Yeah lots of selective memory bias with how models performed in the past. There’s not some magical thing that’s made them worse in the past few years. They have gotten slightly better over the time. For example take a look at the GFS v16. It has been regularly outperforming the standard GFS so far the past few months