> [porn] is very easily classifiable with a list of the physical features of actors (with maybe a few behavioral distinctions), the actors in the video, how they are matched ("how" doing a lot of heavy lifting here) and possibly director, producer, and age of content, it would be easy for them to have a very specific dossier on all users. Moreover, it would be financially beneficial, because it'd be easy to maximize engagement with that stuff and a past record of engagement time

See, this was my thinking when I created an account and tried to train the recommendation algorithm.

I was sorely disappointed when the secret sauce turned out to be nothing more than:

    for vid1 in watch_history[:-10]:
        for vid2 in all_videos: // Notice how it does not exclude watch_history
            best[Levenshtein_distance(vid1.title, vid2.title) + Math.random()] = vid2
    return ksort(best)

I would conclude their engagement maximizer is meant to give me the worst possible recommendations just so I spend more time on the site and load more pages (ad banners) before I find something, but then it would have taken the highest Levenshtein distance instead of the lowest. I'm not sure what the sibling comment from jedberg ("They’re doing all that stuff and have been for many years.") is on about, because they are definitely not matching the video content. I'm guessing it's still too computationally expensive. Reverse psychology (letting you think that you are getting personalized hits but giving you unrelated stuff instead) seems a bit too much like a conspiracy theory. And the theory of giving you something somewhat-good does not explain why it always matches the title and never anything else.