This seems heavily impacted by the specific proxy being used for "engagement". In fact the article notes "In 2012 YouTube switched from maximizing clicks to ma

This seems heavily impacted by the specific proxy being used for "engagement". In fact the article notes "In 2012 YouTube switched from maximizing clicks to maximizing watch-time. They found it led to a short-term decrease in clicks but a long-term increase in retention. I believe Netflix similarly has invested a lot of time in developing “deep” measures of engagement." If you were to rewrite this article to not use the word engagement and say "user clicks" it would ring a lot more true, but the attempt to generalize from clicks to things like likes, etc, actually muddies the issue and impliesthis issue is more fundamental than it is. In many ways, RL promises to solve this problem by letting us optimize metrics like user retention directly, even when there is not a way to decompose it into a per-item score, but RL is hard to deploy efficiently and has not seen huge takeup. But to go back to the article, I think it's argument that estimating quality is important is clearly true, though often you can implicitly estimate quality from some other metric (e.g. aggregate watch time), but the article implicitly argues that you have to start from estimating clicks and everything else is downstream of that, which is not fundamentally true.