[wongarsu]

Which at its core is probably an alignment problem in the way the models are evaluated: they are measured on their short-term effects, and there exploitation rules. But if you look at the long-term effect of recommendations you really need a healthy dose of exploration to keep your users around.