Someone told me about a tech talk from someone on TikTok's data science team.

TikTok's recommendation algorithm apparently works by creating a precise embedding of a given user, and embedding videos into that same embedding space.

This is a pretty typical modern recommender approach, I believe.

Apparently TikTok has found that the embedding is so predictive that if you select other users near a given user and look at their profile pictures, they look the same.

Same style of outfits, same basic facial shape, etc.

It kind of reminds me of studies about identical twins separated at birth.

They often share eerily similar behavioral quirks down to small details.

But if there are enough people in the world, the likelihood someone just like you--a non-twin "twin"--exists, just by happenstance.

A digital doppelganger.

And apparently TikTok's algorithm is powerful enough to find them.