It was never really clearly communicated to the end-users, and there’s a confusion between “recordsize” (a dataset property), “block”, and what I like to call “block-on-disk”, which also applies to the ARC.
I touched on it here.
Then there’s also the performance gains of parallelization, where if you go too large with a recordsize (above 4MB), you might actually lose some of the performance benefits that were gained from fewer metadata operations. The sweet spot might be around 1 - 4MB recordsize.
There’s a fun discussion about it here, where someone used clickbait in the topic. 