It's definitely improving rapidly.
My primary use case is for increasing my skills. I am self studying a lot of highly technical math from some graduate texts. The texts have a lot of exercises and I do roughly 80% of them which is probably too much. I write my solutions on my iPad with the Ipencil and take screenshots to upload to ChatGPT. It does find any errors I make and often confirms my solutions while suggesting alternative approaches. I always double check what it’s saying, but it’s correct the vast majority of the time on substantive proofs and calculations. When I first started do this, it made more errors. In the last two month it has only made one error. I kept hammering it and explaining that I couldn’t understand why the F they were right. What was I missing. I claimed the exercise must have been flawed and had an error in the statement. It claimed that was highly unlikely, but maybe I should google for errata…sure enough, it was a known error. I uploaded that to ChatGPT and I asked if it used a strong prior that the book was correct and if that was why it was wrong….it said yes. So it did believe it was correct bc the book was assumed to be correct.
I imagine those kinds of errors will become less common as time goes by. The product is getting better and better.
There is no point in heroically trying to fight this. Who knows…there will come point where the tech companies will need to create their own power. Maybe this will finally force green energy bc there just isn’t enough supply of non green energy.