Thanks for sharing these initiatives.

It's informative for outsiders to take note of the absence of traditional AI causality tools in this portfolio. For instance, reasoning through ontologies would facilitate finer-grained recommendations. Maybe there is an intentional steering-away from mention of AI.

Speaking from my personal recommendations offered by Netflix, the quality hasn't improved noticeably. This may be due to the lack of deeper declarative semantics, as well, one can infer, as limited fidelity and quantity of user interactions with content. The obvious: casual inference is limited by the types and amount of data available.

It's anecdotal, but one suspects that this comparatively shallow interaction with titles could, over the long run, tend to drive subscribers away from channel subscriptions and toward title-by-title rentals (and potential churn).

Perhaps there's a related, unstated, tendency to limit causal inference hypotheses to the current content "push" model, instead of a more interactive landscape. Customers might stratify around their types and amount of engagement with content, and with the decider choices so limited (not just at Netflix, of course). Without understanding the behaviors and preferences of these different classes of users, it seems fair to question how much can be learned from "hold back" analysis.

Just a few thoughts from the outside. To repeat myself: really enjoy glimpsing how things work on the inside.