You know they run facial recognition on every photo, right? That makes doing any statistical analysis of "face", "nut", and "hard" look very, very cheap. Also

You know they run facial recognition on every photo, right? That makes doing any statistical analysis of "face", "nut", and "hard" look very, very cheap. Also, while it is totally true that image recognition software does chew computational resources, FB (and Google, and the other web-scale companies) have computational resources up the wazoo. When you've got 100,000 servers and could sic an Ivy league worth of vision research PhDs at the problem, this issue (like many others) gets slightly easier. Besides, it scales better than people, and the alternative is tasking 1/2 your CS team (Facebooks has several hundred people IIRC, might be larger now) on what is, approximately, the worst job you could imagine.