Why does changing the reference level wildly change the predicted probabilities?

Interfaces brms
, ,
2

Might be related to this: Problem in multilevel (hierarchical) multinomial logistic regression (with brms)