im noticing a very weird pattern with the modern models, where the more the model reasons – the worse the outputs are.
which is like, really weird to think about. reasoning is supposed to improve the performance, but in my experience it literally just destroys it, in every way