Perhaps most importantly, human bias is not deterministic either. Everything seems consistent with people getting excited by the hype (or just the novelty) when a new model is released, and then coming back to reality a little while later.

If they're messing with models like this, surely they'd at least occasionally make a tweak that improves intelligence rather than degrading it. But I never see anyone say Model 4.2 got smarter a week ago or whatever.