More than a feeling.

Pretty much all the machine learning recommendation engines that emerged in the Netflix era were doomed to collapse under their own weight over time for non-mainstream users because the some limited number of mainstream modes dominate as most statistically "optimal" across the total user pool. These algorithms are best in the early days, when they're still exploring the content space for good novel fits but eventually get trapped into deep, boring grooves that work really well for tons of non-discriminating users with similar tastes.

Separately, in real commercial terms, they're all fundamentally poisoned by business model objectives of highlighting cheap content or servicing partnership/advertising deals, etc. And that problem also becomes more and more prominent as the companies running them grow and become more influential and as they need to squeeze harder and harder for revenue growth.

It was basically just a long, winding, wildly expensive road back to broadcast radio programming.

It was a good run for a while, but we're long due for a new model.