing a bubble around under a rug, see [3] Structurally inadequate for many tasks such as sorting (that N log N thing) and chess (look at a few million moves with alpha-beta and you can beat the worst player at the chess club, look at a few billion you beat grandmasters)
They never cared about tackling performance, just making money. They deserve everything coming their way (market loss, lack of trust).
Let's see how things unfold.
