Speaking entirely anecdotally, I suspect having a bunch of people working on an ML algorithm actually does more harm than good. It's more important that everyone in the product stack is on the same page about the point.
The Facebook Newsfeed UI, for example, has to balance so many different signals. Compare that with TikTok's UI, which is very efficient at collecting a clear reinforcement signal and only has to deal with one homogenous kind of data. It's like the entire TikTok team, not just the data scientists, focused in on doing ML really well.
I'm really impressed and kind of scared whenever I use TikTok. It really does feel like a whole different level of ML-powered addictiveness, designed from the ground up rather than having the ML part added on.
