3 ms·

You’re only going to get suggestions from hot memory for whatever is cached on the nearby server cluster… …cross referenced with advertising interests unless y

You’re only going to get suggestions from hot memory for whatever is cached on the nearby server cluster… …cross referenced with advertising interests unless you are paying for Premium. If you have very particular and/or static tastes then a lot of what is nearby on the network will be outside your interests or something you’ve already seen. I mean once the model is built the recommendation engine doesn’t hill climb and the “Throbbing Gristle” auto generated playlist runs downhill to Donovan’s “Season of the Witch” because that’s less unique. Machine learning works by cutting off the long tail. Good luck.

thenaturalist 5y ago
> You’re only going to get suggestions from hot memory for whatever is cached on the nearby server cluster… …cross referenced with advertising interests unless you are paying for Premium. That sounds realistic. Do you have a source for this?
- brudgers 5y ago
  In the big picture: It's the only way to engineer a system where search responses are never slow except for last mile network issues. Which of course means, never "hitting disk." And not worrying too much about quality. Of course Premium doesn't change the caching infrastructure, just the economic calculus of competing preferences between the viewer and the advertisers when it comes to content recommendation. Which means that paying for Premium is in part paying Google to do less computation...it's not just offsetting lost revenue, it's also reducing Google's overhead.