I work directly on it, its been my exact job, training and building models exactly like what you are using. They have not increased in capability at all, the only thing increasing is adoption and thereby the amount of training data for it to pattern match.

Thats all its doing. And its now to the point where it already peaked... about a year ago I'd say. Up to that point it was reading mostly good data, produced by humans with full logic and human thought. Now tons of new data being injected was created by itself... and is subtly flawed. Little hallucinations here and there, and they gradually degrade the models, and the more human produced training data disappears the worse it will get.

You'll see.